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 std::collections::hash_map::Entry;
16 : use std::collections::{BTreeMap, HashMap, HashSet};
17 : use std::fmt::{Debug, Display};
18 : use std::fs::File;
19 : use std::future::Future;
20 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
21 : use std::sync::{Arc, Mutex, Weak};
22 : use std::time::{Duration, Instant, SystemTime};
23 : use std::{fmt, fs};
24 :
25 : use anyhow::{Context, bail};
26 : use arc_swap::ArcSwap;
27 : use camino::{Utf8Path, Utf8PathBuf};
28 : use chrono::NaiveDateTime;
29 : use enumset::EnumSet;
30 : use futures::StreamExt;
31 : use futures::stream::FuturesUnordered;
32 : use itertools::Itertools as _;
33 : use once_cell::sync::Lazy;
34 : use pageserver_api::models;
35 : pub use pageserver_api::models::TenantState;
36 : use pageserver_api::models::{
37 : CompactInfoResponse, LsnLease, TimelineArchivalState, TimelineState, TopTenantShardItem,
38 : WalRedoManagerStatus,
39 : };
40 : use pageserver_api::shard::{ShardIdentity, ShardStripeSize, TenantShardId};
41 : use remote_storage::{DownloadError, GenericRemoteStorage, TimeoutOrCancel};
42 : use remote_timeline_client::index::GcCompactionState;
43 : use remote_timeline_client::manifest::{
44 : LATEST_TENANT_MANIFEST_VERSION, OffloadedTimelineManifest, TenantManifest,
45 : };
46 : use remote_timeline_client::{
47 : FAILED_REMOTE_OP_RETRIES, FAILED_UPLOAD_WARN_THRESHOLD, UploadQueueNotReadyError,
48 : };
49 : use secondary::heatmap::{HeatMapTenant, HeatMapTimeline};
50 : use storage_broker::BrokerClientChannel;
51 : use timeline::compaction::{CompactionOutcome, GcCompactionQueue};
52 : use timeline::offload::{OffloadError, offload_timeline};
53 : use timeline::{
54 : CompactFlags, CompactOptions, CompactionError, PreviousHeatmap, ShutdownMode, import_pgdata,
55 : };
56 : use tokio::io::BufReader;
57 : use tokio::sync::{Notify, Semaphore, watch};
58 : use tokio::task::JoinSet;
59 : use tokio_util::sync::CancellationToken;
60 : use tracing::*;
61 : use upload_queue::NotInitialized;
62 : use utils::circuit_breaker::CircuitBreaker;
63 : use utils::crashsafe::path_with_suffix_extension;
64 : use utils::sync::gate::{Gate, GateGuard};
65 : use utils::timeout::{TimeoutCancellableError, timeout_cancellable};
66 : use utils::try_rcu::ArcSwapExt;
67 : use utils::zstd::{create_zst_tarball, extract_zst_tarball};
68 : use utils::{backoff, completion, failpoint_support, fs_ext, pausable_failpoint};
69 :
70 : use self::config::{AttachedLocationConfig, AttachmentMode, LocationConf, TenantConf};
71 : use self::metadata::TimelineMetadata;
72 : use self::mgr::{GetActiveTenantError, GetTenantError};
73 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
74 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
75 : use self::timeline::uninit::{TimelineCreateGuard, TimelineExclusionError, UninitializedTimeline};
76 : use self::timeline::{
77 : EvictionTaskTenantState, GcCutoffs, TimelineDeleteProgress, TimelineResources, WaitLsnError,
78 : };
79 : use crate::config::PageServerConf;
80 : use crate::context::{DownloadBehavior, RequestContext};
81 : use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
82 : use crate::l0_flush::L0FlushGlobalState;
83 : use crate::metrics::{
84 : BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
85 : INITDB_RUN_TIME, INITDB_SEMAPHORE_ACQUISITION_TIME, TENANT, TENANT_STATE_METRIC,
86 : TENANT_SYNTHETIC_SIZE_METRIC, remove_tenant_metrics,
87 : };
88 : use crate::task_mgr::TaskKind;
89 : use crate::tenant::config::{LocationMode, TenantConfOpt};
90 : use crate::tenant::gc_result::GcResult;
91 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
92 : use crate::tenant::remote_timeline_client::{
93 : INITDB_PATH, MaybeDeletedIndexPart, remote_initdb_archive_path,
94 : };
95 : use crate::tenant::storage_layer::{DeltaLayer, ImageLayer};
96 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
97 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
98 : use crate::virtual_file::VirtualFile;
99 : use crate::walingest::WalLagCooldown;
100 : use crate::walredo::PostgresRedoManager;
101 : use crate::{InitializationOrder, TEMP_FILE_SUFFIX, import_datadir, span, task_mgr, walredo};
102 :
103 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
