Line data Source code
1 : //!
2 : //! Timeline repository implementation that keeps old data in files on disk, and
3 : //! the recent changes in memory. See tenant/*_layer.rs files.
4 : //! The functions here are responsible for locating the correct layer for the
5 : //! get/put call, walking back the timeline branching history as needed.
6 : //!
7 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
8 : //! directory. See docs/pageserver-storage.md for how the files are managed.
9 : //! In addition to the layer files, there is a metadata file in the same
10 : //! directory that contains information about the timeline, in particular its
11 : //! parent timeline, and the last LSN that has been written to disk.
12 : //!
13 :
14 : use anyhow::{bail, Context};
15 : use arc_swap::ArcSwap;
16 : use camino::Utf8Path;
17 : use camino::Utf8PathBuf;
18 : use enumset::EnumSet;
19 : use futures::stream::FuturesUnordered;
20 : use futures::FutureExt;
21 : use futures::StreamExt;
22 : use pageserver_api::models;
23 : use pageserver_api::models::AuxFilePolicy;
24 : use pageserver_api::models::TimelineArchivalState;
25 : use pageserver_api::models::TimelineState;
26 : use pageserver_api::models::TopTenantShardItem;
27 : use pageserver_api::models::WalRedoManagerStatus;
28 : use pageserver_api::shard::ShardIdentity;
29 : use pageserver_api::shard::ShardStripeSize;
30 : use pageserver_api::shard::TenantShardId;
31 : use remote_storage::DownloadError;
32 : use remote_storage::GenericRemoteStorage;
33 : use remote_storage::TimeoutOrCancel;
34 : use std::collections::BTreeMap;
35 : use std::fmt;
36 : use std::sync::Weak;
37 : use std::time::SystemTime;
38 : use storage_broker::BrokerClientChannel;
39 : use tokio::io::BufReader;
40 : use tokio::sync::watch;
41 : use tokio::task::JoinSet;
42 : use tokio_util::sync::CancellationToken;
43 : use tracing::*;
44 : use upload_queue::NotInitialized;
45 : use utils::backoff;
46 : use utils::circuit_breaker::CircuitBreaker;
47 : use utils::completion;
48 : use utils::crashsafe::path_with_suffix_extension;
49 : use utils::failpoint_support;
50 : use utils::fs_ext;
51 : use utils::pausable_failpoint;
52 : use utils::sync::gate::Gate;
53 : use utils::sync::gate::GateGuard;
54 : use utils::timeout::timeout_cancellable;
55 : use utils::timeout::TimeoutCancellableError;
56 : use utils::zstd::create_zst_tarball;
57 : use utils::zstd::extract_zst_tarball;
58 :
59 : use self::config::AttachedLocationConfig;
60 : use self::config::AttachmentMode;
61 : use self::config::LocationConf;
62 : use self::config::TenantConf;
63 : use self::metadata::TimelineMetadata;
64 : use self::mgr::GetActiveTenantError;
65 : use self::mgr::GetTenantError;
66 : use self::remote_timeline_client::upload::upload_index_part;
67 : use self::remote_timeline_client::RemoteTimelineClient;
68 : use self::timeline::uninit::TimelineCreateGuard;
69 : use self::timeline::uninit::TimelineExclusionError;
70 : use self::timeline::uninit::UninitializedTimeline;
71 : use self::timeline::EvictionTaskTenantState;
72 : use self::timeline::GcCutoffs;
73 : use self::timeline::TimelineResources;
74 : use self::timeline::WaitLsnError;
75 : use crate::config::PageServerConf;
76 : use crate::context::{DownloadBehavior, RequestContext};
77 : use crate::deletion_queue::DeletionQueueClient;
78 : use crate::deletion_queue::DeletionQueueError;
79 : use crate::import_datadir;
80 : use crate::is_uninit_mark;
81 : use crate::l0_flush::L0FlushGlobalState;
82 : use crate::metrics::TENANT;
83 : use crate::metrics::{
84 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
85 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
86 : };
87 : use crate::repository::GcResult;
88 : use crate::task_mgr;
89 : use crate::task_mgr::TaskKind;
90 : use crate::tenant::config::LocationMode;
91 : use crate::tenant::config::TenantConfOpt;
92 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
93 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
94 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
95 : use crate::tenant::remote_timeline_client::INITDB_PATH;
96 : use crate::tenant::storage_layer::DeltaLayer;
97 : use crate::tenant::storage_layer::ImageLayer;
98 : use crate::walredo;
99 : use crate::InitializationOrder;
100 : use std::collections::hash_map::Entry;
101 : use std::collections::HashMap;
102 : use std::collections::HashSet;
103 : use std::fmt::Debug;
104 : use std::fmt::Display;
105 : use std::fs;
106 : use std::fs::File;
107 : use std::sync::atomic::{AtomicU64, Ordering};
108 : use std::sync::Arc;
109 : use std::sync::Mutex;
110 : use std::time::{Duration, Instant};
111 :
112 : use crate::span;
113 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
114 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
115 : use crate::virtual_file::VirtualFile;
116 : use crate::walredo::PostgresRedoManager;
117 : use crate::TEMP_FILE_SUFFIX;
118 : use once_cell::sync::Lazy;
119 : pub use pageserver_api::models::TenantState;
120 : use tokio::sync::Semaphore;
121 :
122 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
123 : use utils::{
124 : crashsafe,
125 : generation::Generation,
126 : id::TimelineId,
127 : lsn::{Lsn, RecordLsn},
128 : };
129 :
130 : pub mod blob_io;
131 : pub mod block_io;
132 : pub mod vectored_blob_io;
133 :
134 : pub mod disk_btree;
135 : pub(crate) mod ephemeral_file;
136 : pub mod layer_map;
137 :
138 : pub mod metadata;
139 : pub mod remote_timeline_client;
140 : pub mod storage_layer;
141 :
142 : pub mod config;
143 : pub mod mgr;
144 : pub mod secondary;
145 : pub mod tasks;
146 : pub mod upload_queue;
147 :
148 : pub(crate) mod timeline;
149 :
150 : pub mod size;
151 :
152 : mod gc_block;
153 : pub(crate) mod throttle;
154 :
155 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
156 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
157 :
158 : // re-export for use in walreceiver
159 : pub use crate::tenant::timeline::WalReceiverInfo;
160 :
161 : /// The "tenants" part of `tenants/<tenant>/timelines...`
162 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
163 :
164 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
165 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
166 :
167 : /// References to shared objects that are passed into each tenant, such
168 : /// as the shared remote storage client and process initialization state.
169 : #[derive(Clone)]
170 : pub struct TenantSharedResources {
171 : pub broker_client: storage_broker::BrokerClientChannel,
172 : pub remote_storage: GenericRemoteStorage,
173 : pub deletion_queue_client: DeletionQueueClient,
174 : pub l0_flush_global_state: L0FlushGlobalState,
175 : }
176 :
177 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
178 : /// for an attached tenant is a subset of the [`LocationConf`], represented
179 : /// in this struct.
180 : pub(super) struct AttachedTenantConf {
181 : tenant_conf: TenantConfOpt,
182 : location: AttachedLocationConfig,
183 : }
184 :
185 : impl AttachedTenantConf {
186 0 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
187 0 : Self {
188 0 : tenant_conf,
189 0 : location,
190 0 : }
191 0 : }
192 :
193 190 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
194 190 : match &location_conf.mode {
195 190 : LocationMode::Attached(attach_conf) => Ok(Self {
196 190 : tenant_conf: location_conf.tenant_conf,
197 190 : location: *attach_conf,
198 190 : }),
199 : LocationMode::Secondary(_) => {
200 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
201 : }
202 : }
203 190 : }
204 : }
205 : struct TimelinePreload {
206 : timeline_id: TimelineId,
207 : client: RemoteTimelineClient,
208 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
209 : }
210 :
211 : pub(crate) struct TenantPreload {
212 : timelines: HashMap<TimelineId, TimelinePreload>,
213 : }
214 :
215 : /// When we spawn a tenant, there is a special mode for tenant creation that
216 : /// avoids trying to read anything from remote storage.
217 : pub(crate) enum SpawnMode {
218 : /// Activate as soon as possible
219 : Eager,
220 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
221 : Lazy,
222 : }
223 :
224 : ///
225 : /// Tenant consists of multiple timelines. Keep them in a hash table.
226 : ///
227 : pub struct Tenant {
228 : // Global pageserver config parameters
229 : pub conf: &'static PageServerConf,
230 :
231 : /// The value creation timestamp, used to measure activation delay, see:
232 : /// <https://github.com/neondatabase/neon/issues/4025>
233 : constructed_at: Instant,
234 :
235 : state: watch::Sender<TenantState>,
236 :
237 : // Overridden tenant-specific config parameters.
238 : // We keep TenantConfOpt sturct here to preserve the information
239 : // about parameters that are not set.
240 : // This is necessary to allow global config updates.
241 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
242 :
243 : tenant_shard_id: TenantShardId,
244 :
245 : // The detailed sharding information, beyond the number/count in tenant_shard_id
246 : shard_identity: ShardIdentity,
247 :
248 : /// The remote storage generation, used to protect S3 objects from split-brain.
249 : /// Does not change over the lifetime of the [`Tenant`] object.
250 : ///
251 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
252 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
253 : generation: Generation,
254 :
255 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
256 :
257 : /// During timeline creation, we first insert the TimelineId to the
258 : /// creating map, then `timelines`, then remove it from the creating map.
259 : /// **Lock order**: if acquring both, acquire`timelines` before `timelines_creating`
260 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
261 :
262 : // This mutex prevents creation of new timelines during GC.
263 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
264 : // `timelines` mutex during all GC iteration
265 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
266 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
267 : // timeout...
268 : gc_cs: tokio::sync::Mutex<()>,
269 : walredo_mgr: Option<Arc<WalRedoManager>>,
270 :
271 : // provides access to timeline data sitting in the remote storage
272 : pub(crate) remote_storage: GenericRemoteStorage,
273 :
274 : // Access to global deletion queue for when this tenant wants to schedule a deletion
275 : deletion_queue_client: DeletionQueueClient,
276 :
277 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
278 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
279 : cached_synthetic_tenant_size: Arc<AtomicU64>,
280 :
281 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
282 :
283 : /// Track repeated failures to compact, so that we can back off.
284 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
285 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
286 :
287 : /// If the tenant is in Activating state, notify this to encourage it
288 : /// to proceed to Active as soon as possible, rather than waiting for lazy
289 : /// background warmup.
290 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
291 :
292 : // Cancellation token fires when we have entered shutdown(). This is a parent of
293 : // Timelines' cancellation token.
294 : pub(crate) cancel: CancellationToken,
295 :
296 : // Users of the Tenant such as the page service must take this Gate to avoid
297 : // trying to use a Tenant which is shutting down.
298 : pub(crate) gate: Gate,
299 :
300 : /// Throttle applied at the top of [`Timeline::get`].
301 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
302 : pub(crate) timeline_get_throttle:
303 : Arc<throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>>,
304 :
305 : /// An ongoing timeline detach concurrency limiter.
306 : ///
307 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
308 : /// to have two running at the same time. A different one can be started if an earlier one
309 : /// has failed for whatever reason.
310 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
311 :
312 : /// `index_part.json` based gc blocking reason tracking.
313 : ///
314 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
315 : /// proceeding.
316 : pub(crate) gc_block: gc_block::GcBlock,
317 :
318 : l0_flush_global_state: L0FlushGlobalState,
319 : }
320 :
321 : impl std::fmt::Debug for Tenant {
322 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
323 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
324 0 : }
325 : }
326 :
327 : pub(crate) enum WalRedoManager {
328 : Prod(WalredoManagerId, PostgresRedoManager),
329 : #[cfg(test)]
330 : Test(harness::TestRedoManager),
331 : }
332 :
333 0 : #[derive(thiserror::Error, Debug)]
334 : #[error("pageserver is shutting down")]
335 : pub(crate) struct GlobalShutDown;
336 :
337 : impl WalRedoManager {
338 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
339 0 : let id = WalredoManagerId::next();
340 0 : let arc = Arc::new(Self::Prod(id, mgr));
341 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
342 0 : match &mut *guard {
343 0 : Some(map) => {
344 0 : map.insert(id, Arc::downgrade(&arc));
345 0 : Ok(arc)
346 : }
347 0 : None => Err(GlobalShutDown),
348 : }
349 0 : }
350 : }
351 :
352 : impl Drop for WalRedoManager {
353 10 : fn drop(&mut self) {
354 10 : match self {
355 0 : Self::Prod(id, _) => {
356 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
357 0 : if let Some(map) = &mut *guard {
358 0 : map.remove(id).expect("new() registers, drop() unregisters");
359 0 : }
360 : }
361 : #[cfg(test)]
362 10 : Self::Test(_) => {
363 10 : // Not applicable to test redo manager
364 10 : }
365 : }
366 10 : }
367 : }
368 :
369 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
370 : /// the walredo processes outside of the regular order.
371 : ///
372 : /// This is necessary to work around a systemd bug where it freezes if there are
373 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
374 : #[allow(clippy::type_complexity)]
375 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
376 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
377 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
378 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
379 : pub(crate) struct WalredoManagerId(u64);
380 : impl WalredoManagerId {
381 0 : pub fn next() -> Self {
382 0 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
383 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
384 0 : if id == 0 {
385 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
386 0 : }
387 0 : Self(id)
388 0 : }
389 : }
390 :
391 : #[cfg(test)]
392 : impl From<harness::TestRedoManager> for WalRedoManager {
393 182 : fn from(mgr: harness::TestRedoManager) -> Self {
394 182 : Self::Test(mgr)
395 182 : }
396 : }
397 :
398 : impl WalRedoManager {
399 6 : pub(crate) async fn shutdown(&self) -> bool {
400 6 : match self {
401 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
402 : #[cfg(test)]
403 : Self::Test(_) => {
404 : // Not applicable to test redo manager
405 6 : true
406 : }
407 : }
408 6 : }
409 :
410 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
411 0 : match self {
412 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
413 0 : #[cfg(test)]
414 0 : Self::Test(_) => {
415 0 : // Not applicable to test redo manager
416 0 : }
417 0 : }
418 0 : }
419 :
420 : /// # Cancel-Safety
421 : ///
422 : /// This method is cancellation-safe.
423 296 : pub async fn request_redo(
424 296 : &self,
425 296 : key: crate::repository::Key,
426 296 : lsn: Lsn,
427 296 : base_img: Option<(Lsn, bytes::Bytes)>,
428 296 : records: Vec<(Lsn, crate::walrecord::NeonWalRecord)>,
429 296 : pg_version: u32,
430 296 : ) -> Result<bytes::Bytes, walredo::Error> {
431 296 : match self {
432 0 : Self::Prod(_, mgr) => {
433 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
434 0 : .await
435 : }
436 : #[cfg(test)]
437 296 : Self::Test(mgr) => {
438 296 : mgr.request_redo(key, lsn, base_img, records, pg_version)
439 0 : .await
440 : }
441 : }
442 296 : }
443 :
444 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
445 0 : match self {
446 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
447 0 : #[cfg(test)]
448 0 : WalRedoManager::Test(_) => None,
449 0 : }
450 0 : }
451 : }
452 :
453 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
454 : pub enum GetTimelineError {
455 : #[error("Timeline is shutting down")]
456 : ShuttingDown,
457 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
458 : NotActive {
459 : tenant_id: TenantShardId,
460 : timeline_id: TimelineId,
461 : state: TimelineState,
462 : },
463 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
464 : NotFound {
465 : tenant_id: TenantShardId,
466 : timeline_id: TimelineId,
467 : },
468 : }
469 :
470 0 : #[derive(Debug, thiserror::Error)]
471 : pub enum LoadLocalTimelineError {
472 : #[error("FailedToLoad")]
473 : Load(#[source] anyhow::Error),
474 : #[error("FailedToResumeDeletion")]
475 : ResumeDeletion(#[source] anyhow::Error),
476 : }
477 :
478 0 : #[derive(thiserror::Error)]
479 : pub enum DeleteTimelineError {
480 : #[error("NotFound")]
481 : NotFound,
482 :
483 : #[error("HasChildren")]
484 : HasChildren(Vec<TimelineId>),
485 :
486 : #[error("Timeline deletion is already in progress")]
487 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
488 :
489 : #[error(transparent)]
490 : Other(#[from] anyhow::Error),
491 : }
492 :
493 : impl Debug for DeleteTimelineError {
494 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
495 0 : match self {
496 0 : Self::NotFound => write!(f, "NotFound"),
497 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
498 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
499 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
500 : }
501 0 : }
502 : }
503 :
504 : pub enum SetStoppingError {
505 : AlreadyStopping(completion::Barrier),
506 : Broken,
507 : }
508 :
509 : impl Debug for SetStoppingError {
510 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
511 0 : match self {
512 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
513 0 : Self::Broken => write!(f, "Broken"),
514 : }
515 0 : }
516 : }
517 :
518 0 : #[derive(thiserror::Error, Debug)]
519 : pub enum CreateTimelineError {
520 : #[error("creation of timeline with the given ID is in progress")]
521 : AlreadyCreating,
522 : #[error("timeline already exists with different parameters")]
523 : Conflict,
524 : #[error(transparent)]
525 : AncestorLsn(anyhow::Error),
526 : #[error("ancestor timeline is not active")]
527 : AncestorNotActive,
528 : #[error("tenant shutting down")]
529 : ShuttingDown,
530 : #[error(transparent)]
531 : Other(#[from] anyhow::Error),
532 : }
533 :
534 : #[derive(thiserror::Error, Debug)]
535 : enum InitdbError {
536 : Other(anyhow::Error),
537 : Cancelled,
538 : Spawn(std::io::Result<()>),
539 : Failed(std::process::ExitStatus, Vec<u8>),
540 : }
541 :
542 : impl fmt::Display for InitdbError {
543 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
544 0 : match self {
545 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
546 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
547 0 : InitdbError::Failed(status, stderr) => write!(
548 0 : f,
549 0 : "Command failed with status {:?}: {}",
550 0 : status,
551 0 : String::from_utf8_lossy(stderr)
552 0 : ),
553 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
554 : }
555 0 : }
556 : }
557 :
558 : impl From<std::io::Error> for InitdbError {
559 0 : fn from(error: std::io::Error) -> Self {
560 0 : InitdbError::Spawn(Err(error))
561 0 : }
562 : }
563 :
564 : enum CreateTimelineCause {
565 : Load,
566 : Delete,
567 : }
568 :
569 0 : #[derive(thiserror::Error, Debug)]
570 : pub(crate) enum GcError {
571 : // The tenant is shutting down
572 : #[error("tenant shutting down")]
573 : TenantCancelled,
574 :
575 : // The tenant is shutting down
576 : #[error("timeline shutting down")]
577 : TimelineCancelled,
578 :
579 : // The tenant is in a state inelegible to run GC
580 : #[error("not active")]
581 : NotActive,
582 :
583 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
584 : #[error("not active")]
585 : BadLsn { why: String },
586 :
587 : // A remote storage error while scheduling updates after compaction
588 : #[error(transparent)]
589 : Remote(anyhow::Error),
590 :
591 : // An error reading while calculating GC cutoffs
592 : #[error(transparent)]
593 : GcCutoffs(PageReconstructError),
594 :
595 : // If GC was invoked for a particular timeline, this error means it didn't exist
596 : #[error("timeline not found")]
597 : TimelineNotFound,
598 : }
599 :
600 : impl From<PageReconstructError> for GcError {
601 0 : fn from(value: PageReconstructError) -> Self {
602 0 : match value {
603 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
604 0 : other => Self::GcCutoffs(other),
605 : }
606 0 : }
607 : }
608 :
609 : impl From<NotInitialized> for GcError {
610 0 : fn from(value: NotInitialized) -> Self {
611 0 : match value {
612 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
613 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
614 : }
615 0 : }
616 : }
617 :
618 : impl From<timeline::layer_manager::Shutdown> for GcError {
619 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
620 0 : GcError::TimelineCancelled
621 0 : }
622 : }
623 :
624 0 : #[derive(thiserror::Error, Debug)]
625 : pub(crate) enum LoadConfigError {
626 : #[error("TOML deserialization error: '{0}'")]
627 : DeserializeToml(#[from] toml_edit::de::Error),
628 :
629 : #[error("Config not found at {0}")]
630 : NotFound(Utf8PathBuf),
631 : }
632 :
633 : impl Tenant {
634 : /// Yet another helper for timeline initialization.
635 : ///
636 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
637 : /// - Scans the local timeline directory for layer files and builds the layer map
638 : /// - Downloads remote index file and adds remote files to the layer map
639 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
640 : ///
641 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
642 : /// it is marked as Active.
643 : #[allow(clippy::too_many_arguments)]
644 6 : async fn timeline_init_and_sync(
645 6 : &self,
646 6 : timeline_id: TimelineId,
647 6 : resources: TimelineResources,
648 6 : index_part: Option<IndexPart>,
649 6 : metadata: TimelineMetadata,
650 6 : ancestor: Option<Arc<Timeline>>,
651 6 : last_aux_file_policy: Option<AuxFilePolicy>,
652 6 : _ctx: &RequestContext,
653 6 : ) -> anyhow::Result<()> {
654 6 : let tenant_id = self.tenant_shard_id;
655 :
656 6 : let timeline = self.create_timeline_struct(
657 6 : timeline_id,
658 6 : &metadata,
659 6 : ancestor.clone(),
660 6 : resources,
661 6 : CreateTimelineCause::Load,
662 6 : // This could be derived from ancestor branch + index part. Though the only caller of `timeline_init_and_sync` is `load_remote_timeline`,
663 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.
664 6 : // Therefore, we pass this field explicitly for now, and remove it once we fully migrate to aux file v2.
665 6 : last_aux_file_policy,
666 6 : )?;
667 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
668 6 : anyhow::ensure!(
669 6 : disk_consistent_lsn.is_valid(),
670 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
671 : );
672 6 : assert_eq!(
673 6 : disk_consistent_lsn,
674 6 : metadata.disk_consistent_lsn(),
675 0 : "these are used interchangeably"
676 : );
677 :
678 6 : if let Some(index_part) = index_part.as_ref() {
679 6 : timeline.remote_client.init_upload_queue(index_part)?;
680 :
681 6 : timeline
682 6 : .last_aux_file_policy
683 6 : .store(index_part.last_aux_file_policy());
684 : } else {
685 : // No data on the remote storage, but we have local metadata file. We can end up
686 : // here with timeline_create being interrupted before finishing index part upload.
687 : // By doing what we do here, the index part upload is retried.
688 : // If control plane retries timeline creation in the meantime, the mgmt API handler
689 : // for timeline creation will coalesce on the upload we queue here.
690 :
691 : // FIXME: this branch should be dead code as we no longer write local metadata.
692 :
693 0 : timeline
694 0 : .remote_client
695 0 : .init_upload_queue_for_empty_remote(&metadata)?;
696 0 : timeline
697 0 : .remote_client
698 0 : .schedule_index_upload_for_full_metadata_update(&metadata)?;
699 : }
700 :
701 6 : timeline
702 6 : .load_layer_map(disk_consistent_lsn, index_part)
703 6 : .await
704 6 : .with_context(|| {
705 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
706 6 : })?;
707 :
708 : {
709 : // avoiding holding it across awaits
710 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
711 6 : match timelines_accessor.entry(timeline_id) {
712 : // We should never try and load the same timeline twice during startup
713 : Entry::Occupied(_) => {
714 0 : unreachable!(
715 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
716 0 : );
717 : }
718 6 : Entry::Vacant(v) => {
719 6 : v.insert(Arc::clone(&timeline));
720 6 : timeline.maybe_spawn_flush_loop();
721 6 : }
722 6 : }
723 6 : };
724 6 :
725 6 : // Sanity check: a timeline should have some content.
726 6 : anyhow::ensure!(
727 6 : ancestor.is_some()
728 4 : || timeline
729 4 : .layers
730 4 : .read()
731 0 : .await
732 4 : .layer_map()
733 4 : .expect("currently loading, layer manager cannot be shutdown already")
734 4 : .iter_historic_layers()
735 4 : .next()
736 4 : .is_some(),
737 0 : "Timeline has no ancestor and no layer files"
738 : );
739 :
740 6 : Ok(())
741 6 : }
742 :
743 : /// Attach a tenant that's available in cloud storage.
744 : ///
745 : /// This returns quickly, after just creating the in-memory object
746 : /// Tenant struct and launching a background task to download
747 : /// the remote index files. On return, the tenant is most likely still in
748 : /// Attaching state, and it will become Active once the background task
749 : /// finishes. You can use wait_until_active() to wait for the task to
750 : /// complete.
751 : ///
752 : #[allow(clippy::too_many_arguments)]
753 0 : pub(crate) fn spawn(
754 0 : conf: &'static PageServerConf,
755 0 : tenant_shard_id: TenantShardId,
756 0 : resources: TenantSharedResources,
757 0 : attached_conf: AttachedTenantConf,
758 0 : shard_identity: ShardIdentity,
759 0 : init_order: Option<InitializationOrder>,
760 0 : mode: SpawnMode,
761 0 : ctx: &RequestContext,
762 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
763 0 : let wal_redo_manager =
764 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
765 :
766 : let TenantSharedResources {
767 0 : broker_client,
768 0 : remote_storage,
769 0 : deletion_queue_client,
770 0 : l0_flush_global_state,
771 0 : } = resources;
772 0 :
773 0 : let attach_mode = attached_conf.location.attach_mode;
774 0 : let generation = attached_conf.location.generation;
775 0 :
776 0 : let tenant = Arc::new(Tenant::new(
777 0 : TenantState::Attaching,
778 0 : conf,
779 0 : attached_conf,
780 0 : shard_identity,
781 0 : Some(wal_redo_manager),
782 0 : tenant_shard_id,
783 0 : remote_storage.clone(),
784 0 : deletion_queue_client,
785 0 : l0_flush_global_state,
786 0 : ));
787 0 :
788 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
789 0 : // we shut down while attaching.
790 0 : let attach_gate_guard = tenant
791 0 : .gate
792 0 : .enter()
793 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
794 0 :
795 0 : // Do all the hard work in the background
796 0 : let tenant_clone = Arc::clone(&tenant);
797 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
798 0 : task_mgr::spawn(
799 0 : &tokio::runtime::Handle::current(),
800 0 : TaskKind::Attach,
801 0 : tenant_shard_id,
802 0 : None,
803 0 : "attach tenant",
804 0 : async move {
805 0 :
806 0 : info!(
807 : ?attach_mode,
808 0 : "Attaching tenant"
809 : );
810 :
811 0 : let _gate_guard = attach_gate_guard;
812 0 :
813 0 : // Is this tenant being spawned as part of process startup?
814 0 : let starting_up = init_order.is_some();
815 : scopeguard::defer! {
816 : if starting_up {
817 : TENANT.startup_complete.inc();
818 : }
819 : }
820 :
821 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
822 : enum BrokenVerbosity {
823 : Error,
824 : Info
825 : }
826 0 : let make_broken =
827 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
828 0 : match verbosity {
829 : BrokenVerbosity::Info => {
830 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
831 : },
832 : BrokenVerbosity::Error => {
833 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
834 : }
835 : }
836 0 : t.state.send_modify(|state| {
837 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
838 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
839 0 : assert!(
840 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
841 0 : "the attach task owns the tenant state until activation is complete"
842 : );
843 :
844 0 : *state = TenantState::broken_from_reason(err.to_string());
845 0 : });
846 0 : };
847 :
848 0 : let mut init_order = init_order;
849 0 : // take the completion because initial tenant loading will complete when all of
850 0 : // these tasks complete.
851 0 : let _completion = init_order
852 0 : .as_mut()
853 0 : .and_then(|x| x.initial_tenant_load.take());
854 0 : let remote_load_completion = init_order
855 0 : .as_mut()
856 0 : .and_then(|x| x.initial_tenant_load_remote.take());
857 :
858 : enum AttachType<'a> {
859 : /// We are attaching this tenant lazily in the background.
860 : Warmup {
861 : _permit: tokio::sync::SemaphorePermit<'a>,
862 : during_startup: bool
863 : },
864 : /// We are attaching this tenant as soon as we can, because for example an
865 : /// endpoint tried to access it.
866 : OnDemand,
867 : /// During normal operations after startup, we are attaching a tenant, and
868 : /// eager attach was requested.
869 : Normal,
870 : }
871 :
872 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
873 : // Before doing any I/O, wait for at least one of:
874 : // - A client attempting to access to this tenant (on-demand loading)
875 : // - A permit becoming available in the warmup semaphore (background warmup)
876 :
877 : tokio::select!(
878 : permit = tenant_clone.activate_now_sem.acquire() => {
879 : let _ = permit.expect("activate_now_sem is never closed");
880 : tracing::info!("Activating tenant (on-demand)");
881 : AttachType::OnDemand
882 : },
883 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
884 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
885 : tracing::info!("Activating tenant (warmup)");
886 : AttachType::Warmup {
887 : _permit,
888 : during_startup: init_order.is_some()
889 : }
890 : }
891 : _ = tenant_clone.cancel.cancelled() => {
892 : // This is safe, but should be pretty rare: it is interesting if a tenant
893 : // stayed in Activating for such a long time that shutdown found it in
894 : // that state.
895 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
896 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
897 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
898 : // just shutting down), but ensures progress.
899 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
900 : return Ok(());
901 : },
902 : )
903 : } else {
904 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
905 : // concurrent_tenant_warmup queue
906 0 : AttachType::Normal
907 : };
908 :
909 0 : let preload = match &mode {
910 : SpawnMode::Eager | SpawnMode::Lazy => {
911 0 : let _preload_timer = TENANT.preload.start_timer();
912 0 : let res = tenant_clone
913 0 : .preload(&remote_storage, task_mgr::shutdown_token())
914 0 : .await;
915 0 : match res {
916 0 : Ok(p) => Some(p),
917 0 : Err(e) => {
918 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
919 0 : return Ok(());
920 : }
921 : }
922 : }
923 :
924 : };
925 :
926 : // Remote preload is complete.
927 0 : drop(remote_load_completion);
928 :
929 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
930 0 : let attached = {
931 0 : let _attach_timer = Some(TENANT.attach.start_timer());
932 0 : tenant_clone.attach(preload, &ctx).await
933 : };
934 :
935 0 : match attached {
936 : Ok(()) => {
937 0 : info!("attach finished, activating");
938 0 : tenant_clone.activate(broker_client, None, &ctx);
939 : }
940 0 : Err(e) => {
941 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
942 0 : }
943 : }
944 :
945 : // If we are doing an opportunistic warmup attachment at startup, initialize
946 : // logical size at the same time. This is better than starting a bunch of idle tenants
947 : // with cold caches and then coming back later to initialize their logical sizes.
948 : //
949 : // It also prevents the warmup proccess competing with the concurrency limit on
950 : // logical size calculations: if logical size calculation semaphore is saturated,
951 : // then warmup will wait for that before proceeding to the next tenant.
952 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
953 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
954 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
955 0 : while futs.next().await.is_some() {}
956 0 : tracing::info!("Warm-up complete");
957 0 : }
958 :
959 0 : Ok(())
960 0 : }
961 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
962 : );
963 0 : Ok(tenant)
964 0 : }
965 :
966 364 : #[instrument(skip_all)]
967 : pub(crate) async fn preload(
968 : self: &Arc<Self>,
969 : remote_storage: &GenericRemoteStorage,
970 : cancel: CancellationToken,
971 : ) -> anyhow::Result<TenantPreload> {
972 : span::debug_assert_current_span_has_tenant_id();
973 : // Get list of remote timelines
974 : // download index files for every tenant timeline
975 : info!("listing remote timelines");
976 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
977 : remote_storage,
978 : self.tenant_shard_id,
979 : cancel.clone(),
980 : )
981 : .await?;
982 :
983 : info!("found {} timelines", remote_timeline_ids.len(),);
984 :
985 : for k in other_keys {
986 : warn!("Unexpected non timeline key {k}");
987 : }
988 :
989 : Ok(TenantPreload {
990 : timelines: Self::load_timeline_metadata(
991 : self,
992 : remote_timeline_ids,
993 : remote_storage,
994 : cancel,
995 : )
996 : .await?,
997 : })
998 : }
999 :
1000 : ///
1001 : /// Background task that downloads all data for a tenant and brings it to Active state.
1002 : ///
1003 : /// No background tasks are started as part of this routine.
1004 : ///
1005 182 : async fn attach(
1006 182 : self: &Arc<Tenant>,
1007 182 : preload: Option<TenantPreload>,
1008 182 : ctx: &RequestContext,
1009 182 : ) -> anyhow::Result<()> {
1010 182 : span::debug_assert_current_span_has_tenant_id();
1011 182 :
1012 182 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1013 :
1014 182 : let Some(preload) = preload else {
1015 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1016 : };
1017 :
1018 182 : let mut timelines_to_resume_deletions = vec![];
1019 182 :
1020 182 : let mut remote_index_and_client = HashMap::new();
1021 182 : let mut timeline_ancestors = HashMap::new();
1022 182 : let mut existent_timelines = HashSet::new();
1023 188 : for (timeline_id, preload) in preload.timelines {
1024 6 : let index_part = match preload.index_part {
1025 6 : Ok(i) => {
1026 6 : debug!("remote index part exists for timeline {timeline_id}");
1027 : // We found index_part on the remote, this is the standard case.
1028 6 : existent_timelines.insert(timeline_id);
1029 6 : i
1030 : }
1031 : Err(DownloadError::NotFound) => {
1032 : // There is no index_part on the remote. We only get here
1033 : // if there is some prefix for the timeline in the remote storage.
1034 : // This can e.g. be the initdb.tar.zst archive, maybe a
1035 : // remnant from a prior incomplete creation or deletion attempt.
1036 : // Delete the local directory as the deciding criterion for a
1037 : // timeline's existence is presence of index_part.
1038 0 : info!(%timeline_id, "index_part not found on remote");
1039 0 : continue;
1040 : }
1041 0 : Err(e) => {
1042 0 : // Some (possibly ephemeral) error happened during index_part download.
