Line data Source code
1 : //! Timeline repository implementation that keeps old data in layer files, and
2 : //! the recent changes in ephemeral files.
3 : //!
4 : //! See tenant/*_layer.rs files. The functions here are responsible for locating
5 : //! the correct layer for the get/put call, walking back the timeline branching
6 : //! history as needed.
7 : //!
8 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
9 : //! directory. See docs/pageserver-storage.md for how the files are managed.
10 : //! In addition to the layer files, there is a metadata file in the same
11 : //! directory that contains information about the timeline, in particular its
12 : //! parent timeline, and the last LSN that has been written to disk.
13 : //!
14 :
15 : use anyhow::{bail, Context};
16 : use arc_swap::ArcSwap;
17 : use camino::Utf8Path;
18 : use camino::Utf8PathBuf;
19 : use chrono::NaiveDateTime;
20 : use enumset::EnumSet;
21 : use futures::stream::FuturesUnordered;
22 : use futures::StreamExt;
23 : use pageserver_api::models;
24 : use pageserver_api::models::LsnLease;
25 : use pageserver_api::models::TimelineArchivalState;
26 : use pageserver_api::models::TimelineState;
27 : use pageserver_api::models::TopTenantShardItem;
28 : use pageserver_api::models::WalRedoManagerStatus;
29 : use pageserver_api::shard::ShardIdentity;
30 : use pageserver_api::shard::ShardStripeSize;
31 : use pageserver_api::shard::TenantShardId;
32 : use remote_storage::DownloadError;
33 : use remote_storage::GenericRemoteStorage;
34 : use remote_storage::TimeoutOrCancel;
35 : use remote_timeline_client::manifest::{
36 : OffloadedTimelineManifest, TenantManifest, LATEST_TENANT_MANIFEST_VERSION,
37 : };
38 : use remote_timeline_client::UploadQueueNotReadyError;
39 : use std::collections::BTreeMap;
40 : use std::fmt;
41 : use std::future::Future;
42 : use std::sync::atomic::AtomicBool;
43 : use std::sync::Weak;
44 : use std::time::SystemTime;
45 : use storage_broker::BrokerClientChannel;
46 : use timeline::import_pgdata;
47 : use timeline::offload::offload_timeline;
48 : use timeline::ShutdownMode;
49 : use tokio::io::BufReader;
50 : use tokio::sync::watch;
51 : use tokio::task::JoinSet;
52 : use tokio_util::sync::CancellationToken;
53 : use tracing::*;
54 : use upload_queue::NotInitialized;
55 : use utils::backoff;
56 : use utils::circuit_breaker::CircuitBreaker;
57 : use utils::completion;
58 : use utils::crashsafe::path_with_suffix_extension;
59 : use utils::failpoint_support;
60 : use utils::fs_ext;
61 : use utils::pausable_failpoint;
62 : use utils::sync::gate::Gate;
63 : use utils::sync::gate::GateGuard;
64 : use utils::timeout::timeout_cancellable;
65 : use utils::timeout::TimeoutCancellableError;
66 : use utils::zstd::create_zst_tarball;
67 : use utils::zstd::extract_zst_tarball;
68 :
69 : use self::config::AttachedLocationConfig;
70 : use self::config::AttachmentMode;
71 : use self::config::LocationConf;
72 : use self::config::TenantConf;
73 : use self::metadata::TimelineMetadata;
74 : use self::mgr::GetActiveTenantError;
75 : use self::mgr::GetTenantError;
76 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
77 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
78 : use self::timeline::uninit::TimelineCreateGuard;
79 : use self::timeline::uninit::TimelineExclusionError;
80 : use self::timeline::uninit::UninitializedTimeline;
81 : use self::timeline::EvictionTaskTenantState;
82 : use self::timeline::GcCutoffs;
83 : use self::timeline::TimelineDeleteProgress;
84 : use self::timeline::TimelineResources;
85 : use self::timeline::WaitLsnError;
86 : use crate::config::PageServerConf;
87 : use crate::context::{DownloadBehavior, RequestContext};
88 : use crate::deletion_queue::DeletionQueueClient;
89 : use crate::deletion_queue::DeletionQueueError;
90 : use crate::import_datadir;
91 : use crate::is_uninit_mark;
92 : use crate::l0_flush::L0FlushGlobalState;
93 : use crate::metrics::TENANT;
94 : use crate::metrics::{
95 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
96 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
97 : };
98 : use crate::task_mgr;
99 : use crate::task_mgr::TaskKind;
100 : use crate::tenant::config::LocationMode;
101 : use crate::tenant::config::TenantConfOpt;
102 : use crate::tenant::gc_result::GcResult;
103 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
104 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
105 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
106 : use crate::tenant::remote_timeline_client::INITDB_PATH;
107 : use crate::tenant::storage_layer::DeltaLayer;
108 : use crate::tenant::storage_layer::ImageLayer;
109 : use crate::walingest::WalLagCooldown;
110 : use crate::walredo;
111 : use crate::InitializationOrder;
112 : use std::collections::hash_map::Entry;
113 : use std::collections::HashMap;
114 : use std::collections::HashSet;
115 : use std::fmt::Debug;
116 : use std::fmt::Display;
117 : use std::fs;
118 : use std::fs::File;
119 : use std::sync::atomic::{AtomicU64, Ordering};
120 : use std::sync::Arc;
121 : use std::sync::Mutex;
122 : use std::time::{Duration, Instant};
123 :
124 : use crate::span;
125 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
126 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
127 : use crate::virtual_file::VirtualFile;
128 : use crate::walredo::PostgresRedoManager;
129 : use crate::TEMP_FILE_SUFFIX;
130 : use once_cell::sync::Lazy;
131 : pub use pageserver_api::models::TenantState;
132 : use tokio::sync::Semaphore;
133 :
134 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
135 : use utils::{
136 : crashsafe,
137 : generation::Generation,
138 : id::TimelineId,
139 : lsn::{Lsn, RecordLsn},
140 : };
141 :
142 : pub mod blob_io;
143 : pub mod block_io;
144 : pub mod vectored_blob_io;
145 :
146 : pub mod disk_btree;
147 : pub(crate) mod ephemeral_file;
148 : pub mod layer_map;
149 :
150 : pub mod metadata;
151 : pub mod remote_timeline_client;
152 : pub mod storage_layer;
153 :
154 : pub mod checks;
155 : pub mod config;
156 : pub mod mgr;
157 : pub mod secondary;
158 : pub mod tasks;
159 : pub mod upload_queue;
160 :
161 : pub(crate) mod timeline;
162 :
163 : pub mod size;
164 :
165 : mod gc_block;
166 : mod gc_result;
167 : pub(crate) mod throttle;
168 :
169 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
170 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
171 :
172 : // re-export for use in walreceiver
173 : pub use crate::tenant::timeline::WalReceiverInfo;
174 :
175 : /// The "tenants" part of `tenants/<tenant>/timelines...`
176 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
177 :
178 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
179 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
180 :
181 : /// References to shared objects that are passed into each tenant, such
182 : /// as the shared remote storage client and process initialization state.
183 : #[derive(Clone)]
184 : pub struct TenantSharedResources {
185 : pub broker_client: storage_broker::BrokerClientChannel,
186 : pub remote_storage: GenericRemoteStorage,
187 : pub deletion_queue_client: DeletionQueueClient,
188 : pub l0_flush_global_state: L0FlushGlobalState,
189 : }
190 :
191 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
192 : /// for an attached tenant is a subset of the [`LocationConf`], represented
193 : /// in this struct.
194 : #[derive(Clone)]
195 : pub(super) struct AttachedTenantConf {
196 : tenant_conf: TenantConfOpt,
197 : location: AttachedLocationConfig,
198 : /// The deadline before which we are blocked from GC so that
199 : /// leases have a chance to be renewed.
200 : lsn_lease_deadline: Option<tokio::time::Instant>,
201 : }
202 :
203 : impl AttachedTenantConf {
204 192 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
205 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
206 : //
207 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
208 : // length, we guarantee that all the leases we granted before will have a chance to renew
209 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
210 192 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
211 192 : Some(
212 192 : tokio::time::Instant::now()
213 192 : + tenant_conf
214 192 : .lsn_lease_length
215 192 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
216 192 : )
217 : } else {
218 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
219 : // because we don't do GC in these modes.
220 0 : None
221 : };
222 :
223 192 : Self {
224 192 : tenant_conf,
225 192 : location,
226 192 : lsn_lease_deadline,
227 192 : }
228 192 : }
229 :
230 192 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
231 192 : match &location_conf.mode {
232 192 : LocationMode::Attached(attach_conf) => {
233 192 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
234 : }
235 : LocationMode::Secondary(_) => {
236 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
237 : }
238 : }
239 192 : }
240 :
241 762 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
242 762 : self.lsn_lease_deadline
243 762 : .map(|d| tokio::time::Instant::now() < d)
244 762 : .unwrap_or(false)
245 762 : }
246 : }
247 : struct TimelinePreload {
248 : timeline_id: TimelineId,
249 : client: RemoteTimelineClient,
250 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
251 : }
252 :
253 : pub(crate) struct TenantPreload {
254 : tenant_manifest: TenantManifest,
255 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
256 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
257 : }
258 :
259 : /// When we spawn a tenant, there is a special mode for tenant creation that
260 : /// avoids trying to read anything from remote storage.
261 : pub(crate) enum SpawnMode {
262 : /// Activate as soon as possible
263 : Eager,
264 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
265 : Lazy,
266 : }
267 :
268 : ///
269 : /// Tenant consists of multiple timelines. Keep them in a hash table.
270 : ///
271 : pub struct Tenant {
272 : // Global pageserver config parameters
273 : pub conf: &'static PageServerConf,
274 :
275 : /// The value creation timestamp, used to measure activation delay, see:
276 : /// <https://github.com/neondatabase/neon/issues/4025>
277 : constructed_at: Instant,
278 :
279 : state: watch::Sender<TenantState>,
280 :
281 : // Overridden tenant-specific config parameters.
282 : // We keep TenantConfOpt sturct here to preserve the information
283 : // about parameters that are not set.
284 : // This is necessary to allow global config updates.
285 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
286 :
287 : tenant_shard_id: TenantShardId,
288 :
289 : // The detailed sharding information, beyond the number/count in tenant_shard_id
290 : shard_identity: ShardIdentity,
291 :
292 : /// The remote storage generation, used to protect S3 objects from split-brain.
293 : /// Does not change over the lifetime of the [`Tenant`] object.
294 : ///
295 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
296 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
297 : generation: Generation,
298 :
299 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
300 :
301 : /// During timeline creation, we first insert the TimelineId to the
302 : /// creating map, then `timelines`, then remove it from the creating map.
303 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
304 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
305 :
306 : /// Possibly offloaded and archived timelines
307 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
308 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
309 :
310 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
311 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
312 : /// each other (this could be optimized to coalesce writes if necessary).
313 : ///
314 : /// The contents of the Mutex are the last manifest we successfully uploaded
315 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
316 :
317 : // This mutex prevents creation of new timelines during GC.
318 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
319 : // `timelines` mutex during all GC iteration
320 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
321 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
322 : // timeout...
323 : gc_cs: tokio::sync::Mutex<()>,
324 : walredo_mgr: Option<Arc<WalRedoManager>>,
325 :
326 : // provides access to timeline data sitting in the remote storage
327 : pub(crate) remote_storage: GenericRemoteStorage,
328 :
329 : // Access to global deletion queue for when this tenant wants to schedule a deletion
330 : deletion_queue_client: DeletionQueueClient,
331 :
332 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
333 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
334 : cached_synthetic_tenant_size: Arc<AtomicU64>,
335 :
336 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
337 :
338 : /// Track repeated failures to compact, so that we can back off.
339 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
340 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
341 :
342 : /// If the tenant is in Activating state, notify this to encourage it
343 : /// to proceed to Active as soon as possible, rather than waiting for lazy
344 : /// background warmup.
345 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
346 :
347 : /// Time it took for the tenant to activate. Zero if not active yet.
348 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
349 :
350 : // Cancellation token fires when we have entered shutdown(). This is a parent of
351 : // Timelines' cancellation token.
352 : pub(crate) cancel: CancellationToken,
353 :
354 : // Users of the Tenant such as the page service must take this Gate to avoid
355 : // trying to use a Tenant which is shutting down.
356 : pub(crate) gate: Gate,
357 :
358 : /// Throttle applied at the top of [`Timeline::get`].
359 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
360 : pub(crate) timeline_get_throttle:
361 : Arc<throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
362 :
363 : /// An ongoing timeline detach concurrency limiter.
364 : ///
365 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
366 : /// to have two running at the same time. A different one can be started if an earlier one
367 : /// has failed for whatever reason.
368 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
369 :
370 : /// `index_part.json` based gc blocking reason tracking.
371 : ///
372 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
373 : /// proceeding.
374 : pub(crate) gc_block: gc_block::GcBlock,
375 :
376 : l0_flush_global_state: L0FlushGlobalState,
377 : }
378 : impl std::fmt::Debug for Tenant {
379 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
380 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
381 0 : }
382 : }
383 :
384 : pub(crate) enum WalRedoManager {
385 : Prod(WalredoManagerId, PostgresRedoManager),
386 : #[cfg(test)]
387 : Test(harness::TestRedoManager),
388 : }
389 :
390 0 : #[derive(thiserror::Error, Debug)]
391 : #[error("pageserver is shutting down")]
392 : pub(crate) struct GlobalShutDown;
393 :
394 : impl WalRedoManager {
395 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
396 0 : let id = WalredoManagerId::next();
397 0 : let arc = Arc::new(Self::Prod(id, mgr));
398 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
399 0 : match &mut *guard {
400 0 : Some(map) => {
401 0 : map.insert(id, Arc::downgrade(&arc));
402 0 : Ok(arc)
403 : }
404 0 : None => Err(GlobalShutDown),
405 : }
406 0 : }
407 : }
408 :
409 : impl Drop for WalRedoManager {
410 10 : fn drop(&mut self) {
411 10 : match self {
412 0 : Self::Prod(id, _) => {
413 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
414 0 : if let Some(map) = &mut *guard {
415 0 : map.remove(id).expect("new() registers, drop() unregisters");
416 0 : }
417 : }
418 : #[cfg(test)]
419 10 : Self::Test(_) => {
420 10 : // Not applicable to test redo manager
421 10 : }
422 : }
423 10 : }
424 : }
425 :
426 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
427 : /// the walredo processes outside of the regular order.
428 : ///
429 : /// This is necessary to work around a systemd bug where it freezes if there are
430 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
431 : #[allow(clippy::type_complexity)]
432 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
433 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
434 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
435 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
436 : pub(crate) struct WalredoManagerId(u64);
437 : impl WalredoManagerId {
438 0 : pub fn next() -> Self {
439 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
440 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
441 0 : if id == 0 {
442 0 : panic!("WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique");
443 0 : }
444 0 : Self(id)
445 0 : }
446 : }
447 :
448 : #[cfg(test)]
449 : impl From<harness::TestRedoManager> for WalRedoManager {
450 192 : fn from(mgr: harness::TestRedoManager) -> Self {
451 192 : Self::Test(mgr)
452 192 : }
453 : }
454 :
455 : impl WalRedoManager {
456 6 : pub(crate) async fn shutdown(&self) -> bool {
457 6 : match self {
458 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
459 : #[cfg(test)]
460 : Self::Test(_) => {
461 : // Not applicable to test redo manager
462 6 : true
463 : }
464 : }
465 6 : }
466 :
467 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
468 0 : match self {
469 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
470 0 : #[cfg(test)]
471 0 : Self::Test(_) => {
472 0 : // Not applicable to test redo manager
473 0 : }
474 0 : }
475 0 : }
476 :
477 : /// # Cancel-Safety
478 : ///
479 : /// This method is cancellation-safe.
480 410 : pub async fn request_redo(
481 410 : &self,
482 410 : key: pageserver_api::key::Key,
483 410 : lsn: Lsn,
484 410 : base_img: Option<(Lsn, bytes::Bytes)>,
485 410 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
486 410 : pg_version: u32,
487 410 : ) -> Result<bytes::Bytes, walredo::Error> {
488 410 : match self {
489 0 : Self::Prod(_, mgr) => {
490 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
491 0 : .await
492 : }
493 : #[cfg(test)]
494 410 : Self::Test(mgr) => {
495 410 : mgr.request_redo(key, lsn, base_img, records, pg_version)
496 0 : .await
497 : }
498 : }
499 410 : }
500 :
501 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
502 0 : match self {
503 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
504 0 : #[cfg(test)]
505 0 : WalRedoManager::Test(_) => None,
506 0 : }
507 0 : }
508 : }
509 :
510 : /// A very lightweight memory representation of an offloaded timeline.
511 : ///
512 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
513 : /// like unoffloading them, or (at a later date), decide to perform flattening.
514 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
515 : /// more offloaded timelines than we can manage ones that aren't.
516 : pub struct OffloadedTimeline {
517 : pub tenant_shard_id: TenantShardId,
518 : pub timeline_id: TimelineId,
519 : pub ancestor_timeline_id: Option<TimelineId>,
520 : /// Whether to retain the branch lsn at the ancestor or not
521 : pub ancestor_retain_lsn: Option<Lsn>,
522 :
523 : /// When the timeline was archived.
524 : ///
525 : /// Present for future flattening deliberations.
526 : pub archived_at: NaiveDateTime,
527 :
528 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
529 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
530 : pub delete_progress: TimelineDeleteProgress,
531 :
532 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
533 : pub deleted_from_ancestor: AtomicBool,
534 : }
535 :
536 : impl OffloadedTimeline {
537 : /// Obtains an offloaded timeline from a given timeline object.
538 : ///
539 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
540 : /// the timeline is not in a stopped state.
541 : /// Panics if the timeline is not archived.
542 2 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
543 2 : let (ancestor_retain_lsn, ancestor_timeline_id) =
544 2 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
545 2 : let ancestor_lsn = timeline.get_ancestor_lsn();
546 2 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
547 2 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
548 2 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
549 2 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
550 : } else {
551 0 : (None, None)
552 : };
553 2 : let archived_at = timeline
554 2 : .remote_client
555 2 : .archived_at_stopped_queue()?
556 2 : .expect("must be called on an archived timeline");
557 2 : Ok(Self {
558 2 : tenant_shard_id: timeline.tenant_shard_id,
559 2 : timeline_id: timeline.timeline_id,
560 2 : ancestor_timeline_id,
561 2 : ancestor_retain_lsn,
562 2 : archived_at,
563 2 :
564 2 : delete_progress: timeline.delete_progress.clone(),
565 2 : deleted_from_ancestor: AtomicBool::new(false),
566 2 : })
567 2 : }
568 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
569 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
570 0 : // by the `initialize_gc_info` function.
571 0 : let OffloadedTimelineManifest {
572 0 : timeline_id,
573 0 : ancestor_timeline_id,
574 0 : ancestor_retain_lsn,
575 0 : archived_at,
576 0 : } = *manifest;
577 0 : Self {
578 0 : tenant_shard_id,
579 0 : timeline_id,
580 0 : ancestor_timeline_id,
581 0 : ancestor_retain_lsn,
582 0 : archived_at,
583 0 : delete_progress: TimelineDeleteProgress::default(),
584 0 : deleted_from_ancestor: AtomicBool::new(false),
585 0 : }
586 0 : }
587 2 : fn manifest(&self) -> OffloadedTimelineManifest {
588 2 : let Self {
589 2 : timeline_id,
590 2 : ancestor_timeline_id,
591 2 : ancestor_retain_lsn,
592 2 : archived_at,
593 2 : ..
594 2 : } = self;
595 2 : OffloadedTimelineManifest {
596 2 : timeline_id: *timeline_id,
597 2 : ancestor_timeline_id: *ancestor_timeline_id,
598 2 : ancestor_retain_lsn: *ancestor_retain_lsn,
599 2 : archived_at: *archived_at,
600 2 : }
601 2 : }
602 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
603 0 : fn delete_from_ancestor_with_timelines(
604 0 : &self,
605 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
606 0 : ) {
607 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
608 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
609 : {
610 0 : if let Some((_, ancestor_timeline)) = timelines
611 0 : .iter()
612 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
613 : {
614 0 : let removal_happened = ancestor_timeline
615 0 : .gc_info
616 0 : .write()
617 0 : .unwrap()
618 0 : .remove_child_offloaded(self.timeline_id);
619 0 : if !removal_happened {
620 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
621 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
622 0 : }
623 0 : }
624 0 : }
625 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
626 0 : }
627 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
628 : ///
629 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
630 2 : fn defuse_for_tenant_drop(&self) {
631 2 : self.deleted_from_ancestor.store(true, Ordering::Release);
632 2 : }
633 : }
634 :
635 : impl fmt::Debug for OffloadedTimeline {
636 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
637 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
638 0 : }
639 : }
640 :
641 : impl Drop for OffloadedTimeline {
642 2 : fn drop(&mut self) {
643 2 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
644 0 : tracing::warn!(
645 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
646 : self.timeline_id
647 : );
648 2 : }
649 2 : }
650 : }
651 :
652 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
653 : pub enum MaybeOffloaded {
654 : Yes,
655 : No,
656 : }
657 :
658 : #[derive(Clone, Debug)]
659 : pub enum TimelineOrOffloaded {
660 : Timeline(Arc<Timeline>),
661 : Offloaded(Arc<OffloadedTimeline>),
662 : }
663 :
664 : impl TimelineOrOffloaded {
665 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
666 0 : match self {
667 0 : TimelineOrOffloaded::Timeline(timeline) => {
668 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
669 : }
670 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
671 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
672 : }
673 : }
674 0 : }
675 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
676 0 : self.arc_ref().tenant_shard_id()
677 0 : }
678 0 : pub fn timeline_id(&self) -> TimelineId {
679 0 : self.arc_ref().timeline_id()
680 0 : }
681 2 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
682 2 : match self {
683 2 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
684 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
685 : }
686 2 : }
687 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
688 0 : match self {
689 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
690 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
691 : }
692 0 : }
693 : }
694 :
695 : pub enum TimelineOrOffloadedArcRef<'a> {
696 : Timeline(&'a Arc<Timeline>),
697 : Offloaded(&'a Arc<OffloadedTimeline>),
698 : }
699 :
700 : impl TimelineOrOffloadedArcRef<'_> {
701 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
702 0 : match self {
703 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
704 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
705 : }
706 0 : }
707 0 : pub fn timeline_id(&self) -> TimelineId {
708 0 : match self {
709 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
710 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
711 : }
712 0 : }
713 : }
714 :
715 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
716 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
717 0 : Self::Timeline(timeline)
718 0 : }
719 : }
720 :
721 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
722 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
723 0 : Self::Offloaded(timeline)
724 0 : }
725 : }
726 :
727 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
728 : pub enum GetTimelineError {
729 : #[error("Timeline is shutting down")]
730 : ShuttingDown,
731 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
732 : NotActive {
733 : tenant_id: TenantShardId,
734 : timeline_id: TimelineId,
735 : state: TimelineState,
736 : },
737 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
738 : NotFound {
739 : tenant_id: TenantShardId,
740 : timeline_id: TimelineId,
741 : },
742 : }
743 :
744 0 : #[derive(Debug, thiserror::Error)]
745 : pub enum LoadLocalTimelineError {
746 : #[error("FailedToLoad")]
747 : Load(#[source] anyhow::Error),
748 : #[error("FailedToResumeDeletion")]
749 : ResumeDeletion(#[source] anyhow::Error),
750 : }
751 :
752 0 : #[derive(thiserror::Error)]
753 : pub enum DeleteTimelineError {
754 : #[error("NotFound")]
755 : NotFound,
756 :
757 : #[error("HasChildren")]
758 : HasChildren(Vec<TimelineId>),
759 :
760 : #[error("Timeline deletion is already in progress")]
761 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
762 :
763 : #[error("Cancelled")]
764 : Cancelled,
765 :
766 : #[error(transparent)]
767 : Other(#[from] anyhow::Error),
768 : }
769 :
770 : impl Debug for DeleteTimelineError {
771 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
772 0 : match self {
773 0 : Self::NotFound => write!(f, "NotFound"),
774 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
775 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
776 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
777 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
778 : }
779 0 : }
780 : }
781 :
782 0 : #[derive(thiserror::Error)]
783 : pub enum TimelineArchivalError {
784 : #[error("NotFound")]
785 : NotFound,
786 :
787 : #[error("Timeout")]
788 : Timeout,
789 :
790 : #[error("Cancelled")]
791 : Cancelled,
792 :
793 : #[error("ancestor is archived: {}", .0)]
794 : HasArchivedParent(TimelineId),
795 :
796 : #[error("HasUnarchivedChildren")]
797 : HasUnarchivedChildren(Vec<TimelineId>),
798 :
799 : #[error("Timeline archival is already in progress")]
800 : AlreadyInProgress,
801 :
802 : #[error(transparent)]
803 : Other(anyhow::Error),
804 : }
805 :
806 0 : #[derive(thiserror::Error, Debug)]
807 : pub(crate) enum TenantManifestError {
808 : #[error("Remote storage error: {0}")]
809 : RemoteStorage(anyhow::Error),
810 :
811 : #[error("Cancelled")]
812 : Cancelled,
813 : }
814 :
815 : impl From<TenantManifestError> for TimelineArchivalError {
816 0 : fn from(e: TenantManifestError) -> Self {
817 0 : match e {
818 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
819 0 : TenantManifestError::Cancelled => Self::Cancelled,
820 : }
821 0 : }
822 : }
823 :
824 : impl Debug for TimelineArchivalError {
825 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
826 0 : match self {
827 0 : Self::NotFound => write!(f, "NotFound"),
828 0 : Self::Timeout => write!(f, "Timeout"),
829 0 : Self::Cancelled => write!(f, "Cancelled"),
830 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
831 0 : Self::HasUnarchivedChildren(c) => {
832 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
833 : }
834 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
835 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
836 : }
837 0 : }
838 : }
839 :
840 : pub enum SetStoppingError {
841 : AlreadyStopping(completion::Barrier),
842 : Broken,
843 : }
844 :
845 : impl Debug for SetStoppingError {
846 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
847 0 : match self {
848 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
849 0 : Self::Broken => write!(f, "Broken"),
850 : }
851 0 : }
852 : }
853 :
854 : /// Arguments to [`Tenant::create_timeline`].
855 : ///
856 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
857 : /// is `None`, the result of the timeline create call is not deterministic.
858 : ///
859 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
860 : #[derive(Debug)]
861 : pub(crate) enum CreateTimelineParams {
862 : Bootstrap(CreateTimelineParamsBootstrap),
863 : Branch(CreateTimelineParamsBranch),
864 : ImportPgdata(CreateTimelineParamsImportPgdata),
865 : }
866 :
867 : #[derive(Debug)]
868 : pub(crate) struct CreateTimelineParamsBootstrap {
869 : pub(crate) new_timeline_id: TimelineId,
870 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
871 : pub(crate) pg_version: u32,
872 : }
873 :
874 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
875 : #[derive(Debug)]
876 : pub(crate) struct CreateTimelineParamsBranch {
877 : pub(crate) new_timeline_id: TimelineId,
878 : pub(crate) ancestor_timeline_id: TimelineId,
879 : pub(crate) ancestor_start_lsn: Option<Lsn>,
880 : }
881 :
882 : #[derive(Debug)]
883 : pub(crate) struct CreateTimelineParamsImportPgdata {
884 : pub(crate) new_timeline_id: TimelineId,
885 : pub(crate) location: import_pgdata::index_part_format::Location,
886 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
887 : }
888 :
889 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`] in [`Tenant::start_creating_timeline`].
890 : ///
891 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
892 : ///
893 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
894 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
895 : ///
896 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
897 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
898 : ///
899 : /// Notes:
900 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
901 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
902 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
903 : ///
904 : #[derive(Debug, Clone, PartialEq, Eq)]
905 : pub(crate) enum CreateTimelineIdempotency {
906 : /// NB: special treatment, see comment in [`Self`].
907 : FailWithConflict,
908 : Bootstrap {
909 : pg_version: u32,
910 : },
911 : /// NB: branches always have the same `pg_version` as their ancestor.
912 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
913 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
914 : /// determining the child branch pg_version.
915 : Branch {
916 : ancestor_timeline_id: TimelineId,
917 : ancestor_start_lsn: Lsn,
918 : },
919 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
920 : }
921 :
922 : #[derive(Debug, Clone, PartialEq, Eq)]
923 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
924 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
925 : }
926 :
927 : /// What is returned by [`Tenant::start_creating_timeline`].
928 : #[must_use]
929 : enum StartCreatingTimelineResult {
930 : CreateGuard(TimelineCreateGuard),
931 : Idempotent(Arc<Timeline>),
932 : }
933 :
934 : enum TimelineInitAndSyncResult {
935 : ReadyToActivate(Arc<Timeline>),
936 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
937 : }
938 :
939 : impl TimelineInitAndSyncResult {
940 0 : fn ready_to_activate(self) -> Option<Arc<Timeline>> {
941 0 : match self {
942 0 : Self::ReadyToActivate(timeline) => Some(timeline),
943 0 : _ => None,
944 : }
945 0 : }
946 : }
947 :
948 : #[must_use]
949 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
950 : timeline: Arc<Timeline>,
951 : import_pgdata: import_pgdata::index_part_format::Root,
952 : guard: TimelineCreateGuard,
953 : }
954 :
955 : /// What is returned by [`Tenant::create_timeline`].
956 : enum CreateTimelineResult {
957 : Created(Arc<Timeline>),
958 : Idempotent(Arc<Timeline>),
959 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`Tenant::timelines`] when
960 : /// we return this result, nor will this concrete object ever be added there.
961 : /// Cf method comment on [`Tenant::create_timeline_import_pgdata`].
962 : ImportSpawned(Arc<Timeline>),
963 : }
964 :
965 : impl CreateTimelineResult {
966 0 : fn discriminant(&self) -> &'static str {
967 0 : match self {
968 0 : Self::Created(_) => "Created",
969 0 : Self::Idempotent(_) => "Idempotent",
970 0 : Self::ImportSpawned(_) => "ImportSpawned",
971 : }
972 0 : }
973 0 : fn timeline(&self) -> &Arc<Timeline> {
974 0 : match self {
975 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
976 0 : }
977 0 : }
978 : /// Unit test timelines aren't activated, test has to do it if it needs to.
979 : #[cfg(test)]
980 230 : fn into_timeline_for_test(self) -> Arc<Timeline> {
981 230 : match self {
982 230 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
983 230 : }
984 230 : }
985 : }
986 :
987 2 : #[derive(thiserror::Error, Debug)]
988 : pub enum CreateTimelineError {
989 : #[error("creation of timeline with the given ID is in progress")]
990 : AlreadyCreating,
991 : #[error("timeline already exists with different parameters")]
992 : Conflict,
993 : #[error(transparent)]
994 : AncestorLsn(anyhow::Error),
995 : #[error("ancestor timeline is not active")]
996 : AncestorNotActive,
997 : #[error("ancestor timeline is archived")]
998 : AncestorArchived,
999 : #[error("tenant shutting down")]
1000 : ShuttingDown,
1001 : #[error(transparent)]
1002 : Other(#[from] anyhow::Error),
1003 : }
1004 :
1005 0 : #[derive(thiserror::Error, Debug)]
1006 : pub enum InitdbError {
1007 : #[error("Operation was cancelled")]
1008 : Cancelled,
1009 : #[error(transparent)]
1010 : Other(anyhow::Error),
1011 : #[error(transparent)]
1012 : Inner(postgres_initdb::Error),
1013 : }
1014 :
1015 : enum CreateTimelineCause {
1016 : Load,
1017 : Delete,
1018 : }
1019 :
1020 : enum LoadTimelineCause {
1021 : Attach,
1022 : Unoffload,
1023 : ImportPgdata {
1024 : create_guard: TimelineCreateGuard,
1025 : activate: ActivateTimelineArgs,
1026 : },
1027 : }
1028 :
1029 0 : #[derive(thiserror::Error, Debug)]
1030 : pub(crate) enum GcError {
1031 : // The tenant is shutting down
1032 : #[error("tenant shutting down")]
1033 : TenantCancelled,
1034 :
1035 : // The tenant is shutting down
1036 : #[error("timeline shutting down")]
1037 : TimelineCancelled,
1038 :
1039 : // The tenant is in a state inelegible to run GC
1040 : #[error("not active")]
1041 : NotActive,
1042 :
1043 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1044 : #[error("not active")]
1045 : BadLsn { why: String },
1046 :
1047 : // A remote storage error while scheduling updates after compaction
1048 : #[error(transparent)]
1049 : Remote(anyhow::Error),
1050 :
1051 : // An error reading while calculating GC cutoffs
1052 : #[error(transparent)]
1053 : GcCutoffs(PageReconstructError),
1054 :
1055 : // If GC was invoked for a particular timeline, this error means it didn't exist
1056 : #[error("timeline not found")]
1057 : TimelineNotFound,
1058 : }
1059 :
1060 : impl From<PageReconstructError> for GcError {
1061 0 : fn from(value: PageReconstructError) -> Self {
1062 0 : match value {
1063 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1064 0 : other => Self::GcCutoffs(other),
1065 : }
1066 0 : }
1067 : }
1068 :
1069 : impl From<NotInitialized> for GcError {
1070 0 : fn from(value: NotInitialized) -> Self {
1071 0 : match value {
1072 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1073 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1074 : }
1075 0 : }
1076 : }
1077 :
1078 : impl From<timeline::layer_manager::Shutdown> for GcError {
1079 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1080 0 : GcError::TimelineCancelled
1081 0 : }
1082 : }
1083 :
1084 0 : #[derive(thiserror::Error, Debug)]
1085 : pub(crate) enum LoadConfigError {
1086 : #[error("TOML deserialization error: '{0}'")]
1087 : DeserializeToml(#[from] toml_edit::de::Error),
1088 :
1089 : #[error("Config not found at {0}")]
1090 : NotFound(Utf8PathBuf),
1091 : }
1092 :
1093 : impl Tenant {
1094 : /// Yet another helper for timeline initialization.
1095 : ///
1096 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1097 : /// - Scans the local timeline directory for layer files and builds the layer map
1098 : /// - Downloads remote index file and adds remote files to the layer map
1099 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1100 : ///
1101 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1102 : /// it is marked as Active.
1103 : #[allow(clippy::too_many_arguments)]
1104 6 : async fn timeline_init_and_sync(
1105 6 : self: &Arc<Self>,
1106 6 : timeline_id: TimelineId,
1107 6 : resources: TimelineResources,
1108 6 : mut index_part: IndexPart,
1109 6 : metadata: TimelineMetadata,
1110 6 : ancestor: Option<Arc<Timeline>>,
1111 6 : cause: LoadTimelineCause,
1112 6 : ctx: &RequestContext,
1113 6 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1114 6 : let tenant_id = self.tenant_shard_id;
1115 6 :
1116 6 : let import_pgdata = index_part.import_pgdata.take();
1117 6 : let idempotency = match &import_pgdata {
1118 0 : Some(import_pgdata) => {
1119 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1120 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1121 0 : })
1122 : }
1123 : None => {
1124 6 : if metadata.ancestor_timeline().is_none() {
1125 4 : CreateTimelineIdempotency::Bootstrap {
1126 4 : pg_version: metadata.pg_version(),
1127 4 : }
1128 : } else {
1129 2 : CreateTimelineIdempotency::Branch {
1130 2 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1131 2 : ancestor_start_lsn: metadata.ancestor_lsn(),
1132 2 : }
1133 : }
1134 : }
1135 : };
1136 :
1137 6 : let timeline = self.create_timeline_struct(
1138 6 : timeline_id,
1139 6 : &metadata,
1140 6 : ancestor.clone(),
1141 6 : resources,
1142 6 : CreateTimelineCause::Load,
1143 6 : idempotency.clone(),
1144 6 : )?;
1145 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1146 6 : anyhow::ensure!(
1147 6 : disk_consistent_lsn.is_valid(),
1148 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1149 : );
1150 6 : assert_eq!(
1151 6 : disk_consistent_lsn,
1152 6 : metadata.disk_consistent_lsn(),
1153 0 : "these are used interchangeably"
1154 : );
1155 :
1156 6 : timeline.remote_client.init_upload_queue(&index_part)?;
1157 :
1158 6 : timeline
1159 6 : .load_layer_map(disk_consistent_lsn, index_part)
1160 4 : .await
1161 6 : .with_context(|| {
1162 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1163 6 : })?;
1164 :
1165 0 : match import_pgdata {
1166 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1167 0 : match cause {
1168 0 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1169 : LoadTimelineCause::ImportPgdata { .. } => {
1170 0 : unreachable!("ImportPgdata should not be reloading timeline import is done and persisted as such in s3")
1171 : }
1172 : }
1173 0 : let mut guard = self.timelines_creating.lock().unwrap();
1174 0 : if !guard.insert(timeline_id) {
1175 : // We should never try and load the same timeline twice during startup
1176 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1177 0 : }
1178 0 : let timeline_create_guard = TimelineCreateGuard {
1179 0 : _tenant_gate_guard: self.gate.enter()?,
1180 0 : owning_tenant: self.clone(),
1181 0 : timeline_id,
1182 0 : idempotency,
1183 0 : // The users of this specific return value don't need the timline_path in there.
1184 0 : timeline_path: timeline
1185 0 : .conf
1186 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1187 0 : };
1188 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1189 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1190 0 : timeline,
1191 0 : import_pgdata,
1192 0 : guard: timeline_create_guard,
1193 0 : },
1194 0 : ))
1195 : }
1196 : Some(_) | None => {
1197 : {
1198 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
1199 6 : match timelines_accessor.entry(timeline_id) {
1200 : // We should never try and load the same timeline twice during startup
1201 : Entry::Occupied(_) => {
1202 0 : unreachable!(
1203 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1204 0 : );
1205 : }
1206 6 : Entry::Vacant(v) => {
1207 6 : v.insert(Arc::clone(&timeline));
1208 6 : timeline.maybe_spawn_flush_loop();
1209 6 : }
1210 6 : }
1211 6 : }
1212 6 :
1213 6 : // Sanity check: a timeline should have some content.
1214 6 : anyhow::ensure!(
1215 6 : ancestor.is_some()
1216 4 : || timeline
1217 4 : .layers
1218 4 : .read()
1219 0 : .await
1220 4 : .layer_map()
1221 4 : .expect("currently loading, layer manager cannot be shutdown already")
1222 4 : .iter_historic_layers()
1223 4 : .next()
1224 4 : .is_some(),
1225 0 : "Timeline has no ancestor and no layer files"
1226 : );
1227 :
1228 6 : match cause {
1229 6 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1230 : LoadTimelineCause::ImportPgdata {
1231 0 : create_guard,
1232 0 : activate,
1233 0 : } => {
1234 0 : // TODO: see the comment in the task code above how I'm not so certain
1235 0 : // it is safe to activate here because of concurrent shutdowns.
1236 0 : match activate {
1237 0 : ActivateTimelineArgs::Yes { broker_client } => {
1238 0 : info!("activating timeline after reload from pgdata import task");
1239 0 : timeline.activate(self.clone(), broker_client, None, ctx);
1240 : }
1241 0 : ActivateTimelineArgs::No => (),
1242 : }
1243 0 : drop(create_guard);
1244 : }
1245 : }
1246 :
1247 6 : Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
1248 : }
1249 : }
1250 6 : }
1251 :
1252 : /// Attach a tenant that's available in cloud storage.
1253 : ///
1254 : /// This returns quickly, after just creating the in-memory object
1255 : /// Tenant struct and launching a background task to download
1256 : /// the remote index files. On return, the tenant is most likely still in
1257 : /// Attaching state, and it will become Active once the background task
1258 : /// finishes. You can use wait_until_active() to wait for the task to
1259 : /// complete.
1260 : ///
1261 : #[allow(clippy::too_many_arguments)]
1262 0 : pub(crate) fn spawn(
1263 0 : conf: &'static PageServerConf,
1264 0 : tenant_shard_id: TenantShardId,
1265 0 : resources: TenantSharedResources,
1266 0 : attached_conf: AttachedTenantConf,
1267 0 : shard_identity: ShardIdentity,
1268 0 : init_order: Option<InitializationOrder>,
1269 0 : mode: SpawnMode,
1270 0 : ctx: &RequestContext,
1271 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1272 0 : let wal_redo_manager =
1273 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1274 :
1275 : let TenantSharedResources {
1276 0 : broker_client,
1277 0 : remote_storage,
1278 0 : deletion_queue_client,
1279 0 : l0_flush_global_state,
1280 0 : } = resources;
1281 0 :
1282 0 : let attach_mode = attached_conf.location.attach_mode;
1283 0 : let generation = attached_conf.location.generation;
1284 0 :
1285 0 : let tenant = Arc::new(Tenant::new(
1286 0 : TenantState::Attaching,
1287 0 : conf,
1288 0 : attached_conf,
1289 0 : shard_identity,
1290 0 : Some(wal_redo_manager),
1291 0 : tenant_shard_id,
1292 0 : remote_storage.clone(),
1293 0 : deletion_queue_client,
1294 0 : l0_flush_global_state,
1295 0 : ));
1296 0 :
1297 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1298 0 : // we shut down while attaching.