104 : use utils::crashsafe;
105 : use utils::generation::Generation;
106 : use utils::id::TimelineId;
107 : use utils::lsn::{Lsn, RecordLsn};
108 :
109 : pub mod blob_io;
110 : pub mod block_io;
111 : pub mod vectored_blob_io;
112 :
113 : pub mod disk_btree;
114 : pub(crate) mod ephemeral_file;
115 : pub mod layer_map;
116 :
117 : pub mod metadata;
118 : pub mod remote_timeline_client;
119 : pub mod storage_layer;
120 :
121 : pub mod checks;
122 : pub mod config;
123 : pub mod mgr;
124 : pub mod secondary;
125 : pub mod tasks;
126 : pub mod upload_queue;
127 :
128 : pub(crate) mod timeline;
129 :
130 : pub mod size;
131 :
132 : mod gc_block;
133 : mod gc_result;
134 : pub(crate) mod throttle;
135 :
136 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
137 :
138 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
139 : // re-export for use in walreceiver
140 : pub use crate::tenant::timeline::WalReceiverInfo;
141 :
142 : /// The "tenants" part of `tenants/<tenant>/timelines...`
143 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
144 :
145 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
146 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
147 :
148 : /// References to shared objects that are passed into each tenant, such
149 : /// as the shared remote storage client and process initialization state.
150 : #[derive(Clone)]
151 : pub struct TenantSharedResources {
152 : pub broker_client: storage_broker::BrokerClientChannel,
153 : pub remote_storage: GenericRemoteStorage,
154 : pub deletion_queue_client: DeletionQueueClient,
155 : pub l0_flush_global_state: L0FlushGlobalState,
156 : }
157 :
158 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
159 : /// for an attached tenant is a subset of the [`LocationConf`], represented
160 : /// in this struct.
161 : #[derive(Clone)]
162 : pub(super) struct AttachedTenantConf {
163 : tenant_conf: TenantConfOpt,
164 : location: AttachedLocationConfig,
165 : /// The deadline before which we are blocked from GC so that
166 : /// leases have a chance to be renewed.
167 : lsn_lease_deadline: Option<tokio::time::Instant>,
168 : }
169 :
170 : impl AttachedTenantConf {
171 444 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
172 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
173 : //
174 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
175 : // length, we guarantee that all the leases we granted before will have a chance to renew
176 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
177 444 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
178 444 : Some(
179 444 : tokio::time::Instant::now()
180 444 : + tenant_conf
181 444 : .lsn_lease_length
182 444 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
183 444 : )
184 : } else {
185 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
186 : // because we don't do GC in these modes.
187 0 : None
188 : };
189 :
190 444 : Self {
191 444 : tenant_conf,
192 444 : location,
193 444 : lsn_lease_deadline,
194 444 : }
195 444 : }
196 :
197 444 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
198 444 : match &location_conf.mode {
199 444 : LocationMode::Attached(attach_conf) => {
200 444 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
201 : }
202 : LocationMode::Secondary(_) => {
203 0 : anyhow::bail!(
204 0 : "Attempted to construct AttachedTenantConf from a LocationConf in secondary mode"
205 0 : )
206 : }
207 : }
208 444 : }
209 :
210 1524 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
211 1524 : self.lsn_lease_deadline
212 1524 : .map(|d| tokio::time::Instant::now() < d)
213 1524 : .unwrap_or(false)
214 1524 : }
215 : }
216 : struct TimelinePreload {
217 : timeline_id: TimelineId,
218 : client: RemoteTimelineClient,
219 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
220 : previous_heatmap: Option<PreviousHeatmap>,
221 : }
222 :
223 : pub(crate) struct TenantPreload {
224 : tenant_manifest: TenantManifest,
225 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
226 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
227 : }
228 :
229 : /// When we spawn a tenant, there is a special mode for tenant creation that
230 : /// avoids trying to read anything from remote storage.