1043 0 : // Pretend the timeline exists to not delete the timeline directory,
1044 0 : // as it might be a temporary issue and we don't want to re-download
1045 0 : // everything after it resolves.
1046 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1047 :
1048 0 : existent_timelines.insert(timeline_id);
1049 0 : continue;
1050 : }
1051 : };
1052 6 : match index_part {
1053 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1054 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1055 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1056 6 : }
1057 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1058 0 : info!(
1059 0 : "timeline {} is deleted, picking to resume deletion",
1060 : timeline_id
1061 : );
1062 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1063 : }
1064 : }
1065 : }
1066 :
1067 182 : let mut gc_blocks = HashMap::new();
1068 :
1069 : // For every timeline, download the metadata file, scan the local directory,
1070 : // and build a layer map that contains an entry for each remote and local
1071 : // layer file.
1072 182 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1073 188 : for (timeline_id, remote_metadata) in sorted_timelines {
1074 6 : let (index_part, remote_client) = remote_index_and_client
1075 6 : .remove(&timeline_id)
1076 6 : .expect("just put it in above");
1077 :
1078 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1079 : // could just filter these away, but it helps while testing
1080 0 : anyhow::ensure!(
1081 0 : !blocking.reasons.is_empty(),
1082 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1083 : );
1084 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1085 0 : assert!(prev.is_none());
1086 6 : }
1087 :
1088 : // TODO again handle early failure
1089 6 : self.load_remote_timeline(
1090 6 : timeline_id,
1091 6 : index_part,
1092 6 : remote_metadata,
1093 6 : TimelineResources {
1094 6 : remote_client,
1095 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1096 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1097 6 : },
1098 6 : ctx,
1099 6 : )
1100 11 : .await
1101 6 : .with_context(|| {
1102 0 : format!(
1103 0 : "failed to load remote timeline {} for tenant {}",
1104 0 : timeline_id, self.tenant_shard_id
1105 0 : )
1106 6 : })?;
1107 : }
1108 :
1109 : // Walk through deleted timelines, resume deletion
1110 182 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1111 0 : remote_timeline_client
1112 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1113 0 : .context("init queue stopped")
1114 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1115 :
1116 0 : DeleteTimelineFlow::resume_deletion(
1117 0 : Arc::clone(self),
1118 0 : timeline_id,
1119 0 : &index_part.metadata,
1120 0 : remote_timeline_client,
1121 0 : )
1122 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1123 0 : .await
1124 0 : .context("resume_deletion")
1125 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1126 : }
1127 :
1128 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1129 : // IndexPart is the source of truth.
1130 182 : self.clean_up_timelines(&existent_timelines)?;
1131 :
1132 182 : self.gc_block.set_scanned(gc_blocks);
1133 182 :
1134 182 : fail::fail_point!("attach-before-activate", |_| {
1135 0 : anyhow::bail!("attach-before-activate");
1136 182 : });
1137 182 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1138 :
1139 182 : info!("Done");
1140 :
1141 182 : Ok(())
1142 182 : }
1143 :
1144 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1145 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1146 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1147 182 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1148 182 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1149 :
1150 182 : let entries = match timelines_dir.read_dir_utf8() {
1151 182 : Ok(d) => d,
1152 0 : Err(e) => {
1153 0 : if e.kind() == std::io::ErrorKind::NotFound {
1154 0 : return Ok(());
1155 : } else {
1156 0 : return Err(e).context("list timelines directory for tenant");
1157 : }
1158 : }
1159 : };
1160 :
1161 190 : for entry in entries {
1162 8 : let entry = entry.context("read timeline dir entry")?;
1163 8 : let entry_path = entry.path();
1164 :
1165 8 : let purge = if crate::is_temporary(entry_path)
1166 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1167 8 : || is_uninit_mark(entry_path)
1168 8 : || crate::is_delete_mark(entry_path)
1169 : {
1170 0 : true
1171 : } else {
1172 8 : match TimelineId::try_from(entry_path.file_name()) {
1173 8 : Ok(i) => {
1174 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1175 8 : !existent_timelines.contains(&i)
1176 : }
1177 0 : Err(e) => {
1178 0 : tracing::warn!(
1179 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1180 : );
1181 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1182 0 : false
1183 : }
1184 : }
1185 : };
1186 :
1187 8 : if purge {
1188 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1189 2 : if let Err(e) = match entry.file_type() {
1190 2 : Ok(t) => if t.is_dir() {
1191 2 : std::fs::remove_dir_all(entry_path)
1192 : } else {
1193 0 : std::fs::remove_file(entry_path)
1194 : }
1195 2 : .or_else(fs_ext::ignore_not_found),
1196 0 : Err(e) => Err(e),
1197 : } {
1198 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1199 2 : }
1200 6 : }
1201 : }
1202 :
1203 182 : Ok(())
1204 182 : }
1205 :
1206 : /// Get sum of all remote timelines sizes
1207 : ///
1208 : /// This function relies on the index_part instead of listing the remote storage
1209 0 : pub fn remote_size(&self) -> u64 {
1210 0 : let mut size = 0;
1211 :
1212 0 : for timeline in self.list_timelines() {
1213 0 : size += timeline.remote_client.get_remote_physical_size();
1214 0 : }
1215 :
1216 0 : size
1217 0 : }
1218 :
1219 12 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1220 : async fn load_remote_timeline(
1221 : &self,
1222 : timeline_id: TimelineId,
1223 : index_part: IndexPart,
1224 : remote_metadata: TimelineMetadata,
1225 : resources: TimelineResources,
1226 : ctx: &RequestContext,
1227 : ) -> anyhow::Result<()> {
1228 : span::debug_assert_current_span_has_tenant_id();
1229 :
1230 : info!("downloading index file for timeline {}", timeline_id);
1231 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1232 : .await
1233 : .context("Failed to create new timeline directory")?;
1234 :
1235 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1236 : let timelines = self.timelines.lock().unwrap();
1237 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1238 0 : || {
1239 0 : anyhow::anyhow!(
1240 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1241 0 : )
1242 0 : },
1243 : )?))
1244 : } else {
1245 : None
1246 : };
1247 :
1248 : let last_aux_file_policy = index_part.last_aux_file_policy();
1249 :
1250 : self.timeline_init_and_sync(
1251 : timeline_id,
1252 : resources,
1253 : Some(index_part),
1254 : remote_metadata,
1255 : ancestor,
1256 : last_aux_file_policy,
1257 : ctx,
1258 : )
1259 : .await
1260 : }
1261 :
1262 182 : async fn load_timeline_metadata(
1263 182 : self: &Arc<Tenant>,
1264 182 : timeline_ids: HashSet<TimelineId>,
1265 182 : remote_storage: &GenericRemoteStorage,
1266 182 : cancel: CancellationToken,
1267 182 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1268 182 : let mut part_downloads = JoinSet::new();
1269 188 : for timeline_id in timeline_ids {
1270 6 : let client = RemoteTimelineClient::new(
1271 6 : remote_storage.clone(),
1272 6 : self.deletion_queue_client.clone(),
1273 6 : self.conf,
1274 6 : self.tenant_shard_id,
1275 6 : timeline_id,
1276 6 : self.generation,
1277 6 : );
1278 6 : let cancel_clone = cancel.clone();
1279 6 : part_downloads.spawn(
1280 6 : async move {
1281 6 : debug!("starting index part download");
1282 :
1283 20 : let index_part = client.download_index_file(&cancel_clone).await;
1284 :
1285 6 : debug!("finished index part download");
1286 :
1287 6 : Result::<_, anyhow::Error>::Ok(TimelinePreload {
1288 6 : client,
1289 6 : timeline_id,
1290 6 : index_part,
1291 6 : })
1292 6 : }
1293 6 : .map(move |res| {
1294 6 : res.with_context(|| format!("download index part for timeline {timeline_id}"))
1295 6 : })
1296 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1297 : );
1298 : }
1299 :
1300 182 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1301 :
1302 188 : loop {
1303 188 : tokio::select!(
1304 : next = part_downloads.join_next() => {
1305 : match next {
1306 : Some(result) => {
1307 : let preload_result = result.context("join preload task")?;
1308 : let preload = preload_result?;
1309 : timeline_preloads.insert(preload.timeline_id, preload);
1310 : },
1311 : None => {
1312 : break;
1313 : }
1314 : }
1315 : },
1316 : _ = cancel.cancelled() => {
1317 : anyhow::bail!("Cancelled while waiting for remote index download")
1318 : }
1319 188 : )
1320 188 : }
1321 :
1322 182 : Ok(timeline_preloads)
1323 182 : }
1324 :
1325 0 : pub(crate) async fn apply_timeline_archival_config(
1326 0 : &self,
1327 0 : timeline_id: TimelineId,
1328 0 : state: TimelineArchivalState,
1329 0 : ) -> anyhow::Result<()> {
1330 0 : let timeline = self
1331 0 : .get_timeline(timeline_id, false)
1332 0 : .context("Cannot apply timeline archival config to inexistent timeline")?;
1333 :
1334 0 : let upload_needed = timeline
1335 0 : .remote_client
1336 0 : .schedule_index_upload_for_timeline_archival_state(state)?;
1337 :
1338 0 : if upload_needed {
1339 : const MAX_WAIT: Duration = Duration::from_secs(10);
1340 0 : let Ok(v) =
1341 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
1342 : else {
1343 0 : tracing::warn!("reached timeout for waiting on upload queue");
1344 0 : bail!("reached timeout for upload queue flush");
1345 : };
1346 0 : v?;
1347 0 : }
1348 0 : Ok(())
1349 0 : }
1350 :
1351 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
1352 4 : self.tenant_shard_id
1353 4 : }
1354 :
1355 : /// Get Timeline handle for given Neon timeline ID.
1356 : /// This function is idempotent. It doesn't change internal state in any way.
1357 222 : pub fn get_timeline(
1358 222 : &self,
1359 222 : timeline_id: TimelineId,
1360 222 : active_only: bool,
1361 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
1362 222 : let timelines_accessor = self.timelines.lock().unwrap();
1363 222 : let timeline = timelines_accessor
1364 222 : .get(&timeline_id)
1365 222 : .ok_or(GetTimelineError::NotFound {
1366 222 : tenant_id: self.tenant_shard_id,
1367 222 : timeline_id,
1368 222 : })?;
1369 :
1370 220 : if active_only && !timeline.is_active() {
1371 0 : Err(GetTimelineError::NotActive {
1372 0 : tenant_id: self.tenant_shard_id,
1373 0 : timeline_id,
1374 0 : state: timeline.current_state(),
1375 0 : })
1376 : } else {
1377 220 : Ok(Arc::clone(timeline))
1378 : }
1379 222 : }
1380 :
1381 : /// Lists timelines the tenant contains.
1382 : /// Up to tenant's implementation to omit certain timelines that ar not considered ready for use.
1383 8 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
1384 8 : self.timelines
1385 8 : .lock()
1386 8 : .unwrap()
1387 8 : .values()
1388 8 : .map(Arc::clone)
1389 8 : .collect()
1390 8 : }
1391 :
1392 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
1393 0 : self.timelines.lock().unwrap().keys().cloned().collect()
1394 0 : }
1395 :
1396 : /// This is used to create the initial 'main' timeline during bootstrapping,
1397 : /// or when importing a new base backup. The caller is expected to load an
1398 : /// initial image of the datadir to the new timeline after this.
1399 : ///
1400 : /// Until that happens, the on-disk state is invalid (disk_consistent_lsn=Lsn(0))
1401 : /// and the timeline will fail to load at a restart.
1402 : ///
1403 : /// For tests, use `DatadirModification::init_empty_test_timeline` + `commit` to setup the
1404 : /// minimum amount of keys required to get a writable timeline.
1405 : /// (Without it, `put` might fail due to `repartition` failing.)
1406 174 : pub(crate) async fn create_empty_timeline(
1407 174 : &self,
1408 174 : new_timeline_id: TimelineId,
1409 174 : initdb_lsn: Lsn,
1410 174 : pg_version: u32,
1411 174 : _ctx: &RequestContext,
1412 174 : ) -> anyhow::Result<UninitializedTimeline> {
1413 174 : anyhow::ensure!(
1414 174 : self.is_active(),
1415 0 : "Cannot create empty timelines on inactive tenant"
1416 : );
1417 :
1418 : // Protect against concurrent attempts to use this TimelineId
1419 174 : let create_guard = self.create_timeline_create_guard(new_timeline_id)?;
1420 :
1421 172 : let new_metadata = TimelineMetadata::new(
1422 172 : // Initialize disk_consistent LSN to 0, The caller must import some data to
1423 172 : // make it valid, before calling finish_creation()
1424 172 : Lsn(0),
1425 172 : None,
1426 172 : None,
1427 172 : Lsn(0),
1428 172 : initdb_lsn,
1429 172 : initdb_lsn,
1430 172 : pg_version,
1431 172 : );
1432 172 : self.prepare_new_timeline(
1433 172 : new_timeline_id,
1434 172 : &new_metadata,
1435 172 : create_guard,
1436 172 : initdb_lsn,
1437 172 : None,
1438 172 : None,
1439 172 : )
1440 0 : .await
1441 174 : }
1442 :
1443 : /// Helper for unit tests to create an empty timeline.
1444 : ///
1445 : /// The timeline is has state value `Active` but its background loops are not running.
1446 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
1447 : // Our current tests don't need the background loops.
1448 : #[cfg(test)]
1449 164 : pub async fn create_test_timeline(
1450 164 : &self,
1451 164 : new_timeline_id: TimelineId,
1452 164 : initdb_lsn: Lsn,
1453 164 : pg_version: u32,
1454 164 : ctx: &RequestContext,
1455 164 : ) -> anyhow::Result<Arc<Timeline>> {
1456 164 : let uninit_tl = self
1457 164 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1458 0 : .await?;
1459 164 : let tline = uninit_tl.raw_timeline().expect("we just created it");
1460 164 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
1461 :
1462 : // Setup minimum keys required for the timeline to be usable.
1463 164 : let mut modification = tline.begin_modification(initdb_lsn);
1464 164 : modification
1465 164 : .init_empty_test_timeline()
1466 164 : .context("init_empty_test_timeline")?;
1467 164 : modification
1468 164 : .commit(ctx)
1469 158 : .await
1470 164 : .context("commit init_empty_test_timeline modification")?;
1471 :
1472 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
1473 164 : tline.maybe_spawn_flush_loop();
1474 164 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
1475 :
1476 : // Make sure the freeze_and_flush reaches remote storage.
1477 164 : tline.remote_client.wait_completion().await.unwrap();
1478 :
1479 164 : let tl = uninit_tl.finish_creation()?;
1480 : // The non-test code would call tl.activate() here.
1481 164 : tl.set_state(TimelineState::Active);
1482 164 : Ok(tl)
1483 164 : }
1484 :
1485 : /// Helper for unit tests to create a timeline with some pre-loaded states.
1486 : #[cfg(test)]
1487 : #[allow(clippy::too_many_arguments)]
1488 24 : pub async fn create_test_timeline_with_layers(
1489 24 : &self,
1490 24 : new_timeline_id: TimelineId,
1491 24 : initdb_lsn: Lsn,
1492 24 : pg_version: u32,
1493 24 : ctx: &RequestContext,
1494 24 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
1495 24 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
1496 24 : end_lsn: Lsn,
1497 24 : ) -> anyhow::Result<Arc<Timeline>> {
1498 24 : let tline = self
1499 24 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1500 48 : .await?;
1501 24 : tline.force_advance_lsn(end_lsn);
1502 70 : for deltas in delta_layer_desc {
1503 46 : tline
1504 46 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
1505 138 : .await?;
1506 : }
1507 64 : for (lsn, images) in image_layer_desc {
1508 40 : tline
1509 40 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
1510 214 : .await?;
1511 : }
1512 24 : Ok(tline)
1513 24 : }
1514 :
1515 : /// Create a new timeline.
1516 : ///
1517 : /// Returns the new timeline ID and reference to its Timeline object.
1518 : ///
1519 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
1520 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
1521 : #[allow(clippy::too_many_arguments)]
1522 0 : pub(crate) async fn create_timeline(
1523 0 : self: &Arc<Tenant>,
1524 0 : new_timeline_id: TimelineId,
1525 0 : ancestor_timeline_id: Option<TimelineId>,
1526 0 : mut ancestor_start_lsn: Option<Lsn>,
1527 0 : pg_version: u32,
1528 0 : load_existing_initdb: Option<TimelineId>,
1529 0 : broker_client: storage_broker::BrokerClientChannel,
1530 0 : ctx: &RequestContext,
1531 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
1532 0 : if !self.is_active() {
1533 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
1534 0 : return Err(CreateTimelineError::ShuttingDown);
1535 : } else {
1536 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
1537 0 : "Cannot create timelines on inactive tenant"
1538 0 : )));
1539 : }
1540 0 : }
1541 :
1542 0 : let _gate = self
1543 0 : .gate
1544 0 : .enter()
1545 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
1546 :
1547 : // Get exclusive access to the timeline ID: this ensures that it does not already exist,
1548 : // and that no other creation attempts will be allowed in while we are working.
1549 0 : let create_guard = match self.create_timeline_create_guard(new_timeline_id) {
1550 0 : Ok(m) => m,
1551 : Err(TimelineExclusionError::AlreadyCreating) => {
1552 : // Creation is in progress, we cannot create it again, and we cannot
1553 : // check if this request matches the existing one, so caller must try
1554 : // again later.
1555 0 : return Err(CreateTimelineError::AlreadyCreating);
1556 : }
1557 0 : Err(TimelineExclusionError::Other(e)) => {
1558 0 : return Err(CreateTimelineError::Other(e));
1559 : }
1560 0 : Err(TimelineExclusionError::AlreadyExists(existing)) => {
1561 0 : debug!("timeline {new_timeline_id} already exists");
1562 :
1563 : // Idempotency: creating the same timeline twice is not an error, unless
1564 : // the second creation has different parameters.
1565 0 : if existing.get_ancestor_timeline_id() != ancestor_timeline_id
1566 0 : || existing.pg_version != pg_version
1567 0 : || (ancestor_start_lsn.is_some()
1568 0 : && ancestor_start_lsn != Some(existing.get_ancestor_lsn()))
1569 : {
1570 0 : return Err(CreateTimelineError::Conflict);
1571 0 : }
1572 0 :
1573 0 : // Wait for uploads to complete, so that when we return Ok, the timeline
1574 0 : // is known to be durable on remote storage. Just like we do at the end of
1575 0 : // this function, after we have created the timeline ourselves.
1576 0 : //
1577 0 : // We only really care that the initial version of `index_part.json` has
1578 0 : // been uploaded. That's enough to remember that the timeline
1579 0 : // exists. However, there is no function to wait specifically for that so
1580 0 : // we just wait for all in-progress uploads to finish.
1581 0 : existing
1582 0 : .remote_client
1583 0 : .wait_completion()
1584 0 : .await
1585 0 : .context("wait for timeline uploads to complete")?;
1586 :
1587 0 : return Ok(existing);
1588 : }
1589 : };
1590 :
1591 : pausable_failpoint!("timeline-creation-after-uninit");
1592 :
1593 0 : let loaded_timeline = match ancestor_timeline_id {
1594 0 : Some(ancestor_timeline_id) => {
1595 0 : let ancestor_timeline = self
1596 0 : .get_timeline(ancestor_timeline_id, false)
1597 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
1598 :
1599 : // instead of waiting around, just deny the request because ancestor is not yet
1600 : // ready for other purposes either.
1601 0 : if !ancestor_timeline.is_active() {
1602 0 : return Err(CreateTimelineError::AncestorNotActive);
1603 0 : }
1604 :
1605 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
1606 0 : *lsn = lsn.align();
1607 0 :
1608 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
1609 0 : if ancestor_ancestor_lsn > *lsn {
1610 : // can we safely just branch from the ancestor instead?
1611 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
1612 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
1613 0 : lsn,
1614 0 : ancestor_timeline_id,
1615 0 : ancestor_ancestor_lsn,
1616 0 : )));
1617 0 : }
1618 0 :
1619 0 : // Wait for the WAL to arrive and be processed on the parent branch up
1620 0 : // to the requested branch point. The repository code itself doesn't
1621 0 : // require it, but if we start to receive WAL on the new timeline,
1622 0 : // decoding the new WAL might need to look up previous pages, relation
1623 0 : // sizes etc. and that would get confused if the previous page versions
1624 0 : // are not in the repository yet.
1625 0 : ancestor_timeline
1626 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
1627 0 : .await
1628 0 : .map_err(|e| match e {
1629 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
1630 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
1631 : }
1632 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
1633 0 : })?;
1634 0 : }
1635 :
1636 0 : self.branch_timeline(
1637 0 : &ancestor_timeline,
1638 0 : new_timeline_id,
1639 0 : ancestor_start_lsn,
1640 0 : create_guard,
1641 0 : ctx,
1642 0 : )
1643 0 : .await?
1644 : }
1645 : None => {
1646 0 : self.bootstrap_timeline(
1647 0 : new_timeline_id,
1648 0 : pg_version,
1649 0 : load_existing_initdb,
1650 0 : create_guard,
1651 0 : ctx,
1652 0 : )
1653 0 : .await?
1654 : }
1655 : };
1656 :
1657 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
1658 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
1659 : // not send a success to the caller until it is. The same applies to handling retries,
1660 : // see the handling of [`TimelineExclusionError::AlreadyExists`] above.
1661 0 : let kind = ancestor_timeline_id
1662 0 : .map(|_| "branched")
1663 0 : .unwrap_or("bootstrapped");
1664 0 : loaded_timeline
1665 0 : .remote_client
1666 0 : .wait_completion()
1667 0 : .await
1668 0 : .with_context(|| format!("wait for {} timeline initial uploads to complete", kind))?;
1669 :
1670 0 : loaded_timeline.activate(self.clone(), broker_client, None, ctx);
1671 0 :
1672 0 : Ok(loaded_timeline)
1673 0 : }
1674 :
1675 0 : pub(crate) async fn delete_timeline(
1676 0 : self: Arc<Self>,
1677 0 : timeline_id: TimelineId,
1678 0 : ) -> Result<(), DeleteTimelineError> {
1679 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
1680 :
1681 0 : Ok(())
1682 0 : }
1683 :
1684 : /// perform one garbage collection iteration, removing old data files from disk.
1685 : /// this function is periodically called by gc task.
1686 : /// also it can be explicitly requested through page server api 'do_gc' command.
1687 : ///
1688 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
1689 : ///
1690 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
1691 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
1692 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
1693 : /// `pitr` specifies the same as a time difference from the current time. The effective
1694 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
1695 : /// requires more history to be retained.
1696 : //
1697 754 : pub(crate) async fn gc_iteration(
1698 754 : &self,
1699 754 : target_timeline_id: Option<TimelineId>,
1700 754 : horizon: u64,
1701 754 : pitr: Duration,
1702 754 : cancel: &CancellationToken,
1703 754 : ctx: &RequestContext,
1704 754 : ) -> Result<GcResult, GcError> {
1705 754 : // Don't start doing work during shutdown
1706 754 : if let TenantState::Stopping { .. } = self.current_state() {
1707 0 : return Ok(GcResult::default());
1708 754 : }
1709 754 :
1710 754 : // there is a global allowed_error for this
1711 754 : if !self.is_active() {
1712 0 : return Err(GcError::NotActive);
1713 754 : }
1714 754 :
1715 754 : {
1716 754 : let conf = self.tenant_conf.load();
1717 754 :
1718 754 : if !conf.location.may_delete_layers_hint() {
1719 0 : info!("Skipping GC in location state {:?}", conf.location);
1720 0 : return Ok(GcResult::default());
1721 754 : }
1722 : }
1723 :
1724 754 : let _guard = match self.gc_block.start().await {
1725 754 : Ok(guard) => guard,
1726 0 : Err(reasons) => {
1727 0 : info!("Skipping GC: {reasons}");
1728 0 : return Ok(GcResult::default());
1729 : }
1730 : };
1731 :
1732 754 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
1733 713 : .await
1734 754 : }
1735 :
1736 : /// Perform one compaction iteration.
1737 : /// This function is periodically called by compactor task.
1738 : /// Also it can be explicitly requested per timeline through page server
1739 : /// api's 'compact' command.
1740 : ///
1741 : /// Returns whether we have pending compaction task.
1742 0 : async fn compaction_iteration(
1743 0 : &self,
1744 0 : cancel: &CancellationToken,
1745 0 : ctx: &RequestContext,
1746 0 : ) -> Result<bool, timeline::CompactionError> {
1747 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
1748 0 : if !self.is_active() {
1749 0 : return Ok(false);
1750 0 : }
1751 0 :
1752 0 : {
1753 0 : let conf = self.tenant_conf.load();
1754 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
1755 0 : info!("Skipping compaction in location state {:?}", conf.location);
1756 0 : return Ok(false);
1757 0 : }
1758 0 : }
1759 0 :
1760 0 : // Scan through the hashmap and collect a list of all the timelines,
1761 0 : // while holding the lock. Then drop the lock and actually perform the
1762 0 : // compactions. We don't want to block everything else while the
1763 0 : // compaction runs.
1764 0 : let timelines_to_compact = {
1765 0 : let timelines = self.timelines.lock().unwrap();
1766 0 : let timelines_to_compact = timelines
1767 0 : .iter()
1768 0 : .filter_map(|(timeline_id, timeline)| {
1769 0 : if timeline.is_active() {
1770 0 : Some((*timeline_id, timeline.clone()))
1771 : } else {
1772 0 : None
1773 : }
1774 0 : })
1775 0 : .collect::<Vec<_>>();
1776 0 : drop(timelines);
1777 0 : timelines_to_compact
1778 0 : };
1779 0 :
1780 0 : // Before doing any I/O work, check our circuit breaker
1781 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
1782 0 : info!("Skipping compaction due to previous failures");
1783 0 : return Ok(false);
1784 0 : }
1785 0 :
1786 0 : let mut has_pending_task = false;
1787 :
1788 0 : for (timeline_id, timeline) in &timelines_to_compact {
1789 0 : has_pending_task |= timeline
1790 0 : .compact(cancel, EnumSet::empty(), ctx)
1791 0 : .instrument(info_span!("compact_timeline", %timeline_id))
1792 0 : .await
1793 0 : .inspect_err(|e| match e {
1794 0 : timeline::CompactionError::ShuttingDown => (),
1795 0 : timeline::CompactionError::Other(e) => {
1796 0 : self.compaction_circuit_breaker
1797 0 : .lock()
1798 0 : .unwrap()
1799 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
1800 0 : }
1801 0 : })?;
1802 : }
1803 :
1804 0 : self.compaction_circuit_breaker
1805 0 : .lock()
1806 0 : .unwrap()
1807 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
1808 0 :
1809 0 : Ok(has_pending_task)
1810 0 : }
1811 :
1812 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
1813 : // this happens during ingest: this background housekeeping is for freezing layers
1814 : // that are open but haven't been written to for some time.
1815 0 : async fn ingest_housekeeping(&self) {
1816 0 : // Scan through the hashmap and collect a list of all the timelines,
1817 0 : // while holding the lock. Then drop the lock and actually perform the
1818 0 : // compactions. We don't want to block everything else while the
1819 0 : // compaction runs.
1820 0 : let timelines = {
1821 0 : self.timelines
1822 0 : .lock()
1823 0 : .unwrap()
1824 0 : .values()
1825 0 : .filter_map(|timeline| {
1826 0 : if timeline.is_active() {
1827 0 : Some(timeline.clone())
1828 : } else {
1829 0 : None
1830 : }
1831 0 : })
1832 0 : .collect::<Vec<_>>()
1833 : };
1834 :
1835 0 : for timeline in &timelines {
1836 0 : timeline.maybe_freeze_ephemeral_layer().await;
1837 : }
1838 0 : }
1839 :
1840 2442 : pub fn current_state(&self) -> TenantState {
1841 2442 : self.state.borrow().clone()
1842 2442 : }
1843 :
1844 1682 : pub fn is_active(&self) -> bool {
1845 1682 : self.current_state() == TenantState::Active
1846 1682 : }
1847 :
1848 0 : pub fn generation(&self) -> Generation {
1849 0 : self.generation
1850 0 : }
1851 :
1852 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
1853 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
1854 0 : }
1855 :
1856 : /// Changes tenant status to active, unless shutdown was already requested.
1857 : ///
1858 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
1859 : /// to delay background jobs. Background jobs can be started right away when None is given.
1860 0 : fn activate(
1861 0 : self: &Arc<Self>,
1862 0 : broker_client: BrokerClientChannel,
1863 0 : background_jobs_can_start: Option<&completion::Barrier>,
1864 0 : ctx: &RequestContext,
1865 0 : ) {
1866 0 : span::debug_assert_current_span_has_tenant_id();
1867 0 :
1868 0 : let mut activating = false;
1869 0 : self.state.send_modify(|current_state| {
1870 0 : use pageserver_api::models::ActivatingFrom;
1871 0 : match &*current_state {
1872 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
1873 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
1874 : }
1875 0 : TenantState::Loading => {
1876 0 : *current_state = TenantState::Activating(ActivatingFrom::Loading);
1877 0 : }
1878 0 : TenantState::Attaching => {
1879 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
1880 0 : }
1881 : }
1882 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
1883 0 : activating = true;
1884 0 : // Continue outside the closure. We need to grab timelines.lock()
1885 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
1886 0 : });
1887 0 :
1888 0 : if activating {
1889 0 : let timelines_accessor = self.timelines.lock().unwrap();
1890 0 : let timelines_to_activate = timelines_accessor
1891 0 : .values()
1892 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
1893 0 :
1894 0 : // Before activation, populate each Timeline's GcInfo with information about its children
1895 0 : self.initialize_gc_info(&timelines_accessor);
1896 0 :
1897 0 : // Spawn gc and compaction loops. The loops will shut themselves
1898 0 : // down when they notice that the tenant is inactive.
1899 0 : tasks::start_background_loops(self, background_jobs_can_start);
1900 0 :
1901 0 : let mut activated_timelines = 0;
1902 :
1903 0 : for timeline in timelines_to_activate {
1904 0 : timeline.activate(
1905 0 : self.clone(),
1906 0 : broker_client.clone(),
1907 0 : background_jobs_can_start,
1908 0 : ctx,
1909 0 : );
1910 0 : activated_timelines += 1;
1911 0 : }
1912 :
1913 0 : self.state.send_modify(move |current_state| {
1914 0 : assert!(
1915 0 : matches!(current_state, TenantState::Activating(_)),
1916 0 : "set_stopping and set_broken wait for us to leave Activating state",
1917 : );
1918 0 : *current_state = TenantState::Active;
1919 0 :
1920 0 : let elapsed = self.constructed_at.elapsed();
1921 0 : let total_timelines = timelines_accessor.len();
1922 0 :
1923 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
1924 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
1925 0 : info!(
1926 0 : since_creation_millis = elapsed.as_millis(),
1927 0 : tenant_id = %self.tenant_shard_id.tenant_id,
1928 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1929 0 : activated_timelines,
1930 0 : total_timelines,
1931 0 : post_state = <&'static str>::from(&*current_state),
1932 0 : "activation attempt finished"
1933 : );
1934 :
1935 0 : TENANT.activation.observe(elapsed.as_secs_f64());
1936 0 : });
1937 0 : }
1938 0 : }
1939 :
1940 : /// Shutdown the tenant and join all of the spawned tasks.
1941 : ///
1942 : /// The method caters for all use-cases:
1943 : /// - pageserver shutdown (freeze_and_flush == true)
1944 : /// - detach + ignore (freeze_and_flush == false)
1945 : ///
1946 : /// This will attempt to shutdown even if tenant is broken.
1947 : ///
1948 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
1949 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
1950 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
1951 : /// the ongoing shutdown.
1952 6 : async fn shutdown(
1953 6 : &self,
1954 6 : shutdown_progress: completion::Barrier,
1955 6 : shutdown_mode: timeline::ShutdownMode,
1956 6 : ) -> Result<(), completion::Barrier> {
1957 6 : span::debug_assert_current_span_has_tenant_id();
1958 :
1959 : // Set tenant (and its timlines) to Stoppping state.
1960 : //
1961 : // Since we can only transition into Stopping state after activation is complete,
1962 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
1963 : //
1964 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
1965 : // 1. Lock out any new requests to the tenants.
1966 : // 2. Signal cancellation to WAL receivers (we wait on it below).
1967 : // 3. Signal cancellation for other tenant background loops.
1968 : // 4. ???
1969 : //
1970 : // The waiting for the cancellation is not done uniformly.
1971 : // We certainly wait for WAL receivers to shut down.
1972 : // That is necessary so that no new data comes in before the freeze_and_flush.
1973 : // But the tenant background loops are joined-on in our caller.
1974 : // It's mesed up.
1975 : // we just ignore the failure to stop
1976 :
1977 : // If we're still attaching, fire the cancellation token early to drop out: this
1978 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
1979 : // is very slow.
1980 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
1981 0 : self.cancel.cancel();
1982 0 :
1983 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
1984 0 : // are children of ours, so their flush loops will have shut down already
1985 0 : timeline::ShutdownMode::Hard
1986 : } else {
1987 6 : shutdown_mode
1988 : };
1989 :
1990 6 : match self.set_stopping(shutdown_progress, false, false).await {
1991 6 : Ok(()) => {}
1992 0 : Err(SetStoppingError::Broken) => {
1993 0 : // assume that this is acceptable
1994 0 : }
1995 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
1996 0 : // give caller the option to wait for this this shutdown
1997 0 : info!("Tenant::shutdown: AlreadyStopping");
1998 0 : return Err(other);
1999 : }
2000 : };
2001 :
2002 6 : let mut js = tokio::task::JoinSet::new();
2003 6 : {
2004 6 : let timelines = self.timelines.lock().unwrap();
2005 6 : timelines.values().for_each(|timeline| {
2006 6 : let timeline = Arc::clone(timeline);
2007 6 : let timeline_id = timeline.timeline_id;
2008 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
2009 12 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
2010 6 : })
2011 6 : };
2012 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
2013 6 : tracing::info!("Waiting for timelines...");
2014 12 : while let Some(res) = js.join_next().await {
2015 0 : match res {
2016 6 : Ok(()) => {}
2017 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
2018 0 : Err(je) if je.is_panic() => { /* logged already */ }
2019 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
2020 : }
2021 : }
2022 :
2023 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
2024 : // them to continue to do work during their shutdown methods, e.g. flushing data.