1299 0 : let attach_gate_guard = tenant
1300 0 : .gate
1301 0 : .enter()
1302 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1303 0 :
1304 0 : // Do all the hard work in the background
1305 0 : let tenant_clone = Arc::clone(&tenant);
1306 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1307 0 : task_mgr::spawn(
1308 0 : &tokio::runtime::Handle::current(),
1309 0 : TaskKind::Attach,
1310 0 : tenant_shard_id,
1311 0 : None,
1312 0 : "attach tenant",
1313 0 : async move {
1314 0 :
1315 0 : info!(
1316 : ?attach_mode,
1317 0 : "Attaching tenant"
1318 : );
1319 :
1320 0 : let _gate_guard = attach_gate_guard;
1321 0 :
1322 0 : // Is this tenant being spawned as part of process startup?
1323 0 : let starting_up = init_order.is_some();
1324 0 : scopeguard::defer! {
1325 0 : if starting_up {
1326 0 : TENANT.startup_complete.inc();
1327 0 : }
1328 0 : }
1329 :
1330 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1331 : enum BrokenVerbosity {
1332 : Error,
1333 : Info
1334 : }
1335 0 : let make_broken =
1336 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1337 0 : match verbosity {
1338 : BrokenVerbosity::Info => {
1339 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1340 : },
1341 : BrokenVerbosity::Error => {
1342 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1343 : }
1344 : }
1345 0 : t.state.send_modify(|state| {
1346 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1347 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1348 0 : assert!(
1349 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1350 0 : "the attach task owns the tenant state until activation is complete"
1351 : );
1352 :
1353 0 : *state = TenantState::broken_from_reason(err.to_string());
1354 0 : });
1355 0 : };
1356 :
1357 : // TODO: should also be rejecting tenant conf changes that violate this check.
1358 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1359 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1360 0 : return Ok(());
1361 0 : }
1362 0 :
1363 0 : let mut init_order = init_order;
1364 0 : // take the completion because initial tenant loading will complete when all of
1365 0 : // these tasks complete.
1366 0 : let _completion = init_order
1367 0 : .as_mut()
1368 0 : .and_then(|x| x.initial_tenant_load.take());
1369 0 : let remote_load_completion = init_order
1370 0 : .as_mut()
1371 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1372 :
1373 : enum AttachType<'a> {
1374 : /// We are attaching this tenant lazily in the background.
1375 : Warmup {
1376 : _permit: tokio::sync::SemaphorePermit<'a>,
1377 : during_startup: bool
1378 : },
1379 : /// We are attaching this tenant as soon as we can, because for example an
1380 : /// endpoint tried to access it.
1381 : OnDemand,
1382 : /// During normal operations after startup, we are attaching a tenant, and
1383 : /// eager attach was requested.
1384 : Normal,
1385 : }
1386 :
1387 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1388 : // Before doing any I/O, wait for at least one of:
1389 : // - A client attempting to access to this tenant (on-demand loading)
1390 : // - A permit becoming available in the warmup semaphore (background warmup)
1391 :
1392 0 : tokio::select!(
1393 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1394 0 : let _ = permit.expect("activate_now_sem is never closed");
1395 0 : tracing::info!("Activating tenant (on-demand)");
1396 0 : AttachType::OnDemand
1397 : },
1398 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1399 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1400 0 : tracing::info!("Activating tenant (warmup)");
1401 0 : AttachType::Warmup {
1402 0 : _permit,
1403 0 : during_startup: init_order.is_some()
1404 0 : }
1405 : }
1406 0 : _ = tenant_clone.cancel.cancelled() => {
1407 : // This is safe, but should be pretty rare: it is interesting if a tenant
1408 : // stayed in Activating for such a long time that shutdown found it in
1409 : // that state.
1410 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1411 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1412 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1413 : // just shutting down), but ensures progress.
1414 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1415 0 : return Ok(());
1416 : },
1417 : )
1418 : } else {
1419 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1420 : // concurrent_tenant_warmup queue
1421 0 : AttachType::Normal
1422 : };
1423 :
1424 0 : let preload = match &mode {
1425 : SpawnMode::Eager | SpawnMode::Lazy => {
1426 0 : let _preload_timer = TENANT.preload.start_timer();
1427 0 : let res = tenant_clone
1428 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1429 0 : .await;
1430 0 : match res {
1431 0 : Ok(p) => Some(p),
1432 0 : Err(e) => {
1433 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1434 0 : return Ok(());
1435 : }
1436 : }
1437 : }
1438 :
1439 : };
1440 :
1441 : // Remote preload is complete.
1442 0 : drop(remote_load_completion);
1443 0 :
1444 0 :
1445 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1446 0 : let attach_start = std::time::Instant::now();
1447 0 : let attached = {
1448 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1449 0 : tenant_clone.attach(preload, &ctx).await
1450 : };
1451 0 : let attach_duration = attach_start.elapsed();
1452 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1453 0 :
1454 0 : match attached {
1455 : Ok(()) => {
1456 0 : info!("attach finished, activating");
1457 0 : tenant_clone.activate(broker_client, None, &ctx);
1458 : }
1459 0 : Err(e) => {
1460 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1461 0 : }
1462 : }
1463 :
1464 : // If we are doing an opportunistic warmup attachment at startup, initialize
1465 : // logical size at the same time. This is better than starting a bunch of idle tenants
1466 : // with cold caches and then coming back later to initialize their logical sizes.
1467 : //
1468 : // It also prevents the warmup proccess competing with the concurrency limit on
1469 : // logical size calculations: if logical size calculation semaphore is saturated,
1470 : // then warmup will wait for that before proceeding to the next tenant.
1471 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1472 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1473 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1474 0 : while futs.next().await.is_some() {}
1475 0 : tracing::info!("Warm-up complete");
1476 0 : }
1477 :
1478 0 : Ok(())
1479 0 : }
1480 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1481 : );
1482 0 : Ok(tenant)
1483 0 : }
1484 :
1485 192 : #[instrument(skip_all)]
1486 : pub(crate) async fn preload(
1487 : self: &Arc<Self>,
1488 : remote_storage: &GenericRemoteStorage,
1489 : cancel: CancellationToken,
1490 : ) -> anyhow::Result<TenantPreload> {
1491 : span::debug_assert_current_span_has_tenant_id();
1492 : // Get list of remote timelines
1493 : // download index files for every tenant timeline
1494 : info!("listing remote timelines");
1495 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1496 : remote_storage,
1497 : self.tenant_shard_id,
1498 : cancel.clone(),
1499 : )
1500 : .await?;
1501 : let (offloaded_add, tenant_manifest) =
1502 : match remote_timeline_client::download_tenant_manifest(
1503 : remote_storage,
1504 : &self.tenant_shard_id,
1505 : self.generation,
1506 : &cancel,
1507 : )
1508 : .await
1509 : {
1510 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1511 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1512 : tenant_manifest,
1513 : ),
1514 : Err(DownloadError::NotFound) => {
1515 : ("no manifest".to_string(), TenantManifest::empty())
1516 : }
1517 : Err(e) => Err(e)?,
1518 : };
1519 :
1520 : info!(
1521 : "found {} timelines, and {offloaded_add}",
1522 : remote_timeline_ids.len()
1523 : );
1524 :
1525 : for k in other_keys {
1526 : warn!("Unexpected non timeline key {k}");
1527 : }
1528 :
1529 : // Avoid downloading IndexPart of offloaded timelines.
1530 : let mut offloaded_with_prefix = HashSet::new();
1531 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1532 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1533 : offloaded_with_prefix.insert(offloaded.timeline_id);
1534 : } else {
1535 : // We'll take care later of timelines in the manifest without a prefix
1536 : }
1537 : }
1538 :
1539 : let timelines = self
1540 : .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
1541 : .await?;
1542 :
1543 : Ok(TenantPreload {
1544 : tenant_manifest,
1545 : timelines: timelines
1546 : .into_iter()
1547 6 : .map(|(id, tl)| (id, Some(tl)))
1548 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1549 : .collect(),
1550 : })
1551 : }
1552 :
1553 : ///
1554 : /// Background task that downloads all data for a tenant and brings it to Active state.
1555 : ///
1556 : /// No background tasks are started as part of this routine.
1557 : ///
1558 192 : async fn attach(
1559 192 : self: &Arc<Tenant>,
1560 192 : preload: Option<TenantPreload>,
1561 192 : ctx: &RequestContext,
1562 192 : ) -> anyhow::Result<()> {
1563 192 : span::debug_assert_current_span_has_tenant_id();
1564 192 :
1565 192 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1566 :
1567 192 : let Some(preload) = preload else {
1568 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
1569 : };
1570 :
1571 192 : let mut offloaded_timeline_ids = HashSet::new();
1572 192 : let mut offloaded_timelines_list = Vec::new();
1573 192 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1574 0 : let timeline_id = timeline_manifest.timeline_id;
1575 0 : let offloaded_timeline =
1576 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1577 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1578 0 : offloaded_timeline_ids.insert(timeline_id);
1579 0 : }
1580 : // Complete deletions for offloaded timeline id's from manifest.
1581 : // The manifest will be uploaded later in this function.
1582 192 : offloaded_timelines_list
1583 192 : .retain(|(offloaded_id, offloaded)| {
1584 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1585 0 : // If there is dangling references in another location, they need to be cleaned up.
1586 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1587 0 : if delete {
1588 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1589 0 : offloaded.defuse_for_tenant_drop();
1590 0 : }
1591 0 : !delete
1592 192 : });
1593 192 :
1594 192 : let mut timelines_to_resume_deletions = vec![];
1595 192 :
1596 192 : let mut remote_index_and_client = HashMap::new();
1597 192 : let mut timeline_ancestors = HashMap::new();
1598 192 : let mut existent_timelines = HashSet::new();
1599 198 : for (timeline_id, preload) in preload.timelines {
1600 6 : let Some(preload) = preload else { continue };
1601 : // This is an invariant of the `preload` function's API
1602 6 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1603 6 : let index_part = match preload.index_part {
1604 6 : Ok(i) => {
1605 6 : debug!("remote index part exists for timeline {timeline_id}");
1606 : // We found index_part on the remote, this is the standard case.
1607 6 : existent_timelines.insert(timeline_id);
1608 6 : i
1609 : }
1610 : Err(DownloadError::NotFound) => {
1611 : // There is no index_part on the remote. We only get here
1612 : // if there is some prefix for the timeline in the remote storage.
1613 : // This can e.g. be the initdb.tar.zst archive, maybe a
1614 : // remnant from a prior incomplete creation or deletion attempt.
1615 : // Delete the local directory as the deciding criterion for a
1616 : // timeline's existence is presence of index_part.
1617 0 : info!(%timeline_id, "index_part not found on remote");
1618 0 : continue;
1619 : }
1620 0 : Err(DownloadError::Fatal(why)) => {
1621 0 : // If, while loading one remote timeline, we saw an indication that our generation
1622 0 : // number is likely invalid, then we should not load the whole tenant.
1623 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1624 0 : anyhow::bail!(why.to_string());
1625 : }
1626 0 : Err(e) => {
1627 0 : // Some (possibly ephemeral) error happened during index_part download.
1628 0 : // Pretend the timeline exists to not delete the timeline directory,
1629 0 : // as it might be a temporary issue and we don't want to re-download
1630 0 : // everything after it resolves.
1631 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1632 :
1633 0 : existent_timelines.insert(timeline_id);
1634 0 : continue;
1635 : }
1636 : };
1637 6 : match index_part {
1638 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1639 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1640 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
1641 6 : }
1642 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1643 0 : info!(
1644 0 : "timeline {} is deleted, picking to resume deletion",
1645 : timeline_id
1646 : );
1647 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1648 : }
1649 : }
1650 : }
1651 :
1652 192 : let mut gc_blocks = HashMap::new();
1653 :
1654 : // For every timeline, download the metadata file, scan the local directory,
1655 : // and build a layer map that contains an entry for each remote and local
1656 : // layer file.
1657 192 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1658 198 : for (timeline_id, remote_metadata) in sorted_timelines {
1659 6 : let (index_part, remote_client) = remote_index_and_client
1660 6 : .remove(&timeline_id)
1661 6 : .expect("just put it in above");
1662 :
1663 6 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1664 : // could just filter these away, but it helps while testing
1665 0 : anyhow::ensure!(
1666 0 : !blocking.reasons.is_empty(),
1667 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1668 : );
1669 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1670 0 : assert!(prev.is_none());
1671 6 : }
1672 :
1673 : // TODO again handle early failure
1674 6 : let effect = self
1675 6 : .load_remote_timeline(
1676 6 : timeline_id,
1677 6 : index_part,
1678 6 : remote_metadata,
1679 6 : TimelineResources {
1680 6 : remote_client,
1681 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1682 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1683 6 : },
1684 6 : LoadTimelineCause::Attach,
1685 6 : ctx,
1686 6 : )
1687 8 : .await
1688 6 : .with_context(|| {
1689 0 : format!(
1690 0 : "failed to load remote timeline {} for tenant {}",
1691 0 : timeline_id, self.tenant_shard_id
1692 0 : )
1693 6 : })?;
1694 :
1695 6 : match effect {
1696 6 : TimelineInitAndSyncResult::ReadyToActivate(_) => {
1697 6 : // activation happens later, on Tenant::activate
1698 6 : }
1699 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1700 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1701 0 : timeline,
1702 0 : import_pgdata,
1703 0 : guard,
1704 0 : },
1705 0 : ) => {
1706 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1707 0 : timeline,
1708 0 : import_pgdata,
1709 0 : ActivateTimelineArgs::No,
1710 0 : guard,
1711 0 : ));
1712 0 : }
1713 : }
1714 : }
1715 :
1716 : // Walk through deleted timelines, resume deletion
1717 192 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1718 0 : remote_timeline_client
1719 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1720 0 : .context("init queue stopped")
1721 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1722 :
1723 0 : DeleteTimelineFlow::resume_deletion(
1724 0 : Arc::clone(self),
1725 0 : timeline_id,
1726 0 : &index_part.metadata,
1727 0 : remote_timeline_client,
1728 0 : )
1729 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1730 0 : .await
1731 0 : .context("resume_deletion")
1732 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1733 : }
1734 192 : let needs_manifest_upload =
1735 192 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1736 192 : {
1737 192 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1738 192 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1739 192 : }
1740 192 : if needs_manifest_upload {
1741 0 : self.store_tenant_manifest().await?;
1742 192 : }
1743 :
1744 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1745 : // IndexPart is the source of truth.
1746 192 : self.clean_up_timelines(&existent_timelines)?;
1747 :
1748 192 : self.gc_block.set_scanned(gc_blocks);
1749 192 :
1750 192 : fail::fail_point!("attach-before-activate", |_| {
1751 0 : anyhow::bail!("attach-before-activate");
1752 192 : });
1753 192 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1754 :
1755 192 : info!("Done");
1756 :
1757 192 : Ok(())
1758 192 : }
1759 :
1760 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1761 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1762 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1763 192 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1764 192 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1765 :
1766 192 : let entries = match timelines_dir.read_dir_utf8() {
1767 192 : Ok(d) => d,
1768 0 : Err(e) => {
1769 0 : if e.kind() == std::io::ErrorKind::NotFound {
1770 0 : return Ok(());
1771 : } else {
1772 0 : return Err(e).context("list timelines directory for tenant");
1773 : }
1774 : }
1775 : };
1776 :
1777 200 : for entry in entries {
1778 8 : let entry = entry.context("read timeline dir entry")?;
1779 8 : let entry_path = entry.path();
1780 :
1781 8 : let purge = if crate::is_temporary(entry_path)
1782 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1783 8 : || is_uninit_mark(entry_path)
1784 8 : || crate::is_delete_mark(entry_path)
1785 : {
1786 0 : true
1787 : } else {
1788 8 : match TimelineId::try_from(entry_path.file_name()) {
1789 8 : Ok(i) => {
1790 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1791 8 : !existent_timelines.contains(&i)
1792 : }
1793 0 : Err(e) => {
1794 0 : tracing::warn!(
1795 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1796 : );
1797 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1798 0 : false
1799 : }
1800 : }
1801 : };
1802 :
1803 8 : if purge {
1804 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1805 2 : if let Err(e) = match entry.file_type() {
1806 2 : Ok(t) => if t.is_dir() {
1807 2 : std::fs::remove_dir_all(entry_path)
1808 : } else {
1809 0 : std::fs::remove_file(entry_path)
1810 : }
1811 2 : .or_else(fs_ext::ignore_not_found),
1812 0 : Err(e) => Err(e),
1813 : } {
1814 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1815 2 : }
1816 6 : }
1817 : }
1818 :
1819 192 : Ok(())
1820 192 : }
1821 :
1822 : /// Get sum of all remote timelines sizes
1823 : ///
1824 : /// This function relies on the index_part instead of listing the remote storage
1825 0 : pub fn remote_size(&self) -> u64 {
1826 0 : let mut size = 0;
1827 :
1828 0 : for timeline in self.list_timelines() {
1829 0 : size += timeline.remote_client.get_remote_physical_size();
1830 0 : }
1831 :
1832 0 : size
1833 0 : }
1834 :
1835 6 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1836 : async fn load_remote_timeline(
1837 : self: &Arc<Self>,
1838 : timeline_id: TimelineId,
1839 : index_part: IndexPart,
1840 : remote_metadata: TimelineMetadata,
1841 : resources: TimelineResources,
1842 : cause: LoadTimelineCause,
1843 : ctx: &RequestContext,
1844 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1845 : span::debug_assert_current_span_has_tenant_id();
1846 :
1847 : info!("downloading index file for timeline {}", timeline_id);
1848 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1849 : .await
1850 : .context("Failed to create new timeline directory")?;
1851 :
1852 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1853 : let timelines = self.timelines.lock().unwrap();
1854 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1855 0 : || {
1856 0 : anyhow::anyhow!(
1857 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1858 0 : )
1859 0 : },
1860 : )?))
1861 : } else {
1862 : None
1863 : };
1864 :
1865 : self.timeline_init_and_sync(
1866 : timeline_id,
1867 : resources,
1868 : index_part,
1869 : remote_metadata,
1870 : ancestor,
1871 : cause,
1872 : ctx,
1873 : )
1874 : .await
1875 : }
1876 :
1877 192 : async fn load_timelines_metadata(
1878 192 : self: &Arc<Tenant>,
1879 192 : timeline_ids: HashSet<TimelineId>,
1880 192 : remote_storage: &GenericRemoteStorage,
1881 192 : cancel: CancellationToken,
1882 192 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1883 192 : let mut part_downloads = JoinSet::new();
1884 198 : for timeline_id in timeline_ids {
1885 6 : let cancel_clone = cancel.clone();
1886 6 : part_downloads.spawn(
1887 6 : self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
1888 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1889 : );
1890 : }
1891 :
1892 192 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1893 :
1894 : loop {
1895 198 : tokio::select!(
1896 198 : next = part_downloads.join_next() => {
1897 198 : match next {
1898 6 : Some(result) => {
1899 6 : let preload = result.context("join preload task")?;
1900 6 : timeline_preloads.insert(preload.timeline_id, preload);
1901 : },
1902 : None => {
1903 192 : break;
1904 : }
1905 : }
1906 : },
1907 198 : _ = cancel.cancelled() => {
1908 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1909 : }
1910 : )
1911 : }
1912 :
1913 192 : Ok(timeline_preloads)
1914 192 : }
1915 :
1916 6 : fn build_timeline_client(
1917 6 : &self,
1918 6 : timeline_id: TimelineId,
1919 6 : remote_storage: GenericRemoteStorage,
1920 6 : ) -> RemoteTimelineClient {
1921 6 : RemoteTimelineClient::new(
1922 6 : remote_storage.clone(),
1923 6 : self.deletion_queue_client.clone(),
1924 6 : self.conf,
1925 6 : self.tenant_shard_id,
1926 6 : timeline_id,
1927 6 : self.generation,
1928 6 : &self.tenant_conf.load().location,
1929 6 : )
1930 6 : }
1931 :
1932 6 : fn load_timeline_metadata(
1933 6 : self: &Arc<Tenant>,
1934 6 : timeline_id: TimelineId,
1935 6 : remote_storage: GenericRemoteStorage,
1936 6 : cancel: CancellationToken,
1937 6 : ) -> impl Future<Output = TimelinePreload> {
1938 6 : let client = self.build_timeline_client(timeline_id, remote_storage);
1939 6 : async move {
1940 6 : debug_assert_current_span_has_tenant_and_timeline_id();
1941 6 : debug!("starting index part download");
1942 :
1943 10 : let index_part = client.download_index_file(&cancel).await;
1944 :
1945 6 : debug!("finished index part download");
1946 :
1947 6 : TimelinePreload {
1948 6 : client,
1949 6 : timeline_id,
1950 6 : index_part,
1951 6 : }
1952 6 : }
1953 6 : }
1954 :
1955 0 : fn check_to_be_archived_has_no_unarchived_children(
1956 0 : timeline_id: TimelineId,
1957 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1958 0 : ) -> Result<(), TimelineArchivalError> {
1959 0 : let children: Vec<TimelineId> = timelines
1960 0 : .iter()
1961 0 : .filter_map(|(id, entry)| {
1962 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
1963 0 : return None;
1964 0 : }
1965 0 : if entry.is_archived() == Some(true) {
1966 0 : return None;
1967 0 : }
1968 0 : Some(*id)
1969 0 : })
1970 0 : .collect();
1971 0 :
1972 0 : if !children.is_empty() {
1973 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
1974 0 : }
1975 0 : Ok(())
1976 0 : }
1977 :
1978 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
1979 0 : ancestor_timeline_id: TimelineId,
1980 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
1981 0 : offloaded_timelines: &std::sync::MutexGuard<
1982 0 : '_,
1983 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
1984 0 : >,
1985 0 : ) -> Result<(), TimelineArchivalError> {
1986 0 : let has_archived_parent =
1987 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
1988 0 : ancestor_timeline.is_archived() == Some(true)
1989 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
1990 0 : true
1991 : } else {
1992 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
1993 0 : if cfg!(debug_assertions) {
1994 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
1995 0 : }
1996 0 : return Err(TimelineArchivalError::NotFound);
1997 : };
1998 0 : if has_archived_parent {
1999 0 : return Err(TimelineArchivalError::HasArchivedParent(
2000 0 : ancestor_timeline_id,
2001 0 : ));
2002 0 : }
2003 0 : Ok(())
2004 0 : }
2005 :
2006 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2007 0 : timeline: &Arc<Timeline>,
2008 0 : ) -> Result<(), TimelineArchivalError> {
2009 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2010 0 : if ancestor_timeline.is_archived() == Some(true) {
2011 0 : return Err(TimelineArchivalError::HasArchivedParent(
2012 0 : ancestor_timeline.timeline_id,
2013 0 : ));
2014 0 : }
2015 0 : }
2016 0 : Ok(())
2017 0 : }
2018 :
2019 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2020 : ///
2021 : /// Counterpart to [`offload_timeline`].
2022 0 : async fn unoffload_timeline(
2023 0 : self: &Arc<Self>,
2024 0 : timeline_id: TimelineId,
2025 0 : broker_client: storage_broker::BrokerClientChannel,
2026 0 : ctx: RequestContext,
2027 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2028 0 : info!("unoffloading timeline");
2029 :
2030 : // We activate the timeline below manually, so this must be called on an active timeline.
2031 : // We expect callers of this function to ensure this.
2032 0 : match self.current_state() {
2033 : TenantState::Activating { .. }
2034 : | TenantState::Attaching
2035 : | TenantState::Broken { .. } => {
2036 0 : panic!("Timeline expected to be active")
2037 : }
2038 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2039 0 : TenantState::Active => {}
2040 0 : }
2041 0 : let cancel = self.cancel.clone();
2042 0 :
2043 0 : // Protect against concurrent attempts to use this TimelineId
2044 0 : // We don't care much about idempotency, as it's ensured a layer above.
2045 0 : let allow_offloaded = true;
2046 0 : let _create_guard = self
2047 0 : .create_timeline_create_guard(
2048 0 : timeline_id,
2049 0 : CreateTimelineIdempotency::FailWithConflict,
2050 0 : allow_offloaded,
2051 0 : )
2052 0 : .map_err(|err| match err {
2053 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2054 : TimelineExclusionError::AlreadyExists { .. } => {
2055 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2056 : }
2057 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2058 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2059 0 : })?;
2060 :
2061 0 : let timeline_preload = self
2062 0 : .load_timeline_metadata(timeline_id, self.remote_storage.clone(), cancel.clone())
2063 0 : .await;
2064 :
2065 0 : let index_part = match timeline_preload.index_part {
2066 0 : Ok(index_part) => {
2067 0 : debug!("remote index part exists for timeline {timeline_id}");
2068 0 : index_part
2069 : }
2070 : Err(DownloadError::NotFound) => {
2071 0 : error!(%timeline_id, "index_part not found on remote");
2072 0 : return Err(TimelineArchivalError::NotFound);
2073 : }
2074 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2075 0 : Err(e) => {
2076 0 : // Some (possibly ephemeral) error happened during index_part download.
2077 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2078 0 : return Err(TimelineArchivalError::Other(
2079 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2080 0 : ));
2081 : }
2082 : };
2083 0 : let index_part = match index_part {
2084 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2085 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2086 0 : info!("timeline is deleted according to index_part.json");
2087 0 : return Err(TimelineArchivalError::NotFound);
2088 : }
2089 : };
2090 0 : let remote_metadata = index_part.metadata.clone();
2091 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2092 0 : self.load_remote_timeline(
2093 0 : timeline_id,
2094 0 : index_part,
2095 0 : remote_metadata,
2096 0 : timeline_resources,
2097 0 : LoadTimelineCause::Unoffload,
2098 0 : &ctx,
2099 0 : )
2100 0 : .await
2101 0 : .with_context(|| {
2102 0 : format!(
2103 0 : "failed to load remote timeline {} for tenant {}",
2104 0 : timeline_id, self.tenant_shard_id
2105 0 : )
2106 0 : })
2107 0 : .map_err(TimelineArchivalError::Other)?;
2108 :
2109 0 : let timeline = {
2110 0 : let timelines = self.timelines.lock().unwrap();
2111 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2112 0 : warn!("timeline not available directly after attach");
2113 : // This is not a panic because no locks are held between `load_remote_timeline`
2114 : // which puts the timeline into timelines, and our look into the timeline map.
2115 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2116 0 : "timeline not available directly after attach"
2117 0 : )));
2118 : };
2119 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2120 0 : match offloaded_timelines.remove(&timeline_id) {
2121 0 : Some(offloaded) => {
2122 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2123 0 : }
2124 0 : None => warn!("timeline already removed from offloaded timelines"),
2125 : }
2126 :
2127 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2128 0 :
2129 0 : Arc::clone(timeline)
2130 0 : };
2131 0 :
2132 0 : // Upload new list of offloaded timelines to S3
2133 0 : self.store_tenant_manifest().await?;
2134 :
2135 : // Activate the timeline (if it makes sense)
2136 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2137 0 : let background_jobs_can_start = None;
2138 0 : timeline.activate(
2139 0 : self.clone(),
2140 0 : broker_client.clone(),
2141 0 : background_jobs_can_start,
2142 0 : &ctx,
2143 0 : );
2144 0 : }
2145 :
2146 0 : info!("timeline unoffloading complete");
2147 0 : Ok(timeline)
2148 0 : }
2149 :
2150 0 : pub(crate) async fn apply_timeline_archival_config(
2151 0 : self: &Arc<Self>,
2152 0 : timeline_id: TimelineId,
2153 0 : new_state: TimelineArchivalState,
2154 0 : broker_client: storage_broker::BrokerClientChannel,
2155 0 : ctx: RequestContext,
2156 0 : ) -> Result<(), TimelineArchivalError> {
2157 0 : info!("setting timeline archival config");
2158 : // First part: figure out what is needed to do, and do validation
2159 0 : let timeline_or_unarchive_offloaded = 'outer: {
2160 0 : let timelines = self.timelines.lock().unwrap();
2161 :
2162 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2163 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2164 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2165 0 : return Err(TimelineArchivalError::NotFound);
2166 : };
2167 0 : if new_state == TimelineArchivalState::Archived {
2168 : // It's offloaded already, so nothing to do
2169 0 : return Ok(());
2170 0 : }
2171 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2172 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2173 0 : ancestor_timeline_id,
2174 0 : &timelines,
2175 0 : &offloaded_timelines,
2176 0 : )?;
2177 0 : }
2178 0 : break 'outer None;
2179 : };
2180 :
2181 : // Do some validation. We release the timelines lock below, so there is potential
2182 : // for race conditions: these checks are more present to prevent misunderstandings of
2183 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2184 0 : match new_state {
2185 : TimelineArchivalState::Unarchived => {
2186 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2187 : }
2188 : TimelineArchivalState::Archived => {
2189 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2190 : }
2191 : }
2192 0 : Some(Arc::clone(timeline))
2193 : };
2194 :
2195 : // Second part: unoffload timeline (if needed)
2196 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2197 0 : timeline
2198 : } else {
2199 : // Turn offloaded timeline into a non-offloaded one
2200 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2201 0 : .await?
2202 : };
2203 :
2204 : // Third part: upload new timeline archival state and block until it is present in S3
2205 0 : let upload_needed = match timeline
2206 0 : .remote_client
2207 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2208 : {
2209 0 : Ok(upload_needed) => upload_needed,
2210 0 : Err(e) => {
2211 0 : if timeline.cancel.is_cancelled() {
2212 0 : return Err(TimelineArchivalError::Cancelled);
2213 : } else {
2214 0 : return Err(TimelineArchivalError::Other(e));
2215 : }
2216 : }
2217 : };
2218 :
2219 0 : if upload_needed {
2220 0 : info!("Uploading new state");
2221 : const MAX_WAIT: Duration = Duration::from_secs(10);
2222 0 : let Ok(v) =
2223 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2224 : else {
2225 0 : tracing::warn!("reached timeout for waiting on upload queue");
2226 0 : return Err(TimelineArchivalError::Timeout);
2227 : };
2228 0 : v.map_err(|e| match e {
2229 0 : WaitCompletionError::NotInitialized(e) => {
2230 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2231 : }
2232 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2233 0 : TimelineArchivalError::Cancelled
2234 : }
2235 0 : })?;
2236 0 : }
2237 0 : Ok(())
2238 0 : }
2239 :
2240 2 : pub fn get_offloaded_timeline(
2241 2 : &self,
2242 2 : timeline_id: TimelineId,
2243 2 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2244 2 : self.timelines_offloaded
2245 2 : .lock()
2246 2 : .unwrap()
2247 2 : .get(&timeline_id)
2248 2 : .map(Arc::clone)
2249 2 : .ok_or(GetTimelineError::NotFound {
2250 2 : tenant_id: self.tenant_shard_id,
2251 2 : timeline_id,
2252 2 : })
2253 2 : }
2254 :
2255 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2256 4 : self.tenant_shard_id
2257 4 : }
2258 :
2259 : /// Get Timeline handle for given Neon timeline ID.
2260 : /// This function is idempotent. It doesn't change internal state in any way.
2261 222 : pub fn get_timeline(
2262 222 : &self,
2263 222 : timeline_id: TimelineId,
2264 222 : active_only: bool,
2265 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2266 222 : let timelines_accessor = self.timelines.lock().unwrap();
2267 222 : let timeline = timelines_accessor
2268 222 : .get(&timeline_id)
2269 222 : .ok_or(GetTimelineError::NotFound {
2270 222 : tenant_id: self.tenant_shard_id,
2271 222 : timeline_id,
2272 222 : })?;
2273 :
2274 220 : if active_only && !timeline.is_active() {
2275 0 : Err(GetTimelineError::NotActive {
2276 0 : tenant_id: self.tenant_shard_id,
2277 0 : timeline_id,
2278 0 : state: timeline.current_state(),
2279 0 : })
2280 : } else {
2281 220 : Ok(Arc::clone(timeline))
2282 : }
2283 222 : }
2284 :
2285 : /// Lists timelines the tenant contains.
2286 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2287 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2288 0 : self.timelines
2289 0 : .lock()
2290 0 : .unwrap()
2291 0 : .values()
2292 0 : .map(Arc::clone)
2293 0 : .collect()
2294 0 : }
2295 :
2296 : /// Lists timelines the tenant manages, including offloaded ones.
2297 : ///
2298 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2299 0 : pub fn list_timelines_and_offloaded(
2300 0 : &self,
2301 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2302 0 : let timelines = self
2303 0 : .timelines
2304 0 : .lock()
2305 0 : .unwrap()
2306 0 : .values()
2307 0 : .map(Arc::clone)
2308 0 : .collect();
2309 0 : let offloaded = self
2310 0 : .timelines_offloaded
2311 0 : .lock()
2312 0 : .unwrap()
2313 0 : .values()
2314 0 : .map(Arc::clone)
2315 0 : .collect();
2316 0 : (timelines, offloaded)
2317 0 : }
2318 :
2319 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2320 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2321 0 : }
2322 :
2323 : /// This is used by tests & import-from-basebackup.
2324 : ///
2325 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2326 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2327 : ///
2328 : /// The caller is responsible for getting the timeline into a state that will be accepted
2329 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2330 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2331 : /// to the [`Tenant::timelines`].
2332 : ///
2333 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2334 184 : pub(crate) async fn create_empty_timeline(
2335 184 : self: &Arc<Self>,
2336 184 : new_timeline_id: TimelineId,
2337 184 : initdb_lsn: Lsn,
2338 184 : pg_version: u32,
2339 184 : _ctx: &RequestContext,
2340 184 : ) -> anyhow::Result<UninitializedTimeline> {
2341 184 : anyhow::ensure!(
2342 184 : self.is_active(),
2343 0 : "Cannot create empty timelines on inactive tenant"
2344 : );
2345 :
2346 : // Protect against concurrent attempts to use this TimelineId
2347 184 : let create_guard = match self
2348 184 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2349 174 : .await?
2350 : {
2351 182 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2352 : StartCreatingTimelineResult::Idempotent(_) => {
2353 0 : unreachable!("FailWithConflict implies we get an error instead")
2354 : }
2355 : };
2356 :
2357 182 : let new_metadata = TimelineMetadata::new(
2358 182 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2359 182 : // make it valid, before calling finish_creation()
2360 182 : Lsn(0),
2361 182 : None,
2362 182 : None,
2363 182 : Lsn(0),
2364 182 : initdb_lsn,
2365 182 : initdb_lsn,
2366 182 : pg_version,
2367 182 : );
2368 182 : self.prepare_new_timeline(
2369 182 : new_timeline_id,
2370 182 : &new_metadata,
2371 182 : create_guard,
2372 182 : initdb_lsn,
2373 182 : None,
2374 182 : )
2375 0 : .await
2376 184 : }
2377 :
2378 : /// Helper for unit tests to create an empty timeline.
2379 : ///
2380 : /// The timeline is has state value `Active` but its background loops are not running.
2381 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2382 : // Our current tests don't need the background loops.
2383 : #[cfg(test)]
2384 174 : pub async fn create_test_timeline(
2385 174 : self: &Arc<Self>,
2386 174 : new_timeline_id: TimelineId,
2387 174 : initdb_lsn: Lsn,
2388 174 : pg_version: u32,
2389 174 : ctx: &RequestContext,
2390 174 : ) -> anyhow::Result<Arc<Timeline>> {
2391 174 : let uninit_tl = self
2392 174 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2393 166 : .await?;
2394 174 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2395 174 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2396 :
2397 : // Setup minimum keys required for the timeline to be usable.
2398 174 : let mut modification = tline.begin_modification(initdb_lsn);
2399 174 : modification
2400 174 : .init_empty_test_timeline()
2401 174 : .context("init_empty_test_timeline")?;
2402 174 : modification
2403 174 : .commit(ctx)
2404 168 : .await
2405 174 : .context("commit init_empty_test_timeline modification")?;
2406 :
2407 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2408 174 : tline.maybe_spawn_flush_loop();
2409 174 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2410 :
2411 : // Make sure the freeze_and_flush reaches remote storage.
2412 174 : tline.remote_client.wait_completion().await.unwrap();
2413 :
2414 174 : let tl = uninit_tl.finish_creation()?;
2415 : // The non-test code would call tl.activate() here.
2416 174 : tl.set_state(TimelineState::Active);
2417 174 : Ok(tl)
2418 174 : }
2419 :
2420 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2421 : #[cfg(test)]
2422 : #[allow(clippy::too_many_arguments)]
2423 32 : pub async fn create_test_timeline_with_layers(
2424 32 : self: &Arc<Self>,
2425 32 : new_timeline_id: TimelineId,
2426 32 : initdb_lsn: Lsn,
2427 32 : pg_version: u32,
2428 32 : ctx: &RequestContext,
2429 32 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2430 32 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2431 32 : end_lsn: Lsn,
2432 32 : ) -> anyhow::Result<Arc<Timeline>> {
2433 : use checks::check_valid_layermap;
2434 : use itertools::Itertools;
2435 :
2436 32 : let tline = self
2437 32 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2438 94 : .await?;
2439 32 : tline.force_advance_lsn(end_lsn);
2440 100 : for deltas in delta_layer_desc {
2441 68 : tline
2442 68 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2443 204 : .await?;
2444 : }
2445 80 : for (lsn, images) in image_layer_desc {
2446 48 : tline
2447 48 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2448 294 : .await?;
2449 : }
2450 32 : let layer_names = tline
2451 32 : .layers
2452 32 : .read()
2453 0 : .await
2454 32 : .layer_map()
2455 32 : .unwrap()
2456 32 : .iter_historic_layers()
2457 148 : .map(|layer| layer.layer_name())
2458 32 : .collect_vec();
2459 32 : if let Some(err) = check_valid_layermap(&layer_names) {
2460 0 : bail!("invalid layermap: {err}");
2461 32 : }
2462 32 : Ok(tline)
2463 32 : }
2464 :
2465 : /// Create a new timeline.
2466 : ///
2467 : /// Returns the new timeline ID and reference to its Timeline object.
2468 : ///
2469 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2470 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2471 : #[allow(clippy::too_many_arguments)]
2472 0 : pub(crate) async fn create_timeline(
2473 0 : self: &Arc<Tenant>,
2474 0 : params: CreateTimelineParams,
2475 0 : broker_client: storage_broker::BrokerClientChannel,
2476 0 : ctx: &RequestContext,
2477 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2478 0 : if !self.is_active() {
2479 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2480 0 : return Err(CreateTimelineError::ShuttingDown);
2481 : } else {
2482 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2483 0 : "Cannot create timelines on inactive tenant"
2484 0 : )));
2485 : }
2486 0 : }
2487 :
2488 0 : let _gate = self
2489 0 : .gate
2490 0 : .enter()
2491 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2492 :
2493 0 : let result: CreateTimelineResult = match params {
2494 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2495 0 : new_timeline_id,
2496 0 : existing_initdb_timeline_id,
2497 0 : pg_version,
2498 0 : }) => {
2499 0 : self.bootstrap_timeline(
2500 0 : new_timeline_id,
2501 0 : pg_version,
2502 0 : existing_initdb_timeline_id,
2503 0 : ctx,
2504 0 : )
2505 0 : .await?
2506 : }
2507 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2508 0 : new_timeline_id,
2509 0 : ancestor_timeline_id,
2510 0 : mut ancestor_start_lsn,
2511 : }) => {
2512 0 : let ancestor_timeline = self
2513 0 : .get_timeline(ancestor_timeline_id, false)
2514 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2515 :
2516 : // instead of waiting around, just deny the request because ancestor is not yet
2517 : // ready for other purposes either.
2518 0 : if !ancestor_timeline.is_active() {
2519 0 : return Err(CreateTimelineError::AncestorNotActive);
2520 0 : }
2521 0 :
2522 0 : if ancestor_timeline.is_archived() == Some(true) {
2523 0 : info!("tried to branch archived timeline");
2524 0 : return Err(CreateTimelineError::AncestorArchived);
2525 0 : }
2526 :
2527 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2528 0 : *lsn = lsn.align();
2529 0 :
2530 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2531 0 : if ancestor_ancestor_lsn > *lsn {
2532 : // can we safely just branch from the ancestor instead?
2533 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2534 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2535 0 : lsn,
2536 0 : ancestor_timeline_id,
2537 0 : ancestor_ancestor_lsn,
2538 0 : )));
2539 0 : }
2540 0 :
2541 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2542 0 : // to the requested branch point. The repository code itself doesn't
2543 0 : // require it, but if we start to receive WAL on the new timeline,
2544 0 : // decoding the new WAL might need to look up previous pages, relation
2545 0 : // sizes etc. and that would get confused if the previous page versions
2546 0 : // are not in the repository yet.
2547 0 : ancestor_timeline
2548 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
2549 0 : .await
2550 0 : .map_err(|e| match e {
2551 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2552 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2553 : }
2554 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2555 0 : })?;
2556 0 : }
2557 :
2558 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2559 0 : .await?
2560 : }
2561 0 : CreateTimelineParams::ImportPgdata(params) => {
2562 0 : self.create_timeline_import_pgdata(
2563 0 : params,
2564 0 : ActivateTimelineArgs::Yes {
2565 0 : broker_client: broker_client.clone(),
2566 0 : },
2567 0 : ctx,
2568 0 : )
2569 0 : .await?