231 : pub(crate) enum SpawnMode {
232 : /// Activate as soon as possible
233 : Eager,
234 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
235 : Lazy,
236 : }
237 :
238 : ///
239 : /// Tenant consists of multiple timelines. Keep them in a hash table.
240 : ///
241 : pub struct Tenant {
242 : // Global pageserver config parameters
243 : pub conf: &'static PageServerConf,
244 :
245 : /// The value creation timestamp, used to measure activation delay, see:
246 : /// <https://github.com/neondatabase/neon/issues/4025>
247 : constructed_at: Instant,
248 :
249 : state: watch::Sender<TenantState>,
250 :
251 : // Overridden tenant-specific config parameters.
252 : // We keep TenantConfOpt sturct here to preserve the information
253 : // about parameters that are not set.
254 : // This is necessary to allow global config updates.
255 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
256 :
257 : tenant_shard_id: TenantShardId,
258 :
259 : // The detailed sharding information, beyond the number/count in tenant_shard_id
260 : shard_identity: ShardIdentity,
261 :
262 : /// The remote storage generation, used to protect S3 objects from split-brain.
263 : /// Does not change over the lifetime of the [`Tenant`] object.
264 : ///
265 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
266 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
267 : generation: Generation,
268 :
269 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
270 :
271 : /// During timeline creation, we first insert the TimelineId to the
272 : /// creating map, then `timelines`, then remove it from the creating map.
273 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
274 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
275 :
276 : /// Possibly offloaded and archived timelines
277 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
278 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
279 :
280 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
281 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
282 : /// each other (this could be optimized to coalesce writes if necessary).
283 : ///
284 : /// The contents of the Mutex are the last manifest we successfully uploaded
285 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
286 :
287 : // This mutex prevents creation of new timelines during GC.
288 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
289 : // `timelines` mutex during all GC iteration
290 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
291 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
292 : // timeout...
293 : gc_cs: tokio::sync::Mutex<()>,
294 : walredo_mgr: Option<Arc<WalRedoManager>>,
295 :
296 : // provides access to timeline data sitting in the remote storage
297 : pub(crate) remote_storage: GenericRemoteStorage,
298 :
299 : // Access to global deletion queue for when this tenant wants to schedule a deletion
300 : deletion_queue_client: DeletionQueueClient,
301 :
302 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
303 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
304 : cached_synthetic_tenant_size: Arc<AtomicU64>,
305 :
306 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
307 :
308 : /// Track repeated failures to compact, so that we can back off.
309 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
310 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
311 :
312 : /// Signals the tenant compaction loop that there is L0 compaction work to be done.
313 : pub(crate) l0_compaction_trigger: Arc<Notify>,
314 :
315 : /// Scheduled gc-compaction tasks.
316 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
317 :
318 : /// If the tenant is in Activating state, notify this to encourage it
319 : /// to proceed to Active as soon as possible, rather than waiting for lazy
320 : /// background warmup.
321 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
322 :
323 : /// Time it took for the tenant to activate. Zero if not active yet.
324 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
325 :
326 : // Cancellation token fires when we have entered shutdown(). This is a parent of
327 : // Timelines' cancellation token.
328 : pub(crate) cancel: CancellationToken,
329 :
330 : // Users of the Tenant such as the page service must take this Gate to avoid
331 : // trying to use a Tenant which is shutting down.
332 : pub(crate) gate: Gate,
333 :
334 : /// Throttle applied at the top of [`Timeline::get`].
335 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
336 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
337 :
338 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
339 :
340 : /// An ongoing timeline detach concurrency limiter.
341 : ///
342 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
343 : /// to have two running at the same time. A different one can be started if an earlier one
344 : /// has failed for whatever reason.
345 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
346 :
347 : /// `index_part.json` based gc blocking reason tracking.