2025 6 : tracing::debug!("Cancelling CancellationToken");
2026 6 : self.cancel.cancel();
2027 6 :
2028 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
2029 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
2030 6 : //
2031 6 : // this will additionally shutdown and await all timeline tasks.
2032 6 : tracing::debug!("Waiting for tasks...");
2033 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
2034 :
2035 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
2036 6 : walredo_mgr.shutdown().await;
2037 0 : }
2038 :
2039 : // Wait for any in-flight operations to complete
2040 6 : self.gate.close().await;
2041 :
2042 6 : remove_tenant_metrics(&self.tenant_shard_id);
2043 6 :
2044 6 : Ok(())
2045 6 : }
2046 :
2047 : /// Change tenant status to Stopping, to mark that it is being shut down.
2048 : ///
2049 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2050 : ///
2051 : /// This function is not cancel-safe!
2052 : ///
2053 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
2054 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
2055 6 : async fn set_stopping(
2056 6 : &self,
2057 6 : progress: completion::Barrier,
2058 6 : allow_transition_from_loading: bool,
2059 6 : allow_transition_from_attaching: bool,
2060 6 : ) -> Result<(), SetStoppingError> {
2061 6 : let mut rx = self.state.subscribe();
2062 6 :
2063 6 : // cannot stop before we're done activating, so wait out until we're done activating
2064 6 : rx.wait_for(|state| match state {
2065 0 : TenantState::Attaching if allow_transition_from_attaching => true,
2066 : TenantState::Activating(_) | TenantState::Attaching => {
2067 0 : info!(
2068 0 : "waiting for {} to turn Active|Broken|Stopping",
2069 0 : <&'static str>::from(state)
2070 : );
2071 0 : false
2072 : }
2073 0 : TenantState::Loading => allow_transition_from_loading,
2074 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2075 6 : })
2076 0 : .await
2077 6 : .expect("cannot drop self.state while on a &self method");
2078 6 :
2079 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
2080 6 : let mut err = None;
2081 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
2082 : TenantState::Activating(_) => {
2083 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
2084 : }
2085 : TenantState::Attaching => {
2086 0 : if !allow_transition_from_attaching {
2087 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
2088 0 : };
2089 0 : *current_state = TenantState::Stopping { progress };
2090 0 : true
2091 : }
2092 : TenantState::Loading => {
2093 0 : if !allow_transition_from_loading {
2094 0 : unreachable!("3we ensured above that we're done with activation, and, there is no re-activation")
2095 0 : };
2096 0 : *current_state = TenantState::Stopping { progress };
2097 0 : true
2098 : }
2099 : TenantState::Active => {
2100 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
2101 : // are created after the transition to Stopping. That's harmless, as the Timelines
2102 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
2103 6 : *current_state = TenantState::Stopping { progress };
2104 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
2105 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
2106 6 : true
2107 : }
2108 0 : TenantState::Broken { reason, .. } => {
2109 0 : info!(
2110 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
2111 : );
2112 0 : err = Some(SetStoppingError::Broken);
2113 0 : false
2114 : }
2115 0 : TenantState::Stopping { progress } => {
2116 0 : info!("Tenant is already in Stopping state");
2117 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
2118 0 : false
2119 : }
2120 6 : });
2121 6 : match (stopping, err) {
2122 6 : (true, None) => {} // continue
2123 0 : (false, Some(err)) => return Err(err),
2124 0 : (true, Some(_)) => unreachable!(
2125 0 : "send_if_modified closure must error out if not transitioning to Stopping"
2126 0 : ),
2127 0 : (false, None) => unreachable!(
2128 0 : "send_if_modified closure must return true if transitioning to Stopping"
2129 0 : ),
2130 : }
2131 :
2132 6 : let timelines_accessor = self.timelines.lock().unwrap();
2133 6 : let not_broken_timelines = timelines_accessor
2134 6 : .values()
2135 6 : .filter(|timeline| !timeline.is_broken());
2136 12 : for timeline in not_broken_timelines {
2137 6 : timeline.set_state(TimelineState::Stopping);
2138 6 : }
2139 6 : Ok(())
2140 6 : }
2141 :
2142 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2143 : /// `remove_tenant_from_memory`
2144 : ///
2145 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2146 : ///
2147 : /// In tests, we also use this to set tenants to Broken state on purpose.
2148 0 : pub(crate) async fn set_broken(&self, reason: String) {
2149 0 : let mut rx = self.state.subscribe();
2150 0 :
2151 0 : // The load & attach routines own the tenant state until it has reached `Active`.
2152 0 : // So, wait until it's done.
2153 0 : rx.wait_for(|state| match state {
2154 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2155 0 : info!(
2156 0 : "waiting for {} to turn Active|Broken|Stopping",
2157 0 : <&'static str>::from(state)
2158 : );
2159 0 : false
2160 : }
2161 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2162 0 : })
2163 0 : .await
2164 0 : .expect("cannot drop self.state while on a &self method");
2165 0 :
2166 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
2167 0 : self.set_broken_no_wait(reason)
2168 0 : }
2169 :
2170 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
2171 0 : let reason = reason.to_string();
2172 0 : self.state.send_modify(|current_state| {
2173 0 : match *current_state {
2174 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2175 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
2176 : }
2177 : TenantState::Active => {
2178 0 : if cfg!(feature = "testing") {
2179 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
2180 0 : *current_state = TenantState::broken_from_reason(reason);
2181 : } else {
2182 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
2183 : }
2184 : }
2185 : TenantState::Broken { .. } => {
2186 0 : warn!("Tenant is already in Broken state");
2187 : }
2188 : // This is the only "expected" path, any other path is a bug.
2189 : TenantState::Stopping { .. } => {
2190 0 : warn!(
2191 0 : "Marking Stopping tenant as Broken state, reason: {}",
2192 : reason
2193 : );
2194 0 : *current_state = TenantState::broken_from_reason(reason);
2195 : }
2196 : }
2197 0 : });
2198 0 : }
2199 :
2200 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
2201 0 : self.state.subscribe()
2202 0 : }
2203 :
2204 : /// The activate_now semaphore is initialized with zero units. As soon as
2205 : /// we add a unit, waiters will be able to acquire a unit and proceed.
2206 0 : pub(crate) fn activate_now(&self) {
2207 0 : self.activate_now_sem.add_permits(1);
2208 0 : }
2209 :
2210 0 : pub(crate) async fn wait_to_become_active(
2211 0 : &self,
2212 0 : timeout: Duration,
2213 0 : ) -> Result<(), GetActiveTenantError> {
2214 0 : let mut receiver = self.state.subscribe();
2215 0 : loop {
2216 0 : let current_state = receiver.borrow_and_update().clone();
2217 0 : match current_state {
2218 : TenantState::Loading | TenantState::Attaching | TenantState::Activating(_) => {
2219 : // in these states, there's a chance that we can reach ::Active
2220 0 : self.activate_now();
2221 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
2222 0 : Ok(r) => {
2223 0 : r.map_err(
2224 0 : |_e: tokio::sync::watch::error::RecvError|
2225 : // Tenant existed but was dropped: report it as non-existent
2226 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
2227 0 : )?
2228 : }
2229 : Err(TimeoutCancellableError::Cancelled) => {
2230 0 : return Err(GetActiveTenantError::Cancelled);
2231 : }
2232 : Err(TimeoutCancellableError::Timeout) => {
2233 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
2234 0 : latest_state: Some(self.current_state()),
2235 0 : wait_time: timeout,
2236 0 : });
2237 : }
2238 : }
2239 : }
2240 : TenantState::Active { .. } => {
2241 0 : return Ok(());
2242 : }
2243 0 : TenantState::Broken { reason, .. } => {
2244 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
2245 0 : // it's logically a 500 to external API users (broken is always a bug).
2246 0 : return Err(GetActiveTenantError::Broken(reason));
2247 : }
2248 : TenantState::Stopping { .. } => {
2249 : // There's no chance the tenant can transition back into ::Active
2250 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
2251 : }
2252 : }
2253 : }
2254 0 : }
2255 :
2256 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
2257 0 : self.tenant_conf.load().location.attach_mode
2258 0 : }
2259 :
2260 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
2261 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
2262 : /// rare external API calls, like a reconciliation at startup.
2263 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
2264 0 : let conf = self.tenant_conf.load();
2265 :
2266 0 : let location_config_mode = match conf.location.attach_mode {
2267 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
2268 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
2269 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
2270 : };
2271 :
2272 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
2273 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
2274 0 :
2275 0 : models::LocationConfig {
2276 0 : mode: location_config_mode,
2277 0 : generation: self.generation.into(),
2278 0 : secondary_conf: None,
2279 0 : shard_number: self.shard_identity.number.0,
2280 0 : shard_count: self.shard_identity.count.literal(),
2281 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
2282 0 : tenant_conf: tenant_config,
2283 0 : }
2284 0 : }
2285 :
2286 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
2287 0 : &self.tenant_shard_id
2288 0 : }
2289 :
2290 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
2291 0 : self.shard_identity.stripe_size
2292 0 : }
2293 :
2294 0 : pub(crate) fn get_generation(&self) -> Generation {
2295 0 : self.generation
2296 0 : }
2297 :
2298 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
2299 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
2300 : /// resetting this tenant to a valid state if we fail.
2301 0 : pub(crate) async fn split_prepare(
2302 0 : &self,
2303 0 : child_shards: &Vec<TenantShardId>,
2304 0 : ) -> anyhow::Result<()> {
2305 0 : let timelines = self.timelines.lock().unwrap().clone();
2306 0 : for timeline in timelines.values() {
2307 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
2308 : // to ensure that they do not start a split if currently in the process of doing these.
2309 :
2310 : // Upload an index from the parent: this is partly to provide freshness for the
2311 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
2312 : // always be a parent shard index in the same generation as we wrote the child shard index.
2313 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index");
2314 0 : timeline
2315 0 : .remote_client
2316 0 : .schedule_index_upload_for_file_changes()?;
2317 0 : timeline.remote_client.wait_completion().await?;
2318 :
2319 : // Shut down the timeline's remote client: this means that the indices we write
2320 : // for child shards will not be invalidated by the parent shard deleting layers.
2321 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Shutting down remote storage client");
2322 0 : timeline.remote_client.shutdown().await;
2323 :
2324 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
2325 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
2326 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
2327 : // we use here really is the remotely persistent one).
2328 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Downloading index_part from parent");
2329 0 : let result = timeline.remote_client
2330 0 : .download_index_file(&self.cancel)
2331 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))
2332 0 : .await?;
2333 0 : let index_part = match result {
2334 : MaybeDeletedIndexPart::Deleted(_) => {
2335 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
2336 : }
2337 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2338 : };
2339 :
2340 0 : for child_shard in child_shards {
2341 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index_part for child {}", child_shard.to_index());
2342 0 : upload_index_part(
2343 0 : &self.remote_storage,
2344 0 : child_shard,
2345 0 : &timeline.timeline_id,
2346 0 : self.generation,
2347 0 : &index_part,
2348 0 : &self.cancel,
2349 0 : )
2350 0 : .await?;
2351 : }
2352 : }
2353 :
2354 0 : Ok(())
2355 0 : }
2356 :
2357 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
2358 0 : let mut result = TopTenantShardItem {
2359 0 : id: self.tenant_shard_id,
2360 0 : resident_size: 0,
2361 0 : physical_size: 0,
2362 0 : max_logical_size: 0,
2363 0 : };
2364 :
2365 0 : for timeline in self.timelines.lock().unwrap().values() {
2366 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
2367 0 :
2368 0 : result.physical_size += timeline
2369 0 : .remote_client
2370 0 : .metrics
2371 0 : .remote_physical_size_gauge
2372 0 : .get();
2373 0 : result.max_logical_size = std::cmp::max(
2374 0 : result.max_logical_size,
2375 0 : timeline.metrics.current_logical_size_gauge.get(),
2376 0 : );
2377 0 : }
2378 :
2379 0 : result
2380 0 : }
2381 : }
2382 :
2383 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
2384 : /// perform a topological sort, so that the parent of each timeline comes
2385 : /// before the children.
2386 : /// E extracts the ancestor from T
2387 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
2388 182 : fn tree_sort_timelines<T, E>(
2389 182 : timelines: HashMap<TimelineId, T>,
2390 182 : extractor: E,
2391 182 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
2392 182 : where
2393 182 : E: Fn(&T) -> Option<TimelineId>,
2394 182 : {
2395 182 : let mut result = Vec::with_capacity(timelines.len());
2396 182 :
2397 182 : let mut now = Vec::with_capacity(timelines.len());
2398 182 : // (ancestor, children)
2399 182 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
2400 182 : HashMap::with_capacity(timelines.len());
2401 :
2402 188 : for (timeline_id, value) in timelines {
2403 6 : if let Some(ancestor_id) = extractor(&value) {
2404 2 : let children = later.entry(ancestor_id).or_default();
2405 2 : children.push((timeline_id, value));
2406 4 : } else {
2407 4 : now.push((timeline_id, value));
2408 4 : }
2409 : }
2410 :
2411 188 : while let Some((timeline_id, metadata)) = now.pop() {
2412 6 : result.push((timeline_id, metadata));
2413 : // All children of this can be loaded now
2414 6 : if let Some(mut children) = later.remove(&timeline_id) {
2415 2 : now.append(&mut children);
2416 4 : }
2417 : }
2418 :
2419 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
2420 182 : if !later.is_empty() {
2421 0 : for (missing_id, orphan_ids) in later {
2422 0 : for (orphan_id, _) in orphan_ids {
2423 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
2424 : }
2425 : }
2426 0 : bail!("could not load tenant because some timelines are missing ancestors");
2427 182 : }
2428 182 :
2429 182 : Ok(result)
2430 182 : }
2431 :
2432 : impl Tenant {
2433 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
2434 0 : self.tenant_conf.load().tenant_conf.clone()
2435 0 : }
2436 :
2437 0 : pub fn effective_config(&self) -> TenantConf {
2438 0 : self.tenant_specific_overrides()
2439 0 : .merge(self.conf.default_tenant_conf.clone())
2440 0 : }
2441 :
2442 0 : pub fn get_checkpoint_distance(&self) -> u64 {
2443 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2444 0 : tenant_conf
2445 0 : .checkpoint_distance
2446 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2447 0 : }
2448 :
2449 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
2450 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2451 0 : tenant_conf
2452 0 : .checkpoint_timeout
2453 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2454 0 : }
2455 :
2456 0 : pub fn get_compaction_target_size(&self) -> u64 {
2457 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2458 0 : tenant_conf
2459 0 : .compaction_target_size
2460 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2461 0 : }
2462 :
2463 0 : pub fn get_compaction_period(&self) -> Duration {
2464 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2465 0 : tenant_conf
2466 0 : .compaction_period
2467 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2468 0 : }
2469 :
2470 0 : pub fn get_compaction_threshold(&self) -> usize {
2471 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2472 0 : tenant_conf
2473 0 : .compaction_threshold
2474 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2475 0 : }
2476 :
2477 0 : pub fn get_gc_horizon(&self) -> u64 {
2478 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2479 0 : tenant_conf
2480 0 : .gc_horizon
2481 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
2482 0 : }
2483 :
2484 0 : pub fn get_gc_period(&self) -> Duration {
2485 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2486 0 : tenant_conf
2487 0 : .gc_period
2488 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
2489 0 : }
2490 :
2491 0 : pub fn get_image_creation_threshold(&self) -> usize {
2492 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2493 0 : tenant_conf
2494 0 : .image_creation_threshold
2495 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2496 0 : }
2497 :
2498 0 : pub fn get_pitr_interval(&self) -> Duration {
2499 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2500 0 : tenant_conf
2501 0 : .pitr_interval
2502 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2503 0 : }
2504 :
2505 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
2506 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2507 0 : tenant_conf
2508 0 : .min_resident_size_override
2509 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
2510 0 : }
2511 :
2512 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
2513 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2514 0 : let heatmap_period = tenant_conf
2515 0 : .heatmap_period
2516 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
2517 0 : if heatmap_period.is_zero() {
2518 0 : None
2519 : } else {
2520 0 : Some(heatmap_period)
2521 : }
2522 0 : }
2523 :
2524 0 : pub fn get_lsn_lease_length(&self) -> Duration {
2525 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2526 0 : tenant_conf
2527 0 : .lsn_lease_length
2528 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2529 0 : }
2530 :
2531 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
2532 0 : // Use read-copy-update in order to avoid overwriting the location config
2533 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
2534 0 : // this race is not possible if both request types come from the storage
2535 0 : // controller (as they should!) because an exclusive op lock is required
2536 0 : // on the storage controller side.
2537 0 : self.tenant_conf.rcu(|inner| {
2538 0 : Arc::new(AttachedTenantConf {
2539 0 : tenant_conf: new_tenant_conf.clone(),
2540 0 : location: inner.location,
2541 0 : })
2542 0 : });
2543 0 :
2544 0 : self.tenant_conf_updated(&new_tenant_conf);
2545 0 : // Don't hold self.timelines.lock() during the notifies.
2546 0 : // There's no risk of deadlock right now, but there could be if we consolidate
2547 0 : // mutexes in struct Timeline in the future.
2548 0 : let timelines = self.list_timelines();
2549 0 : for timeline in timelines {
2550 0 : timeline.tenant_conf_updated(&new_tenant_conf);
2551 0 : }
2552 0 : }
2553 :
2554 8 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
2555 8 : let new_tenant_conf = new_conf.tenant_conf.clone();
2556 8 :
2557 8 : self.tenant_conf.store(Arc::new(new_conf));
2558 8 :
2559 8 : self.tenant_conf_updated(&new_tenant_conf);
2560 8 : // Don't hold self.timelines.lock() during the notifies.
2561 8 : // There's no risk of deadlock right now, but there could be if we consolidate
2562 8 : // mutexes in struct Timeline in the future.
2563 8 : let timelines = self.list_timelines();
2564 16 : for timeline in timelines {
2565 8 : timeline.tenant_conf_updated(&new_tenant_conf);
2566 8 : }
2567 8 : }
2568 :
2569 190 : fn get_timeline_get_throttle_config(
2570 190 : psconf: &'static PageServerConf,
2571 190 : overrides: &TenantConfOpt,
2572 190 : ) -> throttle::Config {
2573 190 : overrides
2574 190 : .timeline_get_throttle
2575 190 : .clone()
2576 190 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
2577 190 : }
2578 :
2579 8 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2580 8 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
2581 8 : self.timeline_get_throttle.reconfigure(conf)
2582 8 : }
2583 :
2584 : /// Helper function to create a new Timeline struct.
2585 : ///
2586 : /// The returned Timeline is in Loading state. The caller is responsible for
2587 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
2588 : /// map.
2589 : ///
2590 : /// `validate_ancestor == false` is used when a timeline is created for deletion
2591 : /// and we might not have the ancestor present anymore which is fine for to be
2592 : /// deleted timelines.
2593 406 : fn create_timeline_struct(
2594 406 : &self,
2595 406 : new_timeline_id: TimelineId,
2596 406 : new_metadata: &TimelineMetadata,
2597 406 : ancestor: Option<Arc<Timeline>>,
2598 406 : resources: TimelineResources,
2599 406 : cause: CreateTimelineCause,
2600 406 : last_aux_file_policy: Option<AuxFilePolicy>,
2601 406 : ) -> anyhow::Result<Arc<Timeline>> {
2602 406 : let state = match cause {
2603 : CreateTimelineCause::Load => {
2604 406 : let ancestor_id = new_metadata.ancestor_timeline();
2605 406 : anyhow::ensure!(
2606 406 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
2607 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
2608 : );
2609 406 : TimelineState::Loading
2610 : }
2611 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
2612 : };
2613 :
2614 406 : let pg_version = new_metadata.pg_version();
2615 406 :
2616 406 : let timeline = Timeline::new(
2617 406 : self.conf,
2618 406 : Arc::clone(&self.tenant_conf),
2619 406 : new_metadata,
2620 406 : ancestor,
2621 406 : new_timeline_id,
2622 406 : self.tenant_shard_id,
2623 406 : self.generation,
2624 406 : self.shard_identity,
2625 406 : self.walredo_mgr.clone(),
2626 406 : resources,
2627 406 : pg_version,
2628 406 : state,
2629 406 : last_aux_file_policy,
2630 406 : self.cancel.child_token(),
2631 406 : );
2632 406 :
2633 406 : Ok(timeline)
2634 406 : }
2635 :
2636 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
2637 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
2638 : #[allow(clippy::too_many_arguments)]
2639 182 : fn new(
2640 182 : state: TenantState,
2641 182 : conf: &'static PageServerConf,
2642 182 : attached_conf: AttachedTenantConf,
2643 182 : shard_identity: ShardIdentity,
2644 182 : walredo_mgr: Option<Arc<WalRedoManager>>,
2645 182 : tenant_shard_id: TenantShardId,
2646 182 : remote_storage: GenericRemoteStorage,
2647 182 : deletion_queue_client: DeletionQueueClient,
2648 182 : l0_flush_global_state: L0FlushGlobalState,
2649 182 : ) -> Tenant {
2650 182 : debug_assert!(
2651 182 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
2652 : );
2653 :
2654 182 : let (state, mut rx) = watch::channel(state);
2655 182 :
2656 182 : tokio::spawn(async move {
2657 182 : // reflect tenant state in metrics:
2658 182 : // - global per tenant state: TENANT_STATE_METRIC
2659 182 : // - "set" of broken tenants: BROKEN_TENANTS_SET
2660 182 : //
2661 182 : // set of broken tenants should not have zero counts so that it remains accessible for
2662 182 : // alerting.
2663 182 :
2664 182 : let tid = tenant_shard_id.to_string();
2665 182 : let shard_id = tenant_shard_id.shard_slug().to_string();
2666 182 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
2667 182 :
2668 362 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
2669 362 : ([state.into()], matches!(state, TenantState::Broken { .. }))
2670 362 : }
2671 182 :
2672 182 : let mut tuple = inspect_state(&rx.borrow_and_update());
2673 182 :
2674 182 : let is_broken = tuple.1;
2675 182 : let mut counted_broken = if is_broken {
2676 : // add the id to the set right away, there should not be any updates on the channel
2677 : // after before tenant is removed, if ever
2678 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2679 0 : true
2680 : } else {
2681 182 : false
2682 : };
2683 :
2684 362 : loop {
2685 362 : let labels = &tuple.0;
2686 362 : let current = TENANT_STATE_METRIC.with_label_values(labels);
2687 362 : current.inc();
2688 362 :
2689 362 : if rx.changed().await.is_err() {
2690 : // tenant has been dropped
2691 16 : current.dec();
2692 16 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
2693 16 : break;
2694 180 : }
2695 180 :
2696 180 : current.dec();
2697 180 : tuple = inspect_state(&rx.borrow_and_update());
2698 180 :
2699 180 : let is_broken = tuple.1;
2700 180 : if is_broken && !counted_broken {
2701 0 : counted_broken = true;
2702 0 : // insert the tenant_id (back) into the set while avoiding needless counter
2703 0 : // access
2704 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2705 180 : }
2706 : }
2707 182 : });
2708 182 :
2709 182 : Tenant {
2710 182 : tenant_shard_id,
2711 182 : shard_identity,
2712 182 : generation: attached_conf.location.generation,
2713 182 : conf,
2714 182 : // using now here is good enough approximation to catch tenants with really long
2715 182 : // activation times.
2716 182 : constructed_at: Instant::now(),
2717 182 : timelines: Mutex::new(HashMap::new()),
2718 182 : timelines_creating: Mutex::new(HashSet::new()),
2719 182 : gc_cs: tokio::sync::Mutex::new(()),
2720 182 : walredo_mgr,
2721 182 : remote_storage,
2722 182 : deletion_queue_client,
2723 182 : state,
2724 182 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
2725 182 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
2726 182 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
2727 182 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
2728 182 : format!("compaction-{tenant_shard_id}"),
2729 182 : 5,
2730 182 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
2731 182 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
2732 182 : // use an extremely long backoff.
2733 182 : Some(Duration::from_secs(3600 * 24)),
2734 182 : )),
2735 182 : activate_now_sem: tokio::sync::Semaphore::new(0),
2736 182 : cancel: CancellationToken::default(),
2737 182 : gate: Gate::default(),
2738 182 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
2739 182 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
2740 182 : &crate::metrics::tenant_throttling::TIMELINE_GET,
2741 182 : )),
2742 182 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
2743 182 : ongoing_timeline_detach: std::sync::Mutex::default(),
2744 182 : gc_block: Default::default(),
2745 182 : l0_flush_global_state,
2746 182 : }
2747 182 : }
2748 :
2749 : /// Locate and load config
2750 0 : pub(super) fn load_tenant_config(
2751 0 : conf: &'static PageServerConf,
2752 0 : tenant_shard_id: &TenantShardId,
2753 0 : ) -> Result<LocationConf, LoadConfigError> {
2754 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2755 0 :
2756 0 : info!("loading tenant configuration from {config_path}");
2757 :
2758 : // load and parse file
2759 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
2760 0 : match e.kind() {
2761 : std::io::ErrorKind::NotFound => {
2762 : // The config should almost always exist for a tenant directory:
2763 : // - When attaching a tenant, the config is the first thing we write
2764 : // - When detaching a tenant, we atomically move the directory to a tmp location
2765 : // before deleting contents.
2766 : //
2767 : // The very rare edge case that can result in a missing config is if we crash during attach
2768 : // between creating directory and writing config. Callers should handle that as if the
2769 : // directory didn't exist.
2770 :
2771 0 : LoadConfigError::NotFound(config_path)
2772 : }
2773 : _ => {
2774 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
2775 : // that we cannot cleanly recover
2776 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
2777 : }
2778 : }
2779 0 : })?;
2780 :
2781 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
2782 0 : }
2783 :
2784 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2785 : pub(super) async fn persist_tenant_config(
2786 : conf: &'static PageServerConf,
2787 : tenant_shard_id: &TenantShardId,
2788 : location_conf: &LocationConf,
2789 : ) -> std::io::Result<()> {
2790 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2791 :
2792 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
2793 : }
2794 :
2795 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2796 : pub(super) async fn persist_tenant_config_at(
2797 : tenant_shard_id: &TenantShardId,
2798 : config_path: &Utf8Path,
2799 : location_conf: &LocationConf,
2800 : ) -> std::io::Result<()> {
2801 : debug!("persisting tenantconf to {config_path}");
2802 :
2803 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
2804 : # It is read in case of pageserver restart.
2805 : "#
2806 : .to_string();
2807 :
2808 0 : fail::fail_point!("tenant-config-before-write", |_| {
2809 0 : Err(std::io::Error::new(
2810 0 : std::io::ErrorKind::Other,
2811 0 : "tenant-config-before-write",
2812 0 : ))
2813 0 : });
2814 :
2815 : // Convert the config to a toml file.
2816 : conf_content +=
2817 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
2818 :
2819 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
2820 :
2821 : let conf_content = conf_content.into_bytes();
2822 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
2823 : }
2824 :
2825 : //
2826 : // How garbage collection works:
2827 : //
2828 : // +--bar------------->
2829 : // /
2830 : // +----+-----foo---------------->
2831 : // /
2832 : // ----main--+-------------------------->
2833 : // \
2834 : // +-----baz-------->
2835 : //
2836 : //
2837 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
2838 : // `gc_infos` are being refreshed
2839 : // 2. Scan collected timelines, and on each timeline, make note of the
2840 : // all the points where other timelines have been branched off.
2841 : // We will refrain from removing page versions at those LSNs.
2842 : // 3. For each timeline, scan all layer files on the timeline.
2843 : // Remove all files for which a newer file exists and which
2844 : // don't cover any branch point LSNs.
2845 : //
2846 : // TODO:
2847 : // - if a relation has a non-incremental persistent layer on a child branch, then we
2848 : // don't need to keep that in the parent anymore. But currently
2849 : // we do.
2850 754 : async fn gc_iteration_internal(
2851 754 : &self,
2852 754 : target_timeline_id: Option<TimelineId>,
2853 754 : horizon: u64,
2854 754 : pitr: Duration,
2855 754 : cancel: &CancellationToken,
2856 754 : ctx: &RequestContext,
2857 754 : ) -> Result<GcResult, GcError> {
2858 754 : let mut totals: GcResult = Default::default();
2859 754 : let now = Instant::now();
2860 :
2861 754 : let gc_timelines = self
2862 754 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2863 713 : .await?;
2864 :
2865 754 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
2866 :
2867 : // If there is nothing to GC, we don't want any messages in the INFO log.
2868 754 : if !gc_timelines.is_empty() {
2869 754 : info!("{} timelines need GC", gc_timelines.len());
2870 : } else {
2871 0 : debug!("{} timelines need GC", gc_timelines.len());
2872 : }
2873 :
2874 : // Perform GC for each timeline.
2875 : //
2876 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
2877 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
2878 : // with branch creation.
2879 : //
2880 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
2881 : // creation task can run concurrently with timeline's GC iteration.
2882 1508 : for timeline in gc_timelines {
2883 754 : if cancel.is_cancelled() {
2884 : // We were requested to shut down. Stop and return with the progress we
2885 : // made.
2886 0 : break;
2887 754 : }
2888 754 : let result = match timeline.gc().await {
2889 : Err(GcError::TimelineCancelled) => {
2890 0 : if target_timeline_id.is_some() {
2891 : // If we were targetting this specific timeline, surface cancellation to caller
2892 0 : return Err(GcError::TimelineCancelled);
2893 : } else {
2894 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
2895 : // skip past this and proceed to try GC on other timelines.
2896 0 : continue;
2897 : }
2898 : }
2899 754 : r => r?,
2900 : };
2901 754 : totals += result;
2902 : }
2903 :
2904 754 : totals.elapsed = now.elapsed();
2905 754 : Ok(totals)
2906 754 : }
2907 :
2908 : /// Refreshes the Timeline::gc_info for all timelines, returning the
2909 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
2910 : /// [`Tenant::get_gc_horizon`].
2911 : ///
2912 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
2913 0 : pub(crate) async fn refresh_gc_info(
2914 0 : &self,
2915 0 : cancel: &CancellationToken,
2916 0 : ctx: &RequestContext,
2917 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
2918 0 : // since this method can now be called at different rates than the configured gc loop, it
2919 0 : // might be that these configuration values get applied faster than what it was previously,
2920 0 : // since these were only read from the gc task.
2921 0 : let horizon = self.get_gc_horizon();
2922 0 : let pitr = self.get_pitr_interval();
2923 0 :
2924 0 : // refresh all timelines
2925 0 : let target_timeline_id = None;
2926 0 :
2927 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2928 0 : .await
2929 0 : }
2930 :
2931 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
2932 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
2933 : ///
2934 : /// Subsequently, parent-child relationships are updated incrementally during timeline creation/deletion.
2935 0 : fn initialize_gc_info(
2936 0 : &self,
2937 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
2938 0 : ) {
2939 0 : // This function must be called before activation: after activation timeline create/delete operations
2940 0 : // might happen, and this function is not safe to run concurrently with those.
2941 0 : assert!(!self.is_active());
2942 :
2943 : // Scan all timelines. For each timeline, remember the timeline ID and
2944 : // the branch point where it was created.
2945 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId)>> = BTreeMap::new();
2946 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
2947 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
2948 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
2949 0 : ancestor_children.push((timeline_entry.get_ancestor_lsn(), *timeline_id));
2950 0 : }
2951 0 : });
2952 0 :
2953 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
2954 0 : let horizon = self.get_gc_horizon();
2955 :
2956 : // Populate each timeline's GcInfo with information about its child branches
2957 0 : for timeline in timelines.values() {
2958 0 : let mut branchpoints: Vec<(Lsn, TimelineId)> = all_branchpoints
2959 0 : .remove(&timeline.timeline_id)
2960 0 : .unwrap_or_default();
2961 0 :
2962 0 : branchpoints.sort_by_key(|b| b.0);
2963 0 :
2964 0 : let mut target = timeline.gc_info.write().unwrap();
2965 0 :
2966 0 : target.retain_lsns = branchpoints;
2967 0 :
2968 0 : let space_cutoff = timeline
2969 0 : .get_last_record_lsn()
2970 0 : .checked_sub(horizon)
2971 0 : .unwrap_or(Lsn(0));
2972 0 :
2973 0 : target.cutoffs = GcCutoffs {
2974 0 : space: space_cutoff,
2975 0 : time: Lsn::INVALID,
2976 0 : };
2977 0 : }
2978 0 : }
2979 :
2980 754 : async fn refresh_gc_info_internal(
2981 754 : &self,
2982 754 : target_timeline_id: Option<TimelineId>,
2983 754 : horizon: u64,
2984 754 : pitr: Duration,
2985 754 : cancel: &CancellationToken,
2986 754 : ctx: &RequestContext,
2987 754 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
2988 754 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
2989 754 : // currently visible timelines.
2990 754 : let timelines = self
2991 754 : .timelines
2992 754 : .lock()
2993 754 : .unwrap()
2994 754 : .values()
2995 3310 : .filter(|tl| match target_timeline_id.as_ref() {
2996 3310 : Some(target) => &tl.timeline_id == target,
2997 0 : None => true,
2998 3310 : })
2999 754 : .cloned()
3000 754 : .collect::<Vec<_>>();
3001 754 :
3002 754 : if target_timeline_id.is_some() && timelines.is_empty() {
3003 : // We were to act on a particular timeline and it wasn't found
3004 0 : return Err(GcError::TimelineNotFound);
3005 754 : }
3006 754 :
3007 754 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
3008 754 : HashMap::with_capacity(timelines.len());
3009 :
3010 754 : for timeline in timelines.iter() {
3011 754 : let cutoff = timeline
3012 754 : .get_last_record_lsn()
3013 754 : .checked_sub(horizon)
3014 754 : .unwrap_or(Lsn(0));
3015 :
3016 754 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
3017 754 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
3018 754 : assert!(old.is_none());
3019 : }
3020 :
3021 754 : if !self.is_active() || self.cancel.is_cancelled() {
3022 0 : return Err(GcError::TenantCancelled);
3023 754 : }
3024 :
3025 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
3026 : // because that will stall branch creation.