2570 : }
2571 : };
2572 :
2573 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2574 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2575 : // not send a success to the caller until it is. The same applies to idempotent retries.
2576 : //
2577 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2578 : // assume that, because they can see the timeline via API, that the creation is done and
2579 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2580 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2581 : // interacts with UninitializedTimeline and is generally a bit tricky.
2582 : //
2583 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2584 : // creation API until it returns success. Only then is durability guaranteed.
2585 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2586 0 : result
2587 0 : .timeline()
2588 0 : .remote_client
2589 0 : .wait_completion()
2590 0 : .await
2591 0 : .map_err(|e| match e {
2592 : WaitCompletionError::NotInitialized(
2593 0 : e, // If the queue is already stopped, it's a shutdown error.
2594 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2595 0 : e => CreateTimelineError::Other(e.into()),
2596 0 : })
2597 0 : .context("wait for timeline initial uploads to complete")?;
2598 :
2599 : // The creating task is responsible for activating the timeline.
2600 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2601 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2602 0 : let activated_timeline = match result {
2603 0 : CreateTimelineResult::Created(timeline) => {
2604 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2605 0 : timeline
2606 : }
2607 0 : CreateTimelineResult::Idempotent(timeline) => {
2608 0 : info!(
2609 0 : "request was deemed idempotent, activation will be done by the creating task"
2610 : );
2611 0 : timeline
2612 : }
2613 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2614 0 : info!("import task spawned, timeline will become visible and activated once the import is done");
2615 0 : timeline
2616 : }
2617 : };
2618 :
2619 0 : Ok(activated_timeline)
2620 0 : }
2621 :
2622 : /// The returned [`Arc<Timeline>`] is NOT in the [`Tenant::timelines`] map until the import
2623 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2624 : /// [`Tenant::timelines`] map when the import completes.
2625 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2626 : /// for the response.
2627 0 : async fn create_timeline_import_pgdata(
2628 0 : self: &Arc<Tenant>,
2629 0 : params: CreateTimelineParamsImportPgdata,
2630 0 : activate: ActivateTimelineArgs,
2631 0 : ctx: &RequestContext,
2632 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2633 0 : let CreateTimelineParamsImportPgdata {
2634 0 : new_timeline_id,
2635 0 : location,
2636 0 : idempotency_key,
2637 0 : } = params;
2638 0 :
2639 0 : let started_at = chrono::Utc::now().naive_utc();
2640 :
2641 : //
2642 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2643 : // is the canonical way we do it.
2644 : // - create an empty timeline in-memory
2645 : // - use its remote_timeline_client to do the upload
2646 : // - dispose of the uninit timeline
2647 : // - keep the creation guard alive
2648 :
2649 0 : let timeline_create_guard = match self
2650 0 : .start_creating_timeline(
2651 0 : new_timeline_id,
2652 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2653 0 : idempotency_key: idempotency_key.clone(),
2654 0 : }),
2655 0 : )
2656 0 : .await?
2657 : {
2658 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2659 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2660 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
2661 : }
2662 : };
2663 :
2664 0 : let mut uninit_timeline = {
2665 0 : let this = &self;
2666 0 : let initdb_lsn = Lsn(0);
2667 0 : let _ctx = ctx;
2668 0 : async move {
2669 0 : let new_metadata = TimelineMetadata::new(
2670 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2671 0 : // make it valid, before calling finish_creation()
2672 0 : Lsn(0),
2673 0 : None,
2674 0 : None,
2675 0 : Lsn(0),
2676 0 : initdb_lsn,
2677 0 : initdb_lsn,
2678 0 : 15,
2679 0 : );
2680 0 : this.prepare_new_timeline(
2681 0 : new_timeline_id,
2682 0 : &new_metadata,
2683 0 : timeline_create_guard,
2684 0 : initdb_lsn,
2685 0 : None,
2686 0 : )
2687 0 : .await
2688 0 : }
2689 : }
2690 0 : .await?;
2691 :
2692 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2693 0 : idempotency_key,
2694 0 : location,
2695 0 : started_at,
2696 0 : };
2697 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2698 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2699 0 : );
2700 0 : uninit_timeline
2701 0 : .raw_timeline()
2702 0 : .unwrap()
2703 0 : .remote_client
2704 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2705 :
2706 : // wait_completion happens in caller
2707 :
2708 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2709 0 :
2710 0 : tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2711 0 : timeline.clone(),
2712 0 : index_part,
2713 0 : activate,
2714 0 : timeline_create_guard,
2715 0 : ));
2716 0 :
2717 0 : // NB: the timeline doesn't exist in self.timelines at this point
2718 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2719 0 : }
2720 :
2721 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline.timeline_id))]
2722 : async fn create_timeline_import_pgdata_task(
2723 : self: Arc<Tenant>,
2724 : timeline: Arc<Timeline>,
2725 : index_part: import_pgdata::index_part_format::Root,
2726 : activate: ActivateTimelineArgs,
2727 : timeline_create_guard: TimelineCreateGuard,
2728 : ) {
2729 : debug_assert_current_span_has_tenant_and_timeline_id();
2730 : info!("starting");
2731 : scopeguard::defer! {info!("exiting")};
2732 :
2733 : let res = self
2734 : .create_timeline_import_pgdata_task_impl(
2735 : timeline,
2736 : index_part,
2737 : activate,
2738 : timeline_create_guard,
2739 : )
2740 : .await;
2741 : if let Err(err) = &res {
2742 : error!(?err, "task failed");
2743 : // TODO sleep & retry, sensitive to tenant shutdown
2744 : // TODO: allow timeline deletion requests => should cancel the task
2745 : }
2746 : }
2747 :
2748 0 : async fn create_timeline_import_pgdata_task_impl(
2749 0 : self: Arc<Tenant>,
2750 0 : timeline: Arc<Timeline>,
2751 0 : index_part: import_pgdata::index_part_format::Root,
2752 0 : activate: ActivateTimelineArgs,
2753 0 : timeline_create_guard: TimelineCreateGuard,
2754 0 : ) -> Result<(), anyhow::Error> {
2755 0 : let ctx = RequestContext::new(TaskKind::ImportPgdata, DownloadBehavior::Warn);
2756 0 :
2757 0 : info!("importing pgdata");
2758 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2759 0 : .await
2760 0 : .context("import")?;
2761 0 : info!("import done");
2762 :
2763 : //
2764 : // Reload timeline from remote.
2765 : // This proves that the remote state is attachable, and it reuses the code.
2766 : //
2767 : // TODO: think about whether this is safe to do with concurrent Tenant::shutdown.
2768 : // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
2769 : // But our activate() call might launch new background tasks after Tenant::shutdown
2770 : // already went past shutting down the Tenant::timelines, which this timeline here is no part of.
2771 : // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
2772 : // down while bootstrapping/branching + activating), but, the race condition is much more likely
2773 : // to manifest because of the long runtime of this import task.
2774 :
2775 : // in theory this shouldn't even .await anything except for coop yield
2776 0 : info!("shutting down timeline");
2777 0 : timeline.shutdown(ShutdownMode::Hard).await;
2778 0 : info!("timeline shut down, reloading from remote");
2779 : // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
2780 : // let Some(timeline) = Arc::into_inner(timeline) else {
2781 : // anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
2782 : // };
2783 0 : let timeline_id = timeline.timeline_id;
2784 0 :
2785 0 : // load from object storage like Tenant::attach does
2786 0 : let resources = self.build_timeline_resources(timeline_id);
2787 0 : let index_part = resources
2788 0 : .remote_client
2789 0 : .download_index_file(&self.cancel)
2790 0 : .await?;
2791 0 : let index_part = match index_part {
2792 : MaybeDeletedIndexPart::Deleted(_) => {
2793 : // likely concurrent delete call, cplane should prevent this
2794 0 : anyhow::bail!("index part says deleted but we are not done creating yet, this should not happen but")
2795 : }
2796 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2797 0 : };
2798 0 : let metadata = index_part.metadata.clone();
2799 0 : self
2800 0 : .load_remote_timeline(timeline_id, index_part, metadata, resources, LoadTimelineCause::ImportPgdata{
2801 0 : create_guard: timeline_create_guard, activate, }, &ctx)
2802 0 : .await?
2803 0 : .ready_to_activate()
2804 0 : .context("implementation error: reloaded timeline still needs import after import reported success")?;
2805 :
2806 0 : anyhow::Ok(())
2807 0 : }
2808 :
2809 0 : pub(crate) async fn delete_timeline(
2810 0 : self: Arc<Self>,
2811 0 : timeline_id: TimelineId,
2812 0 : ) -> Result<(), DeleteTimelineError> {
2813 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2814 :
2815 0 : Ok(())
2816 0 : }
2817 :
2818 : /// perform one garbage collection iteration, removing old data files from disk.
2819 : /// this function is periodically called by gc task.
2820 : /// also it can be explicitly requested through page server api 'do_gc' command.
2821 : ///
2822 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2823 : ///
2824 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2825 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2826 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2827 : /// `pitr` specifies the same as a time difference from the current time. The effective
2828 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2829 : /// requires more history to be retained.
2830 : //
2831 754 : pub(crate) async fn gc_iteration(
2832 754 : &self,
2833 754 : target_timeline_id: Option<TimelineId>,
2834 754 : horizon: u64,
2835 754 : pitr: Duration,
2836 754 : cancel: &CancellationToken,
2837 754 : ctx: &RequestContext,
2838 754 : ) -> Result<GcResult, GcError> {
2839 754 : // Don't start doing work during shutdown
2840 754 : if let TenantState::Stopping { .. } = self.current_state() {
2841 0 : return Ok(GcResult::default());
2842 754 : }
2843 754 :
2844 754 : // there is a global allowed_error for this
2845 754 : if !self.is_active() {
2846 0 : return Err(GcError::NotActive);
2847 754 : }
2848 754 :
2849 754 : {
2850 754 : let conf = self.tenant_conf.load();
2851 754 :
2852 754 : // If we may not delete layers, then simply skip GC. Even though a tenant
2853 754 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2854 754 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2855 754 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2856 754 : if !conf.location.may_delete_layers_hint() {
2857 0 : info!("Skipping GC in location state {:?}", conf.location);
2858 0 : return Ok(GcResult::default());
2859 754 : }
2860 754 :
2861 754 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2862 750 : info!("Skipping GC because lsn lease deadline is not reached");
2863 750 : return Ok(GcResult::default());
2864 4 : }
2865 : }
2866 :
2867 4 : let _guard = match self.gc_block.start().await {
2868 4 : Ok(guard) => guard,
2869 0 : Err(reasons) => {
2870 0 : info!("Skipping GC: {reasons}");
2871 0 : return Ok(GcResult::default());
2872 : }
2873 : };
2874 :
2875 4 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2876 3 : .await
2877 754 : }
2878 :
2879 : /// Perform one compaction iteration.
2880 : /// This function is periodically called by compactor task.
2881 : /// Also it can be explicitly requested per timeline through page server
2882 : /// api's 'compact' command.
2883 : ///
2884 : /// Returns whether we have pending compaction task.
2885 0 : async fn compaction_iteration(
2886 0 : self: &Arc<Self>,
2887 0 : cancel: &CancellationToken,
2888 0 : ctx: &RequestContext,
2889 0 : ) -> Result<bool, timeline::CompactionError> {
2890 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
2891 0 : if !self.is_active() {
2892 0 : return Ok(false);
2893 0 : }
2894 0 :
2895 0 : {
2896 0 : let conf = self.tenant_conf.load();
2897 0 :
2898 0 : // Note that compaction usually requires deletions, but we don't respect
2899 0 : // may_delete_layers_hint here: that is because tenants in AttachedMulti
2900 0 : // should proceed with compaction even if they can't do deletion, to avoid
2901 0 : // accumulating dangerously deep stacks of L0 layers. Deletions will be
2902 0 : // enqueued inside RemoteTimelineClient, and executed layer if/when we transition
2903 0 : // to AttachedSingle state.
2904 0 : if !conf.location.may_upload_layers_hint() {
2905 0 : info!("Skipping compaction in location state {:?}", conf.location);
2906 0 : return Ok(false);
2907 0 : }
2908 0 : }
2909 0 :
2910 0 : // Scan through the hashmap and collect a list of all the timelines,
2911 0 : // while holding the lock. Then drop the lock and actually perform the
2912 0 : // compactions. We don't want to block everything else while the
2913 0 : // compaction runs.
2914 0 : let timelines_to_compact_or_offload;
2915 0 : {
2916 0 : let timelines = self.timelines.lock().unwrap();
2917 0 : timelines_to_compact_or_offload = timelines
2918 0 : .iter()
2919 0 : .filter_map(|(timeline_id, timeline)| {
2920 0 : let (is_active, (can_offload, _)) =
2921 0 : (timeline.is_active(), timeline.can_offload());
2922 0 : let has_no_unoffloaded_children = {
2923 0 : !timelines
2924 0 : .iter()
2925 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
2926 : };
2927 0 : let config_allows_offload = self.conf.timeline_offloading
2928 0 : || self
2929 0 : .tenant_conf
2930 0 : .load()
2931 0 : .tenant_conf
2932 0 : .timeline_offloading
2933 0 : .unwrap_or_default();
2934 0 : let can_offload =
2935 0 : can_offload && has_no_unoffloaded_children && config_allows_offload;
2936 0 : if (is_active, can_offload) == (false, false) {
2937 0 : None
2938 : } else {
2939 0 : Some((*timeline_id, timeline.clone(), (is_active, can_offload)))
2940 : }
2941 0 : })
2942 0 : .collect::<Vec<_>>();
2943 0 : drop(timelines);
2944 0 : }
2945 0 :
2946 0 : // Before doing any I/O work, check our circuit breaker
2947 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2948 0 : info!("Skipping compaction due to previous failures");
2949 0 : return Ok(false);
2950 0 : }
2951 0 :
2952 0 : let mut has_pending_task = false;
2953 :
2954 0 : for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
2955 : {
2956 0 : let pending_task_left = if *can_compact {
2957 : Some(
2958 0 : timeline
2959 0 : .compact(cancel, EnumSet::empty(), ctx)
2960 0 : .instrument(info_span!("compact_timeline", %timeline_id))
2961 0 : .await
2962 0 : .inspect_err(|e| match e {
2963 0 : timeline::CompactionError::ShuttingDown => (),
2964 0 : timeline::CompactionError::Offload(_) => {
2965 0 : // Failures to offload timelines do not trip the circuit breaker, because
2966 0 : // they do not do lots of writes the way compaction itself does: it is cheap
2967 0 : // to retry, and it would be bad to stop all compaction because of an issue with offloading.
2968 0 : }
2969 0 : timeline::CompactionError::Other(e) => {
2970 0 : self.compaction_circuit_breaker
2971 0 : .lock()
2972 0 : .unwrap()
2973 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, e);
2974 0 : }
2975 0 : })?,
2976 : )
2977 : } else {
2978 0 : None
2979 : };
2980 0 : has_pending_task |= pending_task_left.unwrap_or(false);
2981 0 : if pending_task_left == Some(false) && *can_offload {
2982 0 : offload_timeline(self, timeline)
2983 0 : .instrument(info_span!("offload_timeline", %timeline_id))
2984 0 : .await?;
2985 0 : }
2986 : }
2987 :
2988 0 : self.compaction_circuit_breaker
2989 0 : .lock()
2990 0 : .unwrap()
2991 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
2992 0 :
2993 0 : Ok(has_pending_task)
2994 0 : }
2995 :
2996 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
2997 : // this happens during ingest: this background housekeeping is for freezing layers
2998 : // that are open but haven't been written to for some time.
2999 0 : async fn ingest_housekeeping(&self) {
3000 0 : // Scan through the hashmap and collect a list of all the timelines,
3001 0 : // while holding the lock. Then drop the lock and actually perform the
3002 0 : // compactions. We don't want to block everything else while the
3003 0 : // compaction runs.
3004 0 : let timelines = {
3005 0 : self.timelines
3006 0 : .lock()
3007 0 : .unwrap()
3008 0 : .values()
3009 0 : .filter_map(|timeline| {
3010 0 : if timeline.is_active() {
3011 0 : Some(timeline.clone())
3012 : } else {
3013 0 : None
3014 : }
3015 0 : })
3016 0 : .collect::<Vec<_>>()
3017 : };
3018 :
3019 0 : for timeline in &timelines {
3020 0 : timeline.maybe_freeze_ephemeral_layer().await;
3021 : }
3022 0 : }
3023 :
3024 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3025 0 : let timelines = self.timelines.lock().unwrap();
3026 0 : !timelines
3027 0 : .iter()
3028 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3029 0 : }
3030 :
3031 1702 : pub fn current_state(&self) -> TenantState {
3032 1702 : self.state.borrow().clone()
3033 1702 : }
3034 :
3035 942 : pub fn is_active(&self) -> bool {
3036 942 : self.current_state() == TenantState::Active
3037 942 : }
3038 :
3039 0 : pub fn generation(&self) -> Generation {
3040 0 : self.generation
3041 0 : }
3042 :
3043 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3044 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3045 0 : }
3046 :
3047 : /// Changes tenant status to active, unless shutdown was already requested.
3048 : ///
3049 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3050 : /// to delay background jobs. Background jobs can be started right away when None is given.
3051 0 : fn activate(
3052 0 : self: &Arc<Self>,
3053 0 : broker_client: BrokerClientChannel,
3054 0 : background_jobs_can_start: Option<&completion::Barrier>,
3055 0 : ctx: &RequestContext,
3056 0 : ) {
3057 0 : span::debug_assert_current_span_has_tenant_id();
3058 0 :
3059 0 : let mut activating = false;
3060 0 : self.state.send_modify(|current_state| {
3061 : use pageserver_api::models::ActivatingFrom;
3062 0 : match &*current_state {
3063 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3064 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3065 : }
3066 0 : TenantState::Attaching => {
3067 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3068 0 : }
3069 0 : }
3070 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3071 0 : activating = true;
3072 0 : // Continue outside the closure. We need to grab timelines.lock()
3073 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3074 0 : });
3075 0 :
3076 0 : if activating {
3077 0 : let timelines_accessor = self.timelines.lock().unwrap();
3078 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3079 0 : let timelines_to_activate = timelines_accessor
3080 0 : .values()
3081 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3082 0 :
3083 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3084 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3085 0 :
3086 0 : // Spawn gc and compaction loops. The loops will shut themselves
3087 0 : // down when they notice that the tenant is inactive.
3088 0 : tasks::start_background_loops(self, background_jobs_can_start);
3089 0 :
3090 0 : let mut activated_timelines = 0;
3091 :
3092 0 : for timeline in timelines_to_activate {
3093 0 : timeline.activate(
3094 0 : self.clone(),
3095 0 : broker_client.clone(),
3096 0 : background_jobs_can_start,
3097 0 : ctx,
3098 0 : );
3099 0 : activated_timelines += 1;
3100 0 : }
3101 :
3102 0 : self.state.send_modify(move |current_state| {
3103 0 : assert!(
3104 0 : matches!(current_state, TenantState::Activating(_)),
3105 0 : "set_stopping and set_broken wait for us to leave Activating state",
3106 : );
3107 0 : *current_state = TenantState::Active;
3108 0 :
3109 0 : let elapsed = self.constructed_at.elapsed();
3110 0 : let total_timelines = timelines_accessor.len();
3111 0 :
3112 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3113 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3114 0 : info!(
3115 0 : since_creation_millis = elapsed.as_millis(),
3116 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3117 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3118 0 : activated_timelines,
3119 0 : total_timelines,
3120 0 : post_state = <&'static str>::from(&*current_state),
3121 0 : "activation attempt finished"
3122 : );
3123 :
3124 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3125 0 : });
3126 0 : }
3127 0 : }
3128 :
3129 : /// Shutdown the tenant and join all of the spawned tasks.
3130 : ///
3131 : /// The method caters for all use-cases:
3132 : /// - pageserver shutdown (freeze_and_flush == true)
3133 : /// - detach + ignore (freeze_and_flush == false)
3134 : ///
3135 : /// This will attempt to shutdown even if tenant is broken.
3136 : ///
3137 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3138 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3139 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3140 : /// the ongoing shutdown.
3141 6 : async fn shutdown(
3142 6 : &self,
3143 6 : shutdown_progress: completion::Barrier,
3144 6 : shutdown_mode: timeline::ShutdownMode,
3145 6 : ) -> Result<(), completion::Barrier> {
3146 6 : span::debug_assert_current_span_has_tenant_id();
3147 :
3148 : // Set tenant (and its timlines) to Stoppping state.
3149 : //
3150 : // Since we can only transition into Stopping state after activation is complete,
3151 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3152 : //
3153 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3154 : // 1. Lock out any new requests to the tenants.
3155 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3156 : // 3. Signal cancellation for other tenant background loops.
3157 : // 4. ???
3158 : //
3159 : // The waiting for the cancellation is not done uniformly.
3160 : // We certainly wait for WAL receivers to shut down.
3161 : // That is necessary so that no new data comes in before the freeze_and_flush.
3162 : // But the tenant background loops are joined-on in our caller.
3163 : // It's mesed up.
3164 : // we just ignore the failure to stop
3165 :
3166 : // If we're still attaching, fire the cancellation token early to drop out: this
3167 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3168 : // is very slow.
3169 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3170 0 : self.cancel.cancel();
3171 0 :
3172 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3173 0 : // are children of ours, so their flush loops will have shut down already
3174 0 : timeline::ShutdownMode::Hard
3175 : } else {
3176 6 : shutdown_mode
3177 : };
3178 :
3179 6 : match self.set_stopping(shutdown_progress, false, false).await {
3180 6 : Ok(()) => {}
3181 0 : Err(SetStoppingError::Broken) => {
3182 0 : // assume that this is acceptable
3183 0 : }
3184 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3185 0 : // give caller the option to wait for this this shutdown
3186 0 : info!("Tenant::shutdown: AlreadyStopping");
3187 0 : return Err(other);
3188 : }
3189 : };
3190 :
3191 6 : let mut js = tokio::task::JoinSet::new();
3192 6 : {
3193 6 : let timelines = self.timelines.lock().unwrap();
3194 6 : timelines.values().for_each(|timeline| {
3195 6 : let timeline = Arc::clone(timeline);
3196 6 : let timeline_id = timeline.timeline_id;
3197 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3198 12 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3199 6 : });
3200 6 : }
3201 6 : {
3202 6 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3203 6 : timelines_offloaded.values().for_each(|timeline| {
3204 0 : timeline.defuse_for_tenant_drop();
3205 6 : });
3206 6 : }
3207 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3208 6 : tracing::info!("Waiting for timelines...");
3209 12 : while let Some(res) = js.join_next().await {
3210 0 : match res {
3211 6 : Ok(()) => {}
3212 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3213 0 : Err(je) if je.is_panic() => { /* logged already */ }
3214 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3215 : }
3216 : }
3217 :
3218 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3219 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3220 6 : tracing::debug!("Cancelling CancellationToken");
3221 6 : self.cancel.cancel();
3222 6 :
3223 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3224 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3225 6 : //
3226 6 : // this will additionally shutdown and await all timeline tasks.
3227 6 : tracing::debug!("Waiting for tasks...");
3228 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3229 :
3230 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3231 6 : walredo_mgr.shutdown().await;
3232 0 : }
3233 :
3234 : // Wait for any in-flight operations to complete
3235 6 : self.gate.close().await;
3236 :
3237 6 : remove_tenant_metrics(&self.tenant_shard_id);
3238 6 :
3239 6 : Ok(())
3240 6 : }
3241 :
3242 : /// Change tenant status to Stopping, to mark that it is being shut down.
3243 : ///
3244 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3245 : ///
3246 : /// This function is not cancel-safe!
3247 : ///
3248 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
3249 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
3250 6 : async fn set_stopping(
3251 6 : &self,
3252 6 : progress: completion::Barrier,
3253 6 : _allow_transition_from_loading: bool,
3254 6 : allow_transition_from_attaching: bool,
3255 6 : ) -> Result<(), SetStoppingError> {
3256 6 : let mut rx = self.state.subscribe();
3257 6 :
3258 6 : // cannot stop before we're done activating, so wait out until we're done activating
3259 6 : rx.wait_for(|state| match state {
3260 0 : TenantState::Attaching if allow_transition_from_attaching => true,
3261 : TenantState::Activating(_) | TenantState::Attaching => {
3262 0 : info!(
3263 0 : "waiting for {} to turn Active|Broken|Stopping",
3264 0 : <&'static str>::from(state)
3265 : );
3266 0 : false
3267 : }
3268 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3269 6 : })
3270 0 : .await
3271 6 : .expect("cannot drop self.state while on a &self method");
3272 6 :
3273 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3274 6 : let mut err = None;
3275 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3276 : TenantState::Activating(_) => {
3277 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
3278 : }
3279 : TenantState::Attaching => {
3280 0 : if !allow_transition_from_attaching {
3281 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
3282 0 : };
3283 0 : *current_state = TenantState::Stopping { progress };
3284 0 : true
3285 : }
3286 : TenantState::Active => {
3287 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3288 : // are created after the transition to Stopping. That's harmless, as the Timelines
3289 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3290 6 : *current_state = TenantState::Stopping { progress };
3291 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
3292 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3293 6 : true
3294 : }
3295 0 : TenantState::Broken { reason, .. } => {
3296 0 : info!(
3297 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3298 : );
3299 0 : err = Some(SetStoppingError::Broken);
3300 0 : false
3301 : }
3302 0 : TenantState::Stopping { progress } => {
3303 0 : info!("Tenant is already in Stopping state");
3304 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3305 0 : false
3306 : }
3307 6 : });
3308 6 : match (stopping, err) {
3309 6 : (true, None) => {} // continue
3310 0 : (false, Some(err)) => return Err(err),
3311 0 : (true, Some(_)) => unreachable!(
3312 0 : "send_if_modified closure must error out if not transitioning to Stopping"
3313 0 : ),
3314 0 : (false, None) => unreachable!(
3315 0 : "send_if_modified closure must return true if transitioning to Stopping"
3316 0 : ),
3317 : }
3318 :
3319 6 : let timelines_accessor = self.timelines.lock().unwrap();
3320 6 : let not_broken_timelines = timelines_accessor
3321 6 : .values()
3322 6 : .filter(|timeline| !timeline.is_broken());
3323 12 : for timeline in not_broken_timelines {
3324 6 : timeline.set_state(TimelineState::Stopping);
3325 6 : }
3326 6 : Ok(())
3327 6 : }
3328 :
3329 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3330 : /// `remove_tenant_from_memory`
3331 : ///
3332 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3333 : ///
3334 : /// In tests, we also use this to set tenants to Broken state on purpose.
3335 0 : pub(crate) async fn set_broken(&self, reason: String) {
3336 0 : let mut rx = self.state.subscribe();
3337 0 :
3338 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3339 0 : // So, wait until it's done.
3340 0 : rx.wait_for(|state| match state {
3341 : TenantState::Activating(_) | TenantState::Attaching => {
3342 0 : info!(
3343 0 : "waiting for {} to turn Active|Broken|Stopping",
3344 0 : <&'static str>::from(state)
3345 : );
3346 0 : false
3347 : }
3348 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3349 0 : })
3350 0 : .await
3351 0 : .expect("cannot drop self.state while on a &self method");
3352 0 :
3353 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3354 0 : self.set_broken_no_wait(reason)
3355 0 : }
3356 :
3357 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3358 0 : let reason = reason.to_string();
3359 0 : self.state.send_modify(|current_state| {
3360 0 : match *current_state {
3361 : TenantState::Activating(_) | TenantState::Attaching => {
3362 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3363 : }
3364 : TenantState::Active => {
3365 0 : if cfg!(feature = "testing") {
3366 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3367 0 : *current_state = TenantState::broken_from_reason(reason);
3368 : } else {
3369 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3370 : }
3371 : }
3372 : TenantState::Broken { .. } => {
3373 0 : warn!("Tenant is already in Broken state");
3374 : }
3375 : // This is the only "expected" path, any other path is a bug.
3376 : TenantState::Stopping { .. } => {
3377 0 : warn!(
3378 0 : "Marking Stopping tenant as Broken state, reason: {}",
3379 : reason
3380 : );
3381 0 : *current_state = TenantState::broken_from_reason(reason);
3382 : }
3383 : }
3384 0 : });
3385 0 : }
3386 :
3387 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3388 0 : self.state.subscribe()
3389 0 : }
3390 :
3391 : /// The activate_now semaphore is initialized with zero units. As soon as
3392 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3393 0 : pub(crate) fn activate_now(&self) {
3394 0 : self.activate_now_sem.add_permits(1);
3395 0 : }
3396 :
3397 0 : pub(crate) async fn wait_to_become_active(
3398 0 : &self,
3399 0 : timeout: Duration,
3400 0 : ) -> Result<(), GetActiveTenantError> {
3401 0 : let mut receiver = self.state.subscribe();
3402 : loop {
3403 0 : let current_state = receiver.borrow_and_update().clone();
3404 0 : match current_state {
3405 : TenantState::Attaching | TenantState::Activating(_) => {
3406 : // in these states, there's a chance that we can reach ::Active
3407 0 : self.activate_now();
3408 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3409 0 : Ok(r) => {
3410 0 : r.map_err(
3411 0 : |_e: tokio::sync::watch::error::RecvError|
3412 : // Tenant existed but was dropped: report it as non-existent
3413 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
3414 0 : )?
3415 : }
3416 : Err(TimeoutCancellableError::Cancelled) => {
3417 0 : return Err(GetActiveTenantError::Cancelled);
3418 : }
3419 : Err(TimeoutCancellableError::Timeout) => {
3420 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3421 0 : latest_state: Some(self.current_state()),
3422 0 : wait_time: timeout,
3423 0 : });
3424 : }
3425 : }
3426 : }
3427 : TenantState::Active { .. } => {
3428 0 : return Ok(());
3429 : }
3430 0 : TenantState::Broken { reason, .. } => {
3431 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3432 0 : // it's logically a 500 to external API users (broken is always a bug).
3433 0 : return Err(GetActiveTenantError::Broken(reason));
3434 : }
3435 : TenantState::Stopping { .. } => {
3436 : // There's no chance the tenant can transition back into ::Active
3437 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3438 : }
3439 : }
3440 : }
3441 0 : }
3442 :
3443 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3444 0 : self.tenant_conf.load().location.attach_mode
3445 0 : }
3446 :
3447 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3448 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3449 : /// rare external API calls, like a reconciliation at startup.
3450 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3451 0 : let conf = self.tenant_conf.load();
3452 :
3453 0 : let location_config_mode = match conf.location.attach_mode {
3454 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3455 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3456 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3457 : };
3458 :
3459 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3460 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3461 0 :
3462 0 : models::LocationConfig {
3463 0 : mode: location_config_mode,
3464 0 : generation: self.generation.into(),
3465 0 : secondary_conf: None,
3466 0 : shard_number: self.shard_identity.number.0,
3467 0 : shard_count: self.shard_identity.count.literal(),
3468 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3469 0 : tenant_conf: tenant_config,
3470 0 : }
3471 0 : }
3472 :
3473 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3474 0 : &self.tenant_shard_id
3475 0 : }
3476 :
3477 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3478 0 : self.shard_identity.stripe_size
3479 0 : }
3480 :
3481 0 : pub(crate) fn get_generation(&self) -> Generation {
3482 0 : self.generation
3483 0 : }
3484 :
3485 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3486 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3487 : /// resetting this tenant to a valid state if we fail.
3488 0 : pub(crate) async fn split_prepare(
3489 0 : &self,
3490 0 : child_shards: &Vec<TenantShardId>,
3491 0 : ) -> anyhow::Result<()> {
3492 0 : let (timelines, offloaded) = {
3493 0 : let timelines = self.timelines.lock().unwrap();
3494 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3495 0 : (timelines.clone(), offloaded.clone())
3496 0 : };
3497 0 : let timelines_iter = timelines
3498 0 : .values()
3499 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3500 0 : .chain(
3501 0 : offloaded
3502 0 : .values()
3503 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3504 0 : );
3505 0 : for timeline in timelines_iter {
3506 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3507 : // to ensure that they do not start a split if currently in the process of doing these.
3508 :
3509 0 : let timeline_id = timeline.timeline_id();
3510 :
3511 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3512 : // Upload an index from the parent: this is partly to provide freshness for the
3513 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3514 : // always be a parent shard index in the same generation as we wrote the child shard index.
3515 0 : tracing::info!(%timeline_id, "Uploading index");
3516 0 : timeline
3517 0 : .remote_client
3518 0 : .schedule_index_upload_for_file_changes()?;
3519 0 : timeline.remote_client.wait_completion().await?;
3520 0 : }
3521 :
3522 0 : let remote_client = match timeline {
3523 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3524 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3525 0 : let remote_client = self
3526 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3527 0 : Arc::new(remote_client)
3528 : }
3529 : };
3530 :
3531 : // Shut down the timeline's remote client: this means that the indices we write
3532 : // for child shards will not be invalidated by the parent shard deleting layers.
3533 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3534 0 : remote_client.shutdown().await;
3535 :
3536 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3537 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3538 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3539 : // we use here really is the remotely persistent one).
3540 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3541 0 : let result = remote_client
3542 0 : .download_index_file(&self.cancel)
3543 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))
3544 0 : .await?;
3545 0 : let index_part = match result {
3546 : MaybeDeletedIndexPart::Deleted(_) => {
3547 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3548 : }
3549 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3550 : };
3551 :
3552 0 : for child_shard in child_shards {
3553 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3554 0 : upload_index_part(
3555 0 : &self.remote_storage,
3556 0 : child_shard,
3557 0 : &timeline_id,
3558 0 : self.generation,
3559 0 : &index_part,
3560 0 : &self.cancel,
3561 0 : )
3562 0 : .await?;
3563 : }
3564 : }
3565 :
3566 0 : let tenant_manifest = self.build_tenant_manifest();
3567 0 : for child_shard in child_shards {
3568 0 : tracing::info!(
3569 0 : "Uploading tenant manifest for child {}",
3570 0 : child_shard.to_index()
3571 : );
3572 0 : upload_tenant_manifest(
3573 0 : &self.remote_storage,
3574 0 : child_shard,
3575 0 : self.generation,
3576 0 : &tenant_manifest,
3577 0 : &self.cancel,
3578 0 : )
3579 0 : .await?;
3580 : }
3581 :
3582 0 : Ok(())
3583 0 : }
3584 :
3585 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3586 0 : let mut result = TopTenantShardItem {
3587 0 : id: self.tenant_shard_id,
3588 0 : resident_size: 0,
3589 0 : physical_size: 0,
3590 0 : max_logical_size: 0,
3591 0 : };
3592 :
3593 0 : for timeline in self.timelines.lock().unwrap().values() {
3594 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3595 0 :
3596 0 : result.physical_size += timeline
3597 0 : .remote_client
3598 0 : .metrics
3599 0 : .remote_physical_size_gauge
3600 0 : .get();
3601 0 : result.max_logical_size = std::cmp::max(
3602 0 : result.max_logical_size,
3603 0 : timeline.metrics.current_logical_size_gauge.get(),
3604 0 : );
3605 0 : }
3606 :
3607 0 : result
3608 0 : }
3609 : }
3610 :
3611 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3612 : /// perform a topological sort, so that the parent of each timeline comes
3613 : /// before the children.
3614 : /// E extracts the ancestor from T
3615 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3616 192 : fn tree_sort_timelines<T, E>(
3617 192 : timelines: HashMap<TimelineId, T>,
3618 192 : extractor: E,
3619 192 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3620 192 : where
3621 192 : E: Fn(&T) -> Option<TimelineId>,
3622 192 : {
3623 192 : let mut result = Vec::with_capacity(timelines.len());
3624 192 :
3625 192 : let mut now = Vec::with_capacity(timelines.len());
3626 192 : // (ancestor, children)
3627 192 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3628 192 : HashMap::with_capacity(timelines.len());
3629 :
3630 198 : for (timeline_id, value) in timelines {
3631 6 : if let Some(ancestor_id) = extractor(&value) {
3632 2 : let children = later.entry(ancestor_id).or_default();
3633 2 : children.push((timeline_id, value));
3634 4 : } else {
3635 4 : now.push((timeline_id, value));
3636 4 : }
3637 : }
3638 :
3639 198 : while let Some((timeline_id, metadata)) = now.pop() {
3640 6 : result.push((timeline_id, metadata));
3641 : // All children of this can be loaded now
3642 6 : if let Some(mut children) = later.remove(&timeline_id) {
3643 2 : now.append(&mut children);
3644 4 : }
3645 : }
3646 :
3647 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3648 192 : if !later.is_empty() {
3649 0 : for (missing_id, orphan_ids) in later {
3650 0 : for (orphan_id, _) in orphan_ids {
3651 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
3652 : }
3653 : }
3654 0 : bail!("could not load tenant because some timelines are missing ancestors");
3655 192 : }
3656 192 :
3657 192 : Ok(result)
3658 192 : }
3659 :
3660 : enum ActivateTimelineArgs {
3661 : Yes {
3662 : broker_client: storage_broker::BrokerClientChannel,
3663 : },
3664 : No,
3665 : }
3666 :
3667 : impl Tenant {
3668 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3669 0 : self.tenant_conf.load().tenant_conf.clone()
3670 0 : }
3671 :
3672 0 : pub fn effective_config(&self) -> TenantConf {
3673 0 : self.tenant_specific_overrides()
3674 0 : .merge(self.conf.default_tenant_conf.clone())
3675 0 : }
3676 :
3677 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3678 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3679 0 : tenant_conf
3680 0 : .checkpoint_distance
3681 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3682 0 : }
3683 :
3684 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3685 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3686 0 : tenant_conf
3687 0 : .checkpoint_timeout
3688 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3689 0 : }
3690 :
3691 0 : pub fn get_compaction_target_size(&self) -> u64 {
3692 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3693 0 : tenant_conf
3694 0 : .compaction_target_size
3695 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3696 0 : }
3697 :
3698 0 : pub fn get_compaction_period(&self) -> Duration {
3699 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3700 0 : tenant_conf
3701 0 : .compaction_period
3702 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3703 0 : }
3704 :
3705 0 : pub fn get_compaction_threshold(&self) -> usize {
3706 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3707 0 : tenant_conf
3708 0 : .compaction_threshold
3709 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3710 0 : }
3711 :
3712 0 : pub fn get_gc_horizon(&self) -> u64 {
3713 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3714 0 : tenant_conf
3715 0 : .gc_horizon
3716 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3717 0 : }
3718 :
3719 0 : pub fn get_gc_period(&self) -> Duration {
3720 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3721 0 : tenant_conf
3722 0 : .gc_period
3723 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3724 0 : }
3725 :
3726 0 : pub fn get_image_creation_threshold(&self) -> usize {
3727 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3728 0 : tenant_conf
3729 0 : .image_creation_threshold
3730 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3731 0 : }
3732 :
3733 0 : pub fn get_pitr_interval(&self) -> Duration {
3734 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3735 0 : tenant_conf
3736 0 : .pitr_interval
3737 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3738 0 : }
3739 :
3740 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3741 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3742 0 : tenant_conf
3743 0 : .min_resident_size_override
3744 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3745 0 : }
3746 :
3747 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3748 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3749 0 : let heatmap_period = tenant_conf
3750 0 : .heatmap_period
3751 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
3752 0 : if heatmap_period.is_zero() {
3753 0 : None
3754 : } else {
3755 0 : Some(heatmap_period)
3756 : }
3757 0 : }
3758 :
3759 4 : pub fn get_lsn_lease_length(&self) -> Duration {
3760 4 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3761 4 : tenant_conf
3762 4 : .lsn_lease_length
3763 4 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
3764 4 : }
3765 :
3766 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
3767 2 : fn build_tenant_manifest(&self) -> TenantManifest {
3768 2 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3769 2 :
3770 2 : let mut timeline_manifests = timelines_offloaded
3771 2 : .iter()
3772 2 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
3773 2 : .collect::<Vec<_>>();
3774 2 : // Sort the manifests so that our output is deterministic
3775 2 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
3776 2 :
3777 2 : TenantManifest {
3778 2 : version: LATEST_TENANT_MANIFEST_VERSION,
3779 2 : offloaded_timelines: timeline_manifests,
3780 2 : }
3781 2 : }
3782 :
3783 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
3784 0 : // Use read-copy-update in order to avoid overwriting the location config
3785 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
3786 0 : // this race is not possible if both request types come from the storage
3787 0 : // controller (as they should!) because an exclusive op lock is required
3788 0 : // on the storage controller side.
3789 0 :
3790 0 : self.tenant_conf.rcu(|inner| {
3791 0 : Arc::new(AttachedTenantConf {
3792 0 : tenant_conf: new_tenant_conf.clone(),
3793 0 : location: inner.location,
3794 0 : // Attached location is not changed, no need to update lsn lease deadline.