348 : ///
349 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
350 : /// proceeding.
351 : pub(crate) gc_block: gc_block::GcBlock,
352 :
353 : l0_flush_global_state: L0FlushGlobalState,
354 : }
355 : impl std::fmt::Debug for Tenant {
356 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
357 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
358 0 : }
359 : }
360 :
361 : pub(crate) enum WalRedoManager {
362 : Prod(WalredoManagerId, PostgresRedoManager),
363 : #[cfg(test)]
364 : Test(harness::TestRedoManager),
365 : }
366 :
367 : #[derive(thiserror::Error, Debug)]
368 : #[error("pageserver is shutting down")]
369 : pub(crate) struct GlobalShutDown;
370 :
371 : impl WalRedoManager {
372 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
373 0 : let id = WalredoManagerId::next();
374 0 : let arc = Arc::new(Self::Prod(id, mgr));
375 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
376 0 : match &mut *guard {
377 0 : Some(map) => {
378 0 : map.insert(id, Arc::downgrade(&arc));
379 0 : Ok(arc)
380 : }
381 0 : None => Err(GlobalShutDown),
382 : }
383 0 : }
384 : }
385 :
386 : impl Drop for WalRedoManager {
387 20 : fn drop(&mut self) {
388 20 : match self {
389 0 : Self::Prod(id, _) => {
390 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
391 0 : if let Some(map) = &mut *guard {
392 0 : map.remove(id).expect("new() registers, drop() unregisters");
393 0 : }
394 : }
395 : #[cfg(test)]
396 20 : Self::Test(_) => {
397 20 : // Not applicable to test redo manager
398 20 : }
399 : }
400 20 : }
401 : }
402 :
403 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
404 : /// the walredo processes outside of the regular order.
405 : ///
406 : /// This is necessary to work around a systemd bug where it freezes if there are
407 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
408 : #[allow(clippy::type_complexity)]
409 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
410 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
411 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
412 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
413 : pub(crate) struct WalredoManagerId(u64);
414 : impl WalredoManagerId {
415 0 : pub fn next() -> Self {
416 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
417 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
418 0 : if id == 0 {
419 0 : panic!(
420 0 : "WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique"
421 0 : );
422 0 : }
423 0 : Self(id)
424 0 : }
425 : }
426 :
427 : #[cfg(test)]
428 : impl From<harness::TestRedoManager> for WalRedoManager {
429 444 : fn from(mgr: harness::TestRedoManager) -> Self {
430 444 : Self::Test(mgr)
431 444 : }
432 : }
433 :
434 : impl WalRedoManager {
435 12 : pub(crate) async fn shutdown(&self) -> bool {
436 12 : match self {
437 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
438 : #[cfg(test)]
439 : Self::Test(_) => {
440 : // Not applicable to test redo manager
441 12 : true
442 : }
443 : }
444 12 : }
445 :
446 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
447 0 : match self {
448 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
449 0 : #[cfg(test)]
450 0 : Self::Test(_) => {
451 0 : // Not applicable to test redo manager
452 0 : }
453 0 : }
454 0 : }
455 :
456 : /// # Cancel-Safety
457 : ///
458 : /// This method is cancellation-safe.
459 1636 : pub async fn request_redo(
460 1636 : &self,
461 1636 : key: pageserver_api::key::Key,
462 1636 : lsn: Lsn,
463 1636 : base_img: Option<(Lsn, bytes::Bytes)>,
464 1636 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
465 1636 : pg_version: u32,
466 1636 : ) -> Result<bytes::Bytes, walredo::Error> {
467 1636 : match self {
468 0 : Self::Prod(_, mgr) => {
469 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
470 0 : .await
471 : }
472 : #[cfg(test)]
473 1636 : Self::Test(mgr) => {
474 1636 : mgr.request_redo(key, lsn, base_img, records, pg_version)
475 1636 : .await
476 : }
477 : }
478 1636 : }
479 :
480 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
481 0 : match self {
482 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
483 0 : #[cfg(test)]
484 0 : WalRedoManager::Test(_) => None,
485 0 : }
486 0 : }
487 : }
488 :
489 : /// A very lightweight memory representation of an offloaded timeline.