3027 754 : let gc_cs = self.gc_cs.lock().await;
3028 :
3029 : // Ok, we now know all the branch points.
3030 : // Update the GC information for each timeline.
3031 754 : let mut gc_timelines = Vec::with_capacity(timelines.len());
3032 1508 : for timeline in timelines {
3033 : // We filtered the timeline list above
3034 754 : if let Some(target_timeline_id) = target_timeline_id {
3035 754 : assert_eq!(target_timeline_id, timeline.timeline_id);
3036 0 : }
3037 :
3038 : {
3039 754 : let mut target = timeline.gc_info.write().unwrap();
3040 754 :
3041 754 : // Cull any expired leases
3042 754 : let now = SystemTime::now();
3043 754 : target.leases.retain(|_, lease| !lease.is_expired(&now));
3044 754 :
3045 754 : timeline
3046 754 : .metrics
3047 754 : .valid_lsn_lease_count_gauge
3048 754 : .set(target.leases.len() as u64);
3049 :
3050 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
3051 754 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
3052 100 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
3053 0 : target.within_ancestor_pitr =
3054 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
3055 100 : }
3056 654 : }
3057 :
3058 : // Update metrics that depend on GC state
3059 754 : timeline
3060 754 : .metrics
3061 754 : .archival_size
3062 754 : .set(if target.within_ancestor_pitr {
3063 0 : timeline.metrics.current_logical_size_gauge.get()
3064 : } else {
3065 754 : 0
3066 : });
3067 754 : timeline.metrics.pitr_history_size.set(
3068 754 : timeline
3069 754 : .get_last_record_lsn()
3070 754 : .checked_sub(target.cutoffs.time)
3071 754 : .unwrap_or(Lsn(0))
3072 754 : .0,
3073 754 : );
3074 :
3075 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
3076 : // - this timeline was created while we were finding cutoffs
3077 : // - lsn for timestamp search fails for this timeline repeatedly
3078 754 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
3079 754 : target.cutoffs = cutoffs.clone();
3080 754 : }
3081 : }
3082 :
3083 754 : gc_timelines.push(timeline);
3084 : }
3085 754 : drop(gc_cs);
3086 754 : Ok(gc_timelines)
3087 754 : }
3088 :
3089 : /// A substitute for `branch_timeline` for use in unit tests.
3090 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
3091 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
3092 : /// timeline background tasks are launched, except the flush loop.
3093 : #[cfg(test)]
3094 230 : async fn branch_timeline_test(
3095 230 : &self,
3096 230 : src_timeline: &Arc<Timeline>,
3097 230 : dst_id: TimelineId,
3098 230 : ancestor_lsn: Option<Lsn>,
3099 230 : ctx: &RequestContext,
3100 230 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3101 230 : let create_guard = self.create_timeline_create_guard(dst_id).unwrap();
3102 230 : let tl = self
3103 230 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, create_guard, ctx)
3104 4 : .await?;
3105 226 : tl.set_state(TimelineState::Active);
3106 226 : Ok(tl)
3107 230 : }
3108 :
3109 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
3110 : #[cfg(test)]
3111 : #[allow(clippy::too_many_arguments)]
3112 6 : pub async fn branch_timeline_test_with_layers(
3113 6 : &self,
3114 6 : src_timeline: &Arc<Timeline>,
3115 6 : dst_id: TimelineId,
3116 6 : ancestor_lsn: Option<Lsn>,
3117 6 : ctx: &RequestContext,
3118 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
3119 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
3120 6 : end_lsn: Lsn,
3121 6 : ) -> anyhow::Result<Arc<Timeline>> {
3122 6 : let tline = self
3123 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
3124 0 : .await?;
3125 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
3126 6 : ancestor_lsn
3127 : } else {
3128 0 : tline.get_last_record_lsn()
3129 : };
3130 6 : assert!(end_lsn >= ancestor_lsn);
3131 6 : tline.force_advance_lsn(end_lsn);
3132 12 : for deltas in delta_layer_desc {
3133 6 : tline
3134 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
3135 18 : .await?;
3136 : }
3137 10 : for (lsn, images) in image_layer_desc {
3138 4 : tline
3139 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
3140 14 : .await?;
3141 : }
3142 6 : Ok(tline)
3143 6 : }
3144 :
3145 : /// Branch an existing timeline.
3146 : ///
3147 : /// The caller is responsible for activating the returned timeline.
3148 0 : async fn branch_timeline(
3149 0 : &self,
3150 0 : src_timeline: &Arc<Timeline>,
3151 0 : dst_id: TimelineId,
3152 0 : start_lsn: Option<Lsn>,
3153 0 : timeline_create_guard: TimelineCreateGuard<'_>,
3154 0 : ctx: &RequestContext,
3155 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3156 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, timeline_create_guard, ctx)
3157 0 : .await
3158 0 : }
3159 :
3160 230 : async fn branch_timeline_impl(
3161 230 : &self,
3162 230 : src_timeline: &Arc<Timeline>,
3163 230 : dst_id: TimelineId,
3164 230 : start_lsn: Option<Lsn>,
3165 230 : timeline_create_guard: TimelineCreateGuard<'_>,
3166 230 : _ctx: &RequestContext,
3167 230 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3168 230 : let src_id = src_timeline.timeline_id;
3169 :
3170 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
3171 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
3172 : // valid while we are creating the branch.
3173 230 : let _gc_cs = self.gc_cs.lock().await;
3174 :
3175 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
3176 230 : let start_lsn = start_lsn.unwrap_or_else(|| {
3177 2 : let lsn = src_timeline.get_last_record_lsn();
3178 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
3179 2 : lsn
3180 230 : });
3181 230 :
3182 230 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
3183 230 : // horizon on the source timeline
3184 230 : //
3185 230 : // We check it against both the planned GC cutoff stored in 'gc_info',
3186 230 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
3187 230 : // planned GC cutoff in 'gc_info' is normally larger than
3188 230 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
3189 230 : // changed the GC settings for the tenant to make the PITR window
3190 230 : // larger, but some of the data was already removed by an earlier GC
3191 230 : // iteration.
3192 230 :
3193 230 : // check against last actual 'latest_gc_cutoff' first
3194 230 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
3195 230 : src_timeline
3196 230 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
3197 230 : .context(format!(
3198 230 : "invalid branch start lsn: less than latest GC cutoff {}",
3199 230 : *latest_gc_cutoff_lsn,
3200 230 : ))
3201 230 : .map_err(CreateTimelineError::AncestorLsn)?;
3202 :
3203 : // and then the planned GC cutoff
3204 : {
3205 226 : let gc_info = src_timeline.gc_info.read().unwrap();
3206 226 : let cutoff = gc_info.min_cutoff();
3207 226 : if start_lsn < cutoff {
3208 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
3209 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
3210 0 : )));
3211 226 : }
3212 226 : }
3213 226 :
3214 226 : //
3215 226 : // The branch point is valid, and we are still holding the 'gc_cs' lock
3216 226 : // so that GC cannot advance the GC cutoff until we are finished.
3217 226 : // Proceed with the branch creation.
3218 226 : //
3219 226 :
3220 226 : // Determine prev-LSN for the new timeline. We can only determine it if
3221 226 : // the timeline was branched at the current end of the source timeline.
3222 226 : let RecordLsn {
3223 226 : last: src_last,
3224 226 : prev: src_prev,
3225 226 : } = src_timeline.get_last_record_rlsn();
3226 226 : let dst_prev = if src_last == start_lsn {
3227 216 : Some(src_prev)
3228 : } else {
3229 10 : None
3230 : };
3231 :
3232 : // Create the metadata file, noting the ancestor of the new timeline.
3233 : // There is initially no data in it, but all the read-calls know to look
3234 : // into the ancestor.
3235 226 : let metadata = TimelineMetadata::new(
3236 226 : start_lsn,
3237 226 : dst_prev,
3238 226 : Some(src_id),
3239 226 : start_lsn,
3240 226 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
3241 226 : src_timeline.initdb_lsn,
3242 226 : src_timeline.pg_version,
3243 226 : );
3244 :
3245 226 : let uninitialized_timeline = self
3246 226 : .prepare_new_timeline(
3247 226 : dst_id,
3248 226 : &metadata,
3249 226 : timeline_create_guard,
3250 226 : start_lsn + 1,
3251 226 : Some(Arc::clone(src_timeline)),
3252 226 : src_timeline.last_aux_file_policy.load(),
3253 226 : )
3254 0 : .await?;
3255 :
3256 226 : let new_timeline = uninitialized_timeline.finish_creation()?;
3257 :
3258 : // Root timeline gets its layers during creation and uploads them along with the metadata.
3259 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
3260 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
3261 : // could get incorrect information and remove more layers, than needed.
3262 : // See also https://github.com/neondatabase/neon/issues/3865
3263 226 : new_timeline
3264 226 : .remote_client
3265 226 : .schedule_index_upload_for_full_metadata_update(&metadata)
3266 226 : .context("branch initial metadata upload")?;
3267 :
3268 226 : Ok(new_timeline)
3269 230 : }
3270 :
3271 : /// For unit tests, make this visible so that other modules can directly create timelines
3272 : #[cfg(test)]
3273 4 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
3274 : pub(crate) async fn bootstrap_timeline_test(
3275 : &self,
3276 : timeline_id: TimelineId,
3277 : pg_version: u32,
3278 : load_existing_initdb: Option<TimelineId>,
3279 : ctx: &RequestContext,
3280 : ) -> anyhow::Result<Arc<Timeline>> {
3281 : let create_guard = self.create_timeline_create_guard(timeline_id).unwrap();
3282 : self.bootstrap_timeline(
3283 : timeline_id,
3284 : pg_version,
3285 : load_existing_initdb,
3286 : create_guard,
3287 : ctx,
3288 : )
3289 : .await
3290 : }
3291 :
3292 0 : async fn upload_initdb(
3293 0 : &self,
3294 0 : timelines_path: &Utf8PathBuf,
3295 0 : pgdata_path: &Utf8PathBuf,
3296 0 : timeline_id: &TimelineId,
3297 0 : ) -> anyhow::Result<()> {
3298 0 : let temp_path = timelines_path.join(format!(
3299 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
3300 0 : ));
3301 :
3302 : scopeguard::defer! {
3303 : if let Err(e) = fs::remove_file(&temp_path) {
3304 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
3305 : }
3306 : }
3307 :
3308 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
3309 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
3310 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
3311 0 : warn!(
3312 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
3313 : );
3314 0 : }
3315 :
3316 : pausable_failpoint!("before-initdb-upload");
3317 :
3318 0 : backoff::retry(
3319 0 : || async {
3320 0 : self::remote_timeline_client::upload_initdb_dir(
3321 0 : &self.remote_storage,
3322 0 : &self.tenant_shard_id.tenant_id,
3323 0 : timeline_id,
3324 0 : pgdata_zstd.try_clone().await?,
3325 0 : tar_zst_size,
3326 0 : &self.cancel,
3327 0 : )
3328 0 : .await
3329 0 : },
3330 0 : |_| false,
3331 0 : 3,
3332 0 : u32::MAX,
3333 0 : "persist_initdb_tar_zst",
3334 0 : &self.cancel,
3335 0 : )
3336 0 : .await
3337 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
3338 0 : .and_then(|x| x)
3339 0 : }
3340 :
3341 : /// - run initdb to init temporary instance and get bootstrap data
3342 : /// - after initialization completes, tar up the temp dir and upload it to S3.
3343 : ///
3344 : /// The caller is responsible for activating the returned timeline.
3345 2 : async fn bootstrap_timeline(
3346 2 : &self,
3347 2 : timeline_id: TimelineId,
3348 2 : pg_version: u32,
3349 2 : load_existing_initdb: Option<TimelineId>,
3350 2 : timeline_create_guard: TimelineCreateGuard<'_>,
3351 2 : ctx: &RequestContext,
3352 2 : ) -> anyhow::Result<Arc<Timeline>> {
3353 2 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
3354 2 : // temporary directory for basebackup files for the given timeline.
3355 2 :
3356 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
3357 2 : let pgdata_path = path_with_suffix_extension(
3358 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
3359 2 : TEMP_FILE_SUFFIX,
3360 2 : );
3361 2 :
3362 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
3363 2 : // we won't race with other creations or existent timelines with the same path.
3364 2 : if pgdata_path.exists() {
3365 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
3366 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
3367 0 : })?;
3368 2 : }
3369 :
3370 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
3371 : scopeguard::defer! {
3372 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
3373 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
3374 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
3375 : }
3376 : }
3377 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
3378 2 : if existing_initdb_timeline_id != timeline_id {
3379 0 : let source_path = &remote_initdb_archive_path(
3380 0 : &self.tenant_shard_id.tenant_id,
3381 0 : &existing_initdb_timeline_id,
3382 0 : );
3383 0 : let dest_path =
3384 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
3385 0 :
3386 0 : // if this fails, it will get retried by retried control plane requests
3387 0 : self.remote_storage
3388 0 : .copy_object(source_path, dest_path, &self.cancel)
3389 0 : .await
3390 0 : .context("copy initdb tar")?;
3391 2 : }
3392 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
3393 2 : self::remote_timeline_client::download_initdb_tar_zst(
3394 2 : self.conf,
3395 2 : &self.remote_storage,
3396 2 : &self.tenant_shard_id,
3397 2 : &existing_initdb_timeline_id,
3398 2 : &self.cancel,
3399 2 : )
3400 632 : .await
3401 2 : .context("download initdb tar")?;
3402 :
3403 : scopeguard::defer! {
3404 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
3405 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
3406 : }
3407 : }
3408 :
3409 2 : let buf_read =
3410 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
3411 2 : extract_zst_tarball(&pgdata_path, buf_read)
3412 9836 : .await
3413 2 : .context("extract initdb tar")?;
3414 : } else {
3415 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
3416 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel).await?;
3417 :
3418 : // Upload the created data dir to S3
3419 0 : if self.tenant_shard_id().is_shard_zero() {
3420 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
3421 0 : .await?;
3422 0 : }
3423 : }
3424 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
3425 2 :
3426 2 : // Import the contents of the data directory at the initial checkpoint
3427 2 : // LSN, and any WAL after that.
3428 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
3429 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
3430 2 : let new_metadata = TimelineMetadata::new(
3431 2 : Lsn(0),
3432 2 : None,
3433 2 : None,
3434 2 : Lsn(0),
3435 2 : pgdata_lsn,
3436 2 : pgdata_lsn,
3437 2 : pg_version,
3438 2 : );
3439 2 : let raw_timeline = self
3440 2 : .prepare_new_timeline(
3441 2 : timeline_id,
3442 2 : &new_metadata,
3443 2 : timeline_create_guard,
3444 2 : pgdata_lsn,
3445 2 : None,
3446 2 : None,
3447 2 : )
3448 0 : .await?;
3449 :
3450 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
3451 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
3452 :
3453 : // Flush the new layer files to disk, before we make the timeline as available to
3454 : // the outside world.
3455 : //
3456 : // Flush loop needs to be spawned in order to be able to flush.
3457 2 : unfinished_timeline.maybe_spawn_flush_loop();
3458 2 :
3459 2 : import_datadir::import_timeline_from_postgres_datadir(
3460 2 : unfinished_timeline,
3461 2 : &pgdata_path,
3462 2 : pgdata_lsn,
3463 2 : ctx,
3464 2 : )
3465 7876 : .await
3466 2 : .with_context(|| {
3467 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
3468 2 : })?;
3469 :
3470 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
3471 0 : anyhow::bail!("failpoint before-checkpoint-new-timeline");
3472 2 : });
3473 :
3474 2 : unfinished_timeline
3475 2 : .freeze_and_flush()
3476 2 : .await
3477 2 : .with_context(|| {
3478 0 : format!(
3479 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
3480 0 : )
3481 2 : })?;
3482 :
3483 : // All done!
3484 2 : let timeline = raw_timeline.finish_creation()?;
3485 :
3486 2 : Ok(timeline)
3487 2 : }
3488 :
3489 : /// Call this before constructing a timeline, to build its required structures
3490 400 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
3491 400 : let remote_client = RemoteTimelineClient::new(
3492 400 : self.remote_storage.clone(),
3493 400 : self.deletion_queue_client.clone(),
3494 400 : self.conf,
3495 400 : self.tenant_shard_id,
3496 400 : timeline_id,
3497 400 : self.generation,
3498 400 : );
3499 400 : TimelineResources {
3500 400 : remote_client,
3501 400 : timeline_get_throttle: self.timeline_get_throttle.clone(),
3502 400 : l0_flush_global_state: self.l0_flush_global_state.clone(),
3503 400 : }
3504 400 : }
3505 :
3506 : /// Creates intermediate timeline structure and its files.
3507 : ///
3508 : /// An empty layer map is initialized, and new data and WAL can be imported starting
3509 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
3510 : /// `finish_creation` to insert the Timeline into the timelines map.
3511 400 : async fn prepare_new_timeline<'a>(
3512 400 : &'a self,
3513 400 : new_timeline_id: TimelineId,
3514 400 : new_metadata: &TimelineMetadata,
3515 400 : create_guard: TimelineCreateGuard<'a>,
3516 400 : start_lsn: Lsn,
3517 400 : ancestor: Option<Arc<Timeline>>,
3518 400 : last_aux_file_policy: Option<AuxFilePolicy>,
3519 400 : ) -> anyhow::Result<UninitializedTimeline> {
3520 400 : let tenant_shard_id = self.tenant_shard_id;
3521 400 :
3522 400 : let resources = self.build_timeline_resources(new_timeline_id);
3523 400 : resources
3524 400 : .remote_client
3525 400 : .init_upload_queue_for_empty_remote(new_metadata)?;
3526 :
3527 400 : let timeline_struct = self
3528 400 : .create_timeline_struct(
3529 400 : new_timeline_id,
3530 400 : new_metadata,
3531 400 : ancestor,
3532 400 : resources,
3533 400 : CreateTimelineCause::Load,
3534 400 : last_aux_file_policy,
3535 400 : )
3536 400 : .context("Failed to create timeline data structure")?;
3537 :
3538 400 : timeline_struct.init_empty_layer_map(start_lsn);
3539 :
3540 400 : if let Err(e) = self
3541 400 : .create_timeline_files(&create_guard.timeline_path)
3542 0 : .await
3543 : {
3544 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
3545 0 : cleanup_timeline_directory(create_guard);
3546 0 : return Err(e);
3547 400 : }
3548 400 :
3549 400 : debug!(
3550 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
3551 : );
3552 :
3553 400 : Ok(UninitializedTimeline::new(
3554 400 : self,
3555 400 : new_timeline_id,
3556 400 : Some((timeline_struct, create_guard)),
3557 400 : ))
3558 400 : }
3559 :
3560 400 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
3561 400 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
3562 :
3563 400 : fail::fail_point!("after-timeline-dir-creation", |_| {
3564 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
3565 400 : });
3566 :
3567 400 : Ok(())
3568 400 : }
3569 :
3570 : /// Get a guard that provides exclusive access to the timeline directory, preventing
3571 : /// concurrent attempts to create the same timeline.
3572 406 : fn create_timeline_create_guard(
3573 406 : &self,
3574 406 : timeline_id: TimelineId,
3575 406 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
3576 406 : let tenant_shard_id = self.tenant_shard_id;
3577 406 :
3578 406 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
3579 :
3580 406 : let create_guard = TimelineCreateGuard::new(self, timeline_id, timeline_path.clone())?;
3581 :
3582 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
3583 : // for creation.
3584 : // A timeline directory should never exist on disk already:
3585 : // - a previous failed creation would have cleaned up after itself
3586 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
3587 : //
3588 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
3589 : // this error may indicate a bug in cleanup on failed creations.
3590 404 : if timeline_path.exists() {
3591 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
3592 0 : "Timeline directory already exists! This is a bug."
3593 0 : )));
3594 404 : }
3595 404 :
3596 404 : Ok(create_guard)
3597 406 : }
3598 :
3599 : /// Gathers inputs from all of the timelines to produce a sizing model input.
3600 : ///
3601 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
3602 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3603 : pub async fn gather_size_inputs(
3604 : &self,
3605 : // `max_retention_period` overrides the cutoff that is used to calculate the size
3606 : // (only if it is shorter than the real cutoff).
3607 : max_retention_period: Option<u64>,
3608 : cause: LogicalSizeCalculationCause,
3609 : cancel: &CancellationToken,
3610 : ctx: &RequestContext,
3611 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
3612 : let logical_sizes_at_once = self
3613 : .conf
3614 : .concurrent_tenant_size_logical_size_queries
3615 : .inner();
3616 :
3617 : // TODO: Having a single mutex block concurrent reads is not great for performance.
3618 : //
3619 : // But the only case where we need to run multiple of these at once is when we
3620 : // request a size for a tenant manually via API, while another background calculation
3621 : // is in progress (which is not a common case).
3622 : //
3623 : // See more for on the issue #2748 condenced out of the initial PR review.
3624 : let mut shared_cache = tokio::select! {
3625 : locked = self.cached_logical_sizes.lock() => locked,
3626 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3627 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3628 : };
3629 :
3630 : size::gather_inputs(
3631 : self,
3632 : logical_sizes_at_once,
3633 : max_retention_period,
3634 : &mut shared_cache,
3635 : cause,
3636 : cancel,
3637 : ctx,
3638 : )
3639 : .await
3640 : }
3641 :
3642 : /// Calculate synthetic tenant size and cache the result.
3643 : /// This is periodically called by background worker.
3644 : /// result is cached in tenant struct
3645 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3646 : pub async fn calculate_synthetic_size(
3647 : &self,
3648 : cause: LogicalSizeCalculationCause,
3649 : cancel: &CancellationToken,
3650 : ctx: &RequestContext,
3651 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
3652 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
3653 :
3654 : let size = inputs.calculate();
3655 :
3656 : self.set_cached_synthetic_size(size);
3657 :
3658 : Ok(size)
3659 : }
3660 :
3661 : /// Cache given synthetic size and update the metric value
3662 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
3663 0 : self.cached_synthetic_tenant_size
3664 0 : .store(size, Ordering::Relaxed);
3665 0 :
3666 0 : // Only shard zero should be calculating synthetic sizes
3667 0 : debug_assert!(self.shard_identity.is_shard_zero());
3668 :
3669 0 : TENANT_SYNTHETIC_SIZE_METRIC
3670 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
3671 0 : .unwrap()
3672 0 : .set(size);
3673 0 : }
3674 :
3675 0 : pub fn cached_synthetic_size(&self) -> u64 {
3676 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
3677 0 : }
3678 :
3679 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
3680 : ///
3681 : /// This function can take a long time: callers should wrap it in a timeout if calling
3682 : /// from an external API handler.
3683 : ///
3684 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
3685 : /// still bounded by tenant/timeline shutdown.
3686 0 : #[tracing::instrument(skip_all)]
3687 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
3688 : let timelines = self.timelines.lock().unwrap().clone();
3689 :
3690 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
3691 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
3692 0 : timeline.freeze_and_flush().await?;
3693 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
3694 0 : timeline.remote_client.wait_completion().await?;
3695 :
3696 0 : Ok(())
3697 0 : }
3698 :
3699 : // We do not use a JoinSet for these tasks, because we don't want them to be
3700 : // aborted when this function's future is cancelled: they should stay alive
3701 : // holding their GateGuard until they complete, to ensure their I/Os complete
3702 : // before Timeline shutdown completes.
3703 : let mut results = FuturesUnordered::new();
3704 :
3705 : for (_timeline_id, timeline) in timelines {
3706 : // Run each timeline's flush in a task holding the timeline's gate: this
3707 : // means that if this function's future is cancelled, the Timeline shutdown
3708 : // will still wait for any I/O in here to complete.
3709 : let Ok(gate) = timeline.gate.enter() else {
3710 : continue;
3711 : };
3712 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
3713 : results.push(jh);
3714 : }
3715 :
3716 : while let Some(r) = results.next().await {
3717 : if let Err(e) = r {
3718 : if !e.is_cancelled() && !e.is_panic() {
3719 : tracing::error!("unexpected join error: {e:?}");
3720 : }
3721 : }
3722 : }
3723 :
3724 : // The flushes we did above were just writes, but the Tenant might have had
3725 : // pending deletions as well from recent compaction/gc: we want to flush those
3726 : // as well. This requires flushing the global delete queue. This is cheap
3727 : // because it's typically a no-op.
3728 : match self.deletion_queue_client.flush_execute().await {
3729 : Ok(_) => {}
3730 : Err(DeletionQueueError::ShuttingDown) => {}
3731 : }
3732 :
3733 : Ok(())
3734 : }
3735 :
3736 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
3737 0 : self.tenant_conf.load().tenant_conf.clone()
3738 0 : }
3739 :
3740 : /// How much local storage would this tenant like to have? It can cope with
3741 : /// less than this (via eviction and on-demand downloads), but this function enables
3742 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
3743 : /// by keeping important things on local disk.
3744 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
3745 0 : let mut wanted = 0;
3746 0 : let timelines = self.timelines.lock().unwrap();
3747 0 : for timeline in timelines.values() {
3748 0 : wanted += timeline.metrics.visible_physical_size_gauge.get();
3749 0 : }
3750 0 : wanted
3751 0 : }
3752 : }
3753 :
3754 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
3755 : /// to get bootstrap data for timeline initialization.
3756 0 : async fn run_initdb(
3757 0 : conf: &'static PageServerConf,
3758 0 : initdb_target_dir: &Utf8Path,
3759 0 : pg_version: u32,
3760 0 : cancel: &CancellationToken,
3761 0 : ) -> Result<(), InitdbError> {
3762 0 : let initdb_bin_path = conf
3763 0 : .pg_bin_dir(pg_version)
3764 0 : .map_err(InitdbError::Other)?
3765 0 : .join("initdb");
3766 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
3767 0 : info!(
3768 0 : "running {} in {}, libdir: {}",
3769 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
3770 : );
3771 :
3772 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
3773 :
3774 0 : let initdb_command = tokio::process::Command::new(&initdb_bin_path)
3775 0 : .args(["-D", initdb_target_dir.as_ref()])
3776 0 : .args(["-U", &conf.superuser])
3777 0 : .args(["-E", "utf8"])
3778 0 : .arg("--no-instructions")
3779 0 : .arg("--no-sync")
3780 0 : .env_clear()
3781 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
3782 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
3783 0 : .stdin(std::process::Stdio::null())
3784 0 : // stdout invocation produces the same output every time, we don't need it
3785 0 : .stdout(std::process::Stdio::null())
3786 0 : // we would be interested in the stderr output, if there was any
3787 0 : .stderr(std::process::Stdio::piped())
3788 0 : .spawn()?;
3789 :
3790 : // Ideally we'd select here with the cancellation token, but the problem is that
3791 : // we can't safely terminate initdb: it launches processes of its own, and killing
3792 : // initdb doesn't kill them. After we return from this function, we want the target
3793 : // directory to be able to be cleaned up.
3794 : // See https://github.com/neondatabase/neon/issues/6385
3795 0 : let initdb_output = initdb_command.wait_with_output().await?;
3796 0 : if !initdb_output.status.success() {
3797 0 : return Err(InitdbError::Failed(
3798 0 : initdb_output.status,
3799 0 : initdb_output.stderr,
3800 0 : ));
3801 0 : }
3802 0 :
3803 0 : // This isn't true cancellation support, see above. Still return an error to
3804 0 : // excercise the cancellation code path.
3805 0 : if cancel.is_cancelled() {
3806 0 : return Err(InitdbError::Cancelled);
3807 0 : }
3808 0 :
3809 0 : Ok(())
3810 0 : }
3811 :
3812 : /// Dump contents of a layer file to stdout.
3813 0 : pub async fn dump_layerfile_from_path(
3814 0 : path: &Utf8Path,
3815 0 : verbose: bool,
3816 0 : ctx: &RequestContext,
3817 0 : ) -> anyhow::Result<()> {
3818 : use std::os::unix::fs::FileExt;
3819 :
3820 : // All layer files start with a two-byte "magic" value, to identify the kind of
3821 : // file.
3822 0 : let file = File::open(path)?;
3823 0 : let mut header_buf = [0u8; 2];
3824 0 : file.read_exact_at(&mut header_buf, 0)?;
3825 :
3826 0 : match u16::from_be_bytes(header_buf) {
3827 : crate::IMAGE_FILE_MAGIC => {
3828 0 : ImageLayer::new_for_path(path, file)?
3829 0 : .dump(verbose, ctx)
3830 0 : .await?
3831 : }
3832 : crate::DELTA_FILE_MAGIC => {
3833 0 : DeltaLayer::new_for_path(path, file)?
3834 0 : .dump(verbose, ctx)
3835 0 : .await?
3836 : }
3837 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
3838 : }
3839 :
3840 0 : Ok(())
3841 0 : }
3842 :
3843 : #[cfg(test)]
3844 : pub(crate) mod harness {
3845 : use bytes::{Bytes, BytesMut};
3846 : use once_cell::sync::OnceCell;
3847 : use pageserver_api::models::ShardParameters;
3848 : use pageserver_api::shard::ShardIndex;
3849 : use utils::logging;
3850 :
3851 : use crate::deletion_queue::mock::MockDeletionQueue;
3852 : use crate::l0_flush::L0FlushConfig;
3853 : use crate::walredo::apply_neon;
3854 : use crate::{repository::Key, walrecord::NeonWalRecord};
3855 :
3856 : use super::*;
3857 : use hex_literal::hex;
3858 : use utils::id::TenantId;
3859 :
3860 : pub const TIMELINE_ID: TimelineId =
3861 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
3862 : pub const NEW_TIMELINE_ID: TimelineId =
3863 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
3864 :
3865 : /// Convenience function to create a page image with given string as the only content
3866 5028683 : pub fn test_img(s: &str) -> Bytes {
3867 5028683 : let mut buf = BytesMut::new();
3868 5028683 : buf.extend_from_slice(s.as_bytes());
3869 5028683 : buf.resize(64, 0);
3870 5028683 :
3871 5028683 : buf.freeze()
3872 5028683 : }
3873 :
3874 : impl From<TenantConf> for TenantConfOpt {
3875 182 : fn from(tenant_conf: TenantConf) -> Self {
3876 182 : Self {
3877 182 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
3878 182 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
3879 182 : compaction_target_size: Some(tenant_conf.compaction_target_size),
3880 182 : compaction_period: Some(tenant_conf.compaction_period),
3881 182 : compaction_threshold: Some(tenant_conf.compaction_threshold),
3882 182 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
3883 182 : gc_horizon: Some(tenant_conf.gc_horizon),
3884 182 : gc_period: Some(tenant_conf.gc_period),
3885 182 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
3886 182 : pitr_interval: Some(tenant_conf.pitr_interval),
3887 182 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
3888 182 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
3889 182 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
3890 182 : eviction_policy: Some(tenant_conf.eviction_policy),
3891 182 : min_resident_size_override: tenant_conf.min_resident_size_override,
3892 182 : evictions_low_residence_duration_metric_threshold: Some(
3893 182 : tenant_conf.evictions_low_residence_duration_metric_threshold,
3894 182 : ),
3895 182 : heatmap_period: Some(tenant_conf.heatmap_period),
3896 182 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
3897 182 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
3898 182 : image_layer_creation_check_threshold: Some(
3899 182 : tenant_conf.image_layer_creation_check_threshold,
3900 182 : ),
3901 182 : switch_aux_file_policy: Some(tenant_conf.switch_aux_file_policy),
3902 182 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
3903 182 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
3904 182 : }
3905 182 : }
3906 : }
3907 :
3908 : pub struct TenantHarness {
3909 : pub conf: &'static PageServerConf,
3910 : pub tenant_conf: TenantConf,
3911 : pub tenant_shard_id: TenantShardId,
3912 : pub generation: Generation,
3913 : pub shard: ShardIndex,
3914 : pub remote_storage: GenericRemoteStorage,
3915 : pub remote_fs_dir: Utf8PathBuf,
3916 : pub deletion_queue: MockDeletionQueue,
3917 : }
3918 :
3919 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
3920 :
3921 196 : pub(crate) fn setup_logging() {
3922 196 : LOG_HANDLE.get_or_init(|| {
3923 184 : logging::init(
3924 184 : logging::LogFormat::Test,
3925 184 : // enable it in case the tests exercise code paths that use
3926 184 : // debug_assert_current_span_has_tenant_and_timeline_id
3927 184 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
3928 184 : logging::Output::Stdout,
3929 184 : )
3930 184 : .expect("Failed to init test logging")
3931 196 : });
3932 196 : }
3933 :
3934 : impl TenantHarness {
3935 182 : pub async fn create_custom(
3936 182 : test_name: &'static str,
3937 182 : tenant_conf: TenantConf,
3938 182 : tenant_id: TenantId,
3939 182 : shard_identity: ShardIdentity,
3940 182 : generation: Generation,
3941 182 : ) -> anyhow::Result<Self> {
3942 182 : setup_logging();
3943 182 :
3944 182 : let repo_dir = PageServerConf::test_repo_dir(test_name);
3945 182 : let _ = fs::remove_dir_all(&repo_dir);
3946 182 : fs::create_dir_all(&repo_dir)?;
3947 :
3948 182 : let conf = PageServerConf::dummy_conf(repo_dir);
3949 182 : // Make a static copy of the config. This can never be free'd, but that's
3950 182 : // OK in a test.