3795 0 : lsn_lease_deadline: inner.lsn_lease_deadline,
3796 0 : })
3797 0 : });
3798 0 :
3799 0 : let updated = self.tenant_conf.load().clone();
3800 0 :
3801 0 : self.tenant_conf_updated(&new_tenant_conf);
3802 0 : // Don't hold self.timelines.lock() during the notifies.
3803 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3804 0 : // mutexes in struct Timeline in the future.
3805 0 : let timelines = self.list_timelines();
3806 0 : for timeline in timelines {
3807 0 : timeline.tenant_conf_updated(&updated);
3808 0 : }
3809 0 : }
3810 :
3811 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
3812 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
3813 0 :
3814 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
3815 0 :
3816 0 : self.tenant_conf_updated(&new_tenant_conf);
3817 0 : // Don't hold self.timelines.lock() during the notifies.
3818 0 : // There's no risk of deadlock right now, but there could be if we consolidate
3819 0 : // mutexes in struct Timeline in the future.
3820 0 : let timelines = self.list_timelines();
3821 0 : for timeline in timelines {
3822 0 : timeline.tenant_conf_updated(&new_conf);
3823 0 : }
3824 0 : }
3825 :
3826 192 : fn get_timeline_get_throttle_config(
3827 192 : psconf: &'static PageServerConf,
3828 192 : overrides: &TenantConfOpt,
3829 192 : ) -> throttle::Config {
3830 192 : overrides
3831 192 : .timeline_get_throttle
3832 192 : .clone()
3833 192 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
3834 192 : }
3835 :
3836 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
3837 0 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
3838 0 : self.timeline_get_throttle.reconfigure(conf)
3839 0 : }
3840 :
3841 : /// Helper function to create a new Timeline struct.
3842 : ///
3843 : /// The returned Timeline is in Loading state. The caller is responsible for
3844 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
3845 : /// map.
3846 : ///
3847 : /// `validate_ancestor == false` is used when a timeline is created for deletion
3848 : /// and we might not have the ancestor present anymore which is fine for to be
3849 : /// deleted timelines.
3850 : #[allow(clippy::too_many_arguments)]
3851 418 : fn create_timeline_struct(
3852 418 : &self,
3853 418 : new_timeline_id: TimelineId,
3854 418 : new_metadata: &TimelineMetadata,
3855 418 : ancestor: Option<Arc<Timeline>>,
3856 418 : resources: TimelineResources,
3857 418 : cause: CreateTimelineCause,
3858 418 : create_idempotency: CreateTimelineIdempotency,
3859 418 : ) -> anyhow::Result<Arc<Timeline>> {
3860 418 : let state = match cause {
3861 : CreateTimelineCause::Load => {
3862 418 : let ancestor_id = new_metadata.ancestor_timeline();
3863 418 : anyhow::ensure!(
3864 418 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
3865 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
3866 : );
3867 418 : TimelineState::Loading
3868 : }
3869 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
3870 : };
3871 :
3872 418 : let pg_version = new_metadata.pg_version();
3873 418 :
3874 418 : let timeline = Timeline::new(
3875 418 : self.conf,
3876 418 : Arc::clone(&self.tenant_conf),
3877 418 : new_metadata,
3878 418 : ancestor,
3879 418 : new_timeline_id,
3880 418 : self.tenant_shard_id,
3881 418 : self.generation,
3882 418 : self.shard_identity,
3883 418 : self.walredo_mgr.clone(),
3884 418 : resources,
3885 418 : pg_version,
3886 418 : state,
3887 418 : self.attach_wal_lag_cooldown.clone(),
3888 418 : create_idempotency,
3889 418 : self.cancel.child_token(),
3890 418 : );
3891 418 :
3892 418 : Ok(timeline)
3893 418 : }
3894 :
3895 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
3896 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
3897 : #[allow(clippy::too_many_arguments)]
3898 192 : fn new(
3899 192 : state: TenantState,
3900 192 : conf: &'static PageServerConf,
3901 192 : attached_conf: AttachedTenantConf,
3902 192 : shard_identity: ShardIdentity,
3903 192 : walredo_mgr: Option<Arc<WalRedoManager>>,
3904 192 : tenant_shard_id: TenantShardId,
3905 192 : remote_storage: GenericRemoteStorage,
3906 192 : deletion_queue_client: DeletionQueueClient,
3907 192 : l0_flush_global_state: L0FlushGlobalState,
3908 192 : ) -> Tenant {
3909 192 : debug_assert!(
3910 192 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
3911 : );
3912 :
3913 192 : let (state, mut rx) = watch::channel(state);
3914 192 :
3915 192 : tokio::spawn(async move {
3916 192 : // reflect tenant state in metrics:
3917 192 : // - global per tenant state: TENANT_STATE_METRIC
3918 192 : // - "set" of broken tenants: BROKEN_TENANTS_SET
3919 192 : //
3920 192 : // set of broken tenants should not have zero counts so that it remains accessible for
3921 192 : // alerting.
3922 192 :
3923 192 : let tid = tenant_shard_id.to_string();
3924 192 : let shard_id = tenant_shard_id.shard_slug().to_string();
3925 192 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
3926 :
3927 382 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
3928 382 : ([state.into()], matches!(state, TenantState::Broken { .. }))
3929 382 : }
3930 :
3931 192 : let mut tuple = inspect_state(&rx.borrow_and_update());
3932 192 :
3933 192 : let is_broken = tuple.1;
3934 192 : let mut counted_broken = if is_broken {
3935 : // add the id to the set right away, there should not be any updates on the channel
3936 : // after before tenant is removed, if ever
3937 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3938 0 : true
3939 : } else {
3940 192 : false
3941 : };
3942 :
3943 : loop {
3944 382 : let labels = &tuple.0;
3945 382 : let current = TENANT_STATE_METRIC.with_label_values(labels);
3946 382 : current.inc();
3947 382 :
3948 382 : if rx.changed().await.is_err() {
3949 : // tenant has been dropped
3950 14 : current.dec();
3951 14 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
3952 14 : break;
3953 190 : }
3954 190 :
3955 190 : current.dec();
3956 190 : tuple = inspect_state(&rx.borrow_and_update());
3957 190 :
3958 190 : let is_broken = tuple.1;
3959 190 : if is_broken && !counted_broken {
3960 0 : counted_broken = true;
3961 0 : // insert the tenant_id (back) into the set while avoiding needless counter
3962 0 : // access
3963 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
3964 190 : }
3965 : }
3966 192 : });
3967 192 :
3968 192 : Tenant {
3969 192 : tenant_shard_id,
3970 192 : shard_identity,
3971 192 : generation: attached_conf.location.generation,
3972 192 : conf,
3973 192 : // using now here is good enough approximation to catch tenants with really long
3974 192 : // activation times.
3975 192 : constructed_at: Instant::now(),
3976 192 : timelines: Mutex::new(HashMap::new()),
3977 192 : timelines_creating: Mutex::new(HashSet::new()),
3978 192 : timelines_offloaded: Mutex::new(HashMap::new()),
3979 192 : tenant_manifest_upload: Default::default(),
3980 192 : gc_cs: tokio::sync::Mutex::new(()),
3981 192 : walredo_mgr,
3982 192 : remote_storage,
3983 192 : deletion_queue_client,
3984 192 : state,
3985 192 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
3986 192 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
3987 192 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
3988 192 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
3989 192 : format!("compaction-{tenant_shard_id}"),
3990 192 : 5,
3991 192 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
3992 192 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
3993 192 : // use an extremely long backoff.
3994 192 : Some(Duration::from_secs(3600 * 24)),
3995 192 : )),
3996 192 : activate_now_sem: tokio::sync::Semaphore::new(0),
3997 192 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
3998 192 : cancel: CancellationToken::default(),
3999 192 : gate: Gate::default(),
4000 192 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
4001 192 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
4002 192 : crate::metrics::tenant_throttling::TimelineGet::new(&tenant_shard_id),
4003 192 : )),
4004 192 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4005 192 : ongoing_timeline_detach: std::sync::Mutex::default(),
4006 192 : gc_block: Default::default(),
4007 192 : l0_flush_global_state,
4008 192 : }
4009 192 : }
4010 :
4011 : /// Locate and load config
4012 0 : pub(super) fn load_tenant_config(
4013 0 : conf: &'static PageServerConf,
4014 0 : tenant_shard_id: &TenantShardId,
4015 0 : ) -> Result<LocationConf, LoadConfigError> {
4016 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4017 0 :
4018 0 : info!("loading tenant configuration from {config_path}");
4019 :
4020 : // load and parse file
4021 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4022 0 : match e.kind() {
4023 : std::io::ErrorKind::NotFound => {
4024 : // The config should almost always exist for a tenant directory:
4025 : // - When attaching a tenant, the config is the first thing we write
4026 : // - When detaching a tenant, we atomically move the directory to a tmp location
4027 : // before deleting contents.
4028 : //
4029 : // The very rare edge case that can result in a missing config is if we crash during attach
4030 : // between creating directory and writing config. Callers should handle that as if the
4031 : // directory didn't exist.
4032 :
4033 0 : LoadConfigError::NotFound(config_path)
4034 : }
4035 : _ => {
4036 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4037 : // that we cannot cleanly recover
4038 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4039 : }
4040 : }
4041 0 : })?;
4042 :
4043 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4044 0 : }
4045 :
4046 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4047 : pub(super) async fn persist_tenant_config(
4048 : conf: &'static PageServerConf,
4049 : tenant_shard_id: &TenantShardId,
4050 : location_conf: &LocationConf,
4051 : ) -> std::io::Result<()> {
4052 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4053 :
4054 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4055 : }
4056 :
4057 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4058 : pub(super) async fn persist_tenant_config_at(
4059 : tenant_shard_id: &TenantShardId,
4060 : config_path: &Utf8Path,
4061 : location_conf: &LocationConf,
4062 : ) -> std::io::Result<()> {
4063 : debug!("persisting tenantconf to {config_path}");
4064 :
4065 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4066 : # It is read in case of pageserver restart.
4067 : "#
4068 : .to_string();
4069 :
4070 0 : fail::fail_point!("tenant-config-before-write", |_| {
4071 0 : Err(std::io::Error::new(
4072 0 : std::io::ErrorKind::Other,
4073 0 : "tenant-config-before-write",
4074 0 : ))
4075 0 : });
4076 :
4077 : // Convert the config to a toml file.
4078 : conf_content +=
4079 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4080 :
4081 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4082 :
4083 : let conf_content = conf_content.into_bytes();
4084 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4085 : }
4086 :
4087 : //
4088 : // How garbage collection works:
4089 : //
4090 : // +--bar------------->
4091 : // /
4092 : // +----+-----foo---------------->
4093 : // /
4094 : // ----main--+-------------------------->
4095 : // \
4096 : // +-----baz-------->
4097 : //
4098 : //
4099 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4100 : // `gc_infos` are being refreshed
4101 : // 2. Scan collected timelines, and on each timeline, make note of the
4102 : // all the points where other timelines have been branched off.
4103 : // We will refrain from removing page versions at those LSNs.
4104 : // 3. For each timeline, scan all layer files on the timeline.
4105 : // Remove all files for which a newer file exists and which
4106 : // don't cover any branch point LSNs.
4107 : //
4108 : // TODO:
4109 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4110 : // don't need to keep that in the parent anymore. But currently
4111 : // we do.
4112 4 : async fn gc_iteration_internal(
4113 4 : &self,
4114 4 : target_timeline_id: Option<TimelineId>,
4115 4 : horizon: u64,
4116 4 : pitr: Duration,
4117 4 : cancel: &CancellationToken,
4118 4 : ctx: &RequestContext,
4119 4 : ) -> Result<GcResult, GcError> {
4120 4 : let mut totals: GcResult = Default::default();
4121 4 : let now = Instant::now();
4122 :
4123 4 : let gc_timelines = self
4124 4 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4125 3 : .await?;
4126 :
4127 4 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4128 :
4129 : // If there is nothing to GC, we don't want any messages in the INFO log.
4130 4 : if !gc_timelines.is_empty() {
4131 4 : info!("{} timelines need GC", gc_timelines.len());
4132 : } else {
4133 0 : debug!("{} timelines need GC", gc_timelines.len());
4134 : }
4135 :
4136 : // Perform GC for each timeline.
4137 : //
4138 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
4139 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4140 : // with branch creation.
4141 : //
4142 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
4143 : // creation task can run concurrently with timeline's GC iteration.
4144 8 : for timeline in gc_timelines {
4145 4 : if cancel.is_cancelled() {
4146 : // We were requested to shut down. Stop and return with the progress we
4147 : // made.
4148 0 : break;
4149 4 : }
4150 4 : let result = match timeline.gc().await {
4151 : Err(GcError::TimelineCancelled) => {
4152 0 : if target_timeline_id.is_some() {
4153 : // If we were targetting this specific timeline, surface cancellation to caller
4154 0 : return Err(GcError::TimelineCancelled);
4155 : } else {
4156 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4157 : // skip past this and proceed to try GC on other timelines.
4158 0 : continue;
4159 : }
4160 : }
4161 4 : r => r?,
4162 : };
4163 4 : totals += result;
4164 : }
4165 :
4166 4 : totals.elapsed = now.elapsed();
4167 4 : Ok(totals)
4168 4 : }
4169 :
4170 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4171 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4172 : /// [`Tenant::get_gc_horizon`].
4173 : ///
4174 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4175 0 : pub(crate) async fn refresh_gc_info(
4176 0 : &self,
4177 0 : cancel: &CancellationToken,
4178 0 : ctx: &RequestContext,
4179 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4180 0 : // since this method can now be called at different rates than the configured gc loop, it
4181 0 : // might be that these configuration values get applied faster than what it was previously,
4182 0 : // since these were only read from the gc task.
4183 0 : let horizon = self.get_gc_horizon();
4184 0 : let pitr = self.get_pitr_interval();
4185 0 :
4186 0 : // refresh all timelines
4187 0 : let target_timeline_id = None;
4188 0 :
4189 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4190 0 : .await
4191 0 : }
4192 :
4193 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4194 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4195 : ///
4196 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4197 0 : fn initialize_gc_info(
4198 0 : &self,
4199 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4200 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4201 0 : restrict_to_timeline: Option<TimelineId>,
4202 0 : ) {
4203 0 : if restrict_to_timeline.is_none() {
4204 : // This function must be called before activation: after activation timeline create/delete operations
4205 : // might happen, and this function is not safe to run concurrently with those.
4206 0 : assert!(!self.is_active());
4207 0 : }
4208 :
4209 : // Scan all timelines. For each timeline, remember the timeline ID and
4210 : // the branch point where it was created.
4211 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4212 0 : BTreeMap::new();
4213 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4214 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4215 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4216 0 : ancestor_children.push((
4217 0 : timeline_entry.get_ancestor_lsn(),
4218 0 : *timeline_id,
4219 0 : MaybeOffloaded::No,
4220 0 : ));
4221 0 : }
4222 0 : });
4223 0 : timelines_offloaded
4224 0 : .iter()
4225 0 : .for_each(|(timeline_id, timeline_entry)| {
4226 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4227 0 : return;
4228 : };
4229 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4230 0 : return;
4231 : };
4232 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4233 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4234 0 : });
4235 0 :
4236 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4237 0 : let horizon = self.get_gc_horizon();
4238 :
4239 : // Populate each timeline's GcInfo with information about its child branches
4240 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4241 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4242 : } else {
4243 0 : itertools::Either::Right(timelines.values())
4244 : };
4245 0 : for timeline in timelines_to_write {
4246 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4247 0 : .remove(&timeline.timeline_id)
4248 0 : .unwrap_or_default();
4249 0 :
4250 0 : branchpoints.sort_by_key(|b| b.0);
4251 0 :
4252 0 : let mut target = timeline.gc_info.write().unwrap();
4253 0 :
4254 0 : target.retain_lsns = branchpoints;
4255 0 :
4256 0 : let space_cutoff = timeline
4257 0 : .get_last_record_lsn()
4258 0 : .checked_sub(horizon)
4259 0 : .unwrap_or(Lsn(0));
4260 0 :
4261 0 : target.cutoffs = GcCutoffs {
4262 0 : space: space_cutoff,
4263 0 : time: Lsn::INVALID,
4264 0 : };
4265 0 : }
4266 0 : }
4267 :
4268 4 : async fn refresh_gc_info_internal(
4269 4 : &self,
4270 4 : target_timeline_id: Option<TimelineId>,
4271 4 : horizon: u64,
4272 4 : pitr: Duration,
4273 4 : cancel: &CancellationToken,
4274 4 : ctx: &RequestContext,
4275 4 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4276 4 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4277 4 : // currently visible timelines.
4278 4 : let timelines = self
4279 4 : .timelines
4280 4 : .lock()
4281 4 : .unwrap()
4282 4 : .values()
4283 4 : .filter(|tl| match target_timeline_id.as_ref() {
4284 4 : Some(target) => &tl.timeline_id == target,
4285 0 : None => true,
4286 4 : })
4287 4 : .cloned()
4288 4 : .collect::<Vec<_>>();
4289 4 :
4290 4 : if target_timeline_id.is_some() && timelines.is_empty() {
4291 : // We were to act on a particular timeline and it wasn't found
4292 0 : return Err(GcError::TimelineNotFound);
4293 4 : }
4294 4 :
4295 4 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4296 4 : HashMap::with_capacity(timelines.len());
4297 :
4298 4 : for timeline in timelines.iter() {
4299 4 : let cutoff = timeline
4300 4 : .get_last_record_lsn()
4301 4 : .checked_sub(horizon)
4302 4 : .unwrap_or(Lsn(0));
4303 :
4304 4 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
4305 4 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4306 4 : assert!(old.is_none());
4307 : }
4308 :
4309 4 : if !self.is_active() || self.cancel.is_cancelled() {
4310 0 : return Err(GcError::TenantCancelled);
4311 4 : }
4312 :
4313 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4314 : // because that will stall branch creation.
4315 4 : let gc_cs = self.gc_cs.lock().await;
4316 :
4317 : // Ok, we now know all the branch points.
4318 : // Update the GC information for each timeline.
4319 4 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4320 8 : for timeline in timelines {
4321 : // We filtered the timeline list above
4322 4 : if let Some(target_timeline_id) = target_timeline_id {
4323 4 : assert_eq!(target_timeline_id, timeline.timeline_id);
4324 0 : }
4325 :
4326 : {
4327 4 : let mut target = timeline.gc_info.write().unwrap();
4328 4 :
4329 4 : // Cull any expired leases
4330 4 : let now = SystemTime::now();
4331 6 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4332 4 :
4333 4 : timeline
4334 4 : .metrics
4335 4 : .valid_lsn_lease_count_gauge
4336 4 : .set(target.leases.len() as u64);
4337 :
4338 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4339 4 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4340 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4341 0 : target.within_ancestor_pitr =
4342 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4343 0 : }
4344 4 : }
4345 :
4346 : // Update metrics that depend on GC state
4347 4 : timeline
4348 4 : .metrics
4349 4 : .archival_size
4350 4 : .set(if target.within_ancestor_pitr {
4351 0 : timeline.metrics.current_logical_size_gauge.get()
4352 : } else {
4353 4 : 0
4354 : });
4355 4 : timeline.metrics.pitr_history_size.set(
4356 4 : timeline
4357 4 : .get_last_record_lsn()
4358 4 : .checked_sub(target.cutoffs.time)
4359 4 : .unwrap_or(Lsn(0))
4360 4 : .0,
4361 4 : );
4362 :
4363 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4364 : // - this timeline was created while we were finding cutoffs
4365 : // - lsn for timestamp search fails for this timeline repeatedly
4366 4 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4367 4 : target.cutoffs = cutoffs.clone();
4368 4 : }
4369 : }
4370 :
4371 4 : gc_timelines.push(timeline);
4372 : }
4373 4 : drop(gc_cs);
4374 4 : Ok(gc_timelines)
4375 4 : }
4376 :
4377 : /// A substitute for `branch_timeline` for use in unit tests.
4378 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4379 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4380 : /// timeline background tasks are launched, except the flush loop.
4381 : #[cfg(test)]
4382 232 : async fn branch_timeline_test(
4383 232 : self: &Arc<Self>,
4384 232 : src_timeline: &Arc<Timeline>,
4385 232 : dst_id: TimelineId,
4386 232 : ancestor_lsn: Option<Lsn>,
4387 232 : ctx: &RequestContext,
4388 232 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4389 232 : let tl = self
4390 232 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4391 190 : .await?
4392 228 : .into_timeline_for_test();
4393 228 : tl.set_state(TimelineState::Active);
4394 228 : Ok(tl)
4395 232 : }
4396 :
4397 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4398 : #[cfg(test)]
4399 : #[allow(clippy::too_many_arguments)]
4400 6 : pub async fn branch_timeline_test_with_layers(
4401 6 : self: &Arc<Self>,
4402 6 : src_timeline: &Arc<Timeline>,
4403 6 : dst_id: TimelineId,
4404 6 : ancestor_lsn: Option<Lsn>,
4405 6 : ctx: &RequestContext,
4406 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4407 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4408 6 : end_lsn: Lsn,
4409 6 : ) -> anyhow::Result<Arc<Timeline>> {
4410 : use checks::check_valid_layermap;
4411 : use itertools::Itertools;
4412 :
4413 6 : let tline = self
4414 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4415 4 : .await?;
4416 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4417 6 : ancestor_lsn
4418 : } else {
4419 0 : tline.get_last_record_lsn()
4420 : };
4421 6 : assert!(end_lsn >= ancestor_lsn);
4422 6 : tline.force_advance_lsn(end_lsn);
4423 12 : for deltas in delta_layer_desc {
4424 6 : tline
4425 6 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4426 18 : .await?;
4427 : }
4428 10 : for (lsn, images) in image_layer_desc {
4429 4 : tline
4430 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4431 14 : .await?;
4432 : }
4433 6 : let layer_names = tline
4434 6 : .layers
4435 6 : .read()
4436 0 : .await
4437 6 : .layer_map()
4438 6 : .unwrap()
4439 6 : .iter_historic_layers()
4440 10 : .map(|layer| layer.layer_name())
4441 6 : .collect_vec();
4442 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4443 0 : bail!("invalid layermap: {err}");
4444 6 : }
4445 6 : Ok(tline)
4446 6 : }
4447 :
4448 : /// Branch an existing timeline.
4449 0 : async fn branch_timeline(
4450 0 : self: &Arc<Self>,
4451 0 : src_timeline: &Arc<Timeline>,
4452 0 : dst_id: TimelineId,
4453 0 : start_lsn: Option<Lsn>,
4454 0 : ctx: &RequestContext,
4455 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4456 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4457 0 : .await
4458 0 : }
4459 :
4460 232 : async fn branch_timeline_impl(
4461 232 : self: &Arc<Self>,
4462 232 : src_timeline: &Arc<Timeline>,
4463 232 : dst_id: TimelineId,
4464 232 : start_lsn: Option<Lsn>,
4465 232 : _ctx: &RequestContext,
4466 232 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4467 232 : let src_id = src_timeline.timeline_id;
4468 :
4469 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4470 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4471 : // valid while we are creating the branch.
4472 232 : let _gc_cs = self.gc_cs.lock().await;
4473 :
4474 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4475 232 : let start_lsn = start_lsn.unwrap_or_else(|| {
4476 2 : let lsn = src_timeline.get_last_record_lsn();
4477 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4478 2 : lsn
4479 232 : });
4480 :
4481 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4482 232 : let timeline_create_guard = match self
4483 232 : .start_creating_timeline(
4484 232 : dst_id,
4485 232 : CreateTimelineIdempotency::Branch {
4486 232 : ancestor_timeline_id: src_timeline.timeline_id,
4487 232 : ancestor_start_lsn: start_lsn,
4488 232 : },
4489 232 : )
4490 190 : .await?
4491 : {
4492 232 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4493 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4494 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4495 : }
4496 : };
4497 :
4498 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4499 : // horizon on the source timeline
4500 : //
4501 : // We check it against both the planned GC cutoff stored in 'gc_info',
4502 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4503 : // planned GC cutoff in 'gc_info' is normally larger than
4504 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
4505 : // changed the GC settings for the tenant to make the PITR window
4506 : // larger, but some of the data was already removed by an earlier GC
4507 : // iteration.
4508 :
4509 : // check against last actual 'latest_gc_cutoff' first
4510 232 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
4511 232 : src_timeline
4512 232 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
4513 232 : .context(format!(
4514 232 : "invalid branch start lsn: less than latest GC cutoff {}",
4515 232 : *latest_gc_cutoff_lsn,
4516 232 : ))
4517 232 : .map_err(CreateTimelineError::AncestorLsn)?;
4518 :
4519 : // and then the planned GC cutoff
4520 : {
4521 228 : let gc_info = src_timeline.gc_info.read().unwrap();
4522 228 : let cutoff = gc_info.min_cutoff();
4523 228 : if start_lsn < cutoff {
4524 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4525 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
4526 0 : )));
4527 228 : }
4528 228 : }
4529 228 :
4530 228 : //
4531 228 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4532 228 : // so that GC cannot advance the GC cutoff until we are finished.
4533 228 : // Proceed with the branch creation.
4534 228 : //
4535 228 :
4536 228 : // Determine prev-LSN for the new timeline. We can only determine it if
4537 228 : // the timeline was branched at the current end of the source timeline.
4538 228 : let RecordLsn {
4539 228 : last: src_last,
4540 228 : prev: src_prev,
4541 228 : } = src_timeline.get_last_record_rlsn();
4542 228 : let dst_prev = if src_last == start_lsn {
4543 216 : Some(src_prev)
4544 : } else {
4545 12 : None
4546 : };
4547 :
4548 : // Create the metadata file, noting the ancestor of the new timeline.
4549 : // There is initially no data in it, but all the read-calls know to look
4550 : // into the ancestor.
4551 228 : let metadata = TimelineMetadata::new(
4552 228 : start_lsn,
4553 228 : dst_prev,
4554 228 : Some(src_id),
4555 228 : start_lsn,
4556 228 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4557 228 : src_timeline.initdb_lsn,
4558 228 : src_timeline.pg_version,
4559 228 : );
4560 :
4561 228 : let uninitialized_timeline = self
4562 228 : .prepare_new_timeline(
4563 228 : dst_id,
4564 228 : &metadata,
4565 228 : timeline_create_guard,
4566 228 : start_lsn + 1,
4567 228 : Some(Arc::clone(src_timeline)),
4568 228 : )
4569 0 : .await?;
4570 :
4571 228 : let new_timeline = uninitialized_timeline.finish_creation()?;
4572 :
4573 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4574 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4575 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4576 : // could get incorrect information and remove more layers, than needed.
4577 : // See also https://github.com/neondatabase/neon/issues/3865
4578 228 : new_timeline
4579 228 : .remote_client
4580 228 : .schedule_index_upload_for_full_metadata_update(&metadata)
4581 228 : .context("branch initial metadata upload")?;
4582 :
4583 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4584 :
4585 228 : Ok(CreateTimelineResult::Created(new_timeline))
4586 232 : }
4587 :
4588 : /// For unit tests, make this visible so that other modules can directly create timelines
4589 : #[cfg(test)]
4590 2 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4591 : pub(crate) async fn bootstrap_timeline_test(
4592 : self: &Arc<Self>,
4593 : timeline_id: TimelineId,
4594 : pg_version: u32,
4595 : load_existing_initdb: Option<TimelineId>,
4596 : ctx: &RequestContext,
4597 : ) -> anyhow::Result<Arc<Timeline>> {
4598 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4599 : .await
4600 : .map_err(anyhow::Error::new)
4601 2 : .map(|r| r.into_timeline_for_test())
4602 : }
4603 :
4604 : /// Get exclusive access to the timeline ID for creation.
4605 : ///
4606 : /// Timeline-creating code paths must use this function before making changes
4607 : /// to in-memory or persistent state.
4608 : ///
4609 : /// The `state` parameter is a description of the timeline creation operation
4610 : /// we intend to perform.
4611 : /// If the timeline was already created in the meantime, we check whether this
4612 : /// request conflicts or is idempotent , based on `state`.
4613 418 : async fn start_creating_timeline(
4614 418 : self: &Arc<Self>,
4615 418 : new_timeline_id: TimelineId,
4616 418 : idempotency: CreateTimelineIdempotency,
4617 418 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4618 418 : let allow_offloaded = false;
4619 418 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4620 416 : Ok(create_guard) => {
4621 416 : pausable_failpoint!("timeline-creation-after-uninit");
4622 416 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4623 : }
4624 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4625 : Err(TimelineExclusionError::AlreadyCreating) => {
4626 : // Creation is in progress, we cannot create it again, and we cannot
4627 : // check if this request matches the existing one, so caller must try
4628 : // again later.
4629 0 : Err(CreateTimelineError::AlreadyCreating)
4630 : }
4631 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4632 : Err(TimelineExclusionError::AlreadyExists {
4633 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4634 0 : ..
4635 0 : }) => {
4636 0 : info!("timeline already exists but is offloaded");
4637 0 : Err(CreateTimelineError::Conflict)
4638 : }
4639 : Err(TimelineExclusionError::AlreadyExists {
4640 2 : existing: TimelineOrOffloaded::Timeline(existing),
4641 2 : arg,
4642 2 : }) => {
4643 2 : {
4644 2 : let existing = &existing.create_idempotency;
4645 2 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4646 2 : debug!("timeline already exists");
4647 :
4648 2 : match (existing, &arg) {
4649 : // FailWithConflict => no idempotency check
4650 : (CreateTimelineIdempotency::FailWithConflict, _)
4651 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4652 2 : warn!("timeline already exists, failing request");
4653 2 : return Err(CreateTimelineError::Conflict);
4654 : }
4655 : // Idempotent <=> CreateTimelineIdempotency is identical
4656 0 : (x, y) if x == y => {
4657 0 : info!("timeline already exists and idempotency matches, succeeding request");
4658 : // fallthrough
4659 : }
4660 : (_, _) => {
4661 0 : warn!("idempotency conflict, failing request");
4662 0 : return Err(CreateTimelineError::Conflict);
4663 : }
4664 : }
4665 : }
4666 :
4667 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4668 : }
4669 : }
4670 418 : }
4671 :
4672 0 : async fn upload_initdb(
4673 0 : &self,
4674 0 : timelines_path: &Utf8PathBuf,
4675 0 : pgdata_path: &Utf8PathBuf,
4676 0 : timeline_id: &TimelineId,
4677 0 : ) -> anyhow::Result<()> {
4678 0 : let temp_path = timelines_path.join(format!(
4679 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4680 0 : ));
4681 0 :
4682 0 : scopeguard::defer! {
4683 0 : if let Err(e) = fs::remove_file(&temp_path) {
4684 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4685 0 : }
4686 0 : }
4687 :
4688 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4689 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4690 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4691 0 : warn!(
4692 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4693 : );
4694 0 : }
4695 :
4696 0 : pausable_failpoint!("before-initdb-upload");
4697 :
4698 0 : backoff::retry(
4699 0 : || async {
4700 0 : self::remote_timeline_client::upload_initdb_dir(
4701 0 : &self.remote_storage,
4702 0 : &self.tenant_shard_id.tenant_id,
4703 0 : timeline_id,
4704 0 : pgdata_zstd.try_clone().await?,
4705 0 : tar_zst_size,
4706 0 : &self.cancel,
4707 : )
4708 0 : .await
4709 0 : },
4710 0 : |_| false,
4711 0 : 3,
4712 0 : u32::MAX,
4713 0 : "persist_initdb_tar_zst",
4714 0 : &self.cancel,
4715 0 : )
4716 0 : .await
4717 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
4718 0 : .and_then(|x| x)
4719 0 : }
4720 :
4721 : /// - run initdb to init temporary instance and get bootstrap data
4722 : /// - after initialization completes, tar up the temp dir and upload it to S3.
4723 2 : async fn bootstrap_timeline(
4724 2 : self: &Arc<Self>,
4725 2 : timeline_id: TimelineId,
4726 2 : pg_version: u32,
4727 2 : load_existing_initdb: Option<TimelineId>,
4728 2 : ctx: &RequestContext,
4729 2 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4730 2 : let timeline_create_guard = match self
4731 2 : .start_creating_timeline(
4732 2 : timeline_id,
4733 2 : CreateTimelineIdempotency::Bootstrap { pg_version },
4734 2 : )
4735 2 : .await?
4736 : {
4737 2 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4738 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4739 0 : return Ok(CreateTimelineResult::Idempotent(timeline))
4740 : }
4741 : };
4742 :
4743 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
4744 : // temporary directory for basebackup files for the given timeline.
4745 :
4746 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
4747 2 : let pgdata_path = path_with_suffix_extension(
4748 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
4749 2 : TEMP_FILE_SUFFIX,
4750 2 : );
4751 2 :
4752 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
4753 2 : // we won't race with other creations or existent timelines with the same path.
4754 2 : if pgdata_path.exists() {
4755 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
4756 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
4757 0 : })?;
4758 2 : }
4759 :
4760 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
4761 2 : scopeguard::defer! {
4762 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
4763 2 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
4764 2 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
4765 2 : }
4766 2 : }
4767 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
4768 2 : if existing_initdb_timeline_id != timeline_id {
4769 0 : let source_path = &remote_initdb_archive_path(
4770 0 : &self.tenant_shard_id.tenant_id,
4771 0 : &existing_initdb_timeline_id,
4772 0 : );
4773 0 : let dest_path =
4774 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
4775 0 :
4776 0 : // if this fails, it will get retried by retried control plane requests
4777 0 : self.remote_storage
4778 0 : .copy_object(source_path, dest_path, &self.cancel)
4779 0 : .await
4780 0 : .context("copy initdb tar")?;
4781 2 : }
4782 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
4783 2 : self::remote_timeline_client::download_initdb_tar_zst(
4784 2 : self.conf,
4785 2 : &self.remote_storage,
4786 2 : &self.tenant_shard_id,
4787 2 : &existing_initdb_timeline_id,
4788 2 : &self.cancel,
4789 2 : )
4790 675 : .await
4791 2 : .context("download initdb tar")?;
4792 :
4793 2 : scopeguard::defer! {
4794 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
4795 2 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
4796 2 : }
4797 2 : }
4798 2 :
4799 2 : let buf_read =
4800 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
4801 2 : extract_zst_tarball(&pgdata_path, buf_read)
4802 11081 : .await
4803 2 : .context("extract initdb tar")?;
4804 : } else {
4805 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
4806 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
4807 0 : .await
4808 0 : .context("run initdb")?;
4809 :
4810 : // Upload the created data dir to S3
4811 0 : if self.tenant_shard_id().is_shard_zero() {
4812 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
4813 0 : .await?;
4814 0 : }
4815 : }
4816 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
4817 2 :
4818 2 : // Import the contents of the data directory at the initial checkpoint
4819 2 : // LSN, and any WAL after that.
4820 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
4821 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
4822 2 : let new_metadata = TimelineMetadata::new(
4823 2 : Lsn(0),
4824 2 : None,
4825 2 : None,
4826 2 : Lsn(0),
4827 2 : pgdata_lsn,
4828 2 : pgdata_lsn,
4829 2 : pg_version,
4830 2 : );
4831 2 : let raw_timeline = self
4832 2 : .prepare_new_timeline(
4833 2 : timeline_id,
4834 2 : &new_metadata,
4835 2 : timeline_create_guard,
4836 2 : pgdata_lsn,
4837 2 : None,
4838 2 : )
4839 0 : .await?;
4840 :
4841 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
4842 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
4843 :
4844 : // Flush the new layer files to disk, before we make the timeline as available to
4845 : // the outside world.
4846 : //
4847 : // Flush loop needs to be spawned in order to be able to flush.
4848 2 : unfinished_timeline.maybe_spawn_flush_loop();
4849 2 :
4850 2 : import_datadir::import_timeline_from_postgres_datadir(
4851 2 : unfinished_timeline,
4852 2 : &pgdata_path,
4853 2 : pgdata_lsn,
4854 2 : ctx,
4855 2 : )
4856 9381 : .await
4857 2 : .with_context(|| {
4858 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
4859 2 : })?;
4860 :
4861 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
4862 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
4863 0 : "failpoint before-checkpoint-new-timeline"
4864 0 : )))
4865 2 : });
4866 :
4867 2 : unfinished_timeline
4868 2 : .freeze_and_flush()
4869 2 : .await
4870 2 : .with_context(|| {
4871 0 : format!(
4872 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
4873 0 : )
4874 2 : })?;
4875 :
4876 : // All done!
4877 2 : let timeline = raw_timeline.finish_creation()?;
4878 :
4879 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4880 :
4881 2 : Ok(CreateTimelineResult::Created(timeline))
4882 2 : }
4883 :
4884 412 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
4885 412 : RemoteTimelineClient::new(
4886 412 : self.remote_storage.clone(),
4887 412 : self.deletion_queue_client.clone(),
4888 412 : self.conf,
4889 412 : self.tenant_shard_id,
4890 412 : timeline_id,
4891 412 : self.generation,
4892 412 : &self.tenant_conf.load().location,
4893 412 : )
4894 412 : }
4895 :
4896 : /// Call this before constructing a timeline, to build its required structures
4897 412 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
4898 412 : TimelineResources {
4899 412 : remote_client: self.build_timeline_remote_client(timeline_id),
4900 412 : timeline_get_throttle: self.timeline_get_throttle.clone(),
4901 412 : l0_flush_global_state: self.l0_flush_global_state.clone(),
4902 412 : }
4903 412 : }
4904 :
4905 : /// Creates intermediate timeline structure and its files.
4906 : ///
4907 : /// An empty layer map is initialized, and new data and WAL can be imported starting
4908 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
4909 : /// `finish_creation` to insert the Timeline into the timelines map.
4910 412 : async fn prepare_new_timeline<'a>(
4911 412 : &'a self,
4912 412 : new_timeline_id: TimelineId,
4913 412 : new_metadata: &TimelineMetadata,
4914 412 : create_guard: TimelineCreateGuard,
4915 412 : start_lsn: Lsn,
4916 412 : ancestor: Option<Arc<Timeline>>,
4917 412 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
4918 412 : let tenant_shard_id = self.tenant_shard_id;
4919 412 :
4920 412 : let resources = self.build_timeline_resources(new_timeline_id);
4921 412 : resources
4922 412 : .remote_client
4923 412 : .init_upload_queue_for_empty_remote(new_metadata)?;
4924 :
4925 412 : let timeline_struct = self
4926 412 : .create_timeline_struct(
4927 412 : new_timeline_id,
4928 412 : new_metadata,
4929 412 : ancestor,
4930 412 : resources,
4931 412 : CreateTimelineCause::Load,
4932 412 : create_guard.idempotency.clone(),
4933 412 : )
4934 412 : .context("Failed to create timeline data structure")?;
4935 :
4936 412 : timeline_struct.init_empty_layer_map(start_lsn);
4937 :
4938 412 : if let Err(e) = self
4939 412 : .create_timeline_files(&create_guard.timeline_path)
4940 0 : .await
4941 : {
4942 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
4943 0 : cleanup_timeline_directory(create_guard);
4944 0 : return Err(e);
4945 412 : }
4946 412 :
4947 412 : debug!(
4948 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
4949 : );
4950 :
4951 412 : Ok(UninitializedTimeline::new(
4952 412 : self,
4953 412 : new_timeline_id,
4954 412 : Some((timeline_struct, create_guard)),
4955 412 : ))
4956 412 : }
4957 :
4958 412 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
4959 412 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
4960 :
4961 412 : fail::fail_point!("after-timeline-dir-creation", |_| {
4962 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
4963 412 : });
4964 :
4965 412 : Ok(())
4966 412 : }
4967 :
4968 : /// Get a guard that provides exclusive access to the timeline directory, preventing
4969 : /// concurrent attempts to create the same timeline.
4970 : ///
4971 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
4972 : /// offloaded timelines or not.
4973 418 : fn create_timeline_create_guard(
4974 418 : self: &Arc<Self>,
4975 418 : timeline_id: TimelineId,
4976 418 : idempotency: CreateTimelineIdempotency,
4977 418 : allow_offloaded: bool,
4978 418 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
4979 418 : let tenant_shard_id = self.tenant_shard_id;
4980 418 :
4981 418 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
4982 :
4983 418 : let create_guard = TimelineCreateGuard::new(
4984 418 : self,
4985 418 : timeline_id,
4986 418 : timeline_path.clone(),
4987 418 : idempotency,
4988 418 : allow_offloaded,
4989 418 : )?;
4990 :
4991 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
4992 : // for creation.
4993 : // A timeline directory should never exist on disk already:
4994 : // - a previous failed creation would have cleaned up after itself
4995 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
4996 : //
4997 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
4998 : // this error may indicate a bug in cleanup on failed creations.
4999 416 : if timeline_path.exists() {
5000 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5001 0 : "Timeline directory already exists! This is a bug."
5002 0 : )));
5003 416 : }
5004 416 :
5005 416 : Ok(create_guard)
5006 418 : }
5007 :
5008 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5009 : ///
5010 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5011 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5012 : pub async fn gather_size_inputs(
5013 : &self,
5014 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5015 : // (only if it is shorter than the real cutoff).