490 : ///
491 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
492 : /// like unoffloading them, or (at a later date), decide to perform flattening.
493 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
494 : /// more offloaded timelines than we can manage ones that aren't.
495 : pub struct OffloadedTimeline {
496 : pub tenant_shard_id: TenantShardId,
497 : pub timeline_id: TimelineId,
498 : pub ancestor_timeline_id: Option<TimelineId>,
499 : /// Whether to retain the branch lsn at the ancestor or not
500 : pub ancestor_retain_lsn: Option<Lsn>,
501 :
502 : /// When the timeline was archived.
503 : ///
504 : /// Present for future flattening deliberations.
505 : pub archived_at: NaiveDateTime,
506 :
507 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
508 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
509 : pub delete_progress: TimelineDeleteProgress,
510 :
511 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
512 : pub deleted_from_ancestor: AtomicBool,
513 : }
514 :
515 : impl OffloadedTimeline {
516 : /// Obtains an offloaded timeline from a given timeline object.
517 : ///
518 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
519 : /// the timeline is not in a stopped state.
520 : /// Panics if the timeline is not archived.
521 4 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
522 4 : let (ancestor_retain_lsn, ancestor_timeline_id) =
523 4 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
524 4 : let ancestor_lsn = timeline.get_ancestor_lsn();
525 4 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
526 4 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
527 4 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
528 4 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
529 : } else {
530 0 : (None, None)
531 : };
532 4 : let archived_at = timeline
533 4 : .remote_client
534 4 : .archived_at_stopped_queue()?
535 4 : .expect("must be called on an archived timeline");
536 4 : Ok(Self {
537 4 : tenant_shard_id: timeline.tenant_shard_id,
538 4 : timeline_id: timeline.timeline_id,
539 4 : ancestor_timeline_id,
540 4 : ancestor_retain_lsn,
541 4 : archived_at,
542 4 :
543 4 : delete_progress: timeline.delete_progress.clone(),
544 4 : deleted_from_ancestor: AtomicBool::new(false),
545 4 : })
546 4 : }
547 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
548 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
549 0 : // by the `initialize_gc_info` function.
550 0 : let OffloadedTimelineManifest {
551 0 : timeline_id,
552 0 : ancestor_timeline_id,
553 0 : ancestor_retain_lsn,
554 0 : archived_at,
555 0 : } = *manifest;
556 0 : Self {
557 0 : tenant_shard_id,
558 0 : timeline_id,
559 0 : ancestor_timeline_id,
560 0 : ancestor_retain_lsn,
561 0 : archived_at,
562 0 : delete_progress: TimelineDeleteProgress::default(),
563 0 : deleted_from_ancestor: AtomicBool::new(false),
564 0 : }
565 0 : }
566 4 : fn manifest(&self) -> OffloadedTimelineManifest {
567 4 : let Self {
568 4 : timeline_id,
569 4 : ancestor_timeline_id,
570 4 : ancestor_retain_lsn,
571 4 : archived_at,
572 4 : ..