3951 182 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
3952 182 :
3953 182 : let shard = shard_identity.shard_index();
3954 182 : let tenant_shard_id = TenantShardId {
3955 182 : tenant_id,
3956 182 : shard_number: shard.shard_number,
3957 182 : shard_count: shard.shard_count,
3958 182 : };
3959 182 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
3960 182 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
3961 :
3962 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
3963 182 : let remote_fs_dir = conf.workdir.join("localfs");
3964 182 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
3965 182 : let config = RemoteStorageConfig {
3966 182 : storage: RemoteStorageKind::LocalFs {
3967 182 : local_path: remote_fs_dir.clone(),
3968 182 : },
3969 182 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
3970 182 : };
3971 182 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
3972 182 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
3973 182 :
3974 182 : Ok(Self {
3975 182 : conf,
3976 182 : tenant_conf,
3977 182 : tenant_shard_id,
3978 182 : generation,
3979 182 : shard,
3980 182 : remote_storage,
3981 182 : remote_fs_dir,
3982 182 : deletion_queue,
3983 182 : })
3984 182 : }
3985 :
3986 170 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
3987 170 : // Disable automatic GC and compaction to make the unit tests more deterministic.
3988 170 : // The tests perform them manually if needed.
3989 170 : let tenant_conf = TenantConf {
3990 170 : gc_period: Duration::ZERO,
3991 170 : compaction_period: Duration::ZERO,
3992 170 : ..TenantConf::default()
3993 170 : };
3994 170 : let tenant_id = TenantId::generate();
3995 170 : let shard = ShardIdentity::unsharded();
3996 170 : Self::create_custom(
3997 170 : test_name,
3998 170 : tenant_conf,
3999 170 : tenant_id,
4000 170 : shard,
4001 170 : Generation::new(0xdeadbeef),
4002 170 : )
4003 0 : .await
4004 170 : }
4005 :
4006 18 : pub fn span(&self) -> tracing::Span {
4007 18 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
4008 18 : }
4009 :
4010 182 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
4011 182 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
4012 182 : (
4013 182 : self.do_try_load(&ctx)
4014 731 : .await
4015 182 : .expect("failed to load test tenant"),
4016 182 : ctx,
4017 182 : )
4018 182 : }
4019 :
4020 364 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
4021 : pub(crate) async fn do_try_load(
4022 : &self,
4023 : ctx: &RequestContext,
4024 : ) -> anyhow::Result<Arc<Tenant>> {
4025 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
4026 :
4027 : let tenant = Arc::new(Tenant::new(
4028 : TenantState::Loading,
4029 : self.conf,
4030 : AttachedTenantConf::try_from(LocationConf::attached_single(
4031 : TenantConfOpt::from(self.tenant_conf.clone()),
4032 : self.generation,
4033 : &ShardParameters::default(),
4034 : ))
4035 : .unwrap(),
4036 : // This is a legacy/test code path: sharding isn't supported here.
4037 : ShardIdentity::unsharded(),
4038 : Some(walredo_mgr),
4039 : self.tenant_shard_id,
4040 : self.remote_storage.clone(),
4041 : self.deletion_queue.new_client(),
4042 : // TODO: ideally we should run all unit tests with both configs
4043 : L0FlushGlobalState::new(L0FlushConfig::default()),
4044 : ));
4045 :
4046 : let preload = tenant
4047 : .preload(&self.remote_storage, CancellationToken::new())
4048 : .await?;
4049 : tenant.attach(Some(preload), ctx).await?;
4050 :
4051 : tenant.state.send_replace(TenantState::Active);
4052 : for timeline in tenant.timelines.lock().unwrap().values() {
4053 : timeline.set_state(TimelineState::Active);
4054 : }
4055 : Ok(tenant)
4056 : }
4057 :
4058 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
4059 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
4060 2 : }
4061 : }
4062 :
4063 : // Mock WAL redo manager that doesn't do much
4064 : pub(crate) struct TestRedoManager;
4065 :
4066 : impl TestRedoManager {
4067 : /// # Cancel-Safety
4068 : ///
4069 : /// This method is cancellation-safe.
4070 296 : pub async fn request_redo(
4071 296 : &self,
4072 296 : key: Key,
4073 296 : lsn: Lsn,
4074 296 : base_img: Option<(Lsn, Bytes)>,
4075 296 : records: Vec<(Lsn, NeonWalRecord)>,
4076 296 : _pg_version: u32,
4077 296 : ) -> Result<Bytes, walredo::Error> {
4078 378 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
4079 296 : if records_neon {
4080 : // For Neon wal records, we can decode without spawning postgres, so do so.
4081 296 : let base_img = base_img.expect("Neon WAL redo requires base image").1;
4082 296 : let mut page = BytesMut::new();
4083 296 : page.extend_from_slice(&base_img);
4084 674 : for (record_lsn, record) in records {
4085 378 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
4086 : }
4087 296 : Ok(page.freeze())
4088 : } else {
4089 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
4090 0 : let s = format!(
4091 0 : "redo for {} to get to {}, with {} and {} records",
4092 0 : key,
4093 0 : lsn,
4094 0 : if base_img.is_some() {
4095 0 : "base image"
4096 : } else {
4097 0 : "no base image"
4098 : },
4099 0 : records.len()
4100 0 : );
4101 0 : println!("{s}");
4102 0 :
4103 0 : Ok(test_img(&s))
4104 : }
4105 296 : }
4106 : }
4107 : }
4108 :
4109 : #[cfg(test)]
4110 : mod tests {
4111 : use std::collections::{BTreeMap, BTreeSet};
4112 :
4113 : use super::*;
4114 : use crate::keyspace::KeySpaceAccum;
4115 : use crate::pgdatadir_mapping::AuxFilesDirectory;
4116 : use crate::repository::{Key, Value};
4117 : use crate::tenant::harness::*;
4118 : use crate::tenant::timeline::CompactFlags;
4119 : use crate::walrecord::NeonWalRecord;
4120 : use crate::DEFAULT_PG_VERSION;
4121 : use bytes::{Bytes, BytesMut};
4122 : use hex_literal::hex;
4123 : use itertools::Itertools;
4124 : use pageserver_api::key::{AUX_FILES_KEY, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
4125 : use pageserver_api::keyspace::KeySpace;
4126 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
4127 : use rand::{thread_rng, Rng};
4128 : use storage_layer::PersistentLayerKey;
4129 : use tests::storage_layer::ValuesReconstructState;
4130 : use tests::timeline::{GetVectoredError, ShutdownMode};
4131 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
4132 : use timeline::{DeltaLayerTestDesc, GcInfo};
4133 : use utils::bin_ser::BeSer;
4134 : use utils::id::TenantId;
4135 :
4136 : static TEST_KEY: Lazy<Key> =
4137 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
4138 :
4139 : #[tokio::test]
4140 2 : async fn test_basic() -> anyhow::Result<()> {
4141 8 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
4142 2 : let tline = tenant
4143 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4144 4 : .await?;
4145 2 :
4146 2 : let mut writer = tline.writer().await;
4147 2 : writer
4148 2 : .put(
4149 2 : *TEST_KEY,
4150 2 : Lsn(0x10),
4151 2 : &Value::Image(test_img("foo at 0x10")),
4152 2 : &ctx,
4153 2 : )
4154 2 : .await?;
4155 2 : writer.finish_write(Lsn(0x10));
4156 2 : drop(writer);
4157 2 :
4158 2 : let mut writer = tline.writer().await;
4159 2 : writer
4160 2 : .put(
4161 2 : *TEST_KEY,
4162 2 : Lsn(0x20),
4163 2 : &Value::Image(test_img("foo at 0x20")),
4164 2 : &ctx,
4165 2 : )
4166 2 : .await?;
4167 2 : writer.finish_write(Lsn(0x20));
4168 2 : drop(writer);
4169 2 :
4170 2 : assert_eq!(
4171 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4172 2 : test_img("foo at 0x10")
4173 2 : );
4174 2 : assert_eq!(
4175 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4176 2 : test_img("foo at 0x10")
4177 2 : );
4178 2 : assert_eq!(
4179 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4180 2 : test_img("foo at 0x20")
4181 2 : );
4182 2 :
4183 2 : Ok(())
4184 2 : }
4185 :
4186 : #[tokio::test]
4187 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
4188 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
4189 2 : .await?
4190 2 : .load()
4191 8 : .await;
4192 2 : let _ = tenant
4193 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4194 4 : .await?;
4195 2 :
4196 2 : match tenant
4197 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4198 2 : .await
4199 2 : {
4200 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
4201 2 : Err(e) => assert_eq!(e.to_string(), "Already exists".to_string()),
4202 2 : }
4203 2 :
4204 2 : Ok(())
4205 2 : }
4206 :
4207 : /// Convenience function to create a page image with given string as the only content
4208 10 : pub fn test_value(s: &str) -> Value {
4209 10 : let mut buf = BytesMut::new();
4210 10 : buf.extend_from_slice(s.as_bytes());
4211 10 : Value::Image(buf.freeze())
4212 10 : }
4213 :
4214 : ///
4215 : /// Test branch creation
4216 : ///
4217 : #[tokio::test]
4218 2 : async fn test_branch() -> anyhow::Result<()> {
4219 2 : use std::str::from_utf8;
4220 2 :
4221 8 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
4222 2 : let tline = tenant
4223 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4224 4 : .await?;
4225 2 : let mut writer = tline.writer().await;
4226 2 :
4227 2 : #[allow(non_snake_case)]
4228 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
4229 2 : #[allow(non_snake_case)]
4230 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
4231 2 :
4232 2 : // Insert a value on the timeline
4233 2 : writer
4234 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
4235 2 : .await?;
4236 2 : writer
4237 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
4238 2 : .await?;
4239 2 : writer.finish_write(Lsn(0x20));
4240 2 :
4241 2 : writer
4242 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
4243 2 : .await?;
4244 2 : writer.finish_write(Lsn(0x30));
4245 2 : writer
4246 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
4247 2 : .await?;
4248 2 : writer.finish_write(Lsn(0x40));
4249 2 :
4250 2 : //assert_current_logical_size(&tline, Lsn(0x40));
4251 2 :
4252 2 : // Branch the history, modify relation differently on the new timeline
4253 2 : tenant
4254 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
4255 2 : .await?;
4256 2 : let newtline = tenant
4257 2 : .get_timeline(NEW_TIMELINE_ID, true)
4258 2 : .expect("Should have a local timeline");
4259 2 : let mut new_writer = newtline.writer().await;
4260 2 : new_writer
4261 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
4262 2 : .await?;
4263 2 : new_writer.finish_write(Lsn(0x40));
4264 2 :
4265 2 : // Check page contents on both branches
4266 2 : assert_eq!(
4267 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4268 2 : "foo at 0x40"
4269 2 : );
4270 2 : assert_eq!(
4271 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4272 2 : "bar at 0x40"
4273 2 : );
4274 2 : assert_eq!(
4275 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
4276 2 : "foobar at 0x20"
4277 2 : );
4278 2 :
4279 2 : //assert_current_logical_size(&tline, Lsn(0x40));
4280 2 :
4281 2 : Ok(())
4282 2 : }
4283 :
4284 20 : async fn make_some_layers(
4285 20 : tline: &Timeline,
4286 20 : start_lsn: Lsn,
4287 20 : ctx: &RequestContext,
4288 20 : ) -> anyhow::Result<()> {
4289 20 : let mut lsn = start_lsn;
4290 : {
4291 20 : let mut writer = tline.writer().await;
4292 : // Create a relation on the timeline
4293 20 : writer
4294 20 : .put(
4295 20 : *TEST_KEY,
4296 20 : lsn,
4297 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4298 20 : ctx,
4299 20 : )
4300 10 : .await?;
4301 20 : writer.finish_write(lsn);
4302 20 : lsn += 0x10;
4303 20 : writer
4304 20 : .put(
4305 20 : *TEST_KEY,
4306 20 : lsn,
4307 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4308 20 : ctx,
4309 20 : )
4310 0 : .await?;
4311 20 : writer.finish_write(lsn);
4312 20 : lsn += 0x10;
4313 20 : }
4314 20 : tline.freeze_and_flush().await?;
4315 : {
4316 20 : let mut writer = tline.writer().await;
4317 20 : writer
4318 20 : .put(
4319 20 : *TEST_KEY,
4320 20 : lsn,
4321 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4322 20 : ctx,
4323 20 : )
4324 10 : .await?;
4325 20 : writer.finish_write(lsn);
4326 20 : lsn += 0x10;
4327 20 : writer
4328 20 : .put(
4329 20 : *TEST_KEY,
4330 20 : lsn,
4331 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4332 20 : ctx,
4333 20 : )
4334 0 : .await?;
4335 20 : writer.finish_write(lsn);
4336 20 : }
4337 20 : tline.freeze_and_flush().await.map_err(|e| e.into())
4338 20 : }
4339 :
4340 : #[tokio::test]
4341 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
4342 2 : let (tenant, ctx) =
4343 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
4344 2 : .await?
4345 2 : .load()
4346 8 : .await;
4347 2 : let tline = tenant
4348 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4349 4 : .await?;
4350 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4351 2 :
4352 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4353 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
4354 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
4355 2 : // below should fail.
4356 2 : tenant
4357 2 : .gc_iteration(
4358 2 : Some(TIMELINE_ID),
4359 2 : 0x10,
4360 2 : Duration::ZERO,
4361 2 : &CancellationToken::new(),
4362 2 : &ctx,
4363 2 : )
4364 2 : .await?;
4365 2 :
4366 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
4367 2 : match tenant
4368 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4369 2 : .await
4370 2 : {
4371 2 : Ok(_) => panic!("branching should have failed"),
4372 2 : Err(err) => {
4373 2 : let CreateTimelineError::AncestorLsn(err) = err else {
4374 2 : panic!("wrong error type")
4375 2 : };
4376 2 : assert!(err.to_string().contains("invalid branch start lsn"));
4377 2 : assert!(err
4378 2 : .source()
4379 2 : .unwrap()
4380 2 : .to_string()
4381 2 : .contains("we might've already garbage collected needed data"))
4382 2 : }
4383 2 : }
4384 2 :
4385 2 : Ok(())
4386 2 : }
4387 :
4388 : #[tokio::test]
4389 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
4390 2 : let (tenant, ctx) =
4391 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
4392 2 : .await?
4393 2 : .load()
4394 8 : .await;
4395 2 :
4396 2 : let tline = tenant
4397 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
4398 4 : .await?;
4399 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
4400 2 : match tenant
4401 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4402 2 : .await
4403 2 : {
4404 2 : Ok(_) => panic!("branching should have failed"),
4405 2 : Err(err) => {
4406 2 : let CreateTimelineError::AncestorLsn(err) = err else {
4407 2 : panic!("wrong error type");
4408 2 : };
4409 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
4410 2 : assert!(&err
4411 2 : .source()
4412 2 : .unwrap()
4413 2 : .to_string()
4414 2 : .contains("is earlier than latest GC cutoff"));
4415 2 : }
4416 2 : }
4417 2 :
4418 2 : Ok(())
4419 2 : }
4420 :
4421 : /*
4422 : // FIXME: This currently fails to error out. Calling GC doesn't currently
4423 : // remove the old value, we'd need to work a little harder
4424 : #[tokio::test]
4425 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
4426 : let repo =
4427 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
4428 : .load();
4429 :
4430 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
4431 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4432 :
4433 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
4434 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
4435 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
4436 : match tline.get(*TEST_KEY, Lsn(0x25)) {
4437 : Ok(_) => panic!("request for page should have failed"),
4438 : Err(err) => assert!(err.to_string().contains("not found at")),
4439 : }
4440 : Ok(())
4441 : }
4442 : */
4443 :
4444 : #[tokio::test]
4445 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
4446 2 : let (tenant, ctx) =
4447 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
4448 2 : .await?
4449 2 : .load()
4450 8 : .await;
4451 2 : let tline = tenant
4452 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4453 4 : .await?;
4454 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4455 2 :
4456 2 : tenant
4457 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4458 2 : .await?;
4459 2 : let newtline = tenant
4460 2 : .get_timeline(NEW_TIMELINE_ID, true)
4461 2 : .expect("Should have a local timeline");
4462 2 :
4463 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4464 2 :
4465 2 : tline.set_broken("test".to_owned());
4466 2 :
4467 2 : tenant
4468 2 : .gc_iteration(
4469 2 : Some(TIMELINE_ID),
4470 2 : 0x10,
4471 2 : Duration::ZERO,
4472 2 : &CancellationToken::new(),
4473 2 : &ctx,
4474 2 : )
4475 2 : .await?;
4476 2 :
4477 2 : // The branchpoints should contain all timelines, even ones marked
4478 2 : // as Broken.
4479 2 : {
4480 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
4481 2 : assert_eq!(branchpoints.len(), 1);
4482 2 : assert_eq!(branchpoints[0], (Lsn(0x40), NEW_TIMELINE_ID));
4483 2 : }
4484 2 :
4485 2 : // You can read the key from the child branch even though the parent is
4486 2 : // Broken, as long as you don't need to access data from the parent.
4487 2 : assert_eq!(
4488 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
4489 2 : test_img(&format!("foo at {}", Lsn(0x70)))
4490 2 : );
4491 2 :
4492 2 : // This needs to traverse to the parent, and fails.
4493 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
4494 2 : assert!(
4495 2 : err.to_string().starts_with(&format!(
4496 2 : "bad state on timeline {}: Broken",
4497 2 : tline.timeline_id
4498 2 : )),
4499 2 : "{err}"
4500 2 : );
4501 2 :
4502 2 : Ok(())
4503 2 : }
4504 :
4505 : #[tokio::test]
4506 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
4507 2 : let (tenant, ctx) =
4508 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
4509 2 : .await?
4510 2 : .load()
4511 8 : .await;
4512 2 : let tline = tenant
4513 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4514 4 : .await?;
4515 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4516 2 :
4517 2 : tenant
4518 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4519 2 : .await?;
4520 2 : let newtline = tenant
4521 2 : .get_timeline(NEW_TIMELINE_ID, true)
4522 2 : .expect("Should have a local timeline");
4523 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4524 2 : tenant
4525 2 : .gc_iteration(
4526 2 : Some(TIMELINE_ID),
4527 2 : 0x10,
4528 2 : Duration::ZERO,
4529 2 : &CancellationToken::new(),
4530 2 : &ctx,
4531 2 : )
4532 2 : .await?;
4533 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
4534 2 :
4535 2 : Ok(())
4536 2 : }
4537 : #[tokio::test]
4538 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
4539 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
4540 2 : .await?
4541 2 : .load()
4542 8 : .await;
4543 2 : let tline = tenant
4544 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4545 4 : .await?;
4546 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4547 2 :
4548 2 : tenant
4549 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4550 2 : .await?;
4551 2 : let newtline = tenant
4552 2 : .get_timeline(NEW_TIMELINE_ID, true)
4553 2 : .expect("Should have a local timeline");
4554 2 :
4555 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4556 2 :
4557 2 : // run gc on parent
4558 2 : tenant
4559 2 : .gc_iteration(
4560 2 : Some(TIMELINE_ID),
4561 2 : 0x10,
4562 2 : Duration::ZERO,
4563 2 : &CancellationToken::new(),
4564 2 : &ctx,
4565 2 : )
4566 2 : .await?;
4567 2 :
4568 2 : // Check that the data is still accessible on the branch.
4569 2 : assert_eq!(
4570 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
4571 2 : test_img(&format!("foo at {}", Lsn(0x40)))
4572 2 : );
4573 2 :
4574 2 : Ok(())
4575 2 : }
4576 :
4577 : #[tokio::test]
4578 2 : async fn timeline_load() -> anyhow::Result<()> {
4579 2 : const TEST_NAME: &str = "timeline_load";
4580 2 : let harness = TenantHarness::create(TEST_NAME).await?;
4581 2 : {
4582 8 : let (tenant, ctx) = harness.load().await;
4583 2 : let tline = tenant
4584 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
4585 4 : .await?;
4586 6 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
4587 2 : // so that all uploads finish & we can call harness.load() below again
4588 2 : tenant
4589 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4590 2 : .instrument(harness.span())
4591 2 : .await
4592 2 : .ok()
4593 2 : .unwrap();
4594 2 : }
4595 2 :
4596 7 : let (tenant, _ctx) = harness.load().await;
4597 2 : tenant
4598 2 : .get_timeline(TIMELINE_ID, true)
4599 2 : .expect("cannot load timeline");
4600 2 :
4601 2 : Ok(())
4602 2 : }
4603 :
4604 : #[tokio::test]
4605 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
4606 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
4607 2 : let harness = TenantHarness::create(TEST_NAME).await?;
4608 2 : // create two timelines
4609 2 : {
4610 8 : let (tenant, ctx) = harness.load().await;
4611 2 : let tline = tenant
4612 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4613 4 : .await?;
4614 2 :
4615 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4616 2 :
4617 2 : let child_tline = tenant
4618 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4619 2 : .await?;
4620 2 : child_tline.set_state(TimelineState::Active);
4621 2 :
4622 2 : let newtline = tenant
4623 2 : .get_timeline(NEW_TIMELINE_ID, true)
4624 2 : .expect("Should have a local timeline");
4625 2 :
4626 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4627 2 :
4628 2 : // so that all uploads finish & we can call harness.load() below again
4629 2 : tenant
4630 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4631 2 : .instrument(harness.span())
4632 2 : .await
4633 2 : .ok()
4634 2 : .unwrap();
4635 2 : }
4636 2 :
4637 2 : // check that both of them are initially unloaded
4638 12 : let (tenant, _ctx) = harness.load().await;
4639 2 :
4640 2 : // check that both, child and ancestor are loaded
4641 2 : let _child_tline = tenant
4642 2 : .get_timeline(NEW_TIMELINE_ID, true)
4643 2 : .expect("cannot get child timeline loaded");
4644 2 :
4645 2 : let _ancestor_tline = tenant
4646 2 : .get_timeline(TIMELINE_ID, true)
4647 2 : .expect("cannot get ancestor timeline loaded");
4648 2 :
4649 2 : Ok(())
4650 2 : }
4651 :
4652 : #[tokio::test]
4653 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
4654 2 : use storage_layer::AsLayerDesc;
4655 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
4656 2 : .await?
4657 2 : .load()
4658 8 : .await;
4659 2 : let tline = tenant
4660 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4661 4 : .await?;
4662 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4663 2 :
4664 2 : let layer_map = tline.layers.read().await;
4665 2 : let level0_deltas = layer_map
4666 2 : .layer_map()?
4667 2 : .level0_deltas()
4668 2 : .iter()
4669 4 : .map(|desc| layer_map.get_from_desc(desc))
4670 2 : .collect::<Vec<_>>();
4671 2 :
4672 2 : assert!(!level0_deltas.is_empty());
4673 2 :
4674 6 : for delta in level0_deltas {
4675 2 : // Ensure we are dumping a delta layer here
4676 4 : assert!(delta.layer_desc().is_delta);
4677 8 : delta.dump(true, &ctx).await.unwrap();
4678 2 : }
4679 2 :
4680 2 : Ok(())
4681 2 : }
4682 :
4683 : #[tokio::test]
4684 2 : async fn test_images() -> anyhow::Result<()> {
4685 8 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
4686 2 : let tline = tenant
4687 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4688 4 : .await?;
4689 2 :
4690 2 : let mut writer = tline.writer().await;
4691 2 : writer
4692 2 : .put(
4693 2 : *TEST_KEY,
4694 2 : Lsn(0x10),
4695 2 : &Value::Image(test_img("foo at 0x10")),
4696 2 : &ctx,
4697 2 : )
4698 2 : .await?;
4699 2 : writer.finish_write(Lsn(0x10));
4700 2 : drop(writer);
4701 2 :
4702 2 : tline.freeze_and_flush().await?;
4703 2 : tline
4704 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4705 2 : .await?;
4706 2 :
4707 2 : let mut writer = tline.writer().await;
4708 2 : writer
4709 2 : .put(
4710 2 : *TEST_KEY,
4711 2 : Lsn(0x20),
4712 2 : &Value::Image(test_img("foo at 0x20")),
4713 2 : &ctx,
4714 2 : )
4715 2 : .await?;
4716 2 : writer.finish_write(Lsn(0x20));
4717 2 : drop(writer);
4718 2 :
4719 2 : tline.freeze_and_flush().await?;
4720 2 : tline
4721 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4722 2 : .await?;
4723 2 :
4724 2 : let mut writer = tline.writer().await;
4725 2 : writer
4726 2 : .put(
4727 2 : *TEST_KEY,
4728 2 : Lsn(0x30),
4729 2 : &Value::Image(test_img("foo at 0x30")),
4730 2 : &ctx,
4731 2 : )
4732 2 : .await?;
4733 2 : writer.finish_write(Lsn(0x30));
4734 2 : drop(writer);
4735 2 :
4736 2 : tline.freeze_and_flush().await?;
4737 2 : tline
4738 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4739 2 : .await?;
4740 2 :
4741 2 : let mut writer = tline.writer().await;
4742 2 : writer
4743 2 : .put(
4744 2 : *TEST_KEY,
4745 2 : Lsn(0x40),
4746 2 : &Value::Image(test_img("foo at 0x40")),
4747 2 : &ctx,
4748 2 : )
4749 2 : .await?;
4750 2 : writer.finish_write(Lsn(0x40));
4751 2 : drop(writer);
4752 2 :
4753 2 : tline.freeze_and_flush().await?;
4754 2 : tline
4755 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4756 2 : .await?;
4757 2 :
4758 2 : assert_eq!(
4759 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4760 2 : test_img("foo at 0x10")
4761 2 : );
4762 2 : assert_eq!(
4763 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4764 2 : test_img("foo at 0x10")
4765 2 : );
4766 2 : assert_eq!(
4767 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4768 2 : test_img("foo at 0x20")
4769 2 : );
4770 2 : assert_eq!(
4771 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
4772 2 : test_img("foo at 0x30")
4773 2 : );
4774 2 : assert_eq!(
4775 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
4776 2 : test_img("foo at 0x40")
4777 2 : );
4778 2 :
4779 2 : Ok(())
4780 2 : }
4781 :
4782 4 : async fn bulk_insert_compact_gc(
4783 4 : tenant: &Tenant,
4784 4 : timeline: &Arc<Timeline>,
4785 4 : ctx: &RequestContext,
4786 4 : lsn: Lsn,
4787 4 : repeat: usize,
4788 4 : key_count: usize,
4789 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
4790 4 : let compact = true;
4791 75580 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
4792 4 : }
4793 :
4794 8 : async fn bulk_insert_maybe_compact_gc(
4795 8 : tenant: &Tenant,
4796 8 : timeline: &Arc<Timeline>,
4797 8 : ctx: &RequestContext,
4798 8 : mut lsn: Lsn,
4799 8 : repeat: usize,
4800 8 : key_count: usize,
4801 8 : compact: bool,
4802 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
4803 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
4804 8 :
4805 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4806 8 : let mut blknum = 0;
4807 8 :
4808 8 : // Enforce that key range is monotonously increasing
4809 8 : let mut keyspace = KeySpaceAccum::new();
4810 8 :
4811 8 : let cancel = CancellationToken::new();
4812 8 :
4813 8 : for _ in 0..repeat {
4814 400 : for _ in 0..key_count {
4815 4000000 : test_key.field6 = blknum;
4816 4000000 : let mut writer = timeline.writer().await;
4817 4000000 : writer
4818 4000000 : .put(
4819 4000000 : test_key,
4820 4000000 : lsn,
4821 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4822 4000000 : ctx,
4823 4000000 : )
4824 2556 : .await?;
4825 4000000 : inserted.entry(test_key).or_default().insert(lsn);
4826 4000000 : writer.finish_write(lsn);
4827 4000000 : drop(writer);
4828 4000000 :
4829 4000000 : keyspace.add_key(test_key);
4830 4000000 :
4831 4000000 : lsn = Lsn(lsn.0 + 0x10);
4832 4000000 : blknum += 1;
4833 : }
4834 :
4835 400 : timeline.freeze_and_flush().await?;
4836 400 : if compact {
4837 : // this requires timeline to be &Arc<Timeline>
4838 43180 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
4839 200 : }
4840 :
4841 : // this doesn't really need to use the timeline_id target, but it is closer to what it
4842 : // originally was.
4843 400 : let res = tenant
4844 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
4845 399 : .await?;
4846 :
4847 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
4848 : }
4849 :
4850 8 : Ok(inserted)
4851 8 : }
4852 :
4853 : //
4854 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
4855 : // Repeat 50 times.
4856 : //
4857 : #[tokio::test]
4858 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
4859 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
4860 8 : let (tenant, ctx) = harness.load().await;
4861 2 : let tline = tenant
4862 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4863 4 : .await?;
4864 2 :
4865 2 : let lsn = Lsn(0x10);
4866 37790 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4867 2 :
4868 2 : Ok(())
4869 2 : }
4870 :
4871 : // Test the vectored get real implementation against a simple sequential implementation.
4872 : //
4873 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
4874 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
4875 : // grow to the right on the X axis.
4876 : // [Delta]
4877 : // [Delta]
4878 : // [Delta]
4879 : // [Delta]
4880 : // ------------ Image ---------------
4881 : //
4882 : // After layer generation we pick the ranges to query as follows:
4883 : // 1. The beginning of each delta layer
4884 : // 2. At the seam between two adjacent delta layers
4885 : //
4886 : // There's one major downside to this test: delta layers only contains images,
4887 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
4888 : #[tokio::test]
4889 2 : async fn test_get_vectored() -> anyhow::Result<()> {
4890 2 : let harness = TenantHarness::create("test_get_vectored").await?;
4891 8 : let (tenant, ctx) = harness.load().await;
4892 2 : let tline = tenant
4893 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4894 4 : .await?;
4895 2 :
4896 2 : let lsn = Lsn(0x10);
4897 37790 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4898 2 :
4899 2 : let guard = tline.layers.read().await;
4900 2 : let lm = guard.layer_map()?;
4901 2 :
4902 2 : lm.dump(true, &ctx).await?;
4903 2 :
4904 2 : let mut reads = Vec::new();
4905 2 : let mut prev = None;
4906 12 : lm.iter_historic_layers().for_each(|desc| {
4907 12 : if !desc.is_delta() {
4908 2 : prev = Some(desc.clone());
4909 2 : return;
4910 10 : }
4911 10 :
4912 10 : let start = desc.key_range.start;
4913 10 : let end = desc
4914 10 : .key_range
4915 10 : .start
4916 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
4917 10 : reads.push(KeySpace {
4918 10 : ranges: vec![start..end],
4919 10 : });
4920 2 :
4921 10 : if let Some(prev) = &prev {
4922 10 : if !prev.is_delta() {
4923 10 : return;
4924 2 : }
4925 0 :
4926 0 : let first_range = Key {
4927 0 : field6: prev.key_range.end.field6 - 4,
4928 0 : ..prev.key_range.end
4929 0 : }..prev.key_range.end;
4930 0 :
4931 0 : let second_range = desc.key_range.start..Key {
4932 0 : field6: desc.key_range.start.field6 + 4,
4933 0 : ..desc.key_range.start
4934 0 : };
4935 0 :
4936 0 : reads.push(KeySpace {
4937 0 : ranges: vec![first_range, second_range],
4938 0 : });
4939 2 : };
4940 2 :
4941 2 : prev = Some(desc.clone());
4942 12 : });
4943 2 :
4944 2 : drop(guard);
4945 2 :
4946 2 : // Pick a big LSN such that we query over all the changes.
4947 2 : let reads_lsn = Lsn(u64::MAX - 1);
4948 2 :
4949 12 : for read in reads {
4950 10 : info!("Doing vectored read on {:?}", read);
4951 2 :
4952 10 : let vectored_res = tline
4953 10 : .get_vectored_impl(
4954 10 : read.clone(),
4955 10 : reads_lsn,
4956 10 : &mut ValuesReconstructState::new(),
4957 10 : &ctx,
4958 10 : )
4959 25 : .await;
4960 2 :
4961 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
4962 10 : let mut expect_missing = false;
4963 10 : let mut key = read.start().unwrap();
4964 330 : while key != read.end().unwrap() {
4965 320 : if let Some(lsns) = inserted.get(&key) {
4966 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
4967 320 : match expected_lsn {
4968 320 : Some(lsn) => {
4969 320 : expected_lsns.insert(key, *lsn);
4970 320 : }
4971 2 : None => {
4972 2 : expect_missing = true;
4973 0 : break;
4974 2 : }
4975 2 : }
4976 2 : } else {
4977 2 : expect_missing = true;
4978 0 : break;
4979 2 : }
4980 2 :
4981 320 : key = key.next();
4982 2 : }
4983 2 :
4984 10 : if expect_missing {
4985 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
4986 2 : } else {
4987 320 : for (key, image) in vectored_res? {
4988 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
4989 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
4990 320 : assert_eq!(image?, expected_image);
4991 2 : }
4992 2 : }
4993 2 : }
4994 2 :
4995 2 : Ok(())
4996 2 : }
4997 :
4998 : #[tokio::test]
4999 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
5000 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
5001 2 :
5002 8 : let (tenant, ctx) = harness.load().await;
5003 2 : let tline = tenant
5004 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5005 2 : .await?;
5006 2 : let tline = tline.raw_timeline().unwrap();
5007 2 :
5008 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
5009 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
5010 2 : modification.set_lsn(Lsn(0x1008))?;
5011 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
5012 2 : modification.commit(&ctx).await?;
5013 2 :
5014 2 : let child_timeline_id = TimelineId::generate();
5015 2 : tenant
5016 2 : .branch_timeline_test(
5017 2 : tline,
5018 2 : child_timeline_id,
5019 2 : Some(tline.get_last_record_lsn()),
5020 2 : &ctx,
5021 2 : )
5022 2 : .await?;
5023 2 :
5024 2 : let child_timeline = tenant
5025 2 : .get_timeline(child_timeline_id, true)
5026 2 : .expect("Should have the branched timeline");
5027 2 :
5028 2 : let aux_keyspace = KeySpace {
5029 2 : ranges: vec![NON_INHERITED_RANGE],
5030 2 : };
5031 2 : let read_lsn = child_timeline.get_last_record_lsn();
5032 2 :
5033 2 : let vectored_res = child_timeline
5034 2 : .get_vectored_impl(
5035 2 : aux_keyspace.clone(),
5036 2 : read_lsn,
5037 2 : &mut ValuesReconstructState::new(),
5038 2 : &ctx,
5039 2 : )
5040 2 : .await;
5041 2 :
5042 2 : let images = vectored_res?;
5043 2 : assert!(images.is_empty());
5044 2 : Ok(())
5045 2 : }
5046 :
5047 : // Test that vectored get handles layer gaps correctly
5048 : // by advancing into the next ancestor timeline if required.