5016 : max_retention_period: Option<u64>,
5017 : cause: LogicalSizeCalculationCause,
5018 : cancel: &CancellationToken,
5019 : ctx: &RequestContext,
5020 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5021 : let logical_sizes_at_once = self
5022 : .conf
5023 : .concurrent_tenant_size_logical_size_queries
5024 : .inner();
5025 :
5026 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5027 : //
5028 : // But the only case where we need to run multiple of these at once is when we
5029 : // request a size for a tenant manually via API, while another background calculation
5030 : // is in progress (which is not a common case).
5031 : //
5032 : // See more for on the issue #2748 condenced out of the initial PR review.
5033 : let mut shared_cache = tokio::select! {
5034 : locked = self.cached_logical_sizes.lock() => locked,
5035 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5036 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5037 : };
5038 :
5039 : size::gather_inputs(
5040 : self,
5041 : logical_sizes_at_once,
5042 : max_retention_period,
5043 : &mut shared_cache,
5044 : cause,
5045 : cancel,
5046 : ctx,
5047 : )
5048 : .await
5049 : }
5050 :
5051 : /// Calculate synthetic tenant size and cache the result.
5052 : /// This is periodically called by background worker.
5053 : /// result is cached in tenant struct
5054 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5055 : pub async fn calculate_synthetic_size(
5056 : &self,
5057 : cause: LogicalSizeCalculationCause,
5058 : cancel: &CancellationToken,
5059 : ctx: &RequestContext,
5060 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5061 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5062 :
5063 : let size = inputs.calculate();
5064 :
5065 : self.set_cached_synthetic_size(size);
5066 :
5067 : Ok(size)
5068 : }
5069 :
5070 : /// Cache given synthetic size and update the metric value
5071 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5072 0 : self.cached_synthetic_tenant_size
5073 0 : .store(size, Ordering::Relaxed);
5074 0 :
5075 0 : // Only shard zero should be calculating synthetic sizes
5076 0 : debug_assert!(self.shard_identity.is_shard_zero());
5077 :
5078 0 : TENANT_SYNTHETIC_SIZE_METRIC
5079 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5080 0 : .unwrap()
5081 0 : .set(size);
5082 0 : }
5083 :
5084 0 : pub fn cached_synthetic_size(&self) -> u64 {
5085 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5086 0 : }
5087 :
5088 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5089 : ///
5090 : /// This function can take a long time: callers should wrap it in a timeout if calling
5091 : /// from an external API handler.
5092 : ///
5093 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5094 : /// still bounded by tenant/timeline shutdown.
5095 0 : #[tracing::instrument(skip_all)]
5096 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5097 : let timelines = self.timelines.lock().unwrap().clone();
5098 :
5099 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5100 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5101 0 : timeline.freeze_and_flush().await?;
5102 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5103 0 : timeline.remote_client.wait_completion().await?;
5104 :
5105 0 : Ok(())
5106 0 : }
5107 :
5108 : // We do not use a JoinSet for these tasks, because we don't want them to be
5109 : // aborted when this function's future is cancelled: they should stay alive
5110 : // holding their GateGuard until they complete, to ensure their I/Os complete
5111 : // before Timeline shutdown completes.
5112 : let mut results = FuturesUnordered::new();
5113 :
5114 : for (_timeline_id, timeline) in timelines {
5115 : // Run each timeline's flush in a task holding the timeline's gate: this
5116 : // means that if this function's future is cancelled, the Timeline shutdown
5117 : // will still wait for any I/O in here to complete.
5118 : let Ok(gate) = timeline.gate.enter() else {
5119 : continue;
5120 : };
5121 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5122 : results.push(jh);
5123 : }
5124 :
5125 : while let Some(r) = results.next().await {
5126 : if let Err(e) = r {
5127 : if !e.is_cancelled() && !e.is_panic() {
5128 : tracing::error!("unexpected join error: {e:?}");
5129 : }
5130 : }
5131 : }
5132 :
5133 : // The flushes we did above were just writes, but the Tenant might have had
5134 : // pending deletions as well from recent compaction/gc: we want to flush those
5135 : // as well. This requires flushing the global delete queue. This is cheap
5136 : // because it's typically a no-op.
5137 : match self.deletion_queue_client.flush_execute().await {
5138 : Ok(_) => {}
5139 : Err(DeletionQueueError::ShuttingDown) => {}
5140 : }
5141 :
5142 : Ok(())
5143 : }
5144 :
5145 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
5146 0 : self.tenant_conf.load().tenant_conf.clone()
5147 0 : }
5148 :
5149 : /// How much local storage would this tenant like to have? It can cope with
5150 : /// less than this (via eviction and on-demand downloads), but this function enables
5151 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
5152 : /// by keeping important things on local disk.
5153 : ///
5154 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5155 : /// than they report here, due to layer eviction. Tenants with many active branches may
5156 : /// actually use more than they report here.
5157 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5158 0 : let timelines = self.timelines.lock().unwrap();
5159 0 :
5160 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5161 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5162 0 : // of them is used actively enough to occupy space on disk.
5163 0 : timelines
5164 0 : .values()
5165 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5166 0 : .max()
5167 0 : .unwrap_or(0)
5168 0 : }
5169 :
5170 : /// Serialize and write the latest TenantManifest to remote storage.
5171 2 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5172 : // Only one manifest write may be done at at time, and the contents of the manifest
5173 : // must be loaded while holding this lock. This makes it safe to call this function
5174 : // from anywhere without worrying about colliding updates.
5175 2 : let mut guard = tokio::select! {
5176 2 : g = self.tenant_manifest_upload.lock() => {
5177 2 : g
5178 : },
5179 2 : _ = self.cancel.cancelled() => {
5180 0 : return Err(TenantManifestError::Cancelled);
5181 : }
5182 : };
5183 :
5184 2 : let manifest = self.build_tenant_manifest();
5185 2 : if Some(&manifest) == (*guard).as_ref() {
5186 : // Optimisation: skip uploads that don't change anything.
5187 0 : return Ok(());
5188 2 : }
5189 2 :
5190 2 : upload_tenant_manifest(
5191 2 : &self.remote_storage,
5192 2 : &self.tenant_shard_id,
5193 2 : self.generation,
5194 2 : &manifest,
5195 2 : &self.cancel,
5196 2 : )
5197 4 : .await
5198 2 : .map_err(|e| {
5199 0 : if self.cancel.is_cancelled() {
5200 0 : TenantManifestError::Cancelled
5201 : } else {
5202 0 : TenantManifestError::RemoteStorage(e)
5203 : }
5204 2 : })?;
5205 :
5206 : // Store the successfully uploaded manifest, so that future callers can avoid
5207 : // re-uploading the same thing.
5208 2 : *guard = Some(manifest);
5209 2 :
5210 2 : Ok(())
5211 2 : }
5212 : }
5213 :
5214 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5215 : /// to get bootstrap data for timeline initialization.
5216 0 : async fn run_initdb(
5217 0 : conf: &'static PageServerConf,
5218 0 : initdb_target_dir: &Utf8Path,
5219 0 : pg_version: u32,
5220 0 : cancel: &CancellationToken,
5221 0 : ) -> Result<(), InitdbError> {
5222 0 : let initdb_bin_path = conf
5223 0 : .pg_bin_dir(pg_version)
5224 0 : .map_err(InitdbError::Other)?
5225 0 : .join("initdb");
5226 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5227 0 : info!(
5228 0 : "running {} in {}, libdir: {}",
5229 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5230 : );
5231 :
5232 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
5233 :
5234 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5235 0 : superuser: &conf.superuser,
5236 0 : locale: &conf.locale,
5237 0 : initdb_bin: &initdb_bin_path,
5238 0 : pg_version,
5239 0 : library_search_path: &initdb_lib_dir,
5240 0 : pgdata: initdb_target_dir,
5241 0 : })
5242 0 : .await
5243 0 : .map_err(InitdbError::Inner);
5244 0 :
5245 0 : // This isn't true cancellation support, see above. Still return an error to
5246 0 : // excercise the cancellation code path.
5247 0 : if cancel.is_cancelled() {
5248 0 : return Err(InitdbError::Cancelled);
5249 0 : }
5250 0 :
5251 0 : res
5252 0 : }
5253 :
5254 : /// Dump contents of a layer file to stdout.
5255 0 : pub async fn dump_layerfile_from_path(
5256 0 : path: &Utf8Path,
5257 0 : verbose: bool,
5258 0 : ctx: &RequestContext,
5259 0 : ) -> anyhow::Result<()> {
5260 : use std::os::unix::fs::FileExt;
5261 :
5262 : // All layer files start with a two-byte "magic" value, to identify the kind of
5263 : // file.
5264 0 : let file = File::open(path)?;
5265 0 : let mut header_buf = [0u8; 2];
5266 0 : file.read_exact_at(&mut header_buf, 0)?;
5267 :
5268 0 : match u16::from_be_bytes(header_buf) {
5269 : crate::IMAGE_FILE_MAGIC => {
5270 0 : ImageLayer::new_for_path(path, file)?
5271 0 : .dump(verbose, ctx)
5272 0 : .await?
5273 : }
5274 : crate::DELTA_FILE_MAGIC => {
5275 0 : DeltaLayer::new_for_path(path, file)?
5276 0 : .dump(verbose, ctx)
5277 0 : .await?
5278 : }
5279 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5280 : }
5281 :
5282 0 : Ok(())
5283 0 : }
5284 :
5285 : #[cfg(test)]
5286 : pub(crate) mod harness {
5287 : use bytes::{Bytes, BytesMut};
5288 : use once_cell::sync::OnceCell;
5289 : use pageserver_api::models::ShardParameters;
5290 : use pageserver_api::shard::ShardIndex;
5291 : use utils::logging;
5292 :
5293 : use crate::deletion_queue::mock::MockDeletionQueue;
5294 : use crate::l0_flush::L0FlushConfig;
5295 : use crate::walredo::apply_neon;
5296 : use pageserver_api::key::Key;
5297 : use pageserver_api::record::NeonWalRecord;
5298 :
5299 : use super::*;
5300 : use hex_literal::hex;
5301 : use utils::id::TenantId;
5302 :
5303 : pub const TIMELINE_ID: TimelineId =
5304 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5305 : pub const NEW_TIMELINE_ID: TimelineId =
5306 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5307 :
5308 : /// Convenience function to create a page image with given string as the only content
5309 5028721 : pub fn test_img(s: &str) -> Bytes {
5310 5028721 : let mut buf = BytesMut::new();
5311 5028721 : buf.extend_from_slice(s.as_bytes());
5312 5028721 : buf.resize(64, 0);
5313 5028721 :
5314 5028721 : buf.freeze()
5315 5028721 : }
5316 :
5317 : impl From<TenantConf> for TenantConfOpt {
5318 192 : fn from(tenant_conf: TenantConf) -> Self {
5319 192 : Self {
5320 192 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5321 192 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5322 192 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5323 192 : compaction_period: Some(tenant_conf.compaction_period),
5324 192 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5325 192 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5326 192 : gc_horizon: Some(tenant_conf.gc_horizon),
5327 192 : gc_period: Some(tenant_conf.gc_period),
5328 192 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5329 192 : pitr_interval: Some(tenant_conf.pitr_interval),
5330 192 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5331 192 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5332 192 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5333 192 : eviction_policy: Some(tenant_conf.eviction_policy),
5334 192 : min_resident_size_override: tenant_conf.min_resident_size_override,
5335 192 : evictions_low_residence_duration_metric_threshold: Some(
5336 192 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5337 192 : ),
5338 192 : heatmap_period: Some(tenant_conf.heatmap_period),
5339 192 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5340 192 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5341 192 : image_layer_creation_check_threshold: Some(
5342 192 : tenant_conf.image_layer_creation_check_threshold,
5343 192 : ),
5344 192 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5345 192 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5346 192 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5347 192 : }
5348 192 : }
5349 : }
5350 :
5351 : pub struct TenantHarness {
5352 : pub conf: &'static PageServerConf,
5353 : pub tenant_conf: TenantConf,
5354 : pub tenant_shard_id: TenantShardId,
5355 : pub generation: Generation,
5356 : pub shard: ShardIndex,
5357 : pub remote_storage: GenericRemoteStorage,
5358 : pub remote_fs_dir: Utf8PathBuf,
5359 : pub deletion_queue: MockDeletionQueue,
5360 : }
5361 :
5362 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5363 :
5364 208 : pub(crate) fn setup_logging() {
5365 208 : LOG_HANDLE.get_or_init(|| {
5366 196 : logging::init(
5367 196 : logging::LogFormat::Test,
5368 196 : // enable it in case the tests exercise code paths that use
5369 196 : // debug_assert_current_span_has_tenant_and_timeline_id
5370 196 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5371 196 : logging::Output::Stdout,
5372 196 : )
5373 196 : .expect("Failed to init test logging")
5374 208 : });
5375 208 : }
5376 :
5377 : impl TenantHarness {
5378 192 : pub async fn create_custom(
5379 192 : test_name: &'static str,
5380 192 : tenant_conf: TenantConf,
5381 192 : tenant_id: TenantId,
5382 192 : shard_identity: ShardIdentity,
5383 192 : generation: Generation,
5384 192 : ) -> anyhow::Result<Self> {
5385 192 : setup_logging();
5386 192 :
5387 192 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5388 192 : let _ = fs::remove_dir_all(&repo_dir);
5389 192 : fs::create_dir_all(&repo_dir)?;
5390 :
5391 192 : let conf = PageServerConf::dummy_conf(repo_dir);
5392 192 : // Make a static copy of the config. This can never be free'd, but that's
5393 192 : // OK in a test.
5394 192 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5395 192 :
5396 192 : let shard = shard_identity.shard_index();
5397 192 : let tenant_shard_id = TenantShardId {
5398 192 : tenant_id,
5399 192 : shard_number: shard.shard_number,
5400 192 : shard_count: shard.shard_count,
5401 192 : };
5402 192 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5403 192 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5404 :
5405 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5406 192 : let remote_fs_dir = conf.workdir.join("localfs");
5407 192 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5408 192 : let config = RemoteStorageConfig {
5409 192 : storage: RemoteStorageKind::LocalFs {
5410 192 : local_path: remote_fs_dir.clone(),
5411 192 : },
5412 192 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5413 192 : };
5414 192 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5415 192 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5416 192 :
5417 192 : Ok(Self {
5418 192 : conf,
5419 192 : tenant_conf,
5420 192 : tenant_shard_id,
5421 192 : generation,
5422 192 : shard,
5423 192 : remote_storage,
5424 192 : remote_fs_dir,
5425 192 : deletion_queue,
5426 192 : })
5427 192 : }
5428 :
5429 180 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5430 180 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5431 180 : // The tests perform them manually if needed.
5432 180 : let tenant_conf = TenantConf {
5433 180 : gc_period: Duration::ZERO,
5434 180 : compaction_period: Duration::ZERO,
5435 180 : ..TenantConf::default()
5436 180 : };
5437 180 : let tenant_id = TenantId::generate();
5438 180 : let shard = ShardIdentity::unsharded();
5439 180 : Self::create_custom(
5440 180 : test_name,
5441 180 : tenant_conf,
5442 180 : tenant_id,
5443 180 : shard,
5444 180 : Generation::new(0xdeadbeef),
5445 180 : )
5446 0 : .await
5447 180 : }
5448 :
5449 20 : pub fn span(&self) -> tracing::Span {
5450 20 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5451 20 : }
5452 :
5453 192 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5454 192 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5455 192 : (
5456 192 : self.do_try_load(&ctx)
5457 1865 : .await
5458 192 : .expect("failed to load test tenant"),
5459 192 : ctx,
5460 192 : )
5461 192 : }
5462 :
5463 192 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5464 : pub(crate) async fn do_try_load(
5465 : &self,
5466 : ctx: &RequestContext,
5467 : ) -> anyhow::Result<Arc<Tenant>> {
5468 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5469 :
5470 : let tenant = Arc::new(Tenant::new(
5471 : TenantState::Attaching,
5472 : self.conf,
5473 : AttachedTenantConf::try_from(LocationConf::attached_single(
5474 : TenantConfOpt::from(self.tenant_conf.clone()),
5475 : self.generation,
5476 : &ShardParameters::default(),
5477 : ))
5478 : .unwrap(),
5479 : // This is a legacy/test code path: sharding isn't supported here.
5480 : ShardIdentity::unsharded(),
5481 : Some(walredo_mgr),
5482 : self.tenant_shard_id,
5483 : self.remote_storage.clone(),
5484 : self.deletion_queue.new_client(),
5485 : // TODO: ideally we should run all unit tests with both configs
5486 : L0FlushGlobalState::new(L0FlushConfig::default()),
5487 : ));
5488 :
5489 : let preload = tenant
5490 : .preload(&self.remote_storage, CancellationToken::new())
5491 : .await?;
5492 : tenant.attach(Some(preload), ctx).await?;
5493 :
5494 : tenant.state.send_replace(TenantState::Active);
5495 : for timeline in tenant.timelines.lock().unwrap().values() {
5496 : timeline.set_state(TimelineState::Active);
5497 : }
5498 : Ok(tenant)
5499 : }
5500 :
5501 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5502 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5503 2 : }
5504 : }
5505 :
5506 : // Mock WAL redo manager that doesn't do much
5507 : pub(crate) struct TestRedoManager;
5508 :
5509 : impl TestRedoManager {
5510 : /// # Cancel-Safety
5511 : ///
5512 : /// This method is cancellation-safe.
5513 410 : pub async fn request_redo(
5514 410 : &self,
5515 410 : key: Key,
5516 410 : lsn: Lsn,
5517 410 : base_img: Option<(Lsn, Bytes)>,
5518 410 : records: Vec<(Lsn, NeonWalRecord)>,
5519 410 : _pg_version: u32,
5520 410 : ) -> Result<Bytes, walredo::Error> {
5521 570 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5522 410 : if records_neon {
5523 : // For Neon wal records, we can decode without spawning postgres, so do so.
5524 410 : let mut page = match (base_img, records.first()) {
5525 344 : (Some((_lsn, img)), _) => {
5526 344 : let mut page = BytesMut::new();
5527 344 : page.extend_from_slice(&img);
5528 344 : page
5529 : }
5530 66 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5531 : _ => {
5532 0 : panic!("Neon WAL redo requires base image or will init record");
5533 : }
5534 : };
5535 :
5536 980 : for (record_lsn, record) in records {
5537 570 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5538 : }
5539 410 : Ok(page.freeze())
5540 : } else {
5541 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5542 0 : let s = format!(
5543 0 : "redo for {} to get to {}, with {} and {} records",
5544 0 : key,
5545 0 : lsn,
5546 0 : if base_img.is_some() {
5547 0 : "base image"
5548 : } else {
5549 0 : "no base image"
5550 : },
5551 0 : records.len()
5552 0 : );
5553 0 : println!("{s}");
5554 0 :
5555 0 : Ok(test_img(&s))
5556 : }
5557 410 : }
5558 : }
5559 : }
5560 :
5561 : #[cfg(test)]
5562 : mod tests {
5563 : use std::collections::{BTreeMap, BTreeSet};
5564 :
5565 : use super::*;
5566 : use crate::keyspace::KeySpaceAccum;
5567 : use crate::tenant::harness::*;
5568 : use crate::tenant::timeline::CompactFlags;
5569 : use crate::DEFAULT_PG_VERSION;
5570 : use bytes::{Bytes, BytesMut};
5571 : use hex_literal::hex;
5572 : use itertools::Itertools;
5573 : use pageserver_api::key::{Key, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
5574 : use pageserver_api::keyspace::KeySpace;
5575 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5576 : use pageserver_api::value::Value;
5577 : use pageserver_compaction::helpers::overlaps_with;
5578 : use rand::{thread_rng, Rng};
5579 : use storage_layer::PersistentLayerKey;
5580 : use tests::storage_layer::ValuesReconstructState;
5581 : use tests::timeline::{GetVectoredError, ShutdownMode};
5582 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5583 : use utils::id::TenantId;
5584 :
5585 : #[cfg(feature = "testing")]
5586 : use pageserver_api::record::NeonWalRecord;
5587 : #[cfg(feature = "testing")]
5588 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5589 : #[cfg(feature = "testing")]
5590 : use timeline::GcInfo;
5591 :
5592 : static TEST_KEY: Lazy<Key> =
5593 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5594 :
5595 : #[tokio::test]
5596 2 : async fn test_basic() -> anyhow::Result<()> {
5597 20 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5598 2 : let tline = tenant
5599 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5600 6 : .await?;
5601 2 :
5602 2 : let mut writer = tline.writer().await;
5603 2 : writer
5604 2 : .put(
5605 2 : *TEST_KEY,
5606 2 : Lsn(0x10),
5607 2 : &Value::Image(test_img("foo at 0x10")),
5608 2 : &ctx,
5609 2 : )
5610 2 : .await?;
5611 2 : writer.finish_write(Lsn(0x10));
5612 2 : drop(writer);
5613 2 :
5614 2 : let mut writer = tline.writer().await;
5615 2 : writer
5616 2 : .put(
5617 2 : *TEST_KEY,
5618 2 : Lsn(0x20),
5619 2 : &Value::Image(test_img("foo at 0x20")),
5620 2 : &ctx,
5621 2 : )
5622 2 : .await?;
5623 2 : writer.finish_write(Lsn(0x20));
5624 2 : drop(writer);
5625 2 :
5626 2 : assert_eq!(
5627 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5628 2 : test_img("foo at 0x10")
5629 2 : );
5630 2 : assert_eq!(
5631 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5632 2 : test_img("foo at 0x10")
5633 2 : );
5634 2 : assert_eq!(
5635 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5636 2 : test_img("foo at 0x20")
5637 2 : );
5638 2 :
5639 2 : Ok(())
5640 2 : }
5641 :
5642 : #[tokio::test]
5643 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5644 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5645 2 : .await?
5646 2 : .load()
5647 20 : .await;
5648 2 : let _ = tenant
5649 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5650 6 : .await?;
5651 2 :
5652 2 : match tenant
5653 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5654 2 : .await
5655 2 : {
5656 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
5657 2 : Err(e) => assert_eq!(
5658 2 : e.to_string(),
5659 2 : "timeline already exists with different parameters".to_string()
5660 2 : ),
5661 2 : }
5662 2 :
5663 2 : Ok(())
5664 2 : }
5665 :
5666 : /// Convenience function to create a page image with given string as the only content
5667 10 : pub fn test_value(s: &str) -> Value {
5668 10 : let mut buf = BytesMut::new();
5669 10 : buf.extend_from_slice(s.as_bytes());
5670 10 : Value::Image(buf.freeze())
5671 10 : }
5672 :
5673 : ///
5674 : /// Test branch creation
5675 : ///
5676 : #[tokio::test]
5677 2 : async fn test_branch() -> anyhow::Result<()> {
5678 2 : use std::str::from_utf8;
5679 2 :
5680 20 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
5681 2 : let tline = tenant
5682 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5683 6 : .await?;
5684 2 : let mut writer = tline.writer().await;
5685 2 :
5686 2 : #[allow(non_snake_case)]
5687 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
5688 2 : #[allow(non_snake_case)]
5689 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
5690 2 :
5691 2 : // Insert a value on the timeline
5692 2 : writer
5693 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
5694 2 : .await?;
5695 2 : writer
5696 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
5697 2 : .await?;
5698 2 : writer.finish_write(Lsn(0x20));
5699 2 :
5700 2 : writer
5701 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
5702 2 : .await?;
5703 2 : writer.finish_write(Lsn(0x30));
5704 2 : writer
5705 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
5706 2 : .await?;
5707 2 : writer.finish_write(Lsn(0x40));
5708 2 :
5709 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5710 2 :
5711 2 : // Branch the history, modify relation differently on the new timeline
5712 2 : tenant
5713 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
5714 2 : .await?;
5715 2 : let newtline = tenant
5716 2 : .get_timeline(NEW_TIMELINE_ID, true)
5717 2 : .expect("Should have a local timeline");
5718 2 : let mut new_writer = newtline.writer().await;
5719 2 : new_writer
5720 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
5721 2 : .await?;
5722 2 : new_writer.finish_write(Lsn(0x40));
5723 2 :
5724 2 : // Check page contents on both branches
5725 2 : assert_eq!(
5726 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5727 2 : "foo at 0x40"
5728 2 : );
5729 2 : assert_eq!(
5730 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
5731 2 : "bar at 0x40"
5732 2 : );
5733 2 : assert_eq!(
5734 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
5735 2 : "foobar at 0x20"
5736 2 : );
5737 2 :
5738 2 : //assert_current_logical_size(&tline, Lsn(0x40));
5739 2 :
5740 2 : Ok(())
5741 2 : }
5742 :
5743 20 : async fn make_some_layers(
5744 20 : tline: &Timeline,
5745 20 : start_lsn: Lsn,
5746 20 : ctx: &RequestContext,
5747 20 : ) -> anyhow::Result<()> {
5748 20 : let mut lsn = start_lsn;
5749 : {
5750 20 : let mut writer = tline.writer().await;
5751 : // Create a relation on the timeline
5752 20 : writer
5753 20 : .put(
5754 20 : *TEST_KEY,
5755 20 : lsn,
5756 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5757 20 : ctx,
5758 20 : )
5759 10 : .await?;
5760 20 : writer.finish_write(lsn);
5761 20 : lsn += 0x10;
5762 20 : writer
5763 20 : .put(
5764 20 : *TEST_KEY,
5765 20 : lsn,
5766 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5767 20 : ctx,
5768 20 : )
5769 0 : .await?;
5770 20 : writer.finish_write(lsn);
5771 20 : lsn += 0x10;
5772 20 : }
5773 20 : tline.freeze_and_flush().await?;
5774 : {
5775 20 : let mut writer = tline.writer().await;
5776 20 : writer
5777 20 : .put(
5778 20 : *TEST_KEY,
5779 20 : lsn,
5780 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5781 20 : ctx,
5782 20 : )
5783 10 : .await?;
5784 20 : writer.finish_write(lsn);
5785 20 : lsn += 0x10;
5786 20 : writer
5787 20 : .put(
5788 20 : *TEST_KEY,
5789 20 : lsn,
5790 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
5791 20 : ctx,
5792 20 : )
5793 0 : .await?;
5794 20 : writer.finish_write(lsn);
5795 20 : }
5796 21 : tline.freeze_and_flush().await.map_err(|e| e.into())
5797 20 : }
5798 :
5799 : #[tokio::test(start_paused = true)]
5800 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
5801 2 : let (tenant, ctx) =
5802 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
5803 2 : .await?
5804 2 : .load()
5805 12 : .await;
5806 2 : // Advance to the lsn lease deadline so that GC is not blocked by
5807 2 : // initial transition into AttachedSingle.
5808 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
5809 2 : tokio::time::resume();
5810 2 : let tline = tenant
5811 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5812 4 : .await?;
5813 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5814 2 :
5815 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5816 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
5817 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
5818 2 : // below should fail.
5819 2 : tenant
5820 2 : .gc_iteration(
5821 2 : Some(TIMELINE_ID),
5822 2 : 0x10,
5823 2 : Duration::ZERO,
5824 2 : &CancellationToken::new(),
5825 2 : &ctx,
5826 2 : )
5827 2 : .await?;
5828 2 :
5829 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
5830 2 : match tenant
5831 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5832 2 : .await
5833 2 : {
5834 2 : Ok(_) => panic!("branching should have failed"),
5835 2 : Err(err) => {
5836 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5837 2 : panic!("wrong error type")
5838 2 : };
5839 2 : assert!(err.to_string().contains("invalid branch start lsn"));
5840 2 : assert!(err
5841 2 : .source()
5842 2 : .unwrap()
5843 2 : .to_string()
5844 2 : .contains("we might've already garbage collected needed data"))
5845 2 : }
5846 2 : }
5847 2 :
5848 2 : Ok(())
5849 2 : }
5850 :
5851 : #[tokio::test]
5852 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
5853 2 : let (tenant, ctx) =
5854 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
5855 2 : .await?
5856 2 : .load()
5857 20 : .await;
5858 2 :
5859 2 : let tline = tenant
5860 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
5861 6 : .await?;
5862 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
5863 2 : match tenant
5864 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
5865 2 : .await
5866 2 : {
5867 2 : Ok(_) => panic!("branching should have failed"),
5868 2 : Err(err) => {
5869 2 : let CreateTimelineError::AncestorLsn(err) = err else {
5870 2 : panic!("wrong error type");
5871 2 : };
5872 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
5873 2 : assert!(&err
5874 2 : .source()
5875 2 : .unwrap()
5876 2 : .to_string()
5877 2 : .contains("is earlier than latest GC cutoff"));
5878 2 : }
5879 2 : }
5880 2 :
5881 2 : Ok(())
5882 2 : }
5883 :
5884 : /*
5885 : // FIXME: This currently fails to error out. Calling GC doesn't currently
5886 : // remove the old value, we'd need to work a little harder
5887 : #[tokio::test]
5888 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
5889 : let repo =
5890 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
5891 : .load();
5892 :
5893 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
5894 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5895 :
5896 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
5897 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
5898 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
5899 : match tline.get(*TEST_KEY, Lsn(0x25)) {
5900 : Ok(_) => panic!("request for page should have failed"),
5901 : Err(err) => assert!(err.to_string().contains("not found at")),
5902 : }
5903 : Ok(())
5904 : }
5905 : */
5906 :
5907 : #[tokio::test]
5908 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
5909 2 : let (tenant, ctx) =
5910 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
5911 2 : .await?
5912 2 : .load()
5913 20 : .await;
5914 2 : let tline = tenant
5915 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5916 6 : .await?;
5917 7 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5918 2 :
5919 2 : tenant
5920 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5921 2 : .await?;
5922 2 : let newtline = tenant
5923 2 : .get_timeline(NEW_TIMELINE_ID, true)
5924 2 : .expect("Should have a local timeline");
5925 2 :
5926 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
5927 2 :
5928 2 : tline.set_broken("test".to_owned());
5929 2 :
5930 2 : tenant
5931 2 : .gc_iteration(
5932 2 : Some(TIMELINE_ID),
5933 2 : 0x10,
5934 2 : Duration::ZERO,
5935 2 : &CancellationToken::new(),
5936 2 : &ctx,
5937 2 : )
5938 2 : .await?;
5939 2 :
5940 2 : // The branchpoints should contain all timelines, even ones marked
5941 2 : // as Broken.
5942 2 : {
5943 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
5944 2 : assert_eq!(branchpoints.len(), 1);
5945 2 : assert_eq!(
5946 2 : branchpoints[0],
5947 2 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
5948 2 : );
5949 2 : }
5950 2 :
5951 2 : // You can read the key from the child branch even though the parent is
5952 2 : // Broken, as long as you don't need to access data from the parent.
5953 2 : assert_eq!(
5954 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
5955 2 : test_img(&format!("foo at {}", Lsn(0x70)))
5956 2 : );
5957 2 :
5958 2 : // This needs to traverse to the parent, and fails.
5959 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
5960 2 : assert!(
5961 2 : err.to_string().starts_with(&format!(
5962 2 : "bad state on timeline {}: Broken",
5963 2 : tline.timeline_id
5964 2 : )),
5965 2 : "{err}"
5966 2 : );
5967 2 :
5968 2 : Ok(())
5969 2 : }
5970 :
5971 : #[tokio::test]
5972 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
5973 2 : let (tenant, ctx) =
5974 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
5975 2 : .await?
5976 2 : .load()
5977 20 : .await;
5978 2 : let tline = tenant
5979 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5980 6 : .await?;
5981 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
5982 2 :
5983 2 : tenant
5984 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
5985 2 : .await?;
5986 2 : let newtline = tenant
5987 2 : .get_timeline(NEW_TIMELINE_ID, true)
5988 2 : .expect("Should have a local timeline");
5989 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
5990 2 : tenant
5991 2 : .gc_iteration(
5992 2 : Some(TIMELINE_ID),
5993 2 : 0x10,
5994 2 : Duration::ZERO,
5995 2 : &CancellationToken::new(),
5996 2 : &ctx,
5997 2 : )
5998 2 : .await?;
5999 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6000 2 :
6001 2 : Ok(())
6002 2 : }
6003 : #[tokio::test]
6004 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6005 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6006 2 : .await?
6007 2 : .load()
6008 20 : .await;
6009 2 : let tline = tenant
6010 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6011 6 : .await?;
6012 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6013 2 :
6014 2 : tenant
6015 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6016 2 : .await?;
6017 2 : let newtline = tenant
6018 2 : .get_timeline(NEW_TIMELINE_ID, true)
6019 2 : .expect("Should have a local timeline");
6020 2 :
6021 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6022 2 :
6023 2 : // run gc on parent
6024 2 : tenant
6025 2 : .gc_iteration(
6026 2 : Some(TIMELINE_ID),
6027 2 : 0x10,
6028 2 : Duration::ZERO,
6029 2 : &CancellationToken::new(),
6030 2 : &ctx,
6031 2 : )
6032 2 : .await?;
6033 2 :
6034 2 : // Check that the data is still accessible on the branch.
6035 2 : assert_eq!(
6036 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6037 2 : test_img(&format!("foo at {}", Lsn(0x40)))
6038 2 : );
6039 2 :
6040 2 : Ok(())
6041 2 : }
6042 :
6043 : #[tokio::test]
6044 2 : async fn timeline_load() -> anyhow::Result<()> {
6045 2 : const TEST_NAME: &str = "timeline_load";
6046 2 : let harness = TenantHarness::create(TEST_NAME).await?;
6047 2 : {
6048 20 : let (tenant, ctx) = harness.load().await;
6049 2 : let tline = tenant
6050 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6051 6 : .await?;
6052 6 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6053 2 : // so that all uploads finish & we can call harness.load() below again
6054 2 : tenant
6055 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6056 2 : .instrument(harness.span())
6057 2 : .await
6058 2 : .ok()
6059 2 : .unwrap();
6060 2 : }
6061 2 :
6062 15 : let (tenant, _ctx) = harness.load().await;
6063 2 : tenant
6064 2 : .get_timeline(TIMELINE_ID, true)
6065 2 : .expect("cannot load timeline");
6066 2 :
6067 2 : Ok(())
6068 2 : }
6069 :
6070 : #[tokio::test]
6071 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6072 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6073 2 : let harness = TenantHarness::create(TEST_NAME).await?;
6074 2 : // create two timelines
6075 2 : {
6076 20 : let (tenant, ctx) = harness.load().await;
6077 2 : let tline = tenant
6078 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6079 6 : .await?;
6080 2 :
6081 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6082 2 :
6083 2 : let child_tline = tenant
6084 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6085 2 : .await?;
6086 2 : child_tline.set_state(TimelineState::Active);
6087 2 :
6088 2 : let newtline = tenant
6089 2 : .get_timeline(NEW_TIMELINE_ID, true)
6090 2 : .expect("Should have a local timeline");
6091 2 :
6092 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6093 2 :
6094 2 : // so that all uploads finish & we can call harness.load() below again
6095 2 : tenant
6096 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6097 2 : .instrument(harness.span())
6098 2 : .await
6099 2 : .ok()
6100 2 : .unwrap();
6101 2 : }
6102 2 :
6103 2 : // check that both of them are initially unloaded
6104 24 : let (tenant, _ctx) = harness.load().await;
6105 2 :
6106 2 : // check that both, child and ancestor are loaded
6107 2 : let _child_tline = tenant
6108 2 : .get_timeline(NEW_TIMELINE_ID, true)
6109 2 : .expect("cannot get child timeline loaded");
6110 2 :
6111 2 : let _ancestor_tline = tenant
6112 2 : .get_timeline(TIMELINE_ID, true)
6113 2 : .expect("cannot get ancestor timeline loaded");
6114 2 :
6115 2 : Ok(())
6116 2 : }
6117 :
6118 : #[tokio::test]
6119 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6120 2 : use storage_layer::AsLayerDesc;
6121 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6122 2 : .await?
6123 2 : .load()
6124 10 : .await;
6125 2 : let tline = tenant
6126 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6127 5 : .await?;
6128 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6129 2 :
6130 2 : let layer_map = tline.layers.read().await;
6131 2 : let level0_deltas = layer_map
6132 2 : .layer_map()?
6133 2 : .level0_deltas()
6134 2 : .iter()
6135 4 : .map(|desc| layer_map.get_from_desc(desc))
6136 2 : .collect::<Vec<_>>();
6137 2 :
6138 2 : assert!(!level0_deltas.is_empty());
6139 2 :
6140 6 : for delta in level0_deltas {
6141 2 : // Ensure we are dumping a delta layer here
6142 4 : assert!(delta.layer_desc().is_delta);
6143 8 : delta.dump(true, &ctx).await.unwrap();
6144 2 : }
6145 2 :
6146 2 : Ok(())
6147 2 : }
6148 :
6149 : #[tokio::test]
6150 2 : async fn test_images() -> anyhow::Result<()> {
6151 4 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6152 2 : let tline = tenant
6153 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6154 5 : .await?;
6155 2 :
6156 2 : let mut writer = tline.writer().await;
6157 2 : writer
6158 2 : .put(
6159 2 : *TEST_KEY,
6160 2 : Lsn(0x10),
6161 2 : &Value::Image(test_img("foo at 0x10")),
6162 2 : &ctx,
6163 2 : )
6164 2 : .await?;
6165 2 : writer.finish_write(Lsn(0x10));
6166 2 : drop(writer);
6167 2 :
6168 2 : tline.freeze_and_flush().await?;
6169 2 : tline
6170 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6171 2 : .await?;
6172 2 :
6173 2 : let mut writer = tline.writer().await;
6174 2 : writer
6175 2 : .put(
6176 2 : *TEST_KEY,
6177 2 : Lsn(0x20),
6178 2 : &Value::Image(test_img("foo at 0x20")),
6179 2 : &ctx,
6180 2 : )
6181 2 : .await?;
6182 2 : writer.finish_write(Lsn(0x20));
6183 2 : drop(writer);
6184 2 :
6185 2 : tline.freeze_and_flush().await?;
6186 2 : tline
6187 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6188 2 : .await?;
6189 2 :
6190 2 : let mut writer = tline.writer().await;
6191 2 : writer
6192 2 : .put(
6193 2 : *TEST_KEY,
6194 2 : Lsn(0x30),
6195 2 : &Value::Image(test_img("foo at 0x30")),
6196 2 : &ctx,
6197 2 : )
6198 2 : .await?;
6199 2 : writer.finish_write(Lsn(0x30));
6200 2 : drop(writer);
6201 2 :
6202 2 : tline.freeze_and_flush().await?;
6203 2 : tline
6204 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6205 2 : .await?;
6206 2 :
6207 2 : let mut writer = tline.writer().await;
6208 2 : writer
6209 2 : .put(
6210 2 : *TEST_KEY,
6211 2 : Lsn(0x40),
6212 2 : &Value::Image(test_img("foo at 0x40")),
6213 2 : &ctx,
6214 2 : )
6215 2 : .await?;
6216 2 : writer.finish_write(Lsn(0x40));
6217 2 : drop(writer);
6218 2 :
6219 2 : tline.freeze_and_flush().await?;
6220 2 : tline
6221 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6222 2 : .await?;
6223 2 :
6224 2 : assert_eq!(
6225 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6226 2 : test_img("foo at 0x10")
6227 2 : );
6228 2 : assert_eq!(
6229 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6230 2 : test_img("foo at 0x10")
6231 2 : );
6232 2 : assert_eq!(
6233 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6234 2 : test_img("foo at 0x20")
6235 2 : );
6236 2 : assert_eq!(
6237 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6238 2 : test_img("foo at 0x30")
6239 2 : );
6240 2 : assert_eq!(
6241 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6242 2 : test_img("foo at 0x40")
6243 2 : );
6244 2 :
6245 2 : Ok(())
6246 2 : }
6247 :
6248 4 : async fn bulk_insert_compact_gc(
6249 4 : tenant: &Tenant,
6250 4 : timeline: &Arc<Timeline>,
6251 4 : ctx: &RequestContext,
6252 4 : lsn: Lsn,
6253 4 : repeat: usize,
6254 4 : key_count: usize,
6255 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6256 4 : let compact = true;
6257 40718 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6258 4 : }
6259 :
6260 8 : async fn bulk_insert_maybe_compact_gc(
6261 8 : tenant: &Tenant,
6262 8 : timeline: &Arc<Timeline>,
6263 8 : ctx: &RequestContext,
6264 8 : mut lsn: Lsn,
6265 8 : repeat: usize,
6266 8 : key_count: usize,
6267 8 : compact: bool,
6268 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6269 8 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6270 8 :
6271 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6272 8 : let mut blknum = 0;
6273 8 :
6274 8 : // Enforce that key range is monotonously increasing
6275 8 : let mut keyspace = KeySpaceAccum::new();
6276 8 :
6277 8 : let cancel = CancellationToken::new();
6278 8 :
6279 8 : for _ in 0..repeat {
6280 400 : for _ in 0..key_count {
6281 4000000 : test_key.field6 = blknum;
6282 4000000 : let mut writer = timeline.writer().await;
6283 4000000 : writer
6284 4000000 : .put(
6285 4000000 : test_key,
6286 4000000 : lsn,
6287 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6288 4000000 : ctx,
6289 4000000 : )
6290 3426 : .await?;
6291 4000000 : inserted.entry(test_key).or_default().insert(lsn);
6292 4000000 : writer.finish_write(lsn);
6293 4000000 : drop(writer);
6294 4000000 :
6295 4000000 : keyspace.add_key(test_key);
6296 4000000 :
6297 4000000 : lsn = Lsn(lsn.0 + 0x10);
6298 4000000 : blknum += 1;
6299 : }
6300 :
6301 400 : timeline.freeze_and_flush().await?;
6302 400 : if compact {
6303 : // this requires timeline to be &Arc<Timeline>
6304 8618 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
6305 200 : }
6306 :
6307 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6308 : // originally was.
6309 400 : let res = tenant
6310 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
6311 0 : .await?;
6312 :
6313 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6314 : }
6315 :
6316 8 : Ok(inserted)
6317 8 : }
6318 :
6319 : //
6320 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6321 : // Repeat 50 times.
6322 : //
6323 : #[tokio::test]
6324 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
6325 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
6326 20 : let (tenant, ctx) = harness.load().await;
6327 2 : let tline = tenant
6328 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6329 6 : .await?;
6330 2 :
6331 2 : let lsn = Lsn(0x10);
6332 20359 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6333 2 :
6334 2 : Ok(())
6335 2 : }
6336 :
6337 : // Test the vectored get real implementation against a simple sequential implementation.
6338 : //
6339 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6340 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6341 : // grow to the right on the X axis.
6342 : // [Delta]
6343 : // [Delta]
6344 : // [Delta]
6345 : // [Delta]
6346 : // ------------ Image ---------------
6347 : //
6348 : // After layer generation we pick the ranges to query as follows:
6349 : // 1. The beginning of each delta layer
6350 : // 2. At the seam between two adjacent delta layers
6351 : //
6352 : // There's one major downside to this test: delta layers only contains images,
6353 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6354 : #[tokio::test]
6355 2 : async fn test_get_vectored() -> anyhow::Result<()> {
6356 2 : let harness = TenantHarness::create("test_get_vectored").await?;
6357 20 : let (tenant, ctx) = harness.load().await;
6358 2 : let tline = tenant
6359 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6360 6 : .await?;
6361 2 :
6362 2 : let lsn = Lsn(0x10);
6363 20359 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6364 2 :
6365 2 : let guard = tline.layers.read().await;
6366 2 : let lm = guard.layer_map()?;
6367 2 :
6368 2 : lm.dump(true, &ctx).await?;
6369 2 :
6370 2 : let mut reads = Vec::new();
6371 2 : let mut prev = None;
6372 12 : lm.iter_historic_layers().for_each(|desc| {
6373 12 : if !desc.is_delta() {
6374 2 : prev = Some(desc.clone());
6375 2 : return;
6376 10 : }
6377 10 :
6378 10 : let start = desc.key_range.start;
6379 10 : let end = desc
6380 10 : .key_range
6381 10 : .start
6382 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6383 10 : reads.push(KeySpace {
6384 10 : ranges: vec![start..end],
6385 10 : });
6386 2 :
6387 10 : if let Some(prev) = &prev {
6388 10 : if !prev.is_delta() {
6389 10 : return;
6390 2 : }
6391 0 :
6392 0 : let first_range = Key {
6393 0 : field6: prev.key_range.end.field6 - 4,
6394 0 : ..prev.key_range.end
6395 0 : }..prev.key_range.end;
6396 0 :
6397 0 : let second_range = desc.key_range.start..Key {
6398 0 : field6: desc.key_range.start.field6 + 4,
6399 0 : ..desc.key_range.start
6400 0 : };
6401 0 :
6402 0 : reads.push(KeySpace {
6403 0 : ranges: vec![first_range, second_range],
6404 0 : });
6405 2 : };
6406 2 :
6407 2 : prev = Some(desc.clone());
6408 12 : });
6409 2 :
6410 2 : drop(guard);
6411 2 :
6412 2 : // Pick a big LSN such that we query over all the changes.
6413 2 : let reads_lsn = Lsn(u64::MAX - 1);
6414 2 :
6415 12 : for read in reads {
6416 10 : info!("Doing vectored read on {:?}", read);
6417 2 :
6418 10 : let vectored_res = tline
6419 10 : .get_vectored_impl(
6420 10 : read.clone(),
6421 10 : reads_lsn,
6422 10 : &mut ValuesReconstructState::new(),
6423 10 : &ctx,
6424 10 : )
6425 25 : .await;
6426 2 :
6427 10 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6428 10 : let mut expect_missing = false;
6429 10 : let mut key = read.start().unwrap();
6430 330 : while key != read.end().unwrap() {
6431 320 : if let Some(lsns) = inserted.get(&key) {
6432 320 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6433 320 : match expected_lsn {
6434 320 : Some(lsn) => {
6435 320 : expected_lsns.insert(key, *lsn);
6436 320 : }
6437 2 : None => {
6438 2 : expect_missing = true;
6439 0 : break;
6440 2 : }
6441 2 : }
6442 2 : } else {
6443 2 : expect_missing = true;
6444 0 : break;
6445 2 : }
6446 2 :
6447 320 : key = key.next();
6448 2 : }
6449 2 :
6450 10 : if expect_missing {
6451 2 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6452 2 : } else {
6453 320 : for (key, image) in vectored_res? {
6454 320 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6455 320 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6456 320 : assert_eq!(image?, expected_image);
6457 2 : }
6458 2 : }
6459 2 : }
6460 2 :
6461 2 : Ok(())
6462 2 : }
6463 :
6464 : #[tokio::test]
6465 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6466 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6467 2 :
6468 20 : let (tenant, ctx) = harness.load().await;
6469 2 : let tline = tenant
6470 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6471 2 : .await?;
6472 2 : let tline = tline.raw_timeline().unwrap();
6473 2 :
6474 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
6475 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6476 2 : modification.set_lsn(Lsn(0x1008))?;
6477 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6478 2 : modification.commit(&ctx).await?;
6479 2 :
6480 2 : let child_timeline_id = TimelineId::generate();
6481 2 : tenant
6482 2 : .branch_timeline_test(
6483 2 : tline,
6484 2 : child_timeline_id,
6485 2 : Some(tline.get_last_record_lsn()),
6486 2 : &ctx,
6487 2 : )
6488 2 : .await?;
6489 2 :
6490 2 : let child_timeline = tenant
6491 2 : .get_timeline(child_timeline_id, true)
6492 2 : .expect("Should have the branched timeline");
6493 2 :
6494 2 : let aux_keyspace = KeySpace {
6495 2 : ranges: vec![NON_INHERITED_RANGE],
6496 2 : };
6497 2 : let read_lsn = child_timeline.get_last_record_lsn();
6498 2 :
6499 2 : let vectored_res = child_timeline
6500 2 : .get_vectored_impl(
6501 2 : aux_keyspace.clone(),
6502 2 : read_lsn,
6503 2 : &mut ValuesReconstructState::new(),
6504 2 : &ctx,
6505 2 : )
6506 2 : .await;
6507 2 :
6508 2 : let images = vectored_res?;
6509 2 : assert!(images.is_empty());
6510 2 : Ok(())
6511 2 : }
6512 :
6513 : // Test that vectored get handles layer gaps correctly
6514 : // by advancing into the next ancestor timeline if required.
6515 : //
6516 : // The test generates timelines that look like the diagram below.
6517 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6518 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6519 : //
6520 : // ```
6521 : //-------------------------------+
6522 : // ... |
6523 : // [ L1 ] |
6524 : // [ / L1 ] | Child Timeline
6525 : // ... |
6526 : // ------------------------------+
6527 : // [ X L1 ] | Parent Timeline
6528 : // ------------------------------+
6529 : // ```
6530 : #[tokio::test]
6531 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6532 2 : let tenant_conf = TenantConf {
6533 2 : // Make compaction deterministic
6534 2 : gc_period: Duration::ZERO,
6535 2 : compaction_period: Duration::ZERO,
6536 2 : // Encourage creation of L1 layers
6537 2 : checkpoint_distance: 16 * 1024,
6538 2 : compaction_target_size: 8 * 1024,
6539 2 : ..TenantConf::default()
6540 2 : };
6541 2 :
6542 2 : let harness = TenantHarness::create_custom(
6543 2 : "test_get_vectored_key_gap",
6544 2 : tenant_conf,
6545 2 : TenantId::generate(),
6546 2 : ShardIdentity::unsharded(),
6547 2 : Generation::new(0xdeadbeef),
6548 2 : )
6549 2 : .await?;
6550 20 : let (tenant, ctx) = harness.load().await;
6551 2 :
6552 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6553 2 : let gap_at_key = current_key.add(100);
6554 2 : let mut current_lsn = Lsn(0x10);
6555 2 :
6556 2 : const KEY_COUNT: usize = 10_000;
6557 2 :
6558 2 : let timeline_id = TimelineId::generate();
6559 2 : let current_timeline = tenant
6560 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6561 6 : .await?;
6562 2 :
6563 2 : current_lsn += 0x100;
6564 2 :
6565 2 : let mut writer = current_timeline.writer().await;
6566 2 : writer
6567 2 : .put(
6568 2 : gap_at_key,
6569 2 : current_lsn,
6570 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6571 2 : &ctx,
6572 2 : )
6573 2 : .await?;
6574 2 : writer.finish_write(current_lsn);
6575 2 : drop(writer);
6576 2 :
6577 2 : let mut latest_lsns = HashMap::new();
6578 2 : latest_lsns.insert(gap_at_key, current_lsn);
6579 2 :
6580 2 : current_timeline.freeze_and_flush().await?;
6581 2 :
6582 2 : let child_timeline_id = TimelineId::generate();
6583 2 :
6584 2 : tenant
6585 2 : .branch_timeline_test(
6586 2 : ¤t_timeline,
6587 2 : child_timeline_id,
6588 2 : Some(current_lsn),
6589 2 : &ctx,
6590 2 : )
6591 2 : .await?;
6592 2 : let child_timeline = tenant
6593 2 : .get_timeline(child_timeline_id, true)
6594 2 : .expect("Should have the branched timeline");
6595 2 :
6596 20002 : for i in 0..KEY_COUNT {
6597 20000 : if current_key == gap_at_key {
6598 2 : current_key = current_key.next();
6599 2 : continue;
6600 19998 : }
6601 19998 :
6602 19998 : current_lsn += 0x10;
6603 2 :
6604 19998 : let mut writer = child_timeline.writer().await;
6605 19998 : writer
6606 19998 : .put(
6607 19998 : current_key,
6608 19998 : current_lsn,
6609 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6610 19998 : &ctx,
6611 19998 : )
6612 70 : .await?;
6613 19998 : writer.finish_write(current_lsn);
6614 19998 : drop(writer);
6615 19998 :
6616 19998 : latest_lsns.insert(current_key, current_lsn);
6617 19998 : current_key = current_key.next();
6618 19998 :
6619 19998 : // Flush every now and then to encourage layer file creation.
6620 19998 : if i % 500 == 0 {
6621 44 : child_timeline.freeze_and_flush().await?;
6622 19958 : }
6623 2 : }
6624 2 :
6625 2 : child_timeline.freeze_and_flush().await?;
6626 2 : let mut flags = EnumSet::new();
6627 2 : flags.insert(CompactFlags::ForceRepartition);
6628 2 : child_timeline
6629 2 : .compact(&CancellationToken::new(), flags, &ctx)
6630 1757 : .await?;
6631 2 :
6632 2 : let key_near_end = {
6633 2 : let mut tmp = current_key;
6634 2 : tmp.field6 -= 10;
6635 2 : tmp
6636 2 : };
6637 2 :
6638 2 : let key_near_gap = {
6639 2 : let mut tmp = gap_at_key;
6640 2 : tmp.field6 -= 10;
6641 2 : tmp
6642 2 : };
6643 2 :
6644 2 : let read = KeySpace {
6645 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6646 2 : };
6647 2 : let results = child_timeline
6648 2 : .get_vectored_impl(
6649 2 : read.clone(),
6650 2 : current_lsn,
6651 2 : &mut ValuesReconstructState::new(),
6652 2 : &ctx,
6653 2 : )
6654 16 : .await?;
6655 2 :
6656 44 : for (key, img_res) in results {
6657 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6658 42 : assert_eq!(img_res?, expected);
6659 2 : }
6660 2 :
6661 2 : Ok(())
6662 2 : }
6663 :
6664 : // Test that vectored get descends into ancestor timelines correctly and
6665 : // does not return an image that's newer than requested.
6666 : //
6667 : // The diagram below ilustrates an interesting case. We have a parent timeline
6668 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
6669 : // from the child timeline, so the parent timeline must be visited. When advacing into
6670 : // the child timeline, the read path needs to remember what the requested Lsn was in
6671 : // order to avoid returning an image that's too new. The test below constructs such
6672 : // a timeline setup and does a few queries around the Lsn of each page image.
6673 : // ```
6674 : // LSN
6675 : // ^
6676 : // |
6677 : // |
6678 : // 500 | --------------------------------------> branch point
6679 : // 400 | X
6680 : // 300 | X
6681 : // 200 | --------------------------------------> requested lsn
6682 : // 100 | X
6683 : // |---------------------------------------> Key
6684 : // |
6685 : // ------> requested key
6686 : //
6687 : // Legend:
6688 : // * X - page images
6689 : // ```
6690 : #[tokio::test]
6691 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
6692 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
6693 20 : let (tenant, ctx) = harness.load().await;
6694 2 :
6695 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6696 2 : let end_key = start_key.add(1000);
6697 2 : let child_gap_at_key = start_key.add(500);
6698 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
6699 2 :
6700 2 : let mut current_lsn = Lsn(0x10);
6701 2 :
6702 2 : let timeline_id = TimelineId::generate();
6703 2 : let parent_timeline = tenant
6704 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6705 6 : .await?;
6706 2 :
6707 2 : current_lsn += 0x100;
6708 2 :
6709 8 : for _ in 0..3 {
6710 6 : let mut key = start_key;
6711 6006 : while key < end_key {
6712 6000 : current_lsn += 0x10;
6713 6000 :
6714 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
6715 2 :
6716 6000 : let mut writer = parent_timeline.writer().await;
6717 6000 : writer
6718 6000 : .put(
6719 6000 : key,
6720 6000 : current_lsn,
6721 6000 : &Value::Image(test_img(&image_value)),
6722 6000 : &ctx,
6723 6000 : )
6724 6 : .await?;
6725 6000 : writer.finish_write(current_lsn);
6726 6000 :
6727 6000 : if key == child_gap_at_key {
6728 6 : parent_gap_lsns.insert(current_lsn, image_value);
6729 5994 : }
6730 2 :
6731 6000 : key = key.next();
6732 2 : }
6733 2 :
6734 6 : parent_timeline.freeze_and_flush().await?;
6735 2 : }
6736 2 :
6737 2 : let child_timeline_id = TimelineId::generate();
6738 2 :
6739 2 : let child_timeline = tenant
6740 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
6741 2 : .await?;
6742 2 :
6743 2 : let mut key = start_key;
6744 2002 : while key < end_key {
6745 2000 : if key == child_gap_at_key {
6746 2 : key = key.next();
6747 2 : continue;
6748 1998 : }
6749 1998 :
6750 1998 : current_lsn += 0x10;
6751 2 :
6752 1998 : let mut writer = child_timeline.writer().await;
6753 1998 : writer
6754 1998 : .put(
6755 1998 : key,
6756 1998 : current_lsn,
6757 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
6758 1998 : &ctx,
6759 1998 : )
6760 17 : .await?;
6761 1998 : writer.finish_write(current_lsn);
6762 1998 :
6763 1998 : key = key.next();
6764 2 : }
6765 2 :
6766 2 : child_timeline.freeze_and_flush().await?;
6767 2 :
6768 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
6769 2 : let mut query_lsns = Vec::new();
6770 6 : for image_lsn in parent_gap_lsns.keys().rev() {
6771 36 : for offset in lsn_offsets {
6772 30 : query_lsns.push(Lsn(image_lsn
6773 30 : .0
6774 30 : .checked_add_signed(offset)
6775 30 : .expect("Shouldn't overflow")));
6776 30 : }
6777 2 : }
6778 2 :
6779 32 : for query_lsn in query_lsns {
6780 30 : let results = child_timeline
6781 30 : .get_vectored_impl(
6782 30 : KeySpace {
6783 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
6784 30 : },
6785 30 : query_lsn,
6786 30 : &mut ValuesReconstructState::new(),
6787 30 : &ctx,
6788 30 : )
6789 29 : .await;
6790 2 :
6791 30 : let expected_item = parent_gap_lsns
6792 30 : .iter()
6793 30 : .rev()
6794 68 : .find(|(lsn, _)| **lsn <= query_lsn);
6795 30 :
6796 30 : info!(
6797 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
6798 2 : query_lsn, expected_item
6799 2 : );
6800 2 :
6801 30 : match expected_item {
6802 26 : Some((_, img_value)) => {
6803 26 : let key_results = results.expect("No vectored get error expected");
6804 26 : let key_result = &key_results[&child_gap_at_key];
6805 26 : let returned_img = key_result
6806 26 : .as_ref()
6807 26 : .expect("No page reconstruct error expected");
6808 26 :
6809 26 : info!(
6810 2 : "Vectored read at LSN {} returned image {}",
6811 0 : query_lsn,
6812 0 : std::str::from_utf8(returned_img)?
6813 2 : );
6814 26 : assert_eq!(*returned_img, test_img(img_value));
6815 2 : }
6816 2 : None => {
6817 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
6818 2 : }
6819 2 : }
6820 2 : }
6821 2 :
6822 2 : Ok(())
6823 2 : }
6824 :
6825 : #[tokio::test]
6826 2 : async fn test_random_updates() -> anyhow::Result<()> {
6827 2 : let names_algorithms = [
6828 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
6829 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
6830 2 : ];
6831 6 : for (name, algorithm) in names_algorithms {
6832 96212 : test_random_updates_algorithm(name, algorithm).await?;
6833 2 : }
6834 2 : Ok(())
6835 2 : }
6836 :
6837 4 : async fn test_random_updates_algorithm(
6838 4 : name: &'static str,
6839 4 : compaction_algorithm: CompactionAlgorithm,
6840 4 : ) -> anyhow::Result<()> {
6841 4 : let mut harness = TenantHarness::create(name).await?;
6842 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
6843 4 : kind: compaction_algorithm,
6844 4 : };
6845 40 : let (tenant, ctx) = harness.load().await;
6846 4 : let tline = tenant
6847 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6848 11 : .await?;
6849 :
6850 : const NUM_KEYS: usize = 1000;
6851 4 : let cancel = CancellationToken::new();
6852 4 :
6853 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6854 4 : let mut test_key_end = test_key;
6855 4 : test_key_end.field6 = NUM_KEYS as u32;
6856 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
6857 4 :
6858 4 : let mut keyspace = KeySpaceAccum::new();
6859 4 :
6860 4 : // Track when each page was last modified. Used to assert that
6861 4 : // a read sees the latest page version.
6862 4 : let mut updated = [Lsn(0); NUM_KEYS];
6863 4 :
6864 4 : let mut lsn = Lsn(0x10);
6865 : #[allow(clippy::needless_range_loop)]
6866 4004 : for blknum in 0..NUM_KEYS {
6867 4000 : lsn = Lsn(lsn.0 + 0x10);
6868 4000 : test_key.field6 = blknum as u32;
6869 4000 : let mut writer = tline.writer().await;
6870 4000 : writer
6871 4000 : .put(
6872 4000 : test_key,
6873 4000 : lsn,
6874 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6875 4000 : &ctx,
6876 4000 : )
6877 4 : .await?;
6878 4000 : writer.finish_write(lsn);
6879 4000 : updated[blknum] = lsn;
6880 4000 : drop(writer);
6881 4000 :
6882 4000 : keyspace.add_key(test_key);
6883 : }
6884 :
6885 204 : for _ in 0..50 {
6886 200200 : for _ in 0..NUM_KEYS {
6887 200000 : lsn = Lsn(lsn.0 + 0x10);
6888 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6889 200000 : test_key.field6 = blknum as u32;
6890 200000 : let mut writer = tline.writer().await;
6891 200000 : writer
6892 200000 : .put(
6893 200000 : test_key,
6894 200000 : lsn,
6895 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6896 200000 : &ctx,
6897 200000 : )
6898 198 : .await?;
6899 200000 : writer.finish_write(lsn);
6900 200000 : drop(writer);
6901 200000 : updated[blknum] = lsn;
6902 : }
6903 :
6904 : // Read all the blocks
6905 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6906 200000 : test_key.field6 = blknum as u32;
6907 200000 : assert_eq!(
6908 200000 : tline.get(test_key, lsn, &ctx).await?,
6909 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
6910 : );
6911 : }
6912 :
6913 : // Perform a cycle of flush, and GC
6914 203 : tline.freeze_and_flush().await?;
6915 200 : tenant
6916 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
6917 0 : .await?;
6918 : }
6919 :
6920 4 : Ok(())
6921 4 : }
6922 :
6923 : #[tokio::test]
6924 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
6925 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
6926 2 : .await?
6927 2 : .load()
6928 15 : .await;
6929 2 : let mut tline = tenant
6930 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6931 5 : .await?;
6932 2 :
6933 2 : const NUM_KEYS: usize = 1000;
6934 2 :
6935 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6936 2 :
6937 2 : let mut keyspace = KeySpaceAccum::new();
6938 2 :
6939 2 : let cancel = CancellationToken::new();
6940 2 :
6941 2 : // Track when each page was last modified. Used to assert that
6942 2 : // a read sees the latest page version.
6943 2 : let mut updated = [Lsn(0); NUM_KEYS];
6944 2 :
6945 2 : let mut lsn = Lsn(0x10);
6946 2 : #[allow(clippy::needless_range_loop)]
6947 2002 : for blknum in 0..NUM_KEYS {
6948 2000 : lsn = Lsn(lsn.0 + 0x10);
6949 2000 : test_key.field6 = blknum as u32;
6950 2000 : let mut writer = tline.writer().await;
6951 2000 : writer
6952 2000 : .put(
6953 2000 : test_key,
6954 2000 : lsn,
6955 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6956 2000 : &ctx,
6957 2000 : )
6958 2 : .await?;
6959 2000 : writer.finish_write(lsn);
6960 2000 : updated[blknum] = lsn;
6961 2000 : drop(writer);
6962 2000 :
6963 2000 : keyspace.add_key(test_key);
6964 2 : }
6965 2 :
6966 102 : for _ in 0..50 {
6967 100 : let new_tline_id = TimelineId::generate();
6968 100 : tenant
6969 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
6970 92 : .await?;
6971 100 : tline = tenant
6972 100 : .get_timeline(new_tline_id, true)
6973 100 : .expect("Should have the branched timeline");
6974 2 :
6975 100100 : for _ in 0..NUM_KEYS {
6976 100000 : lsn = Lsn(lsn.0 + 0x10);
6977 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6978 100000 : test_key.field6 = blknum as u32;
6979 100000 : let mut writer = tline.writer().await;
6980 100000 : writer
6981 100000 : .put(
6982 100000 : test_key,
6983 100000 : lsn,
6984 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6985 100000 : &ctx,
6986 100000 : )
6987 835 : .await?;
6988 100000 : println!("updating {} at {}", blknum, lsn);
6989 100000 : writer.finish_write(lsn);
6990 100000 : drop(writer);
6991 100000 : updated[blknum] = lsn;
6992 2 : }
6993 2 :
6994 2 : // Read all the blocks
6995 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
6996 100000 : test_key.field6 = blknum as u32;
6997 100000 : assert_eq!(
6998 100000 : tline.get(test_key, lsn, &ctx).await?,
6999 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
7000 2 : );
7001 2 : }
7002 2 :
7003 2 : // Perform a cycle of flush, compact, and GC
7004 102 : tline.freeze_and_flush().await?;
7005 15029 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7006 100 : tenant
7007 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7008 2 : .await?;
7009 2 : }
7010 2 :
7011 2 : Ok(())
7012 2 : }
7013 :
7014 : #[tokio::test]
7015 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7016 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7017 2 : .await?
7018 2 : .load()
7019 20 : .await;
7020 2 : let mut tline = tenant
7021 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7022 6 : .await?;
7023 2 :
7024 2 : const NUM_KEYS: usize = 100;
7025 2 : const NUM_TLINES: usize = 50;
7026 2 :
7027 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7028 2 : // Track page mutation lsns across different timelines.
7029 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7030 2 :
7031 2 : let mut lsn = Lsn(0x10);
7032 2 :
7033 2 : #[allow(clippy::needless_range_loop)]
7034 102 : for idx in 0..NUM_TLINES {
7035 100 : let new_tline_id = TimelineId::generate();
7036 100 : tenant
7037 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7038 70 : .await?;
7039 100 : tline = tenant
7040 100 : .get_timeline(new_tline_id, true)
7041 100 : .expect("Should have the branched timeline");
7042 2 :
7043 10100 : for _ in 0..NUM_KEYS {
7044 10000 : lsn = Lsn(lsn.0 + 0x10);
7045 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7046 10000 : test_key.field6 = blknum as u32;
7047 10000 : let mut writer = tline.writer().await;
7048 10000 : writer
7049 10000 : .put(
7050 10000 : test_key,
7051 10000 : lsn,
7052 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7053 10000 : &ctx,
7054 10000 : )
7055 107 : .await?;
7056 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7057 10000 : writer.finish_write(lsn);
7058 10000 : drop(writer);
7059 10000 : updated[idx][blknum] = lsn;
7060 2 : }
7061 2 : }
7062 2 :
7063 2 : // Read pages from leaf timeline across all ancestors.
7064 100 : for (idx, lsns) in updated.iter().enumerate() {
7065 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
7066 2 : // Skip empty mutations.
7067 10000 : if lsn.0 == 0 {
7068 3673 : continue;
7069 6327 : }
7070 6327 : println!("checking [{idx}][{blknum}] at {lsn}");
7071 6327 : test_key.field6 = blknum as u32;
7072 6327 : assert_eq!(
7073 6327 : tline.get(test_key, *lsn, &ctx).await?,
7074 6327 : test_img(&format!("{idx} {blknum} at {lsn}"))
7075 2 : );
7076 2 : }
7077 2 : }
7078 2 : Ok(())
7079 2 : }
7080 :
7081 : #[tokio::test]
7082 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7083 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7084 2 : .await?
7085 2 : .load()
7086 20 : .await;
7087 2 :
7088 2 : let initdb_lsn = Lsn(0x20);
7089 2 : let utline = tenant
7090 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7091 2 : .await?;
7092 2 : let tline = utline.raw_timeline().unwrap();
7093 2 :
7094 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7095 2 : tline.maybe_spawn_flush_loop();
7096 2 :
7097 2 : // Make sure the timeline has the minimum set of required keys for operation.
7098 2 : // The only operation you can always do on an empty timeline is to `put` new data.
7099 2 : // Except if you `put` at `initdb_lsn`.
7100 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7101 2 : // It uses `repartition()`, which assumes some keys to be present.
7102 2 : // Let's make sure the test timeline can handle that case.
7103 2 : {
7104 2 : let mut state = tline.flush_loop_state.lock().unwrap();
7105 2 : assert_eq!(
7106 2 : timeline::FlushLoopState::Running {
7107 2 : expect_initdb_optimization: false,
7108 2 : initdb_optimization_count: 0,
7109 2 : },
7110 2 : *state
7111 2 : );
7112 2 : *state = timeline::FlushLoopState::Running {
7113 2 : expect_initdb_optimization: true,
7114 2 : initdb_optimization_count: 0,
7115 2 : };
7116 2 : }
7117 2 :
7118 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7119 2 : // As explained above, the optimization requires some keys to be present.
7120 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7121 2 : // This is what `create_test_timeline` does, by the way.
7122 2 : let mut modification = tline.begin_modification(initdb_lsn);
7123 2 : modification
7124 2 : .init_empty_test_timeline()
7125 2 : .context("init_empty_test_timeline")?;
7126 2 : modification
7127 2 : .commit(&ctx)
7128 2 : .await
7129 2 : .context("commit init_empty_test_timeline modification")?;
7130 2 :
7131 2 : // Do the flush. The flush code will check the expectations that we set above.
7132 2 : tline.freeze_and_flush().await?;
7133 2 :
7134 2 : // assert freeze_and_flush exercised the initdb optimization
7135 2 : {
7136 2 : let state = tline.flush_loop_state.lock().unwrap();
7137 2 : let timeline::FlushLoopState::Running {
7138 2 : expect_initdb_optimization,
7139 2 : initdb_optimization_count,
7140 2 : } = *state
7141 2 : else {
7142 2 : panic!("unexpected state: {:?}", *state);
7143 2 : };
7144 2 : assert!(expect_initdb_optimization);
7145 2 : assert!(initdb_optimization_count > 0);
7146 2 : }
7147 2 : Ok(())
7148 2 : }
7149 :
7150 : #[tokio::test]
7151 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7152 2 : let name = "test_create_guard_crash";
7153 2 : let harness = TenantHarness::create(name).await?;
7154 2 : {
7155 20 : let (tenant, ctx) = harness.load().await;
7156 2 : let tline = tenant
7157 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7158 2 : .await?;
7159 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
7160 2 : let raw_tline = tline.raw_timeline().unwrap();
7161 2 : raw_tline
7162 2 : .shutdown(super::timeline::ShutdownMode::Hard)
7163 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))
7164 2 : .await;
7165 2 : std::mem::forget(tline);
7166 2 : }
7167 2 :
7168 20 : let (tenant, _) = harness.load().await;
7169 2 : match tenant.get_timeline(TIMELINE_ID, false) {
7170 2 : Ok(_) => panic!("timeline should've been removed during load"),
7171 2 : Err(e) => {
7172 2 : assert_eq!(
7173 2 : e,
7174 2 : GetTimelineError::NotFound {
7175 2 : tenant_id: tenant.tenant_shard_id,
7176 2 : timeline_id: TIMELINE_ID,
7177 2 : }
7178 2 : )
7179 2 : }
7180 2 : }
7181 2 :
7182 2 : assert!(!harness
7183 2 : .conf
7184 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7185 2 : .exists());
7186 2 :
7187 2 : Ok(())
7188 2 : }
7189 :
7190 : #[tokio::test]
7191 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7192 2 : let names_algorithms = [
7193 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7194 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7195 2 : ];
7196 6 : for (name, algorithm) in names_algorithms {
7197 32561 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7198 2 : }
7199 2 : Ok(())
7200 2 : }
7201 :
7202 4 : async fn test_read_at_max_lsn_algorithm(
7203 4 : name: &'static str,
7204 4 : compaction_algorithm: CompactionAlgorithm,
7205 4 : ) -> anyhow::Result<()> {
7206 4 : let mut harness = TenantHarness::create(name).await?;
7207 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7208 4 : kind: compaction_algorithm,
7209 4 : };
7210 40 : let (tenant, ctx) = harness.load().await;
7211 4 : let tline = tenant
7212 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7213 11 : .await?;
7214 :
7215 4 : let lsn = Lsn(0x10);
7216 4 : let compact = false;
7217 32100 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7218 :
7219 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7220 4 : let read_lsn = Lsn(u64::MAX - 1);
7221 :
7222 410 : let result = tline.get(test_key, read_lsn, &ctx).await;
7223 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7224 :
7225 4 : Ok(())
7226 4 : }
7227 :
7228 : #[tokio::test]
7229 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
7230 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
7231 20 : let (tenant, ctx) = harness.load().await;
7232 2 : let tline = tenant
7233 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7234 6 : .await?;
7235 2 :
7236 2 : const NUM_KEYS: usize = 1000;
7237 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7238 2 :
7239 2 : let cancel = CancellationToken::new();
7240 2 :
7241 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7242 2 : base_key.field1 = AUX_KEY_PREFIX;
7243 2 : let mut test_key = base_key;
7244 2 :
7245 2 : // Track when each page was last modified. Used to assert that
7246 2 : // a read sees the latest page version.
7247 2 : let mut updated = [Lsn(0); NUM_KEYS];
7248 2 :
7249 2 : let mut lsn = Lsn(0x10);
7250 2 : #[allow(clippy::needless_range_loop)]
7251 2002 : for blknum in 0..NUM_KEYS {
7252 2000 : lsn = Lsn(lsn.0 + 0x10);
7253 2000 : test_key.field6 = (blknum * STEP) as u32;
7254 2000 : let mut writer = tline.writer().await;
7255 2000 : writer
7256 2000 : .put(
7257 2000 : test_key,
7258 2000 : lsn,
7259 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7260 2000 : &ctx,
7261 2000 : )
7262 2 : .await?;
7263 2000 : writer.finish_write(lsn);
7264 2000 : updated[blknum] = lsn;
7265 2000 : drop(writer);
7266 2 : }
7267 2 :
7268 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7269 2 :
7270 24 : for iter in 0..=10 {
7271 2 : // Read all the blocks
7272 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7273 22000 : test_key.field6 = (blknum * STEP) as u32;
7274 22000 : assert_eq!(
7275 22000 : tline.get(test_key, lsn, &ctx).await?,
7276 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
7277 2 : );
7278 2 : }
7279 2 :
7280 22 : let mut cnt = 0;
7281 22000 : for (key, value) in tline
7282 22 : .get_vectored_impl(
7283 22 : keyspace.clone(),
7284 22 : lsn,
7285 22 : &mut ValuesReconstructState::default(),
7286 22 : &ctx,
7287 22 : )
7288 770 : .await?
7289 2 : {
7290 22000 : let blknum = key.field6 as usize;
7291 22000 : let value = value?;
7292 22000 : assert!(blknum % STEP == 0);
7293 22000 : let blknum = blknum / STEP;
7294 22000 : assert_eq!(
7295 22000 : value,
7296 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
7297 22000 : );
7298 22000 : cnt += 1;
7299 2 : }
7300 2 :
7301 22 : assert_eq!(cnt, NUM_KEYS);
7302 2 :
7303 22022 : for _ in 0..NUM_KEYS {
7304 22000 : lsn = Lsn(lsn.0 + 0x10);
7305 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7306 22000 : test_key.field6 = (blknum * STEP) as u32;
7307 22000 : let mut writer = tline.writer().await;
7308 22000 : writer
7309 22000 : .put(
7310 22000 : test_key,
7311 22000 : lsn,
7312 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7313 22000 : &ctx,
7314 22000 : )
7315 111 : .await?;
7316 22000 : writer.finish_write(lsn);
7317 22000 : drop(writer);
7318 22000 : updated[blknum] = lsn;
7319 2 : }
7320 2 :
7321 2 : // Perform two cycles of flush, compact, and GC
7322 66 : for round in 0..2 {
7323 44 : tline.freeze_and_flush().await?;
7324 44 : tline
7325 44 : .compact(
7326 44 : &cancel,
7327 44 : if iter % 5 == 0 && round == 0 {
7328 6 : let mut flags = EnumSet::new();
7329 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
7330 6 : flags.insert(CompactFlags::ForceRepartition);
7331 6 : flags
7332 2 : } else {
7333 38 : EnumSet::empty()
7334 2 : },
7335 44 : &ctx,
7336 2 : )
7337 6689 : .await?;
7338 44 : tenant
7339 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7340 2 : .await?;
7341 2 : }
7342 2 : }
7343 2 :
7344 2 : Ok(())
7345 2 : }
7346 :
7347 : #[tokio::test]
7348 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7349 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7350 20 : let (tenant, ctx) = harness.load().await;
7351 2 : let tline = tenant
7352 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7353 6 : .await?;
7354 2 :
7355 2 : let cancel = CancellationToken::new();
7356 2 :
7357 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7358 2 : base_key.field1 = AUX_KEY_PREFIX;
7359 2 : let test_key = base_key;
7360 2 : let mut lsn = Lsn(0x10);
7361 2 :
7362 42 : for _ in 0..20 {
7363 40 : lsn = Lsn(lsn.0 + 0x10);
7364 40 : let mut writer = tline.writer().await;
7365 40 : writer
7366 40 : .put(
7367 40 : test_key,
7368 40 : lsn,
7369 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7370 40 : &ctx,
7371 40 : )
7372 20 : .await?;
7373 40 : writer.finish_write(lsn);
7374 40 : drop(writer);
7375 41 : tline.freeze_and_flush().await?; // force create a delta layer
7376 2 : }
7377 2 :
7378 2 : let before_num_l0_delta_files =
7379 2 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7380 2 :
7381 110 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7382 2 :
7383 2 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7384 2 :
7385 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}");
7386 2 :
7387 2 : assert_eq!(
7388 4 : tline.get(test_key, lsn, &ctx).await?,
7389 2 : test_img(&format!("{} at {}", 0, lsn))
7390 2 : );
7391 2 :
7392 2 : Ok(())
7393 2 : }
7394 :
7395 : #[tokio::test]
7396 2 : async fn test_aux_file_e2e() {
7397 2 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7398 2 :
7399 20 : let (tenant, ctx) = harness.load().await;
7400 2 :
7401 2 : let mut lsn = Lsn(0x08);
7402 2 :
7403 2 : let tline: Arc<Timeline> = tenant
7404 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7405 6 : .await
7406 2 : .unwrap();
7407 2 :
7408 2 : {
7409 2 : lsn += 8;
7410 2 : let mut modification = tline.begin_modification(lsn);
7411 2 : modification
7412 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7413 2 : .await
7414 2 : .unwrap();
7415 2 : modification.commit(&ctx).await.unwrap();
7416 2 : }
7417 2 :
7418 2 : // we can read everything from the storage
7419 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7420 2 : assert_eq!(
7421 2 : files.get("pg_logical/mappings/test1"),
7422 2 : Some(&bytes::Bytes::from_static(b"first"))
7423 2 : );
7424 2 :
7425 2 : {
7426 2 : lsn += 8;
7427 2 : let mut modification = tline.begin_modification(lsn);
7428 2 : modification
7429 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7430 2 : .await
7431 2 : .unwrap();
7432 2 : modification.commit(&ctx).await.unwrap();
7433 2 : }
7434 2 :
7435 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
7436 2 : assert_eq!(
7437 2 : files.get("pg_logical/mappings/test2"),
7438 2 : Some(&bytes::Bytes::from_static(b"second"))
7439 2 : );
7440 2 :
7441 2 : let child = tenant
7442 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7443 2 : .await
7444 2 : .unwrap();
7445 2 :
7446 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
7447 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7448 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7449 2 : }
7450 :
7451 : #[tokio::test]
7452 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7453 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7454 20 : let (tenant, ctx) = harness.load().await;
7455 2 : let tline = tenant
7456 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7457 6 : .await?;
7458 2 :
7459 2 : const NUM_KEYS: usize = 1000;
7460 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7461 2 :
7462 2 : let cancel = CancellationToken::new();
7463 2 :
7464 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7465 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7466 2 : let mut test_key = base_key;
7467 2 : let mut lsn = Lsn(0x10);
7468 2 :
7469 8 : async fn scan_with_statistics(
7470 8 : tline: &Timeline,
7471 8 : keyspace: &KeySpace,
7472 8 : lsn: Lsn,
7473 8 : ctx: &RequestContext,
7474 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7475 8 : let mut reconstruct_state = ValuesReconstructState::default();
7476 8 : let res = tline
7477 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7478 239 : .await?;
7479 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7480 8 : }
7481 2 :
7482 2 : #[allow(clippy::needless_range_loop)]
7483 2002 : for blknum in 0..NUM_KEYS {
7484 2000 : lsn = Lsn(lsn.0 + 0x10);
7485 2000 : test_key.field6 = (blknum * STEP) as u32;
7486 2000 : let mut writer = tline.writer().await;
7487 2000 : writer
7488 2000 : .put(
7489 2000 : test_key,
7490 2000 : lsn,
7491 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7492 2000 : &ctx,
7493 2000 : )
7494 2 : .await?;
7495 2000 : writer.finish_write(lsn);
7496 2000 : drop(writer);
7497 2 : }
7498 2 :
7499 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7500 2 :
7501 22 : for iter in 1..=10 {
7502 20020 : for _ in 0..NUM_KEYS {
7503 20000 : lsn = Lsn(lsn.0 + 0x10);
7504 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7505 20000 : test_key.field6 = (blknum * STEP) as u32;
7506 20000 : let mut writer = tline.writer().await;
7507 20000 : writer
7508 20000 : .put(
7509 20000 : test_key,
7510 20000 : lsn,
7511 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7512 20000 : &ctx,
7513 20000 : )
7514 35 : .await?;
7515 20000 : writer.finish_write(lsn);
7516 20000 : drop(writer);
7517 2 : }
7518 2 :
7519 20 : tline.freeze_and_flush().await?;
7520 2 :
7521 20 : if iter % 5 == 0 {
7522 4 : let (_, before_delta_file_accessed) =
7523 231 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7524 4 : tline
7525 4 : .compact(
7526 4 : &cancel,
7527 4 : {
7528 4 : let mut flags = EnumSet::new();
7529 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
7530 4 : flags.insert(CompactFlags::ForceRepartition);
7531 4 : flags
7532 4 : },
7533 4 : &ctx,
7534 4 : )
7535 4842 : .await?;
7536 4 : let (_, after_delta_file_accessed) =
7537 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
7538 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}");
7539 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.