5049 : //
5050 : // The test generates timelines that look like the diagram below.
5051 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
5052 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
5053 : //
5054 : // ```
5055 : //-------------------------------+
5056 : // ... |
5057 : // [ L1 ] |
5058 : // [ / L1 ] | Child Timeline
5059 : // ... |
5060 : // ------------------------------+
5061 : // [ X L1 ] | Parent Timeline
5062 : // ------------------------------+
5063 : // ```
5064 : #[tokio::test]
5065 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
5066 2 : let tenant_conf = TenantConf {
5067 2 : // Make compaction deterministic
5068 2 : gc_period: Duration::ZERO,
5069 2 : compaction_period: Duration::ZERO,
5070 2 : // Encourage creation of L1 layers
5071 2 : checkpoint_distance: 16 * 1024,
5072 2 : compaction_target_size: 8 * 1024,
5073 2 : ..TenantConf::default()
5074 2 : };
5075 2 :
5076 2 : let harness = TenantHarness::create_custom(
5077 2 : "test_get_vectored_key_gap",
5078 2 : tenant_conf,
5079 2 : TenantId::generate(),
5080 2 : ShardIdentity::unsharded(),
5081 2 : Generation::new(0xdeadbeef),
5082 2 : )
5083 2 : .await?;
5084 8 : let (tenant, ctx) = harness.load().await;
5085 2 :
5086 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5087 2 : let gap_at_key = current_key.add(100);
5088 2 : let mut current_lsn = Lsn(0x10);
5089 2 :
5090 2 : const KEY_COUNT: usize = 10_000;
5091 2 :
5092 2 : let timeline_id = TimelineId::generate();
5093 2 : let current_timeline = tenant
5094 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5095 4 : .await?;
5096 2 :
5097 2 : current_lsn += 0x100;
5098 2 :
5099 2 : let mut writer = current_timeline.writer().await;
5100 2 : writer
5101 2 : .put(
5102 2 : gap_at_key,
5103 2 : current_lsn,
5104 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
5105 2 : &ctx,
5106 2 : )
5107 2 : .await?;
5108 2 : writer.finish_write(current_lsn);
5109 2 : drop(writer);
5110 2 :
5111 2 : let mut latest_lsns = HashMap::new();
5112 2 : latest_lsns.insert(gap_at_key, current_lsn);
5113 2 :
5114 2 : current_timeline.freeze_and_flush().await?;
5115 2 :
5116 2 : let child_timeline_id = TimelineId::generate();
5117 2 :
5118 2 : tenant
5119 2 : .branch_timeline_test(
5120 2 : ¤t_timeline,
5121 2 : child_timeline_id,
5122 2 : Some(current_lsn),
5123 2 : &ctx,
5124 2 : )
5125 2 : .await?;
5126 2 : let child_timeline = tenant
5127 2 : .get_timeline(child_timeline_id, true)
5128 2 : .expect("Should have the branched timeline");
5129 2 :
5130 20002 : for i in 0..KEY_COUNT {
5131 20000 : if current_key == gap_at_key {
5132 2 : current_key = current_key.next();
5133 2 : continue;
5134 19998 : }
5135 19998 :
5136 19998 : current_lsn += 0x10;
5137 2 :
5138 19998 : let mut writer = child_timeline.writer().await;
5139 19998 : writer
5140 19998 : .put(
5141 19998 : current_key,
5142 19998 : current_lsn,
5143 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
5144 19998 : &ctx,
5145 19998 : )
5146 65 : .await?;
5147 19998 : writer.finish_write(current_lsn);
5148 19998 : drop(writer);
5149 19998 :
5150 19998 : latest_lsns.insert(current_key, current_lsn);
5151 19998 : current_key = current_key.next();
5152 19998 :
5153 19998 : // Flush every now and then to encourage layer file creation.
5154 19998 : if i % 500 == 0 {
5155 42 : child_timeline.freeze_and_flush().await?;
5156 19958 : }
5157 2 : }
5158 2 :
5159 3 : child_timeline.freeze_and_flush().await?;
5160 2 : let mut flags = EnumSet::new();
5161 2 : flags.insert(CompactFlags::ForceRepartition);
5162 2 : child_timeline
5163 2 : .compact(&CancellationToken::new(), flags, &ctx)
5164 2327 : .await?;
5165 2 :
5166 2 : let key_near_end = {
5167 2 : let mut tmp = current_key;
5168 2 : tmp.field6 -= 10;
5169 2 : tmp
5170 2 : };
5171 2 :
5172 2 : let key_near_gap = {
5173 2 : let mut tmp = gap_at_key;
5174 2 : tmp.field6 -= 10;
5175 2 : tmp
5176 2 : };
5177 2 :
5178 2 : let read = KeySpace {
5179 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
5180 2 : };
5181 2 : let results = child_timeline
5182 2 : .get_vectored_impl(
5183 2 : read.clone(),
5184 2 : current_lsn,
5185 2 : &mut ValuesReconstructState::new(),
5186 2 : &ctx,
5187 2 : )
5188 15 : .await?;
5189 2 :
5190 44 : for (key, img_res) in results {
5191 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
5192 42 : assert_eq!(img_res?, expected);
5193 2 : }
5194 2 :
5195 2 : Ok(())
5196 2 : }
5197 :
5198 : // Test that vectored get descends into ancestor timelines correctly and
5199 : // does not return an image that's newer than requested.
5200 : //
5201 : // The diagram below ilustrates an interesting case. We have a parent timeline
5202 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
5203 : // from the child timeline, so the parent timeline must be visited. When advacing into
5204 : // the child timeline, the read path needs to remember what the requested Lsn was in
5205 : // order to avoid returning an image that's too new. The test below constructs such
5206 : // a timeline setup and does a few queries around the Lsn of each page image.
5207 : // ```
5208 : // LSN
5209 : // ^
5210 : // |
5211 : // |
5212 : // 500 | --------------------------------------> branch point
5213 : // 400 | X
5214 : // 300 | X
5215 : // 200 | --------------------------------------> requested lsn
5216 : // 100 | X
5217 : // |---------------------------------------> Key
5218 : // |
5219 : // ------> requested key
5220 : //
5221 : // Legend:
5222 : // * X - page images
5223 : // ```
5224 : #[tokio::test]
5225 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
5226 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
5227 8 : let (tenant, ctx) = harness.load().await;
5228 2 :
5229 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5230 2 : let end_key = start_key.add(1000);
5231 2 : let child_gap_at_key = start_key.add(500);
5232 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
5233 2 :
5234 2 : let mut current_lsn = Lsn(0x10);
5235 2 :
5236 2 : let timeline_id = TimelineId::generate();
5237 2 : let parent_timeline = tenant
5238 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5239 4 : .await?;
5240 2 :
5241 2 : current_lsn += 0x100;
5242 2 :
5243 8 : for _ in 0..3 {
5244 6 : let mut key = start_key;
5245 6006 : while key < end_key {
5246 6000 : current_lsn += 0x10;
5247 6000 :
5248 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
5249 2 :
5250 6000 : let mut writer = parent_timeline.writer().await;
5251 6000 : writer
5252 6000 : .put(
5253 6000 : key,
5254 6000 : current_lsn,
5255 6000 : &Value::Image(test_img(&image_value)),
5256 6000 : &ctx,
5257 6000 : )
5258 6 : .await?;
5259 6000 : writer.finish_write(current_lsn);
5260 6000 :
5261 6000 : if key == child_gap_at_key {
5262 6 : parent_gap_lsns.insert(current_lsn, image_value);
5263 5994 : }
5264 2 :
5265 6000 : key = key.next();
5266 2 : }
5267 2 :
5268 6 : parent_timeline.freeze_and_flush().await?;
5269 2 : }
5270 2 :
5271 2 : let child_timeline_id = TimelineId::generate();
5272 2 :
5273 2 : let child_timeline = tenant
5274 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
5275 2 : .await?;
5276 2 :
5277 2 : let mut key = start_key;
5278 2002 : while key < end_key {
5279 2000 : if key == child_gap_at_key {
5280 2 : key = key.next();
5281 2 : continue;
5282 1998 : }
5283 1998 :
5284 1998 : current_lsn += 0x10;
5285 2 :
5286 1998 : let mut writer = child_timeline.writer().await;
5287 1998 : writer
5288 1998 : .put(
5289 1998 : key,
5290 1998 : current_lsn,
5291 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
5292 1998 : &ctx,
5293 1998 : )
5294 17 : .await?;
5295 1998 : writer.finish_write(current_lsn);
5296 1998 :
5297 1998 : key = key.next();
5298 2 : }
5299 2 :
5300 2 : child_timeline.freeze_and_flush().await?;
5301 2 :
5302 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
5303 2 : let mut query_lsns = Vec::new();
5304 6 : for image_lsn in parent_gap_lsns.keys().rev() {
5305 36 : for offset in lsn_offsets {
5306 30 : query_lsns.push(Lsn(image_lsn
5307 30 : .0
5308 30 : .checked_add_signed(offset)
5309 30 : .expect("Shouldn't overflow")));
5310 30 : }
5311 2 : }
5312 2 :
5313 32 : for query_lsn in query_lsns {
5314 30 : let results = child_timeline
5315 30 : .get_vectored_impl(
5316 30 : KeySpace {
5317 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
5318 30 : },
5319 30 : query_lsn,
5320 30 : &mut ValuesReconstructState::new(),
5321 30 : &ctx,
5322 30 : )
5323 29 : .await;
5324 2 :
5325 30 : let expected_item = parent_gap_lsns
5326 30 : .iter()
5327 30 : .rev()
5328 68 : .find(|(lsn, _)| **lsn <= query_lsn);
5329 30 :
5330 30 : info!(
5331 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
5332 2 : query_lsn, expected_item
5333 2 : );
5334 2 :
5335 30 : match expected_item {
5336 26 : Some((_, img_value)) => {
5337 26 : let key_results = results.expect("No vectored get error expected");
5338 26 : let key_result = &key_results[&child_gap_at_key];
5339 26 : let returned_img = key_result
5340 26 : .as_ref()
5341 26 : .expect("No page reconstruct error expected");
5342 26 :
5343 26 : info!(
5344 2 : "Vectored read at LSN {} returned image {}",
5345 0 : query_lsn,
5346 0 : std::str::from_utf8(returned_img)?
5347 2 : );
5348 26 : assert_eq!(*returned_img, test_img(img_value));
5349 2 : }
5350 2 : None => {
5351 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
5352 2 : }
5353 2 : }
5354 2 : }
5355 2 :
5356 2 : Ok(())
5357 2 : }
5358 :
5359 : #[tokio::test]
5360 2 : async fn test_random_updates() -> anyhow::Result<()> {
5361 2 : let names_algorithms = [
5362 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
5363 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
5364 2 : ];
5365 6 : for (name, algorithm) in names_algorithms {
5366 46484 : test_random_updates_algorithm(name, algorithm).await?;
5367 2 : }
5368 2 : Ok(())
5369 2 : }
5370 :
5371 4 : async fn test_random_updates_algorithm(
5372 4 : name: &'static str,
5373 4 : compaction_algorithm: CompactionAlgorithm,
5374 4 : ) -> anyhow::Result<()> {
5375 4 : let mut harness = TenantHarness::create(name).await?;
5376 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5377 4 : kind: compaction_algorithm,
5378 4 : };
5379 16 : let (tenant, ctx) = harness.load().await;
5380 4 : let tline = tenant
5381 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5382 7 : .await?;
5383 :
5384 : const NUM_KEYS: usize = 1000;
5385 4 : let cancel = CancellationToken::new();
5386 4 :
5387 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5388 4 : let mut test_key_end = test_key;
5389 4 : test_key_end.field6 = NUM_KEYS as u32;
5390 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
5391 4 :
5392 4 : let mut keyspace = KeySpaceAccum::new();
5393 4 :
5394 4 : // Track when each page was last modified. Used to assert that
5395 4 : // a read sees the latest page version.
5396 4 : let mut updated = [Lsn(0); NUM_KEYS];
5397 4 :
5398 4 : let mut lsn = Lsn(0x10);
5399 : #[allow(clippy::needless_range_loop)]
5400 4004 : for blknum in 0..NUM_KEYS {
5401 4000 : lsn = Lsn(lsn.0 + 0x10);
5402 4000 : test_key.field6 = blknum as u32;
5403 4000 : let mut writer = tline.writer().await;
5404 4000 : writer
5405 4000 : .put(
5406 4000 : test_key,
5407 4000 : lsn,
5408 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5409 4000 : &ctx,
5410 4000 : )
5411 4 : .await?;
5412 4000 : writer.finish_write(lsn);
5413 4000 : updated[blknum] = lsn;
5414 4000 : drop(writer);
5415 4000 :
5416 4000 : keyspace.add_key(test_key);
5417 : }
5418 :
5419 204 : for _ in 0..50 {
5420 200200 : for _ in 0..NUM_KEYS {
5421 200000 : lsn = Lsn(lsn.0 + 0x10);
5422 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5423 200000 : test_key.field6 = blknum as u32;
5424 200000 : let mut writer = tline.writer().await;
5425 200000 : writer
5426 200000 : .put(
5427 200000 : test_key,
5428 200000 : lsn,
5429 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5430 200000 : &ctx,
5431 200000 : )
5432 258 : .await?;
5433 200000 : writer.finish_write(lsn);
5434 200000 : drop(writer);
5435 200000 : updated[blknum] = lsn;
5436 : }
5437 :
5438 : // Read all the blocks
5439 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5440 200000 : test_key.field6 = blknum as u32;
5441 200000 : assert_eq!(
5442 200000 : tline.get(test_key, lsn, &ctx).await?,
5443 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
5444 : );
5445 : }
5446 :
5447 : // Perform a cycle of flush, and GC
5448 200 : tline.freeze_and_flush().await?;
5449 200 : tenant
5450 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5451 184 : .await?;
5452 : }
5453 :
5454 4 : Ok(())
5455 4 : }
5456 :
5457 : #[tokio::test]
5458 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
5459 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
5460 2 : .await?
5461 2 : .load()
5462 8 : .await;
5463 2 : let mut tline = tenant
5464 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5465 4 : .await?;
5466 2 :
5467 2 : const NUM_KEYS: usize = 1000;
5468 2 :
5469 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5470 2 :
5471 2 : let mut keyspace = KeySpaceAccum::new();
5472 2 :
5473 2 : let cancel = CancellationToken::new();
5474 2 :
5475 2 : // Track when each page was last modified. Used to assert that
5476 2 : // a read sees the latest page version.
5477 2 : let mut updated = [Lsn(0); NUM_KEYS];
5478 2 :
5479 2 : let mut lsn = Lsn(0x10);
5480 2 : #[allow(clippy::needless_range_loop)]
5481 2002 : for blknum in 0..NUM_KEYS {
5482 2000 : lsn = Lsn(lsn.0 + 0x10);
5483 2000 : test_key.field6 = blknum as u32;
5484 2000 : let mut writer = tline.writer().await;
5485 2000 : writer
5486 2000 : .put(
5487 2000 : test_key,
5488 2000 : lsn,
5489 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5490 2000 : &ctx,
5491 2000 : )
5492 2 : .await?;
5493 2000 : writer.finish_write(lsn);
5494 2000 : updated[blknum] = lsn;
5495 2000 : drop(writer);
5496 2000 :
5497 2000 : keyspace.add_key(test_key);
5498 2 : }
5499 2 :
5500 102 : for _ in 0..50 {
5501 100 : let new_tline_id = TimelineId::generate();
5502 100 : tenant
5503 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5504 2 : .await?;
5505 100 : tline = tenant
5506 100 : .get_timeline(new_tline_id, true)
5507 100 : .expect("Should have the branched timeline");
5508 2 :
5509 100100 : for _ in 0..NUM_KEYS {
5510 100000 : lsn = Lsn(lsn.0 + 0x10);
5511 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5512 100000 : test_key.field6 = blknum as u32;
5513 100000 : let mut writer = tline.writer().await;
5514 100000 : writer
5515 100000 : .put(
5516 100000 : test_key,
5517 100000 : lsn,
5518 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5519 100000 : &ctx,
5520 100000 : )
5521 836 : .await?;
5522 100000 : println!("updating {} at {}", blknum, lsn);
5523 100000 : writer.finish_write(lsn);
5524 100000 : drop(writer);
5525 100000 : updated[blknum] = lsn;
5526 2 : }
5527 2 :
5528 2 : // Read all the blocks
5529 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5530 100000 : test_key.field6 = blknum as u32;
5531 100000 : assert_eq!(
5532 100000 : tline.get(test_key, lsn, &ctx).await?,
5533 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
5534 2 : );
5535 2 : }
5536 2 :
5537 2 : // Perform a cycle of flush, compact, and GC
5538 100 : tline.freeze_and_flush().await?;
5539 15488 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
5540 100 : tenant
5541 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5542 89 : .await?;
5543 2 : }
5544 2 :
5545 2 : Ok(())
5546 2 : }
5547 :
5548 : #[tokio::test]
5549 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
5550 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
5551 2 : .await?
5552 2 : .load()
5553 8 : .await;
5554 2 : let mut tline = tenant
5555 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5556 4 : .await?;
5557 2 :
5558 2 : const NUM_KEYS: usize = 100;
5559 2 : const NUM_TLINES: usize = 50;
5560 2 :
5561 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5562 2 : // Track page mutation lsns across different timelines.
5563 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
5564 2 :
5565 2 : let mut lsn = Lsn(0x10);
5566 2 :
5567 2 : #[allow(clippy::needless_range_loop)]
5568 102 : for idx in 0..NUM_TLINES {
5569 100 : let new_tline_id = TimelineId::generate();
5570 100 : tenant
5571 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5572 2 : .await?;
5573 100 : tline = tenant
5574 100 : .get_timeline(new_tline_id, true)
5575 100 : .expect("Should have the branched timeline");
5576 2 :
5577 10100 : for _ in 0..NUM_KEYS {
5578 10000 : lsn = Lsn(lsn.0 + 0x10);
5579 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5580 10000 : test_key.field6 = blknum as u32;
5581 10000 : let mut writer = tline.writer().await;
5582 10000 : writer
5583 10000 : .put(
5584 10000 : test_key,
5585 10000 : lsn,
5586 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
5587 10000 : &ctx,
5588 10000 : )
5589 88 : .await?;
5590 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
5591 10000 : writer.finish_write(lsn);
5592 10000 : drop(writer);
5593 10000 : updated[idx][blknum] = lsn;
5594 2 : }
5595 2 : }
5596 2 :
5597 2 : // Read pages from leaf timeline across all ancestors.
5598 100 : for (idx, lsns) in updated.iter().enumerate() {
5599 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
5600 2 : // Skip empty mutations.
5601 10000 : if lsn.0 == 0 {
5602 3711 : continue;
5603 6289 : }
5604 6289 : println!("checking [{idx}][{blknum}] at {lsn}");
5605 6289 : test_key.field6 = blknum as u32;
5606 6289 : assert_eq!(
5607 6289 : tline.get(test_key, *lsn, &ctx).await?,
5608 6289 : test_img(&format!("{idx} {blknum} at {lsn}"))
5609 2 : );
5610 2 : }
5611 2 : }
5612 2 : Ok(())
5613 2 : }
5614 :
5615 : #[tokio::test]
5616 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
5617 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
5618 2 : .await?
5619 2 : .load()
5620 8 : .await;
5621 2 :
5622 2 : let initdb_lsn = Lsn(0x20);
5623 2 : let utline = tenant
5624 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
5625 2 : .await?;
5626 2 : let tline = utline.raw_timeline().unwrap();
5627 2 :
5628 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
5629 2 : tline.maybe_spawn_flush_loop();
5630 2 :
5631 2 : // Make sure the timeline has the minimum set of required keys for operation.
5632 2 : // The only operation you can always do on an empty timeline is to `put` new data.
5633 2 : // Except if you `put` at `initdb_lsn`.
5634 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
5635 2 : // It uses `repartition()`, which assumes some keys to be present.
5636 2 : // Let's make sure the test timeline can handle that case.
5637 2 : {
5638 2 : let mut state = tline.flush_loop_state.lock().unwrap();
5639 2 : assert_eq!(
5640 2 : timeline::FlushLoopState::Running {
5641 2 : expect_initdb_optimization: false,
5642 2 : initdb_optimization_count: 0,
5643 2 : },
5644 2 : *state
5645 2 : );
5646 2 : *state = timeline::FlushLoopState::Running {
5647 2 : expect_initdb_optimization: true,
5648 2 : initdb_optimization_count: 0,
5649 2 : };
5650 2 : }
5651 2 :
5652 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
5653 2 : // As explained above, the optimization requires some keys to be present.
5654 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
5655 2 : // This is what `create_test_timeline` does, by the way.
5656 2 : let mut modification = tline.begin_modification(initdb_lsn);
5657 2 : modification
5658 2 : .init_empty_test_timeline()
5659 2 : .context("init_empty_test_timeline")?;
5660 2 : modification
5661 2 : .commit(&ctx)
5662 2 : .await
5663 2 : .context("commit init_empty_test_timeline modification")?;
5664 2 :
5665 2 : // Do the flush. The flush code will check the expectations that we set above.
5666 2 : tline.freeze_and_flush().await?;
5667 2 :
5668 2 : // assert freeze_and_flush exercised the initdb optimization
5669 2 : {
5670 2 : let state = tline.flush_loop_state.lock().unwrap();
5671 2 : let timeline::FlushLoopState::Running {
5672 2 : expect_initdb_optimization,
5673 2 : initdb_optimization_count,
5674 2 : } = *state
5675 2 : else {
5676 2 : panic!("unexpected state: {:?}", *state);
5677 2 : };
5678 2 : assert!(expect_initdb_optimization);
5679 2 : assert!(initdb_optimization_count > 0);
5680 2 : }
5681 2 : Ok(())
5682 2 : }
5683 :
5684 : #[tokio::test]
5685 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
5686 2 : let name = "test_create_guard_crash";
5687 2 : let harness = TenantHarness::create(name).await?;
5688 2 : {
5689 8 : let (tenant, ctx) = harness.load().await;
5690 2 : let tline = tenant
5691 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5692 2 : .await?;
5693 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
5694 2 : let raw_tline = tline.raw_timeline().unwrap();
5695 2 : raw_tline
5696 2 : .shutdown(super::timeline::ShutdownMode::Hard)
5697 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))
5698 2 : .await;
5699 2 : std::mem::forget(tline);
5700 2 : }
5701 2 :
5702 8 : let (tenant, _) = harness.load().await;
5703 2 : match tenant.get_timeline(TIMELINE_ID, false) {
5704 2 : Ok(_) => panic!("timeline should've been removed during load"),
5705 2 : Err(e) => {
5706 2 : assert_eq!(
5707 2 : e,
5708 2 : GetTimelineError::NotFound {
5709 2 : tenant_id: tenant.tenant_shard_id,
5710 2 : timeline_id: TIMELINE_ID,
5711 2 : }
5712 2 : )
5713 2 : }
5714 2 : }
5715 2 :
5716 2 : assert!(!harness
5717 2 : .conf
5718 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
5719 2 : .exists());
5720 2 :
5721 2 : Ok(())
5722 2 : }
5723 :
5724 : #[tokio::test]
5725 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
5726 2 : let names_algorithms = [
5727 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
5728 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
5729 2 : ];
5730 6 : for (name, algorithm) in names_algorithms {
5731 32832 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
5732 2 : }
5733 2 : Ok(())
5734 2 : }
5735 :
5736 4 : async fn test_read_at_max_lsn_algorithm(
5737 4 : name: &'static str,
5738 4 : compaction_algorithm: CompactionAlgorithm,
5739 4 : ) -> anyhow::Result<()> {
5740 4 : let mut harness = TenantHarness::create(name).await?;
5741 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5742 4 : kind: compaction_algorithm,
5743 4 : };
5744 16 : let (tenant, ctx) = harness.load().await;
5745 4 : let tline = tenant
5746 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5747 7 : .await?;
5748 :
5749 4 : let lsn = Lsn(0x10);
5750 4 : let compact = false;
5751 32399 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
5752 :
5753 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5754 4 : let read_lsn = Lsn(u64::MAX - 1);
5755 :
5756 410 : let result = tline.get(test_key, read_lsn, &ctx).await;
5757 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
5758 :
5759 4 : Ok(())
5760 4 : }
5761 :
5762 : #[tokio::test]
5763 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
5764 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
5765 8 : let (tenant, ctx) = harness.load().await;
5766 2 : let tline = tenant
5767 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5768 4 : .await?;
5769 2 :
5770 2 : const NUM_KEYS: usize = 1000;
5771 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
5772 2 :
5773 2 : let cancel = CancellationToken::new();
5774 2 :
5775 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5776 2 : base_key.field1 = AUX_KEY_PREFIX;
5777 2 : let mut test_key = base_key;
5778 2 :
5779 2 : // Track when each page was last modified. Used to assert that
5780 2 : // a read sees the latest page version.
5781 2 : let mut updated = [Lsn(0); NUM_KEYS];
5782 2 :
5783 2 : let mut lsn = Lsn(0x10);
5784 2 : #[allow(clippy::needless_range_loop)]
5785 2002 : for blknum in 0..NUM_KEYS {
5786 2000 : lsn = Lsn(lsn.0 + 0x10);
5787 2000 : test_key.field6 = (blknum * STEP) as u32;
5788 2000 : let mut writer = tline.writer().await;
5789 2000 : writer
5790 2000 : .put(
5791 2000 : test_key,
5792 2000 : lsn,
5793 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5794 2000 : &ctx,
5795 2000 : )
5796 2 : .await?;
5797 2000 : writer.finish_write(lsn);
5798 2000 : updated[blknum] = lsn;
5799 2000 : drop(writer);
5800 2 : }
5801 2 :
5802 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
5803 2 :
5804 24 : for iter in 0..=10 {
5805 2 : // Read all the blocks
5806 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5807 22000 : test_key.field6 = (blknum * STEP) as u32;
5808 22000 : assert_eq!(
5809 22000 : tline.get(test_key, lsn, &ctx).await?,
5810 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
5811 2 : );
5812 2 : }
5813 2 :
5814 22 : let mut cnt = 0;
5815 22000 : for (key, value) in tline
5816 22 : .get_vectored_impl(
5817 22 : keyspace.clone(),
5818 22 : lsn,
5819 22 : &mut ValuesReconstructState::default(),
5820 22 : &ctx,
5821 22 : )
5822 5590 : .await?
5823 2 : {
5824 22000 : let blknum = key.field6 as usize;
5825 22000 : let value = value?;
5826 22000 : assert!(blknum % STEP == 0);
5827 22000 : let blknum = blknum / STEP;
5828 22000 : assert_eq!(
5829 22000 : value,
5830 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
5831 22000 : );
5832 22000 : cnt += 1;
5833 2 : }
5834 2 :
5835 22 : assert_eq!(cnt, NUM_KEYS);
5836 2 :
5837 22022 : for _ in 0..NUM_KEYS {
5838 22000 : lsn = Lsn(lsn.0 + 0x10);
5839 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5840 22000 : test_key.field6 = (blknum * STEP) as u32;
5841 22000 : let mut writer = tline.writer().await;
5842 22000 : writer
5843 22000 : .put(
5844 22000 : test_key,
5845 22000 : lsn,
5846 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5847 22000 : &ctx,
5848 22000 : )
5849 141 : .await?;
5850 22000 : writer.finish_write(lsn);
5851 22000 : drop(writer);
5852 22000 : updated[blknum] = lsn;
5853 2 : }
5854 2 :
5855 2 : // Perform two cycles of flush, compact, and GC
5856 66 : for round in 0..2 {
5857 44 : tline.freeze_and_flush().await?;
5858 44 : tline
5859 44 : .compact(
5860 44 : &cancel,
5861 44 : if iter % 5 == 0 && round == 0 {
5862 6 : let mut flags = EnumSet::new();
5863 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
5864 6 : flags.insert(CompactFlags::ForceRepartition);
5865 6 : flags
5866 2 : } else {
5867 38 : EnumSet::empty()
5868 2 : },
5869 44 : &ctx,
5870 2 : )
5871 8975 : .await?;
5872 44 : tenant
5873 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5874 31 : .await?;
5875 2 : }
5876 2 : }
5877 2 :
5878 2 : Ok(())
5879 2 : }
5880 :
5881 : #[tokio::test]
5882 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
5883 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
5884 8 : let (tenant, ctx) = harness.load().await;
5885 2 : let tline = tenant
5886 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5887 4 : .await?;
5888 2 :
5889 2 : let cancel = CancellationToken::new();
5890 2 :
5891 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5892 2 : base_key.field1 = AUX_KEY_PREFIX;
5893 2 : let test_key = base_key;
5894 2 : let mut lsn = Lsn(0x10);
5895 2 :
5896 42 : for _ in 0..20 {
5897 40 : lsn = Lsn(lsn.0 + 0x10);
5898 40 : let mut writer = tline.writer().await;
5899 40 : writer
5900 40 : .put(
5901 40 : test_key,
5902 40 : lsn,
5903 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
5904 40 : &ctx,
5905 40 : )
5906 20 : .await?;
5907 40 : writer.finish_write(lsn);
5908 40 : drop(writer);
5909 40 : tline.freeze_and_flush().await?; // force create a delta layer
5910 2 : }
5911 2 :
5912 2 : let before_num_l0_delta_files =
5913 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
5914 2 :
5915 150 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
5916 2 :
5917 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
5918 2 :
5919 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}");
5920 2 :
5921 2 : assert_eq!(
5922 4 : tline.get(test_key, lsn, &ctx).await?,
5923 2 : test_img(&format!("{} at {}", 0, lsn))
5924 2 : );
5925 2 :
5926 2 : Ok(())
5927 2 : }
5928 :
5929 : #[tokio::test]
5930 2 : async fn test_branch_copies_dirty_aux_file_flag() {
5931 2 : let harness = TenantHarness::create("test_branch_copies_dirty_aux_file_flag")
5932 2 : .await
5933 2 : .unwrap();
5934 2 :
5935 2 : // the default aux file policy to switch is v1 if not set by the admins
5936 2 : assert_eq!(
5937 2 : harness.tenant_conf.switch_aux_file_policy,
5938 2 : AuxFilePolicy::V1
5939 2 : );
5940 8 : let (tenant, ctx) = harness.load().await;
5941 2 :
5942 2 : let mut lsn = Lsn(0x08);
5943 2 :
5944 2 : let tline: Arc<Timeline> = tenant
5945 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
5946 4 : .await
5947 2 : .unwrap();
5948 2 :
5949 2 : // no aux file is written at this point, so the persistent flag should be unset
5950 2 : assert_eq!(tline.last_aux_file_policy.load(), None);
5951 2 :
5952 2 : {
5953 2 : lsn += 8;
5954 2 : let mut modification = tline.begin_modification(lsn);
5955 2 : modification
5956 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
5957 5 : .await
5958 2 : .unwrap();
5959 2 : modification.commit(&ctx).await.unwrap();
5960 2 : }
5961 2 :
5962 2 : // there is no tenant manager to pass the configuration through, so lets mimic it
5963 2 : tenant.set_new_location_config(
5964 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
5965 2 : TenantConfOpt {
5966 2 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
5967 2 : ..Default::default()
5968 2 : },
5969 2 : tenant.generation,
5970 2 : &pageserver_api::models::ShardParameters::default(),
5971 2 : ))
5972 2 : .unwrap(),
5973 2 : );
5974 2 :
5975 2 : assert_eq!(
5976 2 : tline.get_switch_aux_file_policy(),
5977 2 : AuxFilePolicy::V2,
5978 2 : "wanted state has been updated"
5979 2 : );
5980 2 : assert_eq!(
5981 2 : tline.last_aux_file_policy.load(),
5982 2 : Some(AuxFilePolicy::V1),
5983 2 : "aux file is written with switch_aux_file_policy unset (which is v1), so we should keep v1"
5984 2 : );
5985 2 :
5986 2 : // we can read everything from the storage
5987 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
5988 2 : assert_eq!(
5989 2 : files.get("pg_logical/mappings/test1"),
5990 2 : Some(&bytes::Bytes::from_static(b"first"))
5991 2 : );
5992 2 :
5993 2 : {
5994 2 : lsn += 8;
5995 2 : let mut modification = tline.begin_modification(lsn);
5996 2 : modification
5997 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
5998 2 : .await
5999 2 : .unwrap();
6000 2 : modification.commit(&ctx).await.unwrap();
6001 2 : }
6002 2 :
6003 2 : assert_eq!(
6004 2 : tline.last_aux_file_policy.load(),
6005 2 : Some(AuxFilePolicy::V1),
6006 2 : "keep v1 storage format when new files are written"
6007 2 : );
6008 2 :
6009 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6010 2 : assert_eq!(
6011 2 : files.get("pg_logical/mappings/test2"),
6012 2 : Some(&bytes::Bytes::from_static(b"second"))
6013 2 : );
6014 2 :
6015 2 : let child = tenant
6016 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
6017 2 : .await
6018 2 : .unwrap();
6019 2 :
6020 2 : // child copies the last flag even if that is not on remote storage yet
6021 2 : assert_eq!(child.get_switch_aux_file_policy(), AuxFilePolicy::V2);
6022 2 : assert_eq!(child.last_aux_file_policy.load(), Some(AuxFilePolicy::V1));
6023 2 :
6024 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
6025 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6026 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
6027 2 :
6028 2 : // even if we crash here without flushing parent timeline with it's new
6029 2 : // last_aux_file_policy we are safe, because child was never meant to access ancestor's
6030 2 : // files. the ancestor can even switch back to V1 because of a migration safely.