7540 4 : assert!(
7541 4 : after_delta_file_accessed <= 2,
7542 2 : "after_delta_file_accessed={after_delta_file_accessed}"
7543 2 : );
7544 16 : }
7545 2 : }
7546 2 :
7547 2 : Ok(())
7548 2 : }
7549 :
7550 : #[tokio::test]
7551 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7552 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7553 16 : let (tenant, ctx) = harness.load().await;
7554 2 :
7555 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7556 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7557 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7558 2 :
7559 2 : let tline = tenant
7560 2 : .create_test_timeline_with_layers(
7561 2 : TIMELINE_ID,
7562 2 : Lsn(0x10),
7563 2 : DEFAULT_PG_VERSION,
7564 2 : &ctx,
7565 2 : Vec::new(), // delta layers
7566 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7567 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
7568 2 : )
7569 13 : .await?;
7570 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7571 2 :
7572 2 : let child = tenant
7573 2 : .branch_timeline_test_with_layers(
7574 2 : &tline,
7575 2 : NEW_TIMELINE_ID,
7576 2 : Some(Lsn(0x20)),
7577 2 : &ctx,
7578 2 : Vec::new(), // delta layers
7579 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7580 2 : Lsn(0x30),
7581 2 : )
7582 9 : .await
7583 2 : .unwrap();
7584 2 :
7585 2 : let lsn = Lsn(0x30);
7586 2 :
7587 2 : // test vectored get on parent timeline
7588 2 : assert_eq!(
7589 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7590 2 : Some(test_img("data key 1"))
7591 2 : );
7592 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7593 3 : .await
7594 2 : .unwrap_err()
7595 2 : .is_missing_key_error());
7596 2 : assert!(
7597 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7598 2 : .await
7599 2 : .unwrap_err()
7600 2 : .is_missing_key_error()
7601 2 : );
7602 2 :
7603 2 : // test vectored get on child timeline
7604 2 : assert_eq!(
7605 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7606 2 : Some(test_img("data key 1"))
7607 2 : );
7608 2 : assert_eq!(
7609 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7610 2 : Some(test_img("data key 2"))
7611 2 : );
7612 2 : assert!(
7613 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7614 2 : .await
7615 2 : .unwrap_err()
7616 2 : .is_missing_key_error()
7617 2 : );
7618 2 :
7619 2 : Ok(())
7620 2 : }
7621 :
7622 : #[tokio::test]
7623 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7624 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7625 20 : let (tenant, ctx) = harness.load().await;
7626 2 :
7627 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7628 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7629 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7630 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7631 2 :
7632 2 : let tline = tenant
7633 2 : .create_test_timeline_with_layers(
7634 2 : TIMELINE_ID,
7635 2 : Lsn(0x10),
7636 2 : DEFAULT_PG_VERSION,
7637 2 : &ctx,
7638 2 : Vec::new(), // delta layers
7639 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
7640 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
7641 2 : )
7642 13 : .await?;
7643 2 :
7644 2 : let child = tenant
7645 2 : .branch_timeline_test_with_layers(
7646 2 : &tline,
7647 2 : NEW_TIMELINE_ID,
7648 2 : Some(Lsn(0x20)),
7649 2 : &ctx,
7650 2 : Vec::new(), // delta layers
7651 2 : vec![(
7652 2 : Lsn(0x30),
7653 2 : vec![(base_key_child, test_img("metadata key 2"))],
7654 2 : )], // image layers
7655 2 : Lsn(0x30),
7656 2 : )
7657 8 : .await
7658 2 : .unwrap();
7659 2 :
7660 2 : let lsn = Lsn(0x30);
7661 2 :
7662 2 : // test vectored get on parent timeline
7663 2 : assert_eq!(
7664 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7665 2 : Some(test_img("metadata key 1"))
7666 2 : );
7667 2 : assert_eq!(
7668 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
7669 2 : None
7670 2 : );
7671 2 : assert_eq!(
7672 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
7673 2 : None
7674 2 : );
7675 2 :
7676 2 : // test vectored get on child timeline
7677 2 : assert_eq!(
7678 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7679 2 : None
7680 2 : );
7681 2 : assert_eq!(
7682 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7683 2 : Some(test_img("metadata key 2"))
7684 2 : );
7685 2 : assert_eq!(
7686 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
7687 2 : None
7688 2 : );
7689 2 :
7690 2 : Ok(())
7691 2 : }
7692 :
7693 36 : async fn get_vectored_impl_wrapper(
7694 36 : tline: &Arc<Timeline>,
7695 36 : key: Key,
7696 36 : lsn: Lsn,
7697 36 : ctx: &RequestContext,
7698 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
7699 36 : let mut reconstruct_state = ValuesReconstructState::new();
7700 36 : let mut res = tline
7701 36 : .get_vectored_impl(
7702 36 : KeySpace::single(key..key.next()),
7703 36 : lsn,
7704 36 : &mut reconstruct_state,
7705 36 : ctx,
7706 36 : )
7707 40 : .await?;
7708 30 : Ok(res.pop_last().map(|(k, v)| {
7709 18 : assert_eq!(k, key);
7710 18 : v.unwrap()
7711 30 : }))
7712 36 : }
7713 :
7714 : #[tokio::test]
7715 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
7716 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
7717 20 : let (tenant, ctx) = harness.load().await;
7718 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7719 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7720 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7721 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7722 2 :
7723 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
7724 2 : // Lsn 0x30 key0, key3, no key1+key2
7725 2 : // Lsn 0x20 key1+key2 tomestones
7726 2 : // Lsn 0x10 key1 in image, key2 in delta
7727 2 : let tline = tenant
7728 2 : .create_test_timeline_with_layers(
7729 2 : TIMELINE_ID,
7730 2 : Lsn(0x10),
7731 2 : DEFAULT_PG_VERSION,
7732 2 : &ctx,
7733 2 : // delta layers
7734 2 : vec![
7735 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7736 2 : Lsn(0x10)..Lsn(0x20),
7737 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7738 2 : ),
7739 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7740 2 : Lsn(0x20)..Lsn(0x30),
7741 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7742 2 : ),
7743 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7744 2 : Lsn(0x20)..Lsn(0x30),
7745 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7746 2 : ),
7747 2 : ],
7748 2 : // image layers
7749 2 : vec![
7750 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
7751 2 : (
7752 2 : Lsn(0x30),
7753 2 : vec![
7754 2 : (key0, test_img("metadata key 0")),
7755 2 : (key3, test_img("metadata key 3")),
7756 2 : ],
7757 2 : ),
7758 2 : ],
7759 2 : Lsn(0x30),
7760 2 : )
7761 40 : .await?;
7762 2 :
7763 2 : let lsn = Lsn(0x30);
7764 2 : let old_lsn = Lsn(0x20);
7765 2 :
7766 2 : assert_eq!(
7767 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
7768 2 : Some(test_img("metadata key 0"))
7769 2 : );
7770 2 : assert_eq!(
7771 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
7772 2 : None,
7773 2 : );
7774 2 : assert_eq!(
7775 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
7776 2 : None,
7777 2 : );
7778 2 : assert_eq!(
7779 8 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
7780 2 : Some(Bytes::new()),
7781 2 : );
7782 2 : assert_eq!(
7783 7 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
7784 2 : Some(Bytes::new()),
7785 2 : );
7786 2 : assert_eq!(
7787 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
7788 2 : Some(test_img("metadata key 3"))
7789 2 : );
7790 2 :
7791 2 : Ok(())
7792 2 : }
7793 :
7794 : #[tokio::test]
7795 2 : async fn test_metadata_tombstone_image_creation() {
7796 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
7797 2 : .await
7798 2 : .unwrap();
7799 20 : let (tenant, ctx) = harness.load().await;
7800 2 :
7801 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
7802 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7803 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7804 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
7805 2 :
7806 2 : let tline = tenant
7807 2 : .create_test_timeline_with_layers(
7808 2 : TIMELINE_ID,
7809 2 : Lsn(0x10),
7810 2 : DEFAULT_PG_VERSION,
7811 2 : &ctx,
7812 2 : // delta layers
7813 2 : vec![
7814 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7815 2 : Lsn(0x10)..Lsn(0x20),
7816 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7817 2 : ),
7818 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7819 2 : Lsn(0x20)..Lsn(0x30),
7820 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7821 2 : ),
7822 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7823 2 : Lsn(0x20)..Lsn(0x30),
7824 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7825 2 : ),
7826 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7827 2 : Lsn(0x30)..Lsn(0x40),
7828 2 : vec![
7829 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
7830 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
7831 2 : ],
7832 2 : ),
7833 2 : ],
7834 2 : // image layers
7835 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7836 2 : Lsn(0x40),
7837 2 : )
7838 37 : .await
7839 2 : .unwrap();
7840 2 :
7841 2 : let cancel = CancellationToken::new();
7842 2 :
7843 2 : tline
7844 2 : .compact(
7845 2 : &cancel,
7846 2 : {
7847 2 : let mut flags = EnumSet::new();
7848 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7849 2 : flags.insert(CompactFlags::ForceRepartition);
7850 2 : flags
7851 2 : },
7852 2 : &ctx,
7853 2 : )
7854 62 : .await
7855 2 : .unwrap();
7856 2 :
7857 2 : // Image layers are created at last_record_lsn
7858 2 : let images = tline
7859 2 : .inspect_image_layers(Lsn(0x40), &ctx)
7860 8 : .await
7861 2 : .unwrap()
7862 2 : .into_iter()
7863 18 : .filter(|(k, _)| k.is_metadata_key())
7864 2 : .collect::<Vec<_>>();
7865 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
7866 2 : }
7867 :
7868 : #[tokio::test]
7869 2 : async fn test_metadata_tombstone_empty_image_creation() {
7870 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
7871 2 : .await
7872 2 : .unwrap();
7873 20 : let (tenant, ctx) = harness.load().await;
7874 2 :
7875 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
7876 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
7877 2 :
7878 2 : let tline = tenant
7879 2 : .create_test_timeline_with_layers(
7880 2 : TIMELINE_ID,
7881 2 : Lsn(0x10),
7882 2 : DEFAULT_PG_VERSION,
7883 2 : &ctx,
7884 2 : // delta layers
7885 2 : vec![
7886 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7887 2 : Lsn(0x10)..Lsn(0x20),
7888 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
7889 2 : ),
7890 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7891 2 : Lsn(0x20)..Lsn(0x30),
7892 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
7893 2 : ),
7894 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
7895 2 : Lsn(0x20)..Lsn(0x30),
7896 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
7897 2 : ),
7898 2 : ],
7899 2 : // image layers
7900 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
7901 2 : Lsn(0x30),
7902 2 : )
7903 31 : .await
7904 2 : .unwrap();
7905 2 :
7906 2 : let cancel = CancellationToken::new();
7907 2 :
7908 2 : tline
7909 2 : .compact(
7910 2 : &cancel,
7911 2 : {
7912 2 : let mut flags = EnumSet::new();
7913 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
7914 2 : flags.insert(CompactFlags::ForceRepartition);
7915 2 : flags
7916 2 : },
7917 2 : &ctx,
7918 2 : )
7919 50 : .await
7920 2 : .unwrap();
7921 2 :
7922 2 : // Image layers are created at last_record_lsn
7923 2 : let images = tline
7924 2 : .inspect_image_layers(Lsn(0x30), &ctx)
7925 4 : .await
7926 2 : .unwrap()
7927 2 : .into_iter()
7928 14 : .filter(|(k, _)| k.is_metadata_key())
7929 2 : .collect::<Vec<_>>();
7930 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
7931 2 : }
7932 :
7933 : #[tokio::test]
7934 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
7935 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
7936 20 : let (tenant, ctx) = harness.load().await;
7937 2 :
7938 102 : fn get_key(id: u32) -> Key {
7939 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7940 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7941 102 : key.field6 = id;
7942 102 : key
7943 102 : }
7944 2 :
7945 2 : // We create
7946 2 : // - one bottom-most image layer,
7947 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7948 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7949 2 : // - a delta layer D3 above the horizon.
7950 2 : //
7951 2 : // | D3 |
7952 2 : // | D1 |
7953 2 : // -| |-- gc horizon -----------------
7954 2 : // | | | D2 |
7955 2 : // --------- img layer ------------------
7956 2 : //
7957 2 : // What we should expact from this compaction is:
7958 2 : // | D3 |
7959 2 : // | Part of D1 |
7960 2 : // --------- img layer with D1+D2 at GC horizon------------------
7961 2 :
7962 2 : // img layer at 0x10
7963 2 : let img_layer = (0..10)
7964 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7965 2 : .collect_vec();
7966 2 :
7967 2 : let delta1 = vec![
7968 2 : (
7969 2 : get_key(1),
7970 2 : Lsn(0x20),
7971 2 : Value::Image(Bytes::from("value 1@0x20")),
7972 2 : ),
7973 2 : (
7974 2 : get_key(2),
7975 2 : Lsn(0x30),
7976 2 : Value::Image(Bytes::from("value 2@0x30")),
7977 2 : ),
7978 2 : (
7979 2 : get_key(3),
7980 2 : Lsn(0x40),
7981 2 : Value::Image(Bytes::from("value 3@0x40")),
7982 2 : ),
7983 2 : ];
7984 2 : let delta2 = vec![
7985 2 : (
7986 2 : get_key(5),
7987 2 : Lsn(0x20),
7988 2 : Value::Image(Bytes::from("value 5@0x20")),
7989 2 : ),
7990 2 : (
7991 2 : get_key(6),
7992 2 : Lsn(0x20),
7993 2 : Value::Image(Bytes::from("value 6@0x20")),
7994 2 : ),
7995 2 : ];
7996 2 : let delta3 = vec![
7997 2 : (
7998 2 : get_key(8),
7999 2 : Lsn(0x48),
8000 2 : Value::Image(Bytes::from("value 8@0x48")),
8001 2 : ),
8002 2 : (
8003 2 : get_key(9),
8004 2 : Lsn(0x48),
8005 2 : Value::Image(Bytes::from("value 9@0x48")),
8006 2 : ),
8007 2 : ];
8008 2 :
8009 2 : let tline = tenant
8010 2 : .create_test_timeline_with_layers(
8011 2 : TIMELINE_ID,
8012 2 : Lsn(0x10),
8013 2 : DEFAULT_PG_VERSION,
8014 2 : &ctx,
8015 2 : vec![
8016 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8017 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8018 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8019 2 : ], // delta layers
8020 2 : vec![(Lsn(0x10), img_layer)], // image layers
8021 2 : Lsn(0x50),
8022 2 : )
8023 49 : .await?;
8024 2 : {
8025 2 : // Update GC info
8026 2 : let mut guard = tline.gc_info.write().unwrap();
8027 2 : guard.cutoffs.time = Lsn(0x30);
8028 2 : guard.cutoffs.space = Lsn(0x30);
8029 2 : }
8030 2 :
8031 2 : let expected_result = [
8032 2 : Bytes::from_static(b"value 0@0x10"),
8033 2 : Bytes::from_static(b"value 1@0x20"),
8034 2 : Bytes::from_static(b"value 2@0x30"),
8035 2 : Bytes::from_static(b"value 3@0x40"),
8036 2 : Bytes::from_static(b"value 4@0x10"),
8037 2 : Bytes::from_static(b"value 5@0x20"),
8038 2 : Bytes::from_static(b"value 6@0x20"),
8039 2 : Bytes::from_static(b"value 7@0x10"),
8040 2 : Bytes::from_static(b"value 8@0x48"),
8041 2 : Bytes::from_static(b"value 9@0x48"),
8042 2 : ];
8043 2 :
8044 20 : for (idx, expected) in expected_result.iter().enumerate() {
8045 20 : assert_eq!(
8046 20 : tline
8047 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8048 30 : .await
8049 20 : .unwrap(),
8050 2 : expected
8051 2 : );
8052 2 : }
8053 2 :
8054 2 : let cancel = CancellationToken::new();
8055 2 : tline
8056 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8057 56 : .await
8058 2 : .unwrap();
8059 2 :
8060 20 : for (idx, expected) in expected_result.iter().enumerate() {
8061 20 : assert_eq!(
8062 20 : tline
8063 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8064 20 : .await
8065 20 : .unwrap(),
8066 2 : expected
8067 2 : );
8068 2 : }
8069 2 :
8070 2 : // Check if the image layer at the GC horizon contains exactly what we want
8071 2 : let image_at_gc_horizon = tline
8072 2 : .inspect_image_layers(Lsn(0x30), &ctx)
8073 2 : .await
8074 2 : .unwrap()
8075 2 : .into_iter()
8076 34 : .filter(|(k, _)| k.is_metadata_key())
8077 2 : .collect::<Vec<_>>();
8078 2 :
8079 2 : assert_eq!(image_at_gc_horizon.len(), 10);
8080 2 : let expected_result = [
8081 2 : Bytes::from_static(b"value 0@0x10"),
8082 2 : Bytes::from_static(b"value 1@0x20"),
8083 2 : Bytes::from_static(b"value 2@0x30"),
8084 2 : Bytes::from_static(b"value 3@0x10"),
8085 2 : Bytes::from_static(b"value 4@0x10"),
8086 2 : Bytes::from_static(b"value 5@0x20"),
8087 2 : Bytes::from_static(b"value 6@0x20"),
8088 2 : Bytes::from_static(b"value 7@0x10"),
8089 2 : Bytes::from_static(b"value 8@0x10"),
8090 2 : Bytes::from_static(b"value 9@0x10"),
8091 2 : ];
8092 22 : for idx in 0..10 {
8093 20 : assert_eq!(
8094 20 : image_at_gc_horizon[idx],
8095 20 : (get_key(idx as u32), expected_result[idx].clone())
8096 20 : );
8097 2 : }
8098 2 :
8099 2 : // Check if old layers are removed / new layers have the expected LSN
8100 2 : let all_layers = inspect_and_sort(&tline, None).await;
8101 2 : assert_eq!(
8102 2 : all_layers,
8103 2 : vec![
8104 2 : // Image layer at GC horizon
8105 2 : PersistentLayerKey {
8106 2 : key_range: Key::MIN..Key::MAX,
8107 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
8108 2 : is_delta: false
8109 2 : },
8110 2 : // The delta layer below the horizon
8111 2 : PersistentLayerKey {
8112 2 : key_range: get_key(3)..get_key(4),
8113 2 : lsn_range: Lsn(0x30)..Lsn(0x48),
8114 2 : is_delta: true
8115 2 : },
8116 2 : // The delta3 layer that should not be picked for the compaction
8117 2 : PersistentLayerKey {
8118 2 : key_range: get_key(8)..get_key(10),
8119 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
8120 2 : is_delta: true
8121 2 : }
8122 2 : ]
8123 2 : );
8124 2 :
8125 2 : // increase GC horizon and compact again
8126 2 : {
8127 2 : // Update GC info
8128 2 : let mut guard = tline.gc_info.write().unwrap();
8129 2 : guard.cutoffs.time = Lsn(0x40);
8130 2 : guard.cutoffs.space = Lsn(0x40);
8131 2 : }
8132 2 : tline
8133 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8134 44 : .await
8135 2 : .unwrap();
8136 2 :
8137 2 : Ok(())
8138 2 : }
8139 :
8140 : #[cfg(feature = "testing")]
8141 : #[tokio::test]
8142 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
8143 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
8144 20 : let (tenant, ctx) = harness.load().await;
8145 2 :
8146 24 : fn get_key(id: u32) -> Key {
8147 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8148 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8149 24 : key.field6 = id;
8150 24 : key
8151 24 : }
8152 2 :
8153 2 : let delta1 = vec![
8154 2 : (
8155 2 : get_key(1),
8156 2 : Lsn(0x20),
8157 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8158 2 : ),
8159 2 : (
8160 2 : get_key(1),
8161 2 : Lsn(0x30),
8162 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8163 2 : ),
8164 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
8165 2 : (
8166 2 : get_key(2),
8167 2 : Lsn(0x20),
8168 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8169 2 : ),
8170 2 : (
8171 2 : get_key(2),
8172 2 : Lsn(0x30),
8173 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8174 2 : ),
8175 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
8176 2 : (
8177 2 : get_key(3),
8178 2 : Lsn(0x20),
8179 2 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
8180 2 : ),
8181 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
8182 2 : (
8183 2 : get_key(4),
8184 2 : Lsn(0x20),
8185 2 : Value::WalRecord(NeonWalRecord::wal_init("i")),
8186 2 : ),
8187 2 : ];
8188 2 : let image1 = vec![(get_key(1), "0x10".into())];
8189 2 :
8190 2 : let tline = tenant
8191 2 : .create_test_timeline_with_layers(
8192 2 : TIMELINE_ID,
8193 2 : Lsn(0x10),
8194 2 : DEFAULT_PG_VERSION,
8195 2 : &ctx,
8196 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8197 2 : Lsn(0x10)..Lsn(0x40),
8198 2 : delta1,
8199 2 : )], // delta layers
8200 2 : vec![(Lsn(0x10), image1)], // image layers
8201 2 : Lsn(0x50),
8202 2 : )
8203 19 : .await?;
8204 2 :
8205 2 : assert_eq!(
8206 8 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
8207 2 : Bytes::from_static(b"0x10,0x20,0x30")
8208 2 : );
8209 2 : assert_eq!(
8210 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
8211 2 : Bytes::from_static(b"0x10,0x20,0x30")
8212 2 : );
8213 2 :
8214 2 : // Need to remove the limit of "Neon WAL redo requires base image".
8215 2 :
8216 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
8217 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
8218 2 :
8219 2 : Ok(())
8220 2 : }
8221 :
8222 : #[tokio::test(start_paused = true)]
8223 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
8224 2 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
8225 2 : .await
8226 2 : .unwrap()
8227 2 : .load()
8228 20 : .await;
8229 2 : // Advance to the lsn lease deadline so that GC is not blocked by
8230 2 : // initial transition into AttachedSingle.
8231 2 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
8232 2 : tokio::time::resume();
8233 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8234 2 :
8235 2 : let end_lsn = Lsn(0x100);
8236 2 : let image_layers = (0x20..=0x90)
8237 2 : .step_by(0x10)
8238 16 : .map(|n| {
8239 16 : (
8240 16 : Lsn(n),
8241 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
8242 16 : )
8243 16 : })
8244 2 : .collect();
8245 2 :
8246 2 : let timeline = tenant
8247 2 : .create_test_timeline_with_layers(
8248 2 : TIMELINE_ID,
8249 2 : Lsn(0x10),
8250 2 : DEFAULT_PG_VERSION,
8251 2 : &ctx,
8252 2 : Vec::new(),
8253 2 : image_layers,
8254 2 : end_lsn,
8255 2 : )
8256 61 : .await?;
8257 2 :
8258 2 : let leased_lsns = [0x30, 0x50, 0x70];
8259 2 : let mut leases = Vec::new();
8260 6 : leased_lsns.iter().for_each(|n| {
8261 6 : leases.push(
8262 6 : timeline
8263 6 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
8264 6 : .expect("lease request should succeed"),
8265 6 : );
8266 6 : });
8267 2 :
8268 2 : let updated_lease_0 = timeline
8269 2 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
8270 2 : .expect("lease renewal should succeed");
8271 2 : assert_eq!(
8272 2 : updated_lease_0.valid_until, leases[0].valid_until,
8273 2 : " Renewing with shorter lease should not change the lease."
8274 2 : );
8275 2 :
8276 2 : let updated_lease_1 = timeline
8277 2 : .renew_lsn_lease(
8278 2 : Lsn(leased_lsns[1]),
8279 2 : timeline.get_lsn_lease_length() * 2,
8280 2 : &ctx,
8281 2 : )
8282 2 : .expect("lease renewal should succeed");
8283 2 : assert!(
8284 2 : updated_lease_1.valid_until > leases[1].valid_until,
8285 2 : "Renewing with a long lease should renew lease with later expiration time."
8286 2 : );
8287 2 :
8288 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
8289 2 : info!(
8290 2 : "latest_gc_cutoff_lsn: {}",
8291 0 : *timeline.get_latest_gc_cutoff_lsn()
8292 2 : );
8293 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
8294 2 :
8295 2 : let res = tenant
8296 2 : .gc_iteration(
8297 2 : Some(TIMELINE_ID),
8298 2 : 0,
8299 2 : Duration::ZERO,
8300 2 : &CancellationToken::new(),
8301 2 : &ctx,
8302 2 : )
8303 2 : .await
8304 2 : .unwrap();
8305 2 :
8306 2 : // Keeping everything <= Lsn(0x80) b/c leases:
8307 2 : // 0/10: initdb layer
8308 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
8309 2 : assert_eq!(res.layers_needed_by_leases, 7);
8310 2 : // Keeping 0/90 b/c it is the latest layer.
8311 2 : assert_eq!(res.layers_not_updated, 1);
8312 2 : // Removed 0/80.
8313 2 : assert_eq!(res.layers_removed, 1);
8314 2 :
8315 2 : // Make lease on a already GC-ed LSN.
8316 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8317 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
8318 2 : timeline
8319 2 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8320 2 : .expect_err("lease request on GC-ed LSN should fail");
8321 2 :
8322 2 : // Should still be able to renew a currently valid lease
8323 2 : // Assumption: original lease to is still valid for 0/50.
8324 2 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8325 2 : timeline
8326 2 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8327 2 : .expect("lease renewal with validation should succeed");
8328 2 :
8329 2 : Ok(())
8330 2 : }
8331 :
8332 : #[cfg(feature = "testing")]
8333 : #[tokio::test]
8334 2 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8335 2 : test_simple_bottom_most_compaction_deltas_helper(
8336 2 : "test_simple_bottom_most_compaction_deltas_1",
8337 2 : false,
8338 2 : )
8339 243 : .await
8340 2 : }
8341 :
8342 : #[cfg(feature = "testing")]
8343 : #[tokio::test]
8344 2 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8345 2 : test_simple_bottom_most_compaction_deltas_helper(
8346 2 : "test_simple_bottom_most_compaction_deltas_2",
8347 2 : true,
8348 2 : )
8349 221 : .await
8350 2 : }
8351 :
8352 : #[cfg(feature = "testing")]
8353 4 : async fn test_simple_bottom_most_compaction_deltas_helper(
8354 4 : test_name: &'static str,
8355 4 : use_delta_bottom_layer: bool,
8356 4 : ) -> anyhow::Result<()> {
8357 4 : let harness = TenantHarness::create(test_name).await?;
8358 40 : let (tenant, ctx) = harness.load().await;
8359 :
8360 276 : fn get_key(id: u32) -> Key {
8361 276 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8362 276 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8363 276 : key.field6 = id;
8364 276 : key
8365 276 : }
8366 :
8367 : // We create
8368 : // - one bottom-most image layer,
8369 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8370 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8371 : // - a delta layer D3 above the horizon.
8372 : //
8373 : // | D3 |
8374 : // | D1 |
8375 : // -| |-- gc horizon -----------------
8376 : // | | | D2 |
8377 : // --------- img layer ------------------
8378 : //
8379 : // What we should expact from this compaction is:
8380 : // | D3 |
8381 : // | Part of D1 |
8382 : // --------- img layer with D1+D2 at GC horizon------------------
8383 :
8384 : // img layer at 0x10
8385 4 : let img_layer = (0..10)
8386 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8387 4 : .collect_vec();
8388 4 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8389 4 : let delta4 = (0..10)
8390 40 : .map(|id| {
8391 40 : (
8392 40 : get_key(id),
8393 40 : Lsn(0x08),
8394 40 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8395 40 : )
8396 40 : })
8397 4 : .collect_vec();
8398 4 :
8399 4 : let delta1 = vec![
8400 4 : (
8401 4 : get_key(1),
8402 4 : Lsn(0x20),
8403 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8404 4 : ),
8405 4 : (
8406 4 : get_key(2),
8407 4 : Lsn(0x30),
8408 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8409 4 : ),
8410 4 : (
8411 4 : get_key(3),
8412 4 : Lsn(0x28),
8413 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8414 4 : ),
8415 4 : (
8416 4 : get_key(3),
8417 4 : Lsn(0x30),
8418 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8419 4 : ),
8420 4 : (
8421 4 : get_key(3),
8422 4 : Lsn(0x40),
8423 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8424 4 : ),
8425 4 : ];
8426 4 : let delta2 = vec![
8427 4 : (
8428 4 : get_key(5),
8429 4 : Lsn(0x20),
8430 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8431 4 : ),
8432 4 : (
8433 4 : get_key(6),
8434 4 : Lsn(0x20),
8435 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8436 4 : ),
8437 4 : ];
8438 4 : let delta3 = vec![
8439 4 : (
8440 4 : get_key(8),
8441 4 : Lsn(0x48),
8442 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8443 4 : ),
8444 4 : (
8445 4 : get_key(9),
8446 4 : Lsn(0x48),
8447 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8448 4 : ),
8449 4 : ];
8450 :
8451 4 : let tline = if use_delta_bottom_layer {
8452 2 : tenant
8453 2 : .create_test_timeline_with_layers(
8454 2 : TIMELINE_ID,
8455 2 : Lsn(0x08),
8456 2 : DEFAULT_PG_VERSION,
8457 2 : &ctx,
8458 2 : vec![
8459 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8460 2 : Lsn(0x08)..Lsn(0x10),
8461 2 : delta4,
8462 2 : ),
8463 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8464 2 : Lsn(0x20)..Lsn(0x48),
8465 2 : delta1,
8466 2 : ),
8467 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8468 2 : Lsn(0x20)..Lsn(0x48),
8469 2 : delta2,
8470 2 : ),
8471 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8472 2 : Lsn(0x48)..Lsn(0x50),
8473 2 : delta3,
8474 2 : ),
8475 2 : ], // delta layers
8476 2 : vec![], // image layers
8477 2 : Lsn(0x50),
8478 2 : )
8479 30 : .await?
8480 : } else {
8481 2 : tenant
8482 2 : .create_test_timeline_with_layers(
8483 2 : TIMELINE_ID,
8484 2 : Lsn(0x10),
8485 2 : DEFAULT_PG_VERSION,
8486 2 : &ctx,
8487 2 : vec![
8488 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8489 2 : Lsn(0x10)..Lsn(0x48),
8490 2 : delta1,
8491 2 : ),
8492 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8493 2 : Lsn(0x10)..Lsn(0x48),
8494 2 : delta2,
8495 2 : ),
8496 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
8497 2 : Lsn(0x48)..Lsn(0x50),
8498 2 : delta3,
8499 2 : ),
8500 2 : ], // delta layers
8501 2 : vec![(Lsn(0x10), img_layer)], // image layers
8502 2 : Lsn(0x50),
8503 2 : )
8504 49 : .await?
8505 : };
8506 4 : {
8507 4 : // Update GC info
8508 4 : let mut guard = tline.gc_info.write().unwrap();
8509 4 : *guard = GcInfo {
8510 4 : retain_lsns: vec![],
8511 4 : cutoffs: GcCutoffs {
8512 4 : time: Lsn(0x30),
8513 4 : space: Lsn(0x30),
8514 4 : },
8515 4 : leases: Default::default(),
8516 4 : within_ancestor_pitr: false,
8517 4 : };
8518 4 : }
8519 4 :
8520 4 : let expected_result = [
8521 4 : Bytes::from_static(b"value 0@0x10"),
8522 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8523 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8524 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
8525 4 : Bytes::from_static(b"value 4@0x10"),
8526 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8527 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8528 4 : Bytes::from_static(b"value 7@0x10"),
8529 4 : Bytes::from_static(b"value 8@0x10@0x48"),
8530 4 : Bytes::from_static(b"value 9@0x10@0x48"),
8531 4 : ];
8532 4 :
8533 4 : let expected_result_at_gc_horizon = [
8534 4 : Bytes::from_static(b"value 0@0x10"),
8535 4 : Bytes::from_static(b"value 1@0x10@0x20"),
8536 4 : Bytes::from_static(b"value 2@0x10@0x30"),
8537 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
8538 4 : Bytes::from_static(b"value 4@0x10"),
8539 4 : Bytes::from_static(b"value 5@0x10@0x20"),
8540 4 : Bytes::from_static(b"value 6@0x10@0x20"),
8541 4 : Bytes::from_static(b"value 7@0x10"),
8542 4 : Bytes::from_static(b"value 8@0x10"),
8543 4 : Bytes::from_static(b"value 9@0x10"),
8544 4 : ];
8545 :
8546 44 : for idx in 0..10 {
8547 40 : assert_eq!(
8548 40 : tline
8549 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8550 60 : .await
8551 40 : .unwrap(),
8552 40 : &expected_result[idx]
8553 : );
8554 40 : assert_eq!(
8555 40 : tline
8556 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8557 31 : .await
8558 40 : .unwrap(),
8559 40 : &expected_result_at_gc_horizon[idx]
8560 : );
8561 : }
8562 :
8563 4 : let cancel = CancellationToken::new();
8564 4 : tline
8565 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8566 109 : .await
8567 4 : .unwrap();
8568 :
8569 44 : for idx in 0..10 {
8570 40 : assert_eq!(
8571 40 : tline
8572 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8573 39 : .await
8574 40 : .unwrap(),
8575 40 : &expected_result[idx]
8576 : );
8577 40 : assert_eq!(
8578 40 : tline
8579 40 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
8580 21 : .await
8581 40 : .unwrap(),
8582 40 : &expected_result_at_gc_horizon[idx]
8583 : );
8584 : }
8585 :
8586 : // increase GC horizon and compact again
8587 4 : {
8588 4 : // Update GC info
8589 4 : let mut guard = tline.gc_info.write().unwrap();
8590 4 : guard.cutoffs.time = Lsn(0x40);
8591 4 : guard.cutoffs.space = Lsn(0x40);
8592 4 : }
8593 4 : tline
8594 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8595 85 : .await
8596 4 : .unwrap();
8597 4 :
8598 4 : Ok(())
8599 4 : }
8600 :
8601 : #[cfg(feature = "testing")]
8602 : #[tokio::test]
8603 2 : async fn test_generate_key_retention() -> anyhow::Result<()> {
8604 2 : let harness = TenantHarness::create("test_generate_key_retention").await?;
8605 20 : let (tenant, ctx) = harness.load().await;
8606 2 : let tline = tenant
8607 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8608 6 : .await?;
8609 2 : tline.force_advance_lsn(Lsn(0x70));
8610 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8611 2 : let history = vec![
8612 2 : (
8613 2 : key,
8614 2 : Lsn(0x10),
8615 2 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
8616 2 : ),
8617 2 : (
8618 2 : key,
8619 2 : Lsn(0x20),
8620 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8621 2 : ),
8622 2 : (
8623 2 : key,
8624 2 : Lsn(0x30),
8625 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8626 2 : ),
8627 2 : (
8628 2 : key,
8629 2 : Lsn(0x40),
8630 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8631 2 : ),
8632 2 : (
8633 2 : key,
8634 2 : Lsn(0x50),
8635 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8636 2 : ),
8637 2 : (
8638 2 : key,
8639 2 : Lsn(0x60),
8640 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8641 2 : ),
8642 2 : (
8643 2 : key,
8644 2 : Lsn(0x70),
8645 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8646 2 : ),
8647 2 : (
8648 2 : key,
8649 2 : Lsn(0x80),
8650 2 : Value::Image(Bytes::copy_from_slice(
8651 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8652 2 : )),
8653 2 : ),
8654 2 : (
8655 2 : key,
8656 2 : Lsn(0x90),
8657 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8658 2 : ),
8659 2 : ];
8660 2 : let res = tline
8661 2 : .generate_key_retention(
8662 2 : key,
8663 2 : &history,
8664 2 : Lsn(0x60),
8665 2 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
8666 2 : 3,
8667 2 : None,
8668 2 : )
8669 2 : .await
8670 2 : .unwrap();
8671 2 : let expected_res = KeyHistoryRetention {
8672 2 : below_horizon: vec![
8673 2 : (
8674 2 : Lsn(0x20),
8675 2 : KeyLogAtLsn(vec![(
8676 2 : Lsn(0x20),
8677 2 : Value::Image(Bytes::from_static(b"0x10;0x20")),
8678 2 : )]),
8679 2 : ),
8680 2 : (
8681 2 : Lsn(0x40),
8682 2 : KeyLogAtLsn(vec![
8683 2 : (
8684 2 : Lsn(0x30),
8685 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8686 2 : ),
8687 2 : (
8688 2 : Lsn(0x40),
8689 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8690 2 : ),
8691 2 : ]),
8692 2 : ),
8693 2 : (
8694 2 : Lsn(0x50),
8695 2 : KeyLogAtLsn(vec![(
8696 2 : Lsn(0x50),
8697 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
8698 2 : )]),
8699 2 : ),
8700 2 : (
8701 2 : Lsn(0x60),
8702 2 : KeyLogAtLsn(vec![(
8703 2 : Lsn(0x60),
8704 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8705 2 : )]),
8706 2 : ),
8707 2 : ],
8708 2 : above_horizon: KeyLogAtLsn(vec![
8709 2 : (
8710 2 : Lsn(0x70),
8711 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8712 2 : ),
8713 2 : (
8714 2 : Lsn(0x80),
8715 2 : Value::Image(Bytes::copy_from_slice(
8716 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8717 2 : )),
8718 2 : ),
8719 2 : (
8720 2 : Lsn(0x90),
8721 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8722 2 : ),
8723 2 : ]),
8724 2 : };
8725 2 : assert_eq!(res, expected_res);
8726 2 :
8727 2 : // We expect GC-compaction to run with the original GC. This would create a situation that
8728 2 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
8729 2 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
8730 2 : // For example, we have
8731 2 : // ```plain
8732 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
8733 2 : // ```
8734 2 : // Now the GC horizon moves up, and we have
8735 2 : // ```plain
8736 2 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
8737 2 : // ```
8738 2 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
8739 2 : // We will end up with
8740 2 : // ```plain
8741 2 : // delta @ 0x30, image @ 0x40 (gc_horizon)
8742 2 : // ```
8743 2 : // Now we run the GC-compaction, and this key does not have a full history.
8744 2 : // We should be able to handle this partial history and drop everything before the
8745 2 : // gc_horizon image.
8746 2 :
8747 2 : let history = vec![
8748 2 : (
8749 2 : key,
8750 2 : Lsn(0x20),
8751 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8752 2 : ),
8753 2 : (
8754 2 : key,
8755 2 : Lsn(0x30),
8756 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8757 2 : ),
8758 2 : (
8759 2 : key,
8760 2 : Lsn(0x40),
8761 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8762 2 : ),
8763 2 : (
8764 2 : key,
8765 2 : Lsn(0x50),
8766 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8767 2 : ),
8768 2 : (
8769 2 : key,
8770 2 : Lsn(0x60),
8771 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8772 2 : ),
8773 2 : (
8774 2 : key,
8775 2 : Lsn(0x70),
8776 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8777 2 : ),
8778 2 : (
8779 2 : key,
8780 2 : Lsn(0x80),
8781 2 : Value::Image(Bytes::copy_from_slice(
8782 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8783 2 : )),
8784 2 : ),
8785 2 : (
8786 2 : key,
8787 2 : Lsn(0x90),
8788 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8789 2 : ),
8790 2 : ];
8791 2 : let res = tline
8792 2 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
8793 2 : .await
8794 2 : .unwrap();
8795 2 : let expected_res = KeyHistoryRetention {
8796 2 : below_horizon: vec![
8797 2 : (
8798 2 : Lsn(0x40),
8799 2 : KeyLogAtLsn(vec![(
8800 2 : Lsn(0x40),
8801 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
8802 2 : )]),
8803 2 : ),
8804 2 : (
8805 2 : Lsn(0x50),
8806 2 : KeyLogAtLsn(vec![(
8807 2 : Lsn(0x50),
8808 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
8809 2 : )]),
8810 2 : ),
8811 2 : (
8812 2 : Lsn(0x60),
8813 2 : KeyLogAtLsn(vec![(
8814 2 : Lsn(0x60),
8815 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8816 2 : )]),
8817 2 : ),
8818 2 : ],
8819 2 : above_horizon: KeyLogAtLsn(vec![
8820 2 : (
8821 2 : Lsn(0x70),
8822 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8823 2 : ),
8824 2 : (
8825 2 : Lsn(0x80),
8826 2 : Value::Image(Bytes::copy_from_slice(
8827 2 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
8828 2 : )),
8829 2 : ),
8830 2 : (
8831 2 : Lsn(0x90),
8832 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
8833 2 : ),
8834 2 : ]),
8835 2 : };
8836 2 : assert_eq!(res, expected_res);
8837 2 :
8838 2 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
8839 2 : // the ancestor image in the test case.