6031 2 : }
6032 :
6033 : #[tokio::test]
6034 2 : async fn aux_file_policy_switch() {
6035 2 : let mut harness = TenantHarness::create("aux_file_policy_switch")
6036 2 : .await
6037 2 : .unwrap();
6038 2 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::CrossValidation; // set to cross-validation mode
6039 8 : let (tenant, ctx) = harness.load().await;
6040 2 :
6041 2 : let mut lsn = Lsn(0x08);
6042 2 :
6043 2 : let tline: Arc<Timeline> = tenant
6044 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6045 4 : .await
6046 2 : .unwrap();
6047 2 :
6048 2 : assert_eq!(
6049 2 : tline.last_aux_file_policy.load(),
6050 2 : None,
6051 2 : "no aux file is written so it should be unset"
6052 2 : );
6053 2 :
6054 2 : {
6055 2 : lsn += 8;
6056 2 : let mut modification = tline.begin_modification(lsn);
6057 2 : modification
6058 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6059 5 : .await
6060 2 : .unwrap();
6061 2 : modification.commit(&ctx).await.unwrap();
6062 2 : }
6063 2 :
6064 2 : // there is no tenant manager to pass the configuration through, so lets mimic it
6065 2 : tenant.set_new_location_config(
6066 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
6067 2 : TenantConfOpt {
6068 2 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6069 2 : ..Default::default()
6070 2 : },
6071 2 : tenant.generation,
6072 2 : &pageserver_api::models::ShardParameters::default(),
6073 2 : ))
6074 2 : .unwrap(),
6075 2 : );
6076 2 :
6077 2 : assert_eq!(
6078 2 : tline.get_switch_aux_file_policy(),
6079 2 : AuxFilePolicy::V2,
6080 2 : "wanted state has been updated"
6081 2 : );
6082 2 : assert_eq!(
6083 2 : tline.last_aux_file_policy.load(),
6084 2 : Some(AuxFilePolicy::CrossValidation),
6085 2 : "dirty index_part.json reflected state is yet to be updated"
6086 2 : );
6087 2 :
6088 2 : // we can still read the auxfile v1 before we ingest anything new
6089 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6090 2 : assert_eq!(
6091 2 : files.get("pg_logical/mappings/test1"),
6092 2 : Some(&bytes::Bytes::from_static(b"first"))
6093 2 : );
6094 2 :
6095 2 : {
6096 2 : lsn += 8;
6097 2 : let mut modification = tline.begin_modification(lsn);
6098 2 : modification
6099 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6100 2 : .await
6101 2 : .unwrap();
6102 2 : modification.commit(&ctx).await.unwrap();
6103 2 : }
6104 2 :
6105 2 : assert_eq!(
6106 2 : tline.last_aux_file_policy.load(),
6107 2 : Some(AuxFilePolicy::V2),
6108 2 : "ingesting a file should apply the wanted switch state when applicable"
6109 2 : );
6110 2 :
6111 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6112 2 : assert_eq!(
6113 2 : files.get("pg_logical/mappings/test1"),
6114 2 : Some(&bytes::Bytes::from_static(b"first")),
6115 2 : "cross validation writes to both v1 and v2 so this should be available in v2"
6116 2 : );
6117 2 : assert_eq!(
6118 2 : files.get("pg_logical/mappings/test2"),
6119 2 : Some(&bytes::Bytes::from_static(b"second"))
6120 2 : );
6121 2 :
6122 2 : // mimic again by trying to flip it from V2 to V1 (not switched to while ingesting a file)
6123 2 : tenant.set_new_location_config(
6124 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
6125 2 : TenantConfOpt {
6126 2 : switch_aux_file_policy: Some(AuxFilePolicy::V1),
6127 2 : ..Default::default()
6128 2 : },
6129 2 : tenant.generation,
6130 2 : &pageserver_api::models::ShardParameters::default(),
6131 2 : ))
6132 2 : .unwrap(),
6133 2 : );
6134 2 :
6135 2 : {
6136 2 : lsn += 8;
6137 2 : let mut modification = tline.begin_modification(lsn);
6138 2 : modification
6139 2 : .put_file("pg_logical/mappings/test2", b"third", &ctx)
6140 2 : .await
6141 2 : .unwrap();
6142 2 : modification.commit(&ctx).await.unwrap();
6143 2 : }
6144 2 :
6145 2 : assert_eq!(
6146 2 : tline.get_switch_aux_file_policy(),
6147 2 : AuxFilePolicy::V1,
6148 2 : "wanted state has been updated again, even if invalid request"
6149 2 : );
6150 2 :
6151 2 : assert_eq!(
6152 2 : tline.last_aux_file_policy.load(),
6153 2 : Some(AuxFilePolicy::V2),
6154 2 : "ingesting a file should apply the wanted switch state when applicable"
6155 2 : );
6156 2 :
6157 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6158 2 : assert_eq!(
6159 2 : files.get("pg_logical/mappings/test1"),
6160 2 : Some(&bytes::Bytes::from_static(b"first"))
6161 2 : );
6162 2 : assert_eq!(
6163 2 : files.get("pg_logical/mappings/test2"),
6164 2 : Some(&bytes::Bytes::from_static(b"third"))
6165 2 : );
6166 2 :
6167 2 : // mimic again by trying to flip it from from V1 to V2 (not switched to while ingesting a file)
6168 2 : tenant.set_new_location_config(
6169 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
6170 2 : TenantConfOpt {
6171 2 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6172 2 : ..Default::default()
6173 2 : },
6174 2 : tenant.generation,
6175 2 : &pageserver_api::models::ShardParameters::default(),
6176 2 : ))
6177 2 : .unwrap(),
6178 2 : );
6179 2 :
6180 2 : {
6181 2 : lsn += 8;
6182 2 : let mut modification = tline.begin_modification(lsn);
6183 2 : modification
6184 2 : .put_file("pg_logical/mappings/test3", b"last", &ctx)
6185 2 : .await
6186 2 : .unwrap();
6187 2 : modification.commit(&ctx).await.unwrap();
6188 2 : }
6189 2 :
6190 2 : assert_eq!(tline.get_switch_aux_file_policy(), AuxFilePolicy::V2);
6191 2 :
6192 2 : assert_eq!(tline.last_aux_file_policy.load(), Some(AuxFilePolicy::V2));
6193 2 :
6194 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6195 2 : assert_eq!(
6196 2 : files.get("pg_logical/mappings/test1"),
6197 2 : Some(&bytes::Bytes::from_static(b"first"))
6198 2 : );
6199 2 : assert_eq!(
6200 2 : files.get("pg_logical/mappings/test2"),
6201 2 : Some(&bytes::Bytes::from_static(b"third"))
6202 2 : );
6203 2 : assert_eq!(
6204 2 : files.get("pg_logical/mappings/test3"),
6205 2 : Some(&bytes::Bytes::from_static(b"last"))
6206 2 : );
6207 2 : }
6208 :
6209 : #[tokio::test]
6210 2 : async fn aux_file_policy_force_switch() {
6211 2 : let mut harness = TenantHarness::create("aux_file_policy_force_switch")
6212 2 : .await
6213 2 : .unwrap();
6214 2 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V1;
6215 8 : let (tenant, ctx) = harness.load().await;
6216 2 :
6217 2 : let mut lsn = Lsn(0x08);
6218 2 :
6219 2 : let tline: Arc<Timeline> = tenant
6220 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6221 4 : .await
6222 2 : .unwrap();
6223 2 :
6224 2 : assert_eq!(
6225 2 : tline.last_aux_file_policy.load(),
6226 2 : None,
6227 2 : "no aux file is written so it should be unset"
6228 2 : );
6229 2 :
6230 2 : {
6231 2 : lsn += 8;
6232 2 : let mut modification = tline.begin_modification(lsn);
6233 2 : modification
6234 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6235 5 : .await
6236 2 : .unwrap();
6237 2 : modification.commit(&ctx).await.unwrap();
6238 2 : }
6239 2 :
6240 2 : tline.do_switch_aux_policy(AuxFilePolicy::V2).unwrap();
6241 2 :
6242 2 : assert_eq!(
6243 2 : tline.last_aux_file_policy.load(),
6244 2 : Some(AuxFilePolicy::V2),
6245 2 : "dirty index_part.json reflected state is yet to be updated"
6246 2 : );
6247 2 :
6248 2 : // lose all data from v1
6249 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6250 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6251 2 :
6252 2 : {
6253 2 : lsn += 8;
6254 2 : let mut modification = tline.begin_modification(lsn);
6255 2 : modification
6256 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6257 2 : .await
6258 2 : .unwrap();
6259 2 : modification.commit(&ctx).await.unwrap();
6260 2 : }
6261 2 :
6262 2 : // read data ingested in v2
6263 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6264 2 : assert_eq!(
6265 2 : files.get("pg_logical/mappings/test2"),
6266 2 : Some(&bytes::Bytes::from_static(b"second"))
6267 2 : );
6268 2 : // lose all data from v1
6269 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6270 2 : }
6271 :
6272 : #[tokio::test]
6273 2 : async fn aux_file_policy_auto_detect() {
6274 2 : let mut harness = TenantHarness::create("aux_file_policy_auto_detect")
6275 2 : .await
6276 2 : .unwrap();
6277 2 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V2; // set to cross-validation mode
6278 8 : let (tenant, ctx) = harness.load().await;
6279 2 :
6280 2 : let mut lsn = Lsn(0x08);
6281 2 :
6282 2 : let tline: Arc<Timeline> = tenant
6283 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6284 4 : .await
6285 2 : .unwrap();
6286 2 :
6287 2 : assert_eq!(
6288 2 : tline.last_aux_file_policy.load(),
6289 2 : None,
6290 2 : "no aux file is written so it should be unset"
6291 2 : );
6292 2 :
6293 2 : {
6294 2 : lsn += 8;
6295 2 : let mut modification = tline.begin_modification(lsn);
6296 2 : let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
6297 2 : files: vec![(
6298 2 : "test_file".to_string(),
6299 2 : Bytes::copy_from_slice(b"test_file"),
6300 2 : )]
6301 2 : .into_iter()
6302 2 : .collect(),
6303 2 : })
6304 2 : .unwrap();
6305 2 : modification.put_for_test(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
6306 2 : modification.commit(&ctx).await.unwrap();
6307 2 : }
6308 2 :
6309 2 : {
6310 2 : lsn += 8;
6311 2 : let mut modification = tline.begin_modification(lsn);
6312 2 : modification
6313 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6314 2 : .await
6315 2 : .unwrap();
6316 2 : modification.commit(&ctx).await.unwrap();
6317 2 : }
6318 2 :
6319 2 : assert_eq!(
6320 2 : tline.last_aux_file_policy.load(),
6321 2 : Some(AuxFilePolicy::V1),
6322 2 : "keep using v1 because there are aux files writting with v1"
6323 2 : );
6324 2 :
6325 2 : // we can still read the auxfile v1
6326 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6327 2 : assert_eq!(
6328 2 : files.get("pg_logical/mappings/test1"),
6329 2 : Some(&bytes::Bytes::from_static(b"first"))
6330 2 : );
6331 2 : assert_eq!(
6332 2 : files.get("test_file"),
6333 2 : Some(&bytes::Bytes::from_static(b"test_file"))
6334 2 : );
6335 2 : }
6336 :
6337 : #[tokio::test]
6338 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
6339 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
6340 8 : let (tenant, ctx) = harness.load().await;
6341 2 : let tline = tenant
6342 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6343 4 : .await?;
6344 2 :
6345 2 : const NUM_KEYS: usize = 1000;
6346 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6347 2 :
6348 2 : let cancel = CancellationToken::new();
6349 2 :
6350 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6351 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6352 2 : let mut test_key = base_key;
6353 2 : let mut lsn = Lsn(0x10);
6354 2 :
6355 8 : async fn scan_with_statistics(
6356 8 : tline: &Timeline,
6357 8 : keyspace: &KeySpace,
6358 8 : lsn: Lsn,
6359 8 : ctx: &RequestContext,
6360 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
6361 8 : let mut reconstruct_state = ValuesReconstructState::default();
6362 8 : let res = tline
6363 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
6364 1703 : .await?;
6365 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
6366 8 : }
6367 2 :
6368 2 : #[allow(clippy::needless_range_loop)]
6369 2002 : for blknum in 0..NUM_KEYS {
6370 2000 : lsn = Lsn(lsn.0 + 0x10);
6371 2000 : test_key.field6 = (blknum * STEP) as u32;
6372 2000 : let mut writer = tline.writer().await;
6373 2000 : writer
6374 2000 : .put(
6375 2000 : test_key,
6376 2000 : lsn,
6377 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6378 2000 : &ctx,
6379 2000 : )
6380 2 : .await?;
6381 2000 : writer.finish_write(lsn);
6382 2000 : drop(writer);
6383 2 : }
6384 2 :
6385 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6386 2 :
6387 22 : for iter in 1..=10 {
6388 20020 : for _ in 0..NUM_KEYS {
6389 20000 : lsn = Lsn(lsn.0 + 0x10);
6390 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6391 20000 : test_key.field6 = (blknum * STEP) as u32;
6392 20000 : let mut writer = tline.writer().await;
6393 20000 : writer
6394 20000 : .put(
6395 20000 : test_key,
6396 20000 : lsn,
6397 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6398 20000 : &ctx,
6399 20000 : )
6400 19 : .await?;
6401 20000 : writer.finish_write(lsn);
6402 20000 : drop(writer);
6403 2 : }
6404 2 :
6405 20 : tline.freeze_and_flush().await?;
6406 2 :
6407 20 : if iter % 5 == 0 {
6408 4 : let (_, before_delta_file_accessed) =
6409 1695 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6410 4 : tline
6411 4 : .compact(
6412 4 : &cancel,
6413 4 : {
6414 4 : let mut flags = EnumSet::new();
6415 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
6416 4 : flags.insert(CompactFlags::ForceRepartition);
6417 4 : flags
6418 4 : },
6419 4 : &ctx,
6420 4 : )
6421 6530 : .await?;
6422 4 : let (_, after_delta_file_accessed) =
6423 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6424 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}");
6425 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.
6426 4 : assert!(
6427 4 : after_delta_file_accessed <= 2,
6428 2 : "after_delta_file_accessed={after_delta_file_accessed}"
6429 2 : );
6430 16 : }
6431 2 : }
6432 2 :
6433 2 : Ok(())
6434 2 : }
6435 :
6436 : #[tokio::test]
6437 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
6438 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
6439 8 : let (tenant, ctx) = harness.load().await;
6440 2 :
6441 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6442 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
6443 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
6444 2 :
6445 2 : let tline = tenant
6446 2 : .create_test_timeline_with_layers(
6447 2 : TIMELINE_ID,
6448 2 : Lsn(0x10),
6449 2 : DEFAULT_PG_VERSION,
6450 2 : &ctx,
6451 2 : Vec::new(), // delta layers
6452 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
6453 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
6454 2 : )
6455 11 : .await?;
6456 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
6457 2 :
6458 2 : let child = tenant
6459 2 : .branch_timeline_test_with_layers(
6460 2 : &tline,
6461 2 : NEW_TIMELINE_ID,
6462 2 : Some(Lsn(0x20)),
6463 2 : &ctx,
6464 2 : Vec::new(), // delta layers
6465 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
6466 2 : Lsn(0x30),
6467 2 : )
6468 7 : .await
6469 2 : .unwrap();
6470 2 :
6471 2 : let lsn = Lsn(0x30);
6472 2 :
6473 2 : // test vectored get on parent timeline
6474 2 : assert_eq!(
6475 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6476 2 : Some(test_img("data key 1"))
6477 2 : );
6478 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
6479 3 : .await
6480 2 : .unwrap_err()
6481 2 : .is_missing_key_error());
6482 2 : assert!(
6483 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
6484 2 : .await
6485 2 : .unwrap_err()
6486 2 : .is_missing_key_error()
6487 2 : );
6488 2 :
6489 2 : // test vectored get on child timeline
6490 2 : assert_eq!(
6491 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6492 2 : Some(test_img("data key 1"))
6493 2 : );
6494 2 : assert_eq!(
6495 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6496 2 : Some(test_img("data key 2"))
6497 2 : );
6498 2 : assert!(
6499 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
6500 2 : .await
6501 2 : .unwrap_err()
6502 2 : .is_missing_key_error()
6503 2 : );
6504 2 :
6505 2 : Ok(())
6506 2 : }
6507 :
6508 : #[tokio::test]
6509 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
6510 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
6511 8 : let (tenant, ctx) = harness.load().await;
6512 2 :
6513 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6514 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
6515 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
6516 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6517 2 :
6518 2 : let tline = tenant
6519 2 : .create_test_timeline_with_layers(
6520 2 : TIMELINE_ID,
6521 2 : Lsn(0x10),
6522 2 : DEFAULT_PG_VERSION,
6523 2 : &ctx,
6524 2 : Vec::new(), // delta layers
6525 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
6526 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
6527 2 : )
6528 11 : .await?;
6529 2 :
6530 2 : let child = tenant
6531 2 : .branch_timeline_test_with_layers(
6532 2 : &tline,
6533 2 : NEW_TIMELINE_ID,
6534 2 : Some(Lsn(0x20)),
6535 2 : &ctx,
6536 2 : Vec::new(), // delta layers
6537 2 : vec![(
6538 2 : Lsn(0x30),
6539 2 : vec![(base_key_child, test_img("metadata key 2"))],
6540 2 : )], // image layers
6541 2 : Lsn(0x30),
6542 2 : )
6543 7 : .await
6544 2 : .unwrap();
6545 2 :
6546 2 : let lsn = Lsn(0x30);
6547 2 :
6548 2 : // test vectored get on parent timeline
6549 2 : assert_eq!(
6550 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6551 2 : Some(test_img("metadata key 1"))
6552 2 : );
6553 2 : assert_eq!(
6554 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
6555 2 : None
6556 2 : );
6557 2 : assert_eq!(
6558 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
6559 2 : None
6560 2 : );
6561 2 :
6562 2 : // test vectored get on child timeline
6563 2 : assert_eq!(
6564 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6565 2 : None
6566 2 : );
6567 2 : assert_eq!(
6568 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6569 2 : Some(test_img("metadata key 2"))
6570 2 : );
6571 2 : assert_eq!(
6572 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
6573 2 : None
6574 2 : );
6575 2 :
6576 2 : Ok(())
6577 2 : }
6578 :
6579 36 : async fn get_vectored_impl_wrapper(
6580 36 : tline: &Arc<Timeline>,
6581 36 : key: Key,
6582 36 : lsn: Lsn,
6583 36 : ctx: &RequestContext,
6584 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
6585 36 : let mut reconstruct_state = ValuesReconstructState::new();
6586 36 : let mut res = tline
6587 36 : .get_vectored_impl(
6588 36 : KeySpace::single(key..key.next()),
6589 36 : lsn,
6590 36 : &mut reconstruct_state,
6591 36 : ctx,
6592 36 : )
6593 33 : .await?;
6594 30 : Ok(res.pop_last().map(|(k, v)| {
6595 18 : assert_eq!(k, key);
6596 18 : v.unwrap()
6597 30 : }))
6598 36 : }
6599 :
6600 : #[tokio::test]
6601 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
6602 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
6603 8 : let (tenant, ctx) = harness.load().await;
6604 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6605 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6606 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6607 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6608 2 :
6609 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
6610 2 : // Lsn 0x30 key0, key3, no key1+key2
6611 2 : // Lsn 0x20 key1+key2 tomestones
6612 2 : // Lsn 0x10 key1 in image, key2 in delta
6613 2 : let tline = tenant
6614 2 : .create_test_timeline_with_layers(
6615 2 : TIMELINE_ID,
6616 2 : Lsn(0x10),
6617 2 : DEFAULT_PG_VERSION,
6618 2 : &ctx,
6619 2 : // delta layers
6620 2 : vec![
6621 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6622 2 : Lsn(0x10)..Lsn(0x20),
6623 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6624 2 : ),
6625 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6626 2 : Lsn(0x20)..Lsn(0x30),
6627 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6628 2 : ),
6629 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6630 2 : Lsn(0x20)..Lsn(0x30),
6631 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6632 2 : ),
6633 2 : ],
6634 2 : // image layers
6635 2 : vec![
6636 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
6637 2 : (
6638 2 : Lsn(0x30),
6639 2 : vec![
6640 2 : (key0, test_img("metadata key 0")),
6641 2 : (key3, test_img("metadata key 3")),
6642 2 : ],
6643 2 : ),
6644 2 : ],
6645 2 : Lsn(0x30),
6646 2 : )
6647 38 : .await?;
6648 2 :
6649 2 : let lsn = Lsn(0x30);
6650 2 : let old_lsn = Lsn(0x20);
6651 2 :
6652 2 : assert_eq!(
6653 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
6654 2 : Some(test_img("metadata key 0"))
6655 2 : );
6656 2 : assert_eq!(
6657 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
6658 2 : None,
6659 2 : );
6660 2 : assert_eq!(
6661 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
6662 2 : None,
6663 2 : );
6664 2 : assert_eq!(
6665 4 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
6666 2 : Some(Bytes::new()),
6667 2 : );
6668 2 : assert_eq!(
6669 4 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
6670 2 : Some(Bytes::new()),
6671 2 : );
6672 2 : assert_eq!(
6673 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
6674 2 : Some(test_img("metadata key 3"))
6675 2 : );
6676 2 :
6677 2 : Ok(())
6678 2 : }
6679 :
6680 : #[tokio::test]
6681 2 : async fn test_metadata_tombstone_image_creation() {
6682 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
6683 2 : .await
6684 2 : .unwrap();
6685 8 : let (tenant, ctx) = harness.load().await;
6686 2 :
6687 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6688 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6689 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6690 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6691 2 :
6692 2 : let tline = tenant
6693 2 : .create_test_timeline_with_layers(
6694 2 : TIMELINE_ID,
6695 2 : Lsn(0x10),
6696 2 : DEFAULT_PG_VERSION,
6697 2 : &ctx,
6698 2 : // delta layers
6699 2 : vec![
6700 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6701 2 : Lsn(0x10)..Lsn(0x20),
6702 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6703 2 : ),
6704 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6705 2 : Lsn(0x20)..Lsn(0x30),
6706 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6707 2 : ),
6708 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6709 2 : Lsn(0x20)..Lsn(0x30),
6710 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6711 2 : ),
6712 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6713 2 : Lsn(0x30)..Lsn(0x40),
6714 2 : vec![
6715 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
6716 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
6717 2 : ],
6718 2 : ),
6719 2 : ],
6720 2 : // image layers
6721 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6722 2 : Lsn(0x40),
6723 2 : )
6724 35 : .await
6725 2 : .unwrap();
6726 2 :
6727 2 : let cancel = CancellationToken::new();
6728 2 :
6729 2 : tline
6730 2 : .compact(
6731 2 : &cancel,
6732 2 : {
6733 2 : let mut flags = EnumSet::new();
6734 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
6735 2 : flags.insert(CompactFlags::ForceRepartition);
6736 2 : flags
6737 2 : },
6738 2 : &ctx,
6739 2 : )
6740 58 : .await
6741 2 : .unwrap();
6742 2 :
6743 2 : // Image layers are created at last_record_lsn
6744 2 : let images = tline
6745 2 : .inspect_image_layers(Lsn(0x40), &ctx)
6746 8 : .await
6747 2 : .unwrap()
6748 2 : .into_iter()
6749 20 : .filter(|(k, _)| k.is_metadata_key())
6750 2 : .collect::<Vec<_>>();
6751 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
6752 2 : }
6753 :
6754 : #[tokio::test]
6755 2 : async fn test_metadata_tombstone_empty_image_creation() {
6756 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
6757 2 : .await
6758 2 : .unwrap();
6759 8 : let (tenant, ctx) = harness.load().await;
6760 2 :
6761 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6762 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6763 2 :
6764 2 : let tline = tenant
6765 2 : .create_test_timeline_with_layers(
6766 2 : TIMELINE_ID,
6767 2 : Lsn(0x10),
6768 2 : DEFAULT_PG_VERSION,
6769 2 : &ctx,
6770 2 : // delta layers
6771 2 : vec![
6772 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6773 2 : Lsn(0x10)..Lsn(0x20),
6774 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6775 2 : ),
6776 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6777 2 : Lsn(0x20)..Lsn(0x30),
6778 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6779 2 : ),
6780 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6781 2 : Lsn(0x20)..Lsn(0x30),
6782 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6783 2 : ),
6784 2 : ],
6785 2 : // image layers
6786 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6787 2 : Lsn(0x30),
6788 2 : )
6789 29 : .await
6790 2 : .unwrap();
6791 2 :
6792 2 : let cancel = CancellationToken::new();
6793 2 :
6794 2 : tline
6795 2 : .compact(
6796 2 : &cancel,
6797 2 : {
6798 2 : let mut flags = EnumSet::new();
6799 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
6800 2 : flags.insert(CompactFlags::ForceRepartition);
6801 2 : flags
6802 2 : },
6803 2 : &ctx,
6804 2 : )
6805 46 : .await
6806 2 : .unwrap();
6807 2 :
6808 2 : // Image layers are created at last_record_lsn
6809 2 : let images = tline
6810 2 : .inspect_image_layers(Lsn(0x30), &ctx)
6811 4 : .await
6812 2 : .unwrap()
6813 2 : .into_iter()
6814 16 : .filter(|(k, _)| k.is_metadata_key())
6815 2 : .collect::<Vec<_>>();
6816 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
6817 2 : }
6818 :
6819 : #[tokio::test]
6820 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
6821 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
6822 8 : let (tenant, ctx) = harness.load().await;
6823 2 :
6824 102 : fn get_key(id: u32) -> Key {
6825 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
6826 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6827 102 : key.field6 = id;
6828 102 : key
6829 102 : }
6830 2 :
6831 2 : // We create
6832 2 : // - one bottom-most image layer,
6833 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
6834 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
6835 2 : // - a delta layer D3 above the horizon.
6836 2 : //
6837 2 : // | D3 |
6838 2 : // | D1 |
6839 2 : // -| |-- gc horizon -----------------
6840 2 : // | | | D2 |
6841 2 : // --------- img layer ------------------
6842 2 : //
6843 2 : // What we should expact from this compaction is:
6844 2 : // | D3 |
6845 2 : // | Part of D1 |
6846 2 : // --------- img layer with D1+D2 at GC horizon------------------
6847 2 :
6848 2 : // img layer at 0x10
6849 2 : let img_layer = (0..10)
6850 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
6851 2 : .collect_vec();
6852 2 :
6853 2 : let delta1 = vec![
6854 2 : (
6855 2 : get_key(1),
6856 2 : Lsn(0x20),
6857 2 : Value::Image(Bytes::from("value 1@0x20")),
6858 2 : ),
6859 2 : (
6860 2 : get_key(2),
6861 2 : Lsn(0x30),
6862 2 : Value::Image(Bytes::from("value 2@0x30")),
6863 2 : ),
6864 2 : (
6865 2 : get_key(3),
6866 2 : Lsn(0x40),
6867 2 : Value::Image(Bytes::from("value 3@0x40")),
6868 2 : ),
6869 2 : ];
6870 2 : let delta2 = vec![
6871 2 : (
6872 2 : get_key(5),
6873 2 : Lsn(0x20),
6874 2 : Value::Image(Bytes::from("value 5@0x20")),
6875 2 : ),
6876 2 : (
6877 2 : get_key(6),
6878 2 : Lsn(0x20),
6879 2 : Value::Image(Bytes::from("value 6@0x20")),
6880 2 : ),
6881 2 : ];
6882 2 : let delta3 = vec![
6883 2 : (
6884 2 : get_key(8),
6885 2 : Lsn(0x48),
6886 2 : Value::Image(Bytes::from("value 8@0x48")),
6887 2 : ),
6888 2 : (
6889 2 : get_key(9),
6890 2 : Lsn(0x48),
6891 2 : Value::Image(Bytes::from("value 9@0x48")),
6892 2 : ),
6893 2 : ];
6894 2 :
6895 2 : let tline = tenant
6896 2 : .create_test_timeline_with_layers(
6897 2 : TIMELINE_ID,
6898 2 : Lsn(0x10),
6899 2 : DEFAULT_PG_VERSION,
6900 2 : &ctx,
6901 2 : vec![
6902 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
6903 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
6904 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
6905 2 : ], // delta layers
6906 2 : vec![(Lsn(0x10), img_layer)], // image layers
6907 2 : Lsn(0x50),
6908 2 : )
6909 47 : .await?;
6910 2 : {
6911 2 : // Update GC info
6912 2 : let mut guard = tline.gc_info.write().unwrap();
6913 2 : guard.cutoffs.time = Lsn(0x30);
6914 2 : guard.cutoffs.space = Lsn(0x30);
6915 2 : }
6916 2 :
6917 2 : let expected_result = [
6918 2 : Bytes::from_static(b"value 0@0x10"),
6919 2 : Bytes::from_static(b"value 1@0x20"),
6920 2 : Bytes::from_static(b"value 2@0x30"),
6921 2 : Bytes::from_static(b"value 3@0x40"),
6922 2 : Bytes::from_static(b"value 4@0x10"),
6923 2 : Bytes::from_static(b"value 5@0x20"),
6924 2 : Bytes::from_static(b"value 6@0x20"),
6925 2 : Bytes::from_static(b"value 7@0x10"),
6926 2 : Bytes::from_static(b"value 8@0x48"),
6927 2 : Bytes::from_static(b"value 9@0x48"),
6928 2 : ];
6929 2 :
6930 20 : for (idx, expected) in expected_result.iter().enumerate() {
6931 20 : assert_eq!(
6932 20 : tline
6933 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
6934 23 : .await
6935 20 : .unwrap(),
6936 2 : expected
6937 2 : );
6938 2 : }
6939 2 :
6940 2 : let cancel = CancellationToken::new();
6941 2 : tline
6942 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
6943 55 : .await
6944 2 : .unwrap();
6945 2 :
6946 20 : for (idx, expected) in expected_result.iter().enumerate() {
6947 20 : assert_eq!(
6948 20 : tline
6949 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
6950 16 : .await
6951 20 : .unwrap(),
6952 2 : expected
6953 2 : );
6954 2 : }
6955 2 :
6956 2 : // Check if the image layer at the GC horizon contains exactly what we want
6957 2 : let image_at_gc_horizon = tline
6958 2 : .inspect_image_layers(Lsn(0x30), &ctx)
6959 2 : .await
6960 2 : .unwrap()
6961 2 : .into_iter()
6962 36 : .filter(|(k, _)| k.is_metadata_key())
6963 2 : .collect::<Vec<_>>();
6964 2 :
6965 2 : assert_eq!(image_at_gc_horizon.len(), 10);
6966 2 : let expected_result = [
6967 2 : Bytes::from_static(b"value 0@0x10"),
6968 2 : Bytes::from_static(b"value 1@0x20"),
6969 2 : Bytes::from_static(b"value 2@0x30"),
6970 2 : Bytes::from_static(b"value 3@0x10"),
6971 2 : Bytes::from_static(b"value 4@0x10"),
6972 2 : Bytes::from_static(b"value 5@0x20"),
6973 2 : Bytes::from_static(b"value 6@0x20"),
6974 2 : Bytes::from_static(b"value 7@0x10"),
6975 2 : Bytes::from_static(b"value 8@0x10"),
6976 2 : Bytes::from_static(b"value 9@0x10"),
6977 2 : ];
6978 22 : for idx in 0..10 {
6979 20 : assert_eq!(
6980 20 : image_at_gc_horizon[idx],
6981 20 : (get_key(idx as u32), expected_result[idx].clone())
6982 20 : );
6983 2 : }
6984 2 :
6985 2 : // Check if old layers are removed / new layers have the expected LSN
6986 2 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
6987 4 : all_layers.sort_by(|k1, k2| {
6988 4 : (
6989 4 : k1.is_delta,
6990 4 : k1.key_range.start,
6991 4 : k1.key_range.end,
6992 4 : k1.lsn_range.start,
6993 4 : k1.lsn_range.end,
6994 4 : )
6995 4 : .cmp(&(
6996 4 : k2.is_delta,
6997 4 : k2.key_range.start,
6998 4 : k2.key_range.end,
6999 4 : k2.lsn_range.start,
7000 4 : k2.lsn_range.end,
7001 4 : ))
7002 4 : });
7003 2 : assert_eq!(
7004 2 : all_layers,
7005 2 : vec![
7006 2 : // Image layer at GC horizon
7007 2 : PersistentLayerKey {
7008 2 : key_range: {
7009 2 : let mut key = Key::MAX;
7010 2 : key.field6 -= 1;
7011 2 : Key::MIN..key
7012 2 : },
7013 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
7014 2 : is_delta: false
7015 2 : },
7016 2 : // The delta layer that is cut in the middle
7017 2 : PersistentLayerKey {
7018 2 : key_range: get_key(3)..get_key(4),
7019 2 : lsn_range: Lsn(0x30)..Lsn(0x41),
7020 2 : is_delta: true
7021 2 : },
7022 2 : // The delta3 layer that should not be picked for the compaction
7023 2 : PersistentLayerKey {
7024 2 : key_range: get_key(8)..get_key(10),
7025 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
7026 2 : is_delta: true
7027 2 : }
7028 2 : ]
7029 2 : );
7030 2 :
7031 2 : // increase GC horizon and compact again
7032 2 : {
7033 2 : // Update GC info
7034 2 : let mut guard = tline.gc_info.write().unwrap();
7035 2 : guard.cutoffs.time = Lsn(0x40);
7036 2 : guard.cutoffs.space = Lsn(0x40);
7037 2 : }
7038 2 : tline
7039 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7040 44 : .await
7041 2 : .unwrap();
7042 2 :
7043 2 : Ok(())
7044 2 : }
7045 :
7046 : #[tokio::test]
7047 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
7048 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
7049 8 : let (tenant, ctx) = harness.load().await;
7050 2 :
7051 24 : fn get_key(id: u32) -> Key {
7052 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7053 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7054 24 : key.field6 = id;
7055 24 : key
7056 24 : }
7057 2 :
7058 2 : let delta1 = vec![
7059 2 : (
7060 2 : get_key(1),
7061 2 : Lsn(0x20),
7062 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7063 2 : ),
7064 2 : (
7065 2 : get_key(1),
7066 2 : Lsn(0x30),
7067 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7068 2 : ),
7069 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
7070 2 : (
7071 2 : get_key(2),
7072 2 : Lsn(0x20),
7073 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
7074 2 : ),
7075 2 : (
7076 2 : get_key(2),
7077 2 : Lsn(0x30),
7078 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
7079 2 : ),
7080 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
7081 2 : (
7082 2 : get_key(3),
7083 2 : Lsn(0x20),
7084 2 : Value::WalRecord(NeonWalRecord::wal_clear()),
7085 2 : ),
7086 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
7087 2 : (
7088 2 : get_key(4),
7089 2 : Lsn(0x20),
7090 2 : Value::WalRecord(NeonWalRecord::wal_init()),
7091 2 : ),
7092 2 : ];
7093 2 : let image1 = vec![(get_key(1), "0x10".into())];
7094 2 :
7095 2 : let tline = tenant
7096 2 : .create_test_timeline_with_layers(
7097 2 : TIMELINE_ID,
7098 2 : Lsn(0x10),
7099 2 : DEFAULT_PG_VERSION,
7100 2 : &ctx,
7101 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
7102 2 : Lsn(0x10)..Lsn(0x40),
7103 2 : delta1,
7104 2 : )], // delta layers
7105 2 : vec![(Lsn(0x10), image1)], // image layers
7106 2 : Lsn(0x50),
7107 2 : )
7108 17 : .await?;
7109 2 :
7110 2 : assert_eq!(
7111 8 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
7112 2 : Bytes::from_static(b"0x10,0x20,0x30")
7113 2 : );
7114 2 : assert_eq!(
7115 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
7116 2 : Bytes::from_static(b"0x10,0x20,0x30")
7117 2 : );
7118 2 :
7119 2 : // Need to remove the limit of "Neon WAL redo requires base image".