8840 2 :
8841 2 : let history = vec![
8842 2 : (
8843 2 : key,
8844 2 : Lsn(0x20),
8845 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8846 2 : ),
8847 2 : (
8848 2 : key,
8849 2 : Lsn(0x30),
8850 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
8851 2 : ),
8852 2 : (
8853 2 : key,
8854 2 : Lsn(0x40),
8855 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8856 2 : ),
8857 2 : (
8858 2 : key,
8859 2 : Lsn(0x70),
8860 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8861 2 : ),
8862 2 : ];
8863 2 : let res = tline
8864 2 : .generate_key_retention(
8865 2 : key,
8866 2 : &history,
8867 2 : Lsn(0x60),
8868 2 : &[],
8869 2 : 3,
8870 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8871 2 : )
8872 2 : .await
8873 2 : .unwrap();
8874 2 : let expected_res = KeyHistoryRetention {
8875 2 : below_horizon: vec![(
8876 2 : Lsn(0x60),
8877 2 : KeyLogAtLsn(vec![(
8878 2 : Lsn(0x60),
8879 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
8880 2 : )]),
8881 2 : )],
8882 2 : above_horizon: KeyLogAtLsn(vec![(
8883 2 : Lsn(0x70),
8884 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8885 2 : )]),
8886 2 : };
8887 2 : assert_eq!(res, expected_res);
8888 2 :
8889 2 : let history = vec![
8890 2 : (
8891 2 : key,
8892 2 : Lsn(0x20),
8893 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8894 2 : ),
8895 2 : (
8896 2 : key,
8897 2 : Lsn(0x40),
8898 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
8899 2 : ),
8900 2 : (
8901 2 : key,
8902 2 : Lsn(0x60),
8903 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
8904 2 : ),
8905 2 : (
8906 2 : key,
8907 2 : Lsn(0x70),
8908 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8909 2 : ),
8910 2 : ];
8911 2 : let res = tline
8912 2 : .generate_key_retention(
8913 2 : key,
8914 2 : &history,
8915 2 : Lsn(0x60),
8916 2 : &[Lsn(0x30)],
8917 2 : 3,
8918 2 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
8919 2 : )
8920 2 : .await
8921 2 : .unwrap();
8922 2 : let expected_res = KeyHistoryRetention {
8923 2 : below_horizon: vec![
8924 2 : (
8925 2 : Lsn(0x30),
8926 2 : KeyLogAtLsn(vec![(
8927 2 : Lsn(0x20),
8928 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
8929 2 : )]),
8930 2 : ),
8931 2 : (
8932 2 : Lsn(0x60),
8933 2 : KeyLogAtLsn(vec![(
8934 2 : Lsn(0x60),
8935 2 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
8936 2 : )]),
8937 2 : ),
8938 2 : ],
8939 2 : above_horizon: KeyLogAtLsn(vec![(
8940 2 : Lsn(0x70),
8941 2 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
8942 2 : )]),
8943 2 : };
8944 2 : assert_eq!(res, expected_res);
8945 2 :
8946 2 : Ok(())
8947 2 : }
8948 :
8949 : #[cfg(feature = "testing")]
8950 : #[tokio::test]
8951 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
8952 2 : let harness =
8953 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
8954 20 : let (tenant, ctx) = harness.load().await;
8955 2 :
8956 518 : fn get_key(id: u32) -> Key {
8957 518 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8958 518 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8959 518 : key.field6 = id;
8960 518 : key
8961 518 : }
8962 2 :
8963 2 : let img_layer = (0..10)
8964 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8965 2 : .collect_vec();
8966 2 :
8967 2 : let delta1 = vec![
8968 2 : (
8969 2 : get_key(1),
8970 2 : Lsn(0x20),
8971 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8972 2 : ),
8973 2 : (
8974 2 : get_key(2),
8975 2 : Lsn(0x30),
8976 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8977 2 : ),
8978 2 : (
8979 2 : get_key(3),
8980 2 : Lsn(0x28),
8981 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8982 2 : ),
8983 2 : (
8984 2 : get_key(3),
8985 2 : Lsn(0x30),
8986 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8987 2 : ),
8988 2 : (
8989 2 : get_key(3),
8990 2 : Lsn(0x40),
8991 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8992 2 : ),
8993 2 : ];
8994 2 : let delta2 = vec![
8995 2 : (
8996 2 : get_key(5),
8997 2 : Lsn(0x20),
8998 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8999 2 : ),
9000 2 : (
9001 2 : get_key(6),
9002 2 : Lsn(0x20),
9003 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9004 2 : ),
9005 2 : ];
9006 2 : let delta3 = vec![
9007 2 : (
9008 2 : get_key(8),
9009 2 : Lsn(0x48),
9010 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9011 2 : ),
9012 2 : (
9013 2 : get_key(9),
9014 2 : Lsn(0x48),
9015 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9016 2 : ),
9017 2 : ];
9018 2 :
9019 2 : let tline = tenant
9020 2 : .create_test_timeline_with_layers(
9021 2 : TIMELINE_ID,
9022 2 : Lsn(0x10),
9023 2 : DEFAULT_PG_VERSION,
9024 2 : &ctx,
9025 2 : vec![
9026 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
9027 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
9028 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9029 2 : ], // delta layers
9030 2 : vec![(Lsn(0x10), img_layer)], // image layers
9031 2 : Lsn(0x50),
9032 2 : )
9033 49 : .await?;
9034 2 : {
9035 2 : // Update GC info
9036 2 : let mut guard = tline.gc_info.write().unwrap();
9037 2 : *guard = GcInfo {
9038 2 : retain_lsns: vec![
9039 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9040 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9041 2 : ],
9042 2 : cutoffs: GcCutoffs {
9043 2 : time: Lsn(0x30),
9044 2 : space: Lsn(0x30),
9045 2 : },
9046 2 : leases: Default::default(),
9047 2 : within_ancestor_pitr: false,
9048 2 : };
9049 2 : }
9050 2 :
9051 2 : let expected_result = [
9052 2 : Bytes::from_static(b"value 0@0x10"),
9053 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9054 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9055 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9056 2 : Bytes::from_static(b"value 4@0x10"),
9057 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9058 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9059 2 : Bytes::from_static(b"value 7@0x10"),
9060 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9061 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9062 2 : ];
9063 2 :
9064 2 : let expected_result_at_gc_horizon = [
9065 2 : Bytes::from_static(b"value 0@0x10"),
9066 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9067 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9068 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9069 2 : Bytes::from_static(b"value 4@0x10"),
9070 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9071 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9072 2 : Bytes::from_static(b"value 7@0x10"),
9073 2 : Bytes::from_static(b"value 8@0x10"),
9074 2 : Bytes::from_static(b"value 9@0x10"),
9075 2 : ];
9076 2 :
9077 2 : let expected_result_at_lsn_20 = [
9078 2 : Bytes::from_static(b"value 0@0x10"),
9079 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9080 2 : Bytes::from_static(b"value 2@0x10"),
9081 2 : Bytes::from_static(b"value 3@0x10"),
9082 2 : Bytes::from_static(b"value 4@0x10"),
9083 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9084 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9085 2 : Bytes::from_static(b"value 7@0x10"),
9086 2 : Bytes::from_static(b"value 8@0x10"),
9087 2 : Bytes::from_static(b"value 9@0x10"),
9088 2 : ];
9089 2 :
9090 2 : let expected_result_at_lsn_10 = [
9091 2 : Bytes::from_static(b"value 0@0x10"),
9092 2 : Bytes::from_static(b"value 1@0x10"),
9093 2 : Bytes::from_static(b"value 2@0x10"),
9094 2 : Bytes::from_static(b"value 3@0x10"),
9095 2 : Bytes::from_static(b"value 4@0x10"),
9096 2 : Bytes::from_static(b"value 5@0x10"),
9097 2 : Bytes::from_static(b"value 6@0x10"),
9098 2 : Bytes::from_static(b"value 7@0x10"),
9099 2 : Bytes::from_static(b"value 8@0x10"),
9100 2 : Bytes::from_static(b"value 9@0x10"),
9101 2 : ];
9102 2 :
9103 12 : let verify_result = || async {
9104 12 : let gc_horizon = {
9105 12 : let gc_info = tline.gc_info.read().unwrap();
9106 12 : gc_info.cutoffs.time
9107 2 : };
9108 132 : for idx in 0..10 {
9109 120 : assert_eq!(
9110 120 : tline
9111 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9112 124 : .await
9113 120 : .unwrap(),
9114 120 : &expected_result[idx]
9115 2 : );
9116 120 : assert_eq!(
9117 120 : tline
9118 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
9119 95 : .await
9120 120 : .unwrap(),
9121 120 : &expected_result_at_gc_horizon[idx]
9122 2 : );
9123 120 : assert_eq!(
9124 120 : tline
9125 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9126 82 : .await
9127 120 : .unwrap(),
9128 120 : &expected_result_at_lsn_20[idx]
9129 2 : );
9130 120 : assert_eq!(
9131 120 : tline
9132 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9133 60 : .await
9134 120 : .unwrap(),
9135 120 : &expected_result_at_lsn_10[idx]
9136 2 : );
9137 2 : }
9138 24 : };
9139 2 :
9140 69 : verify_result().await;
9141 2 :
9142 2 : let cancel = CancellationToken::new();
9143 2 : let mut dryrun_flags = EnumSet::new();
9144 2 : dryrun_flags.insert(CompactFlags::DryRun);
9145 2 :
9146 2 : tline
9147 2 : .compact_with_gc(
9148 2 : &cancel,
9149 2 : CompactOptions {
9150 2 : flags: dryrun_flags,
9151 2 : compact_range: None,
9152 2 : },
9153 2 : &ctx,
9154 2 : )
9155 48 : .await
9156 2 : .unwrap();
9157 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
9158 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9159 57 : verify_result().await;
9160 2 :
9161 2 : tline
9162 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9163 50 : .await
9164 2 : .unwrap();
9165 64 : verify_result().await;
9166 2 :
9167 2 : // compact again
9168 2 : tline
9169 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9170 39 : .await
9171 2 : .unwrap();
9172 57 : verify_result().await;
9173 2 :
9174 2 : // increase GC horizon and compact again
9175 2 : {
9176 2 : // Update GC info
9177 2 : let mut guard = tline.gc_info.write().unwrap();
9178 2 : guard.cutoffs.time = Lsn(0x38);
9179 2 : guard.cutoffs.space = Lsn(0x38);
9180 2 : }
9181 2 : tline
9182 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9183 40 : .await
9184 2 : .unwrap();
9185 57 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
9186 2 :
9187 2 : // not increasing the GC horizon and compact again
9188 2 : tline
9189 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9190 41 : .await
9191 2 : .unwrap();
9192 57 : verify_result().await;
9193 2 :
9194 2 : Ok(())
9195 2 : }
9196 :
9197 : #[cfg(feature = "testing")]
9198 : #[tokio::test]
9199 2 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
9200 2 : {
9201 2 : let harness =
9202 2 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
9203 2 : .await?;
9204 20 : let (tenant, ctx) = harness.load().await;
9205 2 :
9206 352 : fn get_key(id: u32) -> Key {
9207 352 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9208 352 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9209 352 : key.field6 = id;
9210 352 : key
9211 352 : }
9212 2 :
9213 2 : let img_layer = (0..10)
9214 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9215 2 : .collect_vec();
9216 2 :
9217 2 : let delta1 = vec![
9218 2 : (
9219 2 : get_key(1),
9220 2 : Lsn(0x20),
9221 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9222 2 : ),
9223 2 : (
9224 2 : get_key(1),
9225 2 : Lsn(0x28),
9226 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9227 2 : ),
9228 2 : ];
9229 2 : let delta2 = vec![
9230 2 : (
9231 2 : get_key(1),
9232 2 : Lsn(0x30),
9233 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9234 2 : ),
9235 2 : (
9236 2 : get_key(1),
9237 2 : Lsn(0x38),
9238 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
9239 2 : ),
9240 2 : ];
9241 2 : let delta3 = vec![
9242 2 : (
9243 2 : get_key(8),
9244 2 : Lsn(0x48),
9245 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9246 2 : ),
9247 2 : (
9248 2 : get_key(9),
9249 2 : Lsn(0x48),
9250 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9251 2 : ),
9252 2 : ];
9253 2 :
9254 2 : let tline = tenant
9255 2 : .create_test_timeline_with_layers(
9256 2 : TIMELINE_ID,
9257 2 : Lsn(0x10),
9258 2 : DEFAULT_PG_VERSION,
9259 2 : &ctx,
9260 2 : vec![
9261 2 : // delta1 and delta 2 only contain a single key but multiple updates
9262 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
9263 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
9264 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
9265 2 : ], // delta layers
9266 2 : vec![(Lsn(0x10), img_layer)], // image layers
9267 2 : Lsn(0x50),
9268 2 : )
9269 49 : .await?;
9270 2 : {
9271 2 : // Update GC info
9272 2 : let mut guard = tline.gc_info.write().unwrap();
9273 2 : *guard = GcInfo {
9274 2 : retain_lsns: vec![
9275 2 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9276 2 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9277 2 : ],
9278 2 : cutoffs: GcCutoffs {
9279 2 : time: Lsn(0x30),
9280 2 : space: Lsn(0x30),
9281 2 : },
9282 2 : leases: Default::default(),
9283 2 : within_ancestor_pitr: false,
9284 2 : };
9285 2 : }
9286 2 :
9287 2 : let expected_result = [
9288 2 : Bytes::from_static(b"value 0@0x10"),
9289 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
9290 2 : Bytes::from_static(b"value 2@0x10"),
9291 2 : Bytes::from_static(b"value 3@0x10"),
9292 2 : Bytes::from_static(b"value 4@0x10"),
9293 2 : Bytes::from_static(b"value 5@0x10"),
9294 2 : Bytes::from_static(b"value 6@0x10"),
9295 2 : Bytes::from_static(b"value 7@0x10"),
9296 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9297 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9298 2 : ];
9299 2 :
9300 2 : let expected_result_at_gc_horizon = [
9301 2 : Bytes::from_static(b"value 0@0x10"),
9302 2 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
9303 2 : Bytes::from_static(b"value 2@0x10"),
9304 2 : Bytes::from_static(b"value 3@0x10"),
9305 2 : Bytes::from_static(b"value 4@0x10"),
9306 2 : Bytes::from_static(b"value 5@0x10"),
9307 2 : Bytes::from_static(b"value 6@0x10"),
9308 2 : Bytes::from_static(b"value 7@0x10"),
9309 2 : Bytes::from_static(b"value 8@0x10"),
9310 2 : Bytes::from_static(b"value 9@0x10"),
9311 2 : ];
9312 2 :
9313 2 : let expected_result_at_lsn_20 = [
9314 2 : Bytes::from_static(b"value 0@0x10"),
9315 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9316 2 : Bytes::from_static(b"value 2@0x10"),
9317 2 : Bytes::from_static(b"value 3@0x10"),
9318 2 : Bytes::from_static(b"value 4@0x10"),
9319 2 : Bytes::from_static(b"value 5@0x10"),
9320 2 : Bytes::from_static(b"value 6@0x10"),
9321 2 : Bytes::from_static(b"value 7@0x10"),
9322 2 : Bytes::from_static(b"value 8@0x10"),
9323 2 : Bytes::from_static(b"value 9@0x10"),
9324 2 : ];
9325 2 :
9326 2 : let expected_result_at_lsn_10 = [
9327 2 : Bytes::from_static(b"value 0@0x10"),
9328 2 : Bytes::from_static(b"value 1@0x10"),
9329 2 : Bytes::from_static(b"value 2@0x10"),
9330 2 : Bytes::from_static(b"value 3@0x10"),
9331 2 : Bytes::from_static(b"value 4@0x10"),
9332 2 : Bytes::from_static(b"value 5@0x10"),
9333 2 : Bytes::from_static(b"value 6@0x10"),
9334 2 : Bytes::from_static(b"value 7@0x10"),
9335 2 : Bytes::from_static(b"value 8@0x10"),
9336 2 : Bytes::from_static(b"value 9@0x10"),
9337 2 : ];
9338 2 :
9339 8 : let verify_result = || async {
9340 8 : let gc_horizon = {
9341 8 : let gc_info = tline.gc_info.read().unwrap();
9342 8 : gc_info.cutoffs.time
9343 2 : };
9344 88 : for idx in 0..10 {
9345 80 : assert_eq!(
9346 80 : tline
9347 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9348 76 : .await
9349 80 : .unwrap(),
9350 80 : &expected_result[idx]
9351 2 : );
9352 80 : assert_eq!(
9353 80 : tline
9354 80 : .get(get_key(idx as u32), gc_horizon, &ctx)
9355 46 : .await
9356 80 : .unwrap(),
9357 80 : &expected_result_at_gc_horizon[idx]
9358 2 : );
9359 80 : assert_eq!(
9360 80 : tline
9361 80 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9362 48 : .await
9363 80 : .unwrap(),
9364 80 : &expected_result_at_lsn_20[idx]
9365 2 : );
9366 80 : assert_eq!(
9367 80 : tline
9368 80 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9369 41 : .await
9370 80 : .unwrap(),
9371 80 : &expected_result_at_lsn_10[idx]
9372 2 : );
9373 2 : }
9374 16 : };
9375 2 :
9376 61 : verify_result().await;
9377 2 :
9378 2 : let cancel = CancellationToken::new();
9379 2 : let mut dryrun_flags = EnumSet::new();
9380 2 : dryrun_flags.insert(CompactFlags::DryRun);
9381 2 :
9382 2 : tline
9383 2 : .compact_with_gc(
9384 2 : &cancel,
9385 2 : CompactOptions {
9386 2 : flags: dryrun_flags,
9387 2 : compact_range: None,
9388 2 : },
9389 2 : &ctx,
9390 2 : )
9391 49 : .await
9392 2 : .unwrap();
9393 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
9394 2 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9395 49 : verify_result().await;
9396 2 :
9397 2 : tline
9398 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9399 53 : .await
9400 2 : .unwrap();
9401 54 : verify_result().await;
9402 2 :
9403 2 : // compact again
9404 2 : tline
9405 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9406 39 : .await
9407 2 : .unwrap();
9408 47 : verify_result().await;
9409 2 :
9410 2 : Ok(())
9411 2 : }
9412 :
9413 : #[cfg(feature = "testing")]
9414 : #[tokio::test]
9415 2 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9416 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9417 17 : let (tenant, ctx) = harness.load().await;
9418 2 :
9419 126 : fn get_key(id: u32) -> Key {
9420 126 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9421 126 : key.field6 = id;
9422 126 : key
9423 126 : }
9424 2 :
9425 2 : let img_layer = (0..10)
9426 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9427 2 : .collect_vec();
9428 2 :
9429 2 : let delta1 = vec![
9430 2 : (
9431 2 : get_key(1),
9432 2 : Lsn(0x20),
9433 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9434 2 : ),
9435 2 : (
9436 2 : get_key(2),
9437 2 : Lsn(0x30),
9438 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9439 2 : ),
9440 2 : (
9441 2 : get_key(3),
9442 2 : Lsn(0x28),
9443 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9444 2 : ),
9445 2 : (
9446 2 : get_key(3),
9447 2 : Lsn(0x30),
9448 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9449 2 : ),
9450 2 : (
9451 2 : get_key(3),
9452 2 : Lsn(0x40),
9453 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9454 2 : ),
9455 2 : ];
9456 2 : let delta2 = vec![
9457 2 : (
9458 2 : get_key(5),
9459 2 : Lsn(0x20),
9460 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9461 2 : ),
9462 2 : (
9463 2 : get_key(6),
9464 2 : Lsn(0x20),
9465 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9466 2 : ),
9467 2 : ];
9468 2 : let delta3 = vec![
9469 2 : (
9470 2 : get_key(8),
9471 2 : Lsn(0x48),
9472 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9473 2 : ),
9474 2 : (
9475 2 : get_key(9),
9476 2 : Lsn(0x48),
9477 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9478 2 : ),
9479 2 : ];
9480 2 :
9481 2 : let parent_tline = tenant
9482 2 : .create_test_timeline_with_layers(
9483 2 : TIMELINE_ID,
9484 2 : Lsn(0x10),
9485 2 : DEFAULT_PG_VERSION,
9486 2 : &ctx,
9487 2 : vec![], // delta layers
9488 2 : vec![(Lsn(0x18), img_layer)], // image layers
9489 2 : Lsn(0x18),
9490 2 : )
9491 30 : .await?;
9492 2 :
9493 2 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9494 2 :
9495 2 : let branch_tline = tenant
9496 2 : .branch_timeline_test_with_layers(
9497 2 : &parent_tline,
9498 2 : NEW_TIMELINE_ID,
9499 2 : Some(Lsn(0x18)),
9500 2 : &ctx,
9501 2 : vec![
9502 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9503 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9504 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9505 2 : ], // delta layers
9506 2 : vec![], // image layers
9507 2 : Lsn(0x50),
9508 2 : )
9509 19 : .await?;
9510 2 :
9511 2 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
9512 2 :
9513 2 : {
9514 2 : // Update GC info
9515 2 : let mut guard = parent_tline.gc_info.write().unwrap();
9516 2 : *guard = GcInfo {
9517 2 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
9518 2 : cutoffs: GcCutoffs {
9519 2 : time: Lsn(0x10),
9520 2 : space: Lsn(0x10),
9521 2 : },
9522 2 : leases: Default::default(),
9523 2 : within_ancestor_pitr: false,
9524 2 : };
9525 2 : }
9526 2 :
9527 2 : {
9528 2 : // Update GC info
9529 2 : let mut guard = branch_tline.gc_info.write().unwrap();
9530 2 : *guard = GcInfo {
9531 2 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
9532 2 : cutoffs: GcCutoffs {
9533 2 : time: Lsn(0x50),
9534 2 : space: Lsn(0x50),
9535 2 : },
9536 2 : leases: Default::default(),
9537 2 : within_ancestor_pitr: false,
9538 2 : };
9539 2 : }
9540 2 :
9541 2 : let expected_result_at_gc_horizon = [
9542 2 : Bytes::from_static(b"value 0@0x10"),
9543 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9544 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9545 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9546 2 : Bytes::from_static(b"value 4@0x10"),
9547 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9548 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9549 2 : Bytes::from_static(b"value 7@0x10"),
9550 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9551 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9552 2 : ];
9553 2 :
9554 2 : let expected_result_at_lsn_40 = [
9555 2 : Bytes::from_static(b"value 0@0x10"),
9556 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9557 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9558 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9559 2 : Bytes::from_static(b"value 4@0x10"),
9560 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9561 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9562 2 : Bytes::from_static(b"value 7@0x10"),
9563 2 : Bytes::from_static(b"value 8@0x10"),
9564 2 : Bytes::from_static(b"value 9@0x10"),
9565 2 : ];
9566 2 :
9567 4 : let verify_result = || async {
9568 44 : for idx in 0..10 {
9569 40 : assert_eq!(
9570 40 : branch_tline
9571 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9572 50 : .await
9573 40 : .unwrap(),
9574 40 : &expected_result_at_gc_horizon[idx]
9575 2 : );
9576 40 : assert_eq!(
9577 40 : branch_tline
9578 40 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
9579 31 : .await
9580 40 : .unwrap(),
9581 40 : &expected_result_at_lsn_40[idx]
9582 2 : );
9583 2 : }
9584 8 : };
9585 2 :
9586 45 : verify_result().await;
9587 2 :
9588 2 : let cancel = CancellationToken::new();
9589 2 : branch_tline
9590 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9591 17 : .await
9592 2 : .unwrap();
9593 2 :
9594 36 : verify_result().await;
9595 2 :
9596 2 : Ok(())
9597 2 : }
9598 :
9599 : // Regression test for https://github.com/neondatabase/neon/issues/9012
9600 : // Create an image arrangement where we have to read at different LSN ranges
9601 : // from a delta layer. This is achieved by overlapping an image layer on top of
9602 : // a delta layer. Like so:
9603 : //
9604 : // A B
9605 : // +----------------+ -> delta_layer
9606 : // | | ^ lsn
9607 : // | =========|-> nested_image_layer |
9608 : // | C | |
9609 : // +----------------+ |
9610 : // ======== -> baseline_image_layer +-------> key
9611 : //
9612 : //
9613 : // When querying the key range [A, B) we need to read at different LSN ranges
9614 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
9615 : #[cfg(feature = "testing")]
9616 : #[tokio::test]
9617 2 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
9618 2 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
9619 20 : let (tenant, ctx) = harness.load().await;
9620 2 :
9621 2 : let will_init_keys = [2, 6];
9622 44 : fn get_key(id: u32) -> Key {
9623 44 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
9624 44 : key.field6 = id;
9625 44 : key
9626 44 : }
9627 2 :
9628 2 : let mut expected_key_values = HashMap::new();
9629 2 :
9630 2 : let baseline_image_layer_lsn = Lsn(0x10);
9631 2 : let mut baseline_img_layer = Vec::new();
9632 12 : for i in 0..5 {
9633 10 : let key = get_key(i);
9634 10 : let value = format!("value {i}@{baseline_image_layer_lsn}");
9635 10 :
9636 10 : let removed = expected_key_values.insert(key, value.clone());
9637 10 : assert!(removed.is_none());
9638 2 :
9639 10 : baseline_img_layer.push((key, Bytes::from(value)));
9640 2 : }
9641 2 :
9642 2 : let nested_image_layer_lsn = Lsn(0x50);
9643 2 : let mut nested_img_layer = Vec::new();
9644 12 : for i in 5..10 {
9645 10 : let key = get_key(i);
9646 10 : let value = format!("value {i}@{nested_image_layer_lsn}");
9647 10 :
9648 10 : let removed = expected_key_values.insert(key, value.clone());
9649 10 : assert!(removed.is_none());
9650 2 :
9651 10 : nested_img_layer.push((key, Bytes::from(value)));
9652 2 : }
9653 2 :
9654 2 : let mut delta_layer_spec = Vec::default();
9655 2 : let delta_layer_start_lsn = Lsn(0x20);
9656 2 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
9657 2 :
9658 22 : for i in 0..10 {
9659 20 : let key = get_key(i);
9660 20 : let key_in_nested = nested_img_layer
9661 20 : .iter()
9662 80 : .any(|(key_with_img, _)| *key_with_img == key);
9663 20 : let lsn = {
9664 20 : if key_in_nested {
9665 10 : Lsn(nested_image_layer_lsn.0 + 0x10)
9666 2 : } else {
9667 10 : delta_layer_start_lsn
9668 2 : }
9669 2 : };
9670 2 :
9671 20 : let will_init = will_init_keys.contains(&i);
9672 20 : if will_init {
9673 4 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
9674 4 :
9675 4 : expected_key_values.insert(key, "".to_string());
9676 16 : } else {
9677 16 : let delta = format!("@{lsn}");
9678 16 : delta_layer_spec.push((
9679 16 : key,
9680 16 : lsn,
9681 16 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
9682 16 : ));
9683 16 :
9684 16 : expected_key_values
9685 16 : .get_mut(&key)
9686 16 : .expect("An image exists for each key")
9687 16 : .push_str(delta.as_str());
9688 16 : }
9689 20 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
9690 2 : }
9691 2 :
9692 2 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
9693 2 :
9694 2 : assert!(
9695 2 : nested_image_layer_lsn > delta_layer_start_lsn
9696 2 : && nested_image_layer_lsn < delta_layer_end_lsn
9697 2 : );
9698 2 :
9699 2 : let tline = tenant
9700 2 : .create_test_timeline_with_layers(
9701 2 : TIMELINE_ID,
9702 2 : baseline_image_layer_lsn,
9703 2 : DEFAULT_PG_VERSION,
9704 2 : &ctx,
9705 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9706 2 : delta_layer_start_lsn..delta_layer_end_lsn,
9707 2 : delta_layer_spec,
9708 2 : )], // delta layers
9709 2 : vec![
9710 2 : (baseline_image_layer_lsn, baseline_img_layer),
9711 2 : (nested_image_layer_lsn, nested_img_layer),
9712 2 : ], // image layers
9713 2 : delta_layer_end_lsn,
9714 2 : )
9715 42 : .await?;
9716 2 :
9717 2 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
9718 2 : let results = tline
9719 2 : .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
9720 13 : .await
9721 2 : .expect("No vectored errors");
9722 22 : for (key, res) in results {
9723 20 : let value = res.expect("No key errors");
9724 20 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
9725 20 : assert_eq!(value, Bytes::from(expected_value));
9726 2 : }
9727 2 :
9728 2 : Ok(())
9729 2 : }
9730 :
9731 142 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
9732 142 : (
9733 142 : k1.is_delta,
9734 142 : k1.key_range.start,
9735 142 : k1.key_range.end,
9736 142 : k1.lsn_range.start,
9737 142 : k1.lsn_range.end,
9738 142 : )
9739 142 : .cmp(&(
9740 142 : k2.is_delta,
9741 142 : k2.key_range.start,
9742 142 : k2.key_range.end,
9743 142 : k2.lsn_range.start,
9744 142 : k2.lsn_range.end,
9745 142 : ))
9746 142 : }
9747 :
9748 12 : async fn inspect_and_sort(
9749 12 : tline: &Arc<Timeline>,
9750 12 : filter: Option<std::ops::Range<Key>>,
9751 12 : ) -> Vec<PersistentLayerKey> {
9752 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
9753 12 : if let Some(filter) = filter {
9754 64 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
9755 10 : }
9756 12 : all_layers.sort_by(sort_layer_key);
9757 12 : all_layers
9758 12 : }
9759 :
9760 : #[cfg(feature = "testing")]
9761 10 : fn check_layer_map_key_eq(
9762 10 : mut left: Vec<PersistentLayerKey>,
9763 10 : mut right: Vec<PersistentLayerKey>,
9764 10 : ) {
9765 10 : left.sort_by(sort_layer_key);
9766 10 : right.sort_by(sort_layer_key);
9767 10 : if left != right {
9768 0 : eprintln!("---LEFT---");
9769 0 : for left in left.iter() {
9770 0 : eprintln!("{}", left);
9771 0 : }
9772 0 : eprintln!("---RIGHT---");
9773 0 : for right in right.iter() {
9774 0 : eprintln!("{}", right);
9775 0 : }
9776 0 : assert_eq!(left, right);
9777 10 : }
9778 10 : }
9779 :
9780 : #[cfg(feature = "testing")]
9781 : #[tokio::test]
9782 2 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
9783 2 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
9784 20 : let (tenant, ctx) = harness.load().await;
9785 2 :
9786 182 : fn get_key(id: u32) -> Key {
9787 182 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9788 182 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9789 182 : key.field6 = id;
9790 182 : key
9791 182 : }
9792 2 :
9793 2 : // img layer at 0x10
9794 2 : let img_layer = (0..10)
9795 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9796 2 : .collect_vec();
9797 2 :
9798 2 : let delta1 = vec![
9799 2 : (
9800 2 : get_key(1),
9801 2 : Lsn(0x20),
9802 2 : Value::Image(Bytes::from("value 1@0x20")),
9803 2 : ),
9804 2 : (
9805 2 : get_key(2),
9806 2 : Lsn(0x30),
9807 2 : Value::Image(Bytes::from("value 2@0x30")),
9808 2 : ),
9809 2 : (
9810 2 : get_key(3),
9811 2 : Lsn(0x40),
9812 2 : Value::Image(Bytes::from("value 3@0x40")),
9813 2 : ),
9814 2 : ];
9815 2 : let delta2 = vec![
9816 2 : (
9817 2 : get_key(5),
9818 2 : Lsn(0x20),
9819 2 : Value::Image(Bytes::from("value 5@0x20")),
9820 2 : ),
9821 2 : (
9822 2 : get_key(6),
9823 2 : Lsn(0x20),
9824 2 : Value::Image(Bytes::from("value 6@0x20")),
9825 2 : ),
9826 2 : ];
9827 2 : let delta3 = vec![
9828 2 : (
9829 2 : get_key(8),
9830 2 : Lsn(0x48),
9831 2 : Value::Image(Bytes::from("value 8@0x48")),
9832 2 : ),
9833 2 : (
9834 2 : get_key(9),
9835 2 : Lsn(0x48),
9836 2 : Value::Image(Bytes::from("value 9@0x48")),
9837 2 : ),
9838 2 : ];
9839 2 :
9840 2 : let tline = tenant
9841 2 : .create_test_timeline_with_layers(
9842 2 : TIMELINE_ID,
9843 2 : Lsn(0x10),
9844 2 : DEFAULT_PG_VERSION,
9845 2 : &ctx,
9846 2 : vec![
9847 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9848 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9849 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9850 2 : ], // delta layers
9851 2 : vec![(Lsn(0x10), img_layer)], // image layers
9852 2 : Lsn(0x50),
9853 2 : )
9854 49 : .await?;
9855 2 :
9856 2 : {
9857 2 : // Update GC info
9858 2 : let mut guard = tline.gc_info.write().unwrap();
9859 2 : *guard = GcInfo {
9860 2 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
9861 2 : cutoffs: GcCutoffs {
9862 2 : time: Lsn(0x30),
9863 2 : space: Lsn(0x30),
9864 2 : },
9865 2 : leases: Default::default(),
9866 2 : within_ancestor_pitr: false,
9867 2 : };
9868 2 : }
9869 2 :
9870 2 : let cancel = CancellationToken::new();
9871 2 :
9872 2 : // Do a partial compaction on key range 0..2
9873 2 : tline
9874 2 : .partial_compact_with_gc(get_key(0)..get_key(2), &cancel, EnumSet::new(), &ctx)
9875 25 : .await
9876 2 : .unwrap();
9877 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9878 2 : check_layer_map_key_eq(
9879 2 : all_layers,
9880 2 : vec![
9881 2 : // newly-generated image layer for the partial compaction range 0-2
9882 2 : PersistentLayerKey {
9883 2 : key_range: get_key(0)..get_key(2),
9884 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9885 2 : is_delta: false,
9886 2 : },
9887 2 : PersistentLayerKey {
9888 2 : key_range: get_key(0)..get_key(10),
9889 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9890 2 : is_delta: false,
9891 2 : },
9892 2 : // delta1 is split and the second part is rewritten
9893 2 : PersistentLayerKey {
9894 2 : key_range: get_key(2)..get_key(4),
9895 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9896 2 : is_delta: true,
9897 2 : },
9898 2 : PersistentLayerKey {
9899 2 : key_range: get_key(5)..get_key(7),
9900 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9901 2 : is_delta: true,
9902 2 : },
9903 2 : PersistentLayerKey {
9904 2 : key_range: get_key(8)..get_key(10),
9905 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9906 2 : is_delta: true,
9907 2 : },
9908 2 : ],
9909 2 : );
9910 2 :
9911 2 : // Do a partial compaction on key range 2..4
9912 2 : tline
9913 2 : .partial_compact_with_gc(get_key(2)..get_key(4), &cancel, EnumSet::new(), &ctx)
9914 15 : .await
9915 2 : .unwrap();
9916 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9917 2 : check_layer_map_key_eq(
9918 2 : all_layers,
9919 2 : vec![
9920 2 : PersistentLayerKey {
9921 2 : key_range: get_key(0)..get_key(2),
9922 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9923 2 : is_delta: false,
9924 2 : },
9925 2 : PersistentLayerKey {
9926 2 : key_range: get_key(0)..get_key(10),
9927 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9928 2 : is_delta: false,
9929 2 : },
9930 2 : // image layer generated for the compaction range 2-4
9931 2 : PersistentLayerKey {
9932 2 : key_range: get_key(2)..get_key(4),
9933 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9934 2 : is_delta: false,
9935 2 : },
9936 2 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
9937 2 : PersistentLayerKey {
9938 2 : key_range: get_key(2)..get_key(4),
9939 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9940 2 : is_delta: true,
9941 2 : },
9942 2 : PersistentLayerKey {
9943 2 : key_range: get_key(5)..get_key(7),
9944 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9945 2 : is_delta: true,
9946 2 : },
9947 2 : PersistentLayerKey {
9948 2 : key_range: get_key(8)..get_key(10),
9949 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9950 2 : is_delta: true,
9951 2 : },
9952 2 : ],
9953 2 : );
9954 2 :
9955 2 : // Do a partial compaction on key range 4..9
9956 2 : tline
9957 2 : .partial_compact_with_gc(get_key(4)..get_key(9), &cancel, EnumSet::new(), &ctx)
9958 23 : .await
9959 2 : .unwrap();
9960 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
9961 2 : check_layer_map_key_eq(
9962 2 : all_layers,
9963 2 : vec![
9964 2 : PersistentLayerKey {
9965 2 : key_range: get_key(0)..get_key(2),
9966 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9967 2 : is_delta: false,
9968 2 : },
9969 2 : PersistentLayerKey {
9970 2 : key_range: get_key(0)..get_key(10),
9971 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
9972 2 : is_delta: false,
9973 2 : },
9974 2 : PersistentLayerKey {
9975 2 : key_range: get_key(2)..get_key(4),
9976 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9977 2 : is_delta: false,
9978 2 : },
9979 2 : PersistentLayerKey {
9980 2 : key_range: get_key(2)..get_key(4),
9981 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
9982 2 : is_delta: true,
9983 2 : },
9984 2 : // image layer generated for this compaction range
9985 2 : PersistentLayerKey {
9986 2 : key_range: get_key(4)..get_key(9),
9987 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
9988 2 : is_delta: false,
9989 2 : },
9990 2 : PersistentLayerKey {
9991 2 : key_range: get_key(8)..get_key(10),
9992 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
9993 2 : is_delta: true,
9994 2 : },
9995 2 : ],
9996 2 : );
9997 2 :
9998 2 : // Do a partial compaction on key range 9..10
9999 2 : tline
10000 2 : .partial_compact_with_gc(get_key(9)..get_key(10), &cancel, EnumSet::new(), &ctx)
10001 9 : .await
10002 2 : .unwrap();
10003 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10004 2 : check_layer_map_key_eq(
10005 2 : all_layers,
10006 2 : vec![
10007 2 : PersistentLayerKey {
10008 2 : key_range: get_key(0)..get_key(2),
10009 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10010 2 : is_delta: false,
10011 2 : },
10012 2 : PersistentLayerKey {
10013 2 : key_range: get_key(0)..get_key(10),
10014 2 : lsn_range: Lsn(0x10)..Lsn(0x11),
10015 2 : is_delta: false,
10016 2 : },
10017 2 : PersistentLayerKey {
10018 2 : key_range: get_key(2)..get_key(4),
10019 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10020 2 : is_delta: false,
10021 2 : },
10022 2 : PersistentLayerKey {
10023 2 : key_range: get_key(2)..get_key(4),
10024 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10025 2 : is_delta: true,
10026 2 : },
10027 2 : PersistentLayerKey {
10028 2 : key_range: get_key(4)..get_key(9),
10029 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10030 2 : is_delta: false,
10031 2 : },
10032 2 : // image layer generated for the compaction range
10033 2 : PersistentLayerKey {
10034 2 : key_range: get_key(9)..get_key(10),
10035 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10036 2 : is_delta: false,
10037 2 : },
10038 2 : PersistentLayerKey {
10039 2 : key_range: get_key(8)..get_key(10),
10040 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10041 2 : is_delta: true,
10042 2 : },
10043 2 : ],
10044 2 : );
10045 2 :
10046 2 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
10047 2 : tline
10048 2 : .partial_compact_with_gc(get_key(0)..get_key(10), &cancel, EnumSet::new(), &ctx)
10049 48 : .await
10050 2 : .unwrap();
10051 2 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10052 2 : check_layer_map_key_eq(
10053 2 : all_layers,
10054 2 : vec![
10055 2 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
10056 2 : PersistentLayerKey {
10057 2 : key_range: get_key(0)..get_key(10),
10058 2 : lsn_range: Lsn(0x20)..Lsn(0x21),
10059 2 : is_delta: false,
10060 2 : },
10061 2 : PersistentLayerKey {
10062 2 : key_range: get_key(2)..get_key(4),
10063 2 : lsn_range: Lsn(0x20)..Lsn(0x48),
10064 2 : is_delta: true,
10065 2 : },
10066 2 : PersistentLayerKey {
10067 2 : key_range: get_key(8)..get_key(10),
10068 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
10069 2 : is_delta: true,
10070 2 : },
10071 2 : ],
10072 2 : );
10073 2 :
10074 2 : Ok(())
10075 2 : }
10076 :
10077 : #[cfg(feature = "testing")]
10078 : #[tokio::test]
10079 2 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
10080 2 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
10081 2 : .await
10082 2 : .unwrap();
10083 20 : let (tenant, ctx) = harness.load().await;
10084 2 : let tline_parent = tenant
10085 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
10086 6 : .await
10087 2 : .unwrap();
10088 2 : let tline_child = tenant
10089 2 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
10090 2 : .await
10091 2 : .unwrap();
10092 2 : {
10093 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10094 2 : assert_eq!(
10095 2 : gc_info_parent.retain_lsns,
10096 2 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
10097 2 : );
10098 2 : }
10099 2 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
10100 2 : tline_child
10101 2 : .remote_client
10102 2 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
10103 2 : .unwrap();
10104 2 : tline_child.remote_client.wait_completion().await.unwrap();
10105 2 : offload_timeline(&tenant, &tline_child)
10106 2 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
10107 10 : .await.unwrap();
10108 2 : let child_timeline_id = tline_child.timeline_id;
10109 2 : Arc::try_unwrap(tline_child).unwrap();
10110 2 :
10111 2 : {
10112 2 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10113 2 : assert_eq!(
10114 2 : gc_info_parent.retain_lsns,
10115 2 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
10116 2 : );
10117 2 : }
10118 2 :
10119 2 : tenant
10120 2 : .get_offloaded_timeline(child_timeline_id)
10121 2 : .unwrap()
10122 2 : .defuse_for_tenant_drop();
10123 2 :
10124 2 : Ok(())
10125 2 : }
10126 : }
|