7120 2 :
7121 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
7122 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
7123 2 :
7124 2 : Ok(())
7125 2 : }
7126 :
7127 : #[tokio::test]
7128 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
7129 8 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease").await?.load().await;
7130 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7131 2 :
7132 2 : let end_lsn = Lsn(0x100);
7133 2 : let image_layers = (0x20..=0x90)
7134 2 : .step_by(0x10)
7135 16 : .map(|n| {
7136 16 : (
7137 16 : Lsn(n),
7138 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
7139 16 : )
7140 16 : })
7141 2 : .collect();
7142 2 :
7143 2 : let timeline = tenant
7144 2 : .create_test_timeline_with_layers(
7145 2 : TIMELINE_ID,
7146 2 : Lsn(0x10),
7147 2 : DEFAULT_PG_VERSION,
7148 2 : &ctx,
7149 2 : Vec::new(),
7150 2 : image_layers,
7151 2 : end_lsn,
7152 2 : )
7153 60 : .await?;
7154 2 :
7155 2 : let leased_lsns = [0x30, 0x50, 0x70];
7156 2 : let mut leases = Vec::new();
7157 6 : let _: anyhow::Result<_> = leased_lsns.iter().try_for_each(|n| {
7158 6 : leases.push(timeline.make_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)?);
7159 6 : Ok(())
7160 6 : });
7161 2 :
7162 2 : // Renewing with shorter lease should not change the lease.
7163 2 : let updated_lease_0 =
7164 2 : timeline.make_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)?;
7165 2 : assert_eq!(updated_lease_0.valid_until, leases[0].valid_until);
7166 2 :
7167 2 : // Renewing with a long lease should renew lease with later expiration time.
7168 2 : let updated_lease_1 = timeline.make_lsn_lease(
7169 2 : Lsn(leased_lsns[1]),
7170 2 : timeline.get_lsn_lease_length() * 2,
7171 2 : &ctx,
7172 2 : )?;
7173 2 :
7174 2 : assert!(updated_lease_1.valid_until > leases[1].valid_until);
7175 2 :
7176 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
7177 2 : info!(
7178 2 : "latest_gc_cutoff_lsn: {}",
7179 0 : *timeline.get_latest_gc_cutoff_lsn()
7180 2 : );
7181 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
7182 2 :
7183 2 : let res = tenant
7184 2 : .gc_iteration(
7185 2 : Some(TIMELINE_ID),
7186 2 : 0,
7187 2 : Duration::ZERO,
7188 2 : &CancellationToken::new(),
7189 2 : &ctx,
7190 2 : )
7191 2 : .await?;
7192 2 :
7193 2 : // Keeping everything <= Lsn(0x80) b/c leases:
7194 2 : // 0/10: initdb layer
7195 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7196 2 : assert_eq!(res.layers_needed_by_leases, 7);
7197 2 : // Keeping 0/90 b/c it is the latest layer.
7198 2 : assert_eq!(res.layers_not_updated, 1);
7199 2 : // Removed 0/80.
7200 2 : assert_eq!(res.layers_removed, 1);
7201 2 :
7202 2 : // Make lease on a already GC-ed LSN.
7203 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
7204 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
7205 2 : let res = timeline.make_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx);
7206 2 : assert!(res.is_err());
7207 2 :
7208 2 : // Should still be able to renew a currently valid lease
7209 2 : // Assumption: original lease to is still valid for 0/50.
7210 2 : let _ =
7211 2 : timeline.make_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)?;
7212 2 :
7213 2 : Ok(())
7214 2 : }
7215 :
7216 : #[tokio::test]
7217 2 : async fn test_simple_bottom_most_compaction_deltas() -> anyhow::Result<()> {
7218 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_deltas").await?;
7219 8 : let (tenant, ctx) = harness.load().await;
7220 2 :
7221 118 : fn get_key(id: u32) -> Key {
7222 118 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7223 118 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7224 118 : key.field6 = id;
7225 118 : key
7226 118 : }
7227 2 :
7228 2 : // We create
7229 2 : // - one bottom-most image layer,
7230 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7231 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7232 2 : // - a delta layer D3 above the horizon.
7233 2 : //
7234 2 : // | D3 |
7235 2 : // | D1 |
7236 2 : // -| |-- gc horizon -----------------
7237 2 : // | | | D2 |
7238 2 : // --------- img layer ------------------
7239 2 : //
7240 2 : // What we should expact from this compaction is:
7241 2 : // | D3 |
7242 2 : // | Part of D1 |
7243 2 : // --------- img layer with D1+D2 at GC horizon------------------
7244 2 :
7245 2 : // img layer at 0x10
7246 2 : let img_layer = (0..10)
7247 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7248 2 : .collect_vec();
7249 2 :
7250 2 : let delta1 = vec![
7251 2 : (
7252 2 : get_key(1),
7253 2 : Lsn(0x20),
7254 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7255 2 : ),
7256 2 : (
7257 2 : get_key(2),
7258 2 : Lsn(0x30),
7259 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7260 2 : ),
7261 2 : (
7262 2 : get_key(3),
7263 2 : Lsn(0x28),
7264 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7265 2 : ),
7266 2 : (
7267 2 : get_key(3),
7268 2 : Lsn(0x30),
7269 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7270 2 : ),
7271 2 : (
7272 2 : get_key(3),
7273 2 : Lsn(0x40),
7274 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7275 2 : ),
7276 2 : ];
7277 2 : let delta2 = vec![
7278 2 : (
7279 2 : get_key(5),
7280 2 : Lsn(0x20),
7281 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7282 2 : ),
7283 2 : (
7284 2 : get_key(6),
7285 2 : Lsn(0x20),
7286 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7287 2 : ),
7288 2 : ];
7289 2 : let delta3 = vec![
7290 2 : (
7291 2 : get_key(8),
7292 2 : Lsn(0x48),
7293 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7294 2 : ),
7295 2 : (
7296 2 : get_key(9),
7297 2 : Lsn(0x48),
7298 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7299 2 : ),
7300 2 : ];
7301 2 :
7302 2 : let tline = tenant
7303 2 : .create_test_timeline_with_layers(
7304 2 : TIMELINE_ID,
7305 2 : Lsn(0x10),
7306 2 : DEFAULT_PG_VERSION,
7307 2 : &ctx,
7308 2 : vec![
7309 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7310 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7311 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7312 2 : ], // delta layers
7313 2 : vec![(Lsn(0x10), img_layer)], // image layers
7314 2 : Lsn(0x50),
7315 2 : )
7316 47 : .await?;
7317 2 : {
7318 2 : // Update GC info
7319 2 : let mut guard = tline.gc_info.write().unwrap();
7320 2 : *guard = GcInfo {
7321 2 : retain_lsns: vec![],
7322 2 : cutoffs: GcCutoffs {
7323 2 : time: Lsn(0x30),
7324 2 : space: Lsn(0x30),
7325 2 : },
7326 2 : leases: Default::default(),
7327 2 : within_ancestor_pitr: false,
7328 2 : };
7329 2 : }
7330 2 :
7331 2 : let expected_result = [
7332 2 : Bytes::from_static(b"value 0@0x10"),
7333 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7334 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7335 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7336 2 : Bytes::from_static(b"value 4@0x10"),
7337 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7338 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7339 2 : Bytes::from_static(b"value 7@0x10"),
7340 2 : Bytes::from_static(b"value 8@0x10@0x48"),
7341 2 : Bytes::from_static(b"value 9@0x10@0x48"),
7342 2 : ];
7343 2 :
7344 2 : let expected_result_at_gc_horizon = [
7345 2 : Bytes::from_static(b"value 0@0x10"),
7346 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7347 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7348 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7349 2 : Bytes::from_static(b"value 4@0x10"),
7350 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7351 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7352 2 : Bytes::from_static(b"value 7@0x10"),
7353 2 : Bytes::from_static(b"value 8@0x10"),
7354 2 : Bytes::from_static(b"value 9@0x10"),
7355 2 : ];
7356 2 :
7357 22 : for idx in 0..10 {
7358 20 : assert_eq!(
7359 20 : tline
7360 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7361 31 : .await
7362 20 : .unwrap(),
7363 20 : &expected_result[idx]
7364 2 : );
7365 20 : assert_eq!(
7366 20 : tline
7367 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7368 15 : .await
7369 20 : .unwrap(),
7370 20 : &expected_result_at_gc_horizon[idx]
7371 2 : );
7372 2 : }
7373 2 :
7374 2 : let cancel = CancellationToken::new();
7375 2 : tline
7376 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7377 54 : .await
7378 2 : .unwrap();
7379 2 :
7380 22 : for idx in 0..10 {
7381 20 : assert_eq!(
7382 20 : tline
7383 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7384 20 : .await
7385 20 : .unwrap(),
7386 20 : &expected_result[idx]
7387 2 : );
7388 20 : assert_eq!(
7389 20 : tline
7390 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7391 10 : .await
7392 20 : .unwrap(),
7393 20 : &expected_result_at_gc_horizon[idx]
7394 2 : );
7395 2 : }
7396 2 :
7397 2 : // increase GC horizon and compact again
7398 2 : {
7399 2 : // Update GC info
7400 2 : let mut guard = tline.gc_info.write().unwrap();
7401 2 : guard.cutoffs.time = Lsn(0x40);
7402 2 : guard.cutoffs.space = Lsn(0x40);
7403 2 : }
7404 2 : tline
7405 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7406 44 : .await
7407 2 : .unwrap();
7408 2 :
7409 2 : Ok(())
7410 2 : }
7411 :
7412 : #[tokio::test]
7413 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
7414 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
7415 8 : let (tenant, ctx) = harness.load().await;
7416 2 : let tline = tenant
7417 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7418 4 : .await?;
7419 2 : tline.force_advance_lsn(Lsn(0x70));
7420 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7421 2 : let history = vec![
7422 2 : (
7423 2 : key,
7424 2 : Lsn(0x10),
7425 2 : Value::Image(Bytes::copy_from_slice(b"0x10")),
7426 2 : ),
7427 2 : (
7428 2 : key,
7429 2 : Lsn(0x20),
7430 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7431 2 : ),
7432 2 : (
7433 2 : key,
7434 2 : Lsn(0x30),
7435 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7436 2 : ),
7437 2 : (
7438 2 : key,
7439 2 : Lsn(0x40),
7440 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7441 2 : ),
7442 2 : (
7443 2 : key,
7444 2 : Lsn(0x50),
7445 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7446 2 : ),
7447 2 : (
7448 2 : key,
7449 2 : Lsn(0x60),
7450 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7451 2 : ),
7452 2 : (
7453 2 : key,
7454 2 : Lsn(0x70),
7455 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7456 2 : ),
7457 2 : (
7458 2 : key,
7459 2 : Lsn(0x80),
7460 2 : Value::Image(Bytes::copy_from_slice(
7461 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7462 2 : )),
7463 2 : ),
7464 2 : (
7465 2 : key,
7466 2 : Lsn(0x90),
7467 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7468 2 : ),
7469 2 : ];
7470 2 : let res = tline
7471 2 : .generate_key_retention(
7472 2 : key,
7473 2 : &history,
7474 2 : Lsn(0x60),
7475 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
7476 2 : 3,
7477 2 : None,
7478 2 : )
7479 2 : .await
7480 2 : .unwrap();
7481 2 : let expected_res = KeyHistoryRetention {
7482 2 : below_horizon: vec![
7483 2 : (
7484 2 : Lsn(0x20),
7485 2 : KeyLogAtLsn(vec![(
7486 2 : Lsn(0x20),
7487 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20")),
7488 2 : )]),
7489 2 : ),
7490 2 : (
7491 2 : Lsn(0x40),
7492 2 : KeyLogAtLsn(vec![
7493 2 : (
7494 2 : Lsn(0x30),
7495 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7496 2 : ),
7497 2 : (
7498 2 : Lsn(0x40),
7499 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7500 2 : ),
7501 2 : ]),
7502 2 : ),
7503 2 : (
7504 2 : Lsn(0x50),
7505 2 : KeyLogAtLsn(vec![(
7506 2 : Lsn(0x50),
7507 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
7508 2 : )]),
7509 2 : ),
7510 2 : (
7511 2 : Lsn(0x60),
7512 2 : KeyLogAtLsn(vec![(
7513 2 : Lsn(0x60),
7514 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7515 2 : )]),
7516 2 : ),
7517 2 : ],
7518 2 : above_horizon: KeyLogAtLsn(vec![
7519 2 : (
7520 2 : Lsn(0x70),
7521 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7522 2 : ),
7523 2 : (
7524 2 : Lsn(0x80),
7525 2 : Value::Image(Bytes::copy_from_slice(
7526 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7527 2 : )),
7528 2 : ),
7529 2 : (
7530 2 : Lsn(0x90),
7531 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7532 2 : ),
7533 2 : ]),
7534 2 : };
7535 2 : assert_eq!(res, expected_res);
7536 2 :
7537 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
7538 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
7539 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
7540 2 : // For example, we have
7541 2 : // ```plain
7542 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
7543 2 : // ```
7544 2 : // Now the GC horizon moves up, and we have
7545 2 : // ```plain
7546 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
7547 2 : // ```
7548 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
7549 2 : // We will end up with
7550 2 : // ```plain
7551 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
7552 2 : // ```
7553 2 : // Now we run the GC-compaction, and this key does not have a full history.
7554 2 : // We should be able to handle this partial history and drop everything before the
7555 2 : // gc_horizon image.
7556 2 :
7557 2 : let history = vec![
7558 2 : (
7559 2 : key,
7560 2 : Lsn(0x20),
7561 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7562 2 : ),
7563 2 : (
7564 2 : key,
7565 2 : Lsn(0x30),
7566 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7567 2 : ),
7568 2 : (
7569 2 : key,
7570 2 : Lsn(0x40),
7571 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7572 2 : ),
7573 2 : (
7574 2 : key,
7575 2 : Lsn(0x50),
7576 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7577 2 : ),
7578 2 : (
7579 2 : key,
7580 2 : Lsn(0x60),
7581 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7582 2 : ),
7583 2 : (
7584 2 : key,
7585 2 : Lsn(0x70),
7586 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7587 2 : ),
7588 2 : (
7589 2 : key,
7590 2 : Lsn(0x80),
7591 2 : Value::Image(Bytes::copy_from_slice(
7592 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7593 2 : )),
7594 2 : ),
7595 2 : (
7596 2 : key,
7597 2 : Lsn(0x90),
7598 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7599 2 : ),
7600 2 : ];
7601 2 : let res = tline
7602 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
7603 2 : .await
7604 2 : .unwrap();
7605 2 : let expected_res = KeyHistoryRetention {
7606 2 : below_horizon: vec![
7607 2 : (
7608 2 : Lsn(0x40),
7609 2 : KeyLogAtLsn(vec![(
7610 2 : Lsn(0x40),
7611 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
7612 2 : )]),
7613 2 : ),
7614 2 : (
7615 2 : Lsn(0x50),
7616 2 : KeyLogAtLsn(vec![(
7617 2 : Lsn(0x50),
7618 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
7619 2 : )]),
7620 2 : ),
7621 2 : (
7622 2 : Lsn(0x60),
7623 2 : KeyLogAtLsn(vec![(
7624 2 : Lsn(0x60),
7625 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7626 2 : )]),
7627 2 : ),
7628 2 : ],
7629 2 : above_horizon: KeyLogAtLsn(vec![
7630 2 : (
7631 2 : Lsn(0x70),
7632 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7633 2 : ),
7634 2 : (
7635 2 : Lsn(0x80),
7636 2 : Value::Image(Bytes::copy_from_slice(
7637 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
7638 2 : )),
7639 2 : ),
7640 2 : (
7641 2 : Lsn(0x90),
7642 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
7643 2 : ),
7644 2 : ]),
7645 2 : };
7646 2 : assert_eq!(res, expected_res);
7647 2 :
7648 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
7649 2 : // the ancestor image in the test case.
7650 2 :
7651 2 : let history = vec![
7652 2 : (
7653 2 : key,
7654 2 : Lsn(0x20),
7655 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7656 2 : ),
7657 2 : (
7658 2 : key,
7659 2 : Lsn(0x30),
7660 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
7661 2 : ),
7662 2 : (
7663 2 : key,
7664 2 : Lsn(0x40),
7665 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7666 2 : ),
7667 2 : (
7668 2 : key,
7669 2 : Lsn(0x70),
7670 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7671 2 : ),
7672 2 : ];
7673 2 : let res = tline
7674 2 : .generate_key_retention(
7675 2 : key,
7676 2 : &history,
7677 2 : Lsn(0x60),
7678 2 : &[],
7679 2 : 3,
7680 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7681 2 : )
7682 2 : .await
7683 2 : .unwrap();
7684 2 : let expected_res = KeyHistoryRetention {
7685 2 : below_horizon: vec![(
7686 2 : Lsn(0x60),
7687 2 : KeyLogAtLsn(vec![(
7688 2 : Lsn(0x60),
7689 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
7690 2 : )]),
7691 2 : )],
7692 2 : above_horizon: KeyLogAtLsn(vec![(
7693 2 : Lsn(0x70),
7694 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7695 2 : )]),
7696 2 : };
7697 2 : assert_eq!(res, expected_res);
7698 2 :
7699 2 : let history = vec![
7700 2 : (
7701 2 : key,
7702 2 : Lsn(0x20),
7703 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7704 2 : ),
7705 2 : (
7706 2 : key,
7707 2 : Lsn(0x40),
7708 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
7709 2 : ),
7710 2 : (
7711 2 : key,
7712 2 : Lsn(0x60),
7713 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
7714 2 : ),
7715 2 : (
7716 2 : key,
7717 2 : Lsn(0x70),
7718 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7719 2 : ),
7720 2 : ];
7721 2 : let res = tline
7722 2 : .generate_key_retention(
7723 2 : key,
7724 2 : &history,
7725 2 : Lsn(0x60),
7726 2 : &[Lsn(0x30)],
7727 2 : 3,
7728 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
7729 2 : )
7730 2 : .await
7731 2 : .unwrap();
7732 2 : let expected_res = KeyHistoryRetention {
7733 2 : below_horizon: vec![
7734 2 : (
7735 2 : Lsn(0x30),
7736 2 : KeyLogAtLsn(vec![(
7737 2 : Lsn(0x20),
7738 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
7739 2 : )]),
7740 2 : ),
7741 2 : (
7742 2 : Lsn(0x60),
7743 2 : KeyLogAtLsn(vec![(
7744 2 : Lsn(0x60),
7745 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
7746 2 : )]),
7747 2 : ),
7748 2 : ],
7749 2 : above_horizon: KeyLogAtLsn(vec![(
7750 2 : Lsn(0x70),
7751 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
7752 2 : )]),
7753 2 : };
7754 2 : assert_eq!(res, expected_res);
7755 2 :
7756 2 : Ok(())
7757 2 : }
7758 :
7759 : #[tokio::test]
7760 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
7761 2 : let harness =
7762 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
7763 8 : let (tenant, ctx) = harness.load().await;
7764 2 :
7765 518 : fn get_key(id: u32) -> Key {
7766 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7767 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7768 518 : key.field6 = id;
7769 518 : key
7770 518 : }
7771 2 :
7772 2 : let img_layer = (0..10)
7773 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7774 2 : .collect_vec();
7775 2 :
7776 2 : let delta1 = vec![
7777 2 : (
7778 2 : get_key(1),
7779 2 : Lsn(0x20),
7780 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7781 2 : ),
7782 2 : (
7783 2 : get_key(2),
7784 2 : Lsn(0x30),
7785 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7786 2 : ),
7787 2 : (
7788 2 : get_key(3),
7789 2 : Lsn(0x28),
7790 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7791 2 : ),
7792 2 : (
7793 2 : get_key(3),
7794 2 : Lsn(0x30),
7795 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7796 2 : ),
7797 2 : (
7798 2 : get_key(3),
7799 2 : Lsn(0x40),
7800 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7801 2 : ),
7802 2 : ];
7803 2 : let delta2 = vec![
7804 2 : (
7805 2 : get_key(5),
7806 2 : Lsn(0x20),
7807 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7808 2 : ),
7809 2 : (
7810 2 : get_key(6),
7811 2 : Lsn(0x20),
7812 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7813 2 : ),
7814 2 : ];
7815 2 : let delta3 = vec![
7816 2 : (
7817 2 : get_key(8),
7818 2 : Lsn(0x48),
7819 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7820 2 : ),
7821 2 : (
7822 2 : get_key(9),
7823 2 : Lsn(0x48),
7824 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7825 2 : ),
7826 2 : ];
7827 2 :
7828 2 : let tline = tenant
7829 2 : .create_test_timeline_with_layers(
7830 2 : TIMELINE_ID,
7831 2 : Lsn(0x10),
7832 2 : DEFAULT_PG_VERSION,
7833 2 : &ctx,
7834 2 : vec![
7835 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7836 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7837 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7838 2 : ], // delta layers
7839 2 : vec![(Lsn(0x10), img_layer)], // image layers
7840 2 : Lsn(0x50),
7841 2 : )
7842 47 : .await?;
7843 2 : {
7844 2 : // Update GC info
7845 2 : let mut guard = tline.gc_info.write().unwrap();
7846 2 : *guard = GcInfo {
7847 2 : retain_lsns: vec![
7848 2 : (Lsn(0x10), tline.timeline_id),
7849 2 : (Lsn(0x20), tline.timeline_id),
7850 2 : ],
7851 2 : cutoffs: GcCutoffs {
7852 2 : time: Lsn(0x30),
7853 2 : space: Lsn(0x30),
7854 2 : },
7855 2 : leases: Default::default(),
7856 2 : within_ancestor_pitr: false,
7857 2 : };
7858 2 : }
7859 2 :
7860 2 : let expected_result = [
7861 2 : Bytes::from_static(b"value 0@0x10"),
7862 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7863 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7864 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7865 2 : Bytes::from_static(b"value 4@0x10"),
7866 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7867 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7868 2 : Bytes::from_static(b"value 7@0x10"),
7869 2 : Bytes::from_static(b"value 8@0x10@0x48"),
7870 2 : Bytes::from_static(b"value 9@0x10@0x48"),
7871 2 : ];
7872 2 :
7873 2 : let expected_result_at_gc_horizon = [
7874 2 : Bytes::from_static(b"value 0@0x10"),
7875 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7876 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7877 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7878 2 : Bytes::from_static(b"value 4@0x10"),
7879 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7880 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7881 2 : Bytes::from_static(b"value 7@0x10"),
7882 2 : Bytes::from_static(b"value 8@0x10"),
7883 2 : Bytes::from_static(b"value 9@0x10"),
7884 2 : ];
7885 2 :
7886 2 : let expected_result_at_lsn_20 = [
7887 2 : Bytes::from_static(b"value 0@0x10"),
7888 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7889 2 : Bytes::from_static(b"value 2@0x10"),
7890 2 : Bytes::from_static(b"value 3@0x10"),
7891 2 : Bytes::from_static(b"value 4@0x10"),
7892 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7893 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7894 2 : Bytes::from_static(b"value 7@0x10"),
7895 2 : Bytes::from_static(b"value 8@0x10"),
7896 2 : Bytes::from_static(b"value 9@0x10"),
7897 2 : ];
7898 2 :
7899 2 : let expected_result_at_lsn_10 = [
7900 2 : Bytes::from_static(b"value 0@0x10"),
7901 2 : Bytes::from_static(b"value 1@0x10"),
7902 2 : Bytes::from_static(b"value 2@0x10"),
7903 2 : Bytes::from_static(b"value 3@0x10"),
7904 2 : Bytes::from_static(b"value 4@0x10"),
7905 2 : Bytes::from_static(b"value 5@0x10"),
7906 2 : Bytes::from_static(b"value 6@0x10"),
7907 2 : Bytes::from_static(b"value 7@0x10"),
7908 2 : Bytes::from_static(b"value 8@0x10"),
7909 2 : Bytes::from_static(b"value 9@0x10"),
7910 2 : ];
7911 2 :
7912 12 : let verify_result = || async {
7913 12 : let gc_horizon = {
7914 12 : let gc_info = tline.gc_info.read().unwrap();
7915 12 : gc_info.cutoffs.time
7916 12 : };
7917 132 : for idx in 0..10 {
7918 120 : assert_eq!(
7919 120 : tline
7920 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7921 142 : .await
7922 120 : .unwrap(),
7923 120 : &expected_result[idx]
7924 12 : );
7925 120 : assert_eq!(
7926 120 : tline
7927 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
7928 96 : .await
7929 120 : .unwrap(),
7930 120 : &expected_result_at_gc_horizon[idx]
7931 12 : );
7932 120 : assert_eq!(
7933 120 : tline
7934 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
7935 78 : .await
7936 120 : .unwrap(),
7937 120 : &expected_result_at_lsn_20[idx]
7938 12 : );
7939 120 : assert_eq!(
7940 120 : tline
7941 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
7942 61 : .await
7943 120 : .unwrap(),
7944 120 : &expected_result_at_lsn_10[idx]
7945 12 : );
7946 12 : }
7947 12 : };
7948 2 :
7949 70 : verify_result().await;
7950 2 :
7951 2 : let cancel = CancellationToken::new();
7952 2 : let mut dryrun_flags = EnumSet::new();
7953 2 : dryrun_flags.insert(CompactFlags::DryRun);
7954 2 :
7955 2 : tline
7956 2 : .compact_with_gc(&cancel, dryrun_flags, &ctx)
7957 48 : .await
7958 2 : .unwrap();
7959 2 : // We expect layer map to be the same b/c the dry run flag, but we don't know whether there will be other background jobs
7960 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
7961 57 : verify_result().await;
7962 2 :
7963 2 : tline
7964 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7965 71 : .await
7966 2 : .unwrap();
7967 73 : verify_result().await;
7968 2 :
7969 2 : // compact again
7970 2 : tline
7971 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7972 55 : .await
7973 2 : .unwrap();
7974 63 : verify_result().await;
7975 2 :
7976 2 : // increase GC horizon and compact again
7977 2 : {
7978 2 : // Update GC info
7979 2 : let mut guard = tline.gc_info.write().unwrap();
7980 2 : guard.cutoffs.time = Lsn(0x38);
7981 2 : guard.cutoffs.space = Lsn(0x38);
7982 2 : }
7983 2 : tline
7984 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7985 44 : .await
7986 2 : .unwrap();
7987 57 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
7988 2 :
7989 2 : // not increasing the GC horizon and compact again
7990 2 : tline
7991 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
7992 44 : .await
7993 2 : .unwrap();
7994 57 : verify_result().await;
7995 2 :
7996 2 : Ok(())
7997 2 : }
7998 :
7999 : #[tokio::test]
8000 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
8001 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
8002 8 : let (tenant, ctx) = harness.load().await;
8003 2 :
8004 126 : fn get_key(id: u32) -> Key {
8005 126 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8006 126 : key.field6 = id;
8007 126 : key
8008 126 : }
8009 2 :
8010 2 : let img_layer = (0..10)
8011 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8012 2 : .collect_vec();
8013 2 :
8014 2 : let delta1 = vec![
8015 2 : (
8016 2 : get_key(1),
8017 2 : Lsn(0x20),
8018 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8019 2 : ),
8020 2 : (
8021 2 : get_key(2),
8022 2 : Lsn(0x30),
8023 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8024 2 : ),
8025 2 : (
8026 2 : get_key(3),
8027 2 : Lsn(0x28),
8028 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8029 2 : ),
8030 2 : (
8031 2 : get_key(3),
8032 2 : Lsn(0x30),
8033 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8034 2 : ),
8035 2 : (
8036 2 : get_key(3),
8037 2 : Lsn(0x40),
8038 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8039 2 : ),
8040 2 : ];
8041 2 : let delta2 = vec![
8042 2 : (
8043 2 : get_key(5),
8044 2 : Lsn(0x20),
8045 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8046 2 : ),
8047 2 : (
8048 2 : get_key(6),
8049 2 : Lsn(0x20),
8050 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8051 2 : ),
8052 2 : ];
8053 2 : let delta3 = vec![
8054 2 : (
8055 2 : get_key(8),
8056 2 : Lsn(0x48),
8057 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8058 2 : ),
8059 2 : (
8060 2 : get_key(9),
8061 2 : Lsn(0x48),
8062 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8063 2 : ),
8064 2 : ];
8065 2 :
8066 2 : let parent_tline = tenant
8067 2 : .create_test_timeline_with_layers(
8068 2 : TIMELINE_ID,
8069 2 : Lsn(0x10),
8070 2 : DEFAULT_PG_VERSION,
8071 2 : &ctx,
8072 2 : vec![], // delta layers
8073 2 : vec![(Lsn(0x18), img_layer)], // image layers
8074 2 : Lsn(0x18),
8075 2 : )
8076 29 : .await?;
8077 2 :
8078 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8079 2 :
8080 2 : let branch_tline = tenant
8081 2 : .branch_timeline_test_with_layers(
8082 2 : &parent_tline,
8083 2 : NEW_TIMELINE_ID,
8084 2 : Some(Lsn(0x18)),
8085 2 : &ctx,
8086 2 : vec![
8087 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8088 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8089 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8090 2 : ], // delta layers
8091 2 : vec![], // image layers
8092 2 : Lsn(0x50),
8093 2 : )
8094 18 : .await?;
8095 2 :
8096 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
8097 2 :
8098 2 : {
8099 2 : // Update GC info
8100 2 : let mut guard = parent_tline.gc_info.write().unwrap();
8101 2 : *guard = GcInfo {
8102 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id)],
8103 2 : cutoffs: GcCutoffs {
8104 2 : time: Lsn(0x10),
8105 2 : space: Lsn(0x10),
8106 2 : },
8107 2 : leases: Default::default(),
8108 2 : within_ancestor_pitr: false,
8109 2 : };
8110 2 : }
8111 2 :
8112 2 : {
8113 2 : // Update GC info
8114 2 : let mut guard = branch_tline.gc_info.write().unwrap();
8115 2 : *guard = GcInfo {
8116 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id)],
8117 2 : cutoffs: GcCutoffs {
8118 2 : time: Lsn(0x50),
8119 2 : space: Lsn(0x50),
8120 2 : },
8121 2 : leases: Default::default(),
8122 2 : within_ancestor_pitr: false,
8123 2 : };
8124 2 : }
8125 2 :
8126 2 : let expected_result_at_gc_horizon = [
8127 2 : Bytes::from_static(b"value 0@0x10"),
8128 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8129 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8130 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8131 2 : Bytes::from_static(b"value 4@0x10"),
8132 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8133 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8134 2 : Bytes::from_static(b"value 7@0x10"),
8135 2 : Bytes::from_static(b"value 8@0x10@0x48"),
8136 2 : Bytes::from_static(b"value 9@0x10@0x48"),
8137 2 : ];
8138 2 :
8139 2 : let expected_result_at_lsn_40 = [
8140 2 : Bytes::from_static(b"value 0@0x10"),
8141 2 : Bytes::from_static(b"value 1@0x10@0x20"),
8142 2 : Bytes::from_static(b"value 2@0x10@0x30"),
8143 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8144 2 : Bytes::from_static(b"value 4@0x10"),
8145 2 : Bytes::from_static(b"value 5@0x10@0x20"),
8146 2 : Bytes::from_static(b"value 6@0x10@0x20"),
8147 2 : Bytes::from_static(b"value 7@0x10"),
8148 2 : Bytes::from_static(b"value 8@0x10"),
8149 2 : Bytes::from_static(b"value 9@0x10"),
8150 2 : ];
8151 2 :
8152 4 : let verify_result = || async {
8153 44 : for idx in 0..10 {
8154 40 : assert_eq!(
8155 40 : branch_tline
8156 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8157 59 : .await
8158 40 : .unwrap(),
8159 40 : &expected_result_at_gc_horizon[idx]
8160 4 : );
8161 40 : assert_eq!(
8162 40 : branch_tline
8163 40 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
8164 32 : .await
8165 40 : .unwrap(),
8166 40 : &expected_result_at_lsn_40[idx]
8167 4 : );
8168 4 : }
8169 4 : };
8170 2 :
8171 46 : verify_result().await;
8172 2 :
8173 2 : let cancel = CancellationToken::new();
8174 2 : branch_tline
8175 2 : .compact_with_gc(&cancel, EnumSet::new(), &ctx)
8176 35 : .await
8177 2 : .unwrap();
8178 2 :
8179 45 : verify_result().await;
8180 2 :
8181 2 : Ok(())
8182 2 : }
8183 : }
|