573 4 : } = self;
574 4 : OffloadedTimelineManifest {
575 4 : timeline_id: *timeline_id,
576 4 : ancestor_timeline_id: *ancestor_timeline_id,
577 4 : ancestor_retain_lsn: *ancestor_retain_lsn,
578 4 : archived_at: *archived_at,
579 4 : }
580 4 : }
581 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
582 0 : fn delete_from_ancestor_with_timelines(
583 0 : &self,
584 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
585 0 : ) {
586 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
587 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
588 : {
589 0 : if let Some((_, ancestor_timeline)) = timelines
590 0 : .iter()
591 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
592 : {
593 0 : let removal_happened = ancestor_timeline
594 0 : .gc_info
595 0 : .write()
596 0 : .unwrap()
597 0 : .remove_child_offloaded(self.timeline_id);
598 0 : if !removal_happened {
599 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
600 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
601 0 : }
602 0 : }
603 0 : }
604 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
605 0 : }
606 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
607 : ///
608 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
609 4 : fn defuse_for_tenant_drop(&self) {
610 4 : self.deleted_from_ancestor.store(true, Ordering::Release);
611 4 : }
612 : }
613 :
614 : impl fmt::Debug for OffloadedTimeline {
615 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
616 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
617 0 : }
618 : }
619 :
620 : impl Drop for OffloadedTimeline {
621 4 : fn drop(&mut self) {
622 4 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
623 0 : tracing::warn!(
624 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
625 : self.timeline_id
626 : );
627 4 : }
628 4 : }
629 : }
630 :
631 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
632 : pub enum MaybeOffloaded {
633 : Yes,
634 : No,
635 : }
636 :
637 : #[derive(Clone, Debug)]
638 : pub enum TimelineOrOffloaded {
639 : Timeline(Arc<Timeline>),
640 : Offloaded(Arc<OffloadedTimeline>),
641 : }
642 :
643 : impl TimelineOrOffloaded {
644 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
645 0 : match self {
646 0 : TimelineOrOffloaded::Timeline(timeline) => {
647 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
648 : }
649 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
650 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
651 : }
652 : }
653 0 : }
654 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
655 0 : self.arc_ref().tenant_shard_id()
656 0 : }
657 0 : pub fn timeline_id(&self) -> TimelineId {
658 0 : self.arc_ref().timeline_id()
659 0 : }
660 4 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
661 4 : match self {
662 4 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
663 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
664 : }
665 4 : }
666 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
667 0 : match self {
668 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
669 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
670 : }
671 0 : }
672 : }
673 :
674 : pub enum TimelineOrOffloadedArcRef<'a> {
675 : Timeline(&'a Arc<Timeline>),
676 : Offloaded(&'a Arc<OffloadedTimeline>),
677 : }
678 :
679 : impl TimelineOrOffloadedArcRef<'_> {
680 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
681 0 : match self {
682 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
683 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
684 : }
685 0 : }
686 0 : pub fn timeline_id(&self) -> TimelineId {
687 0 : match self {
688 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
689 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
690 : }
691 0 : }
692 : }
693 :
694 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
695 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
696 0 : Self::Timeline(timeline)
697 0 : }
698 : }
699 :
700 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
701 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
702 0 : Self::Offloaded(timeline)
703 0 : }
704 : }
705 :
706 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
707 : pub enum GetTimelineError {
708 : #[error("Timeline is shutting down")]
709 : ShuttingDown,
710 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
711 : NotActive {
712 : tenant_id: TenantShardId,
713 : timeline_id: TimelineId,
714 : state: TimelineState,
715 : },
716 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
717 : NotFound {
718 : tenant_id: TenantShardId,
719 : timeline_id: TimelineId,
720 : },
721 : }
722 :
723 : #[derive(Debug, thiserror::Error)]
724 : pub enum LoadLocalTimelineError {
725 : #[error("FailedToLoad")]
726 : Load(#[source] anyhow::Error),
727 : #[error("FailedToResumeDeletion")]
728 : ResumeDeletion(#[source] anyhow::Error),
729 : }
730 :
731 : #[derive(thiserror::Error)]
732 : pub enum DeleteTimelineError {
733 : #[error("NotFound")]
734 : NotFound,
735 :
736 : #[error("HasChildren")]
737 : HasChildren(Vec<TimelineId>),
738 :
739 : #[error("Timeline deletion is already in progress")]
740 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
741 :
742 : #[error("Cancelled")]
743 : Cancelled,
744 :
745 : #[error(transparent)]
746 : Other(#[from] anyhow::Error),
747 : }
748 :
749 : impl Debug for DeleteTimelineError {
750 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
751 0 : match self {
752 0 : Self::NotFound => write!(f, "NotFound"),
753 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
754 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
755 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
756 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
757 : }
758 0 : }
759 : }
760 :
761 : #[derive(thiserror::Error)]
762 : pub enum TimelineArchivalError {
763 : #[error("NotFound")]
764 : NotFound,
765 :
766 : #[error("Timeout")]
767 : Timeout,
768 :
769 : #[error("Cancelled")]
770 : Cancelled,
771 :
772 : #[error("ancestor is archived: {}", .0)]
773 : HasArchivedParent(TimelineId),
774 :
775 : #[error("HasUnarchivedChildren")]
776 : HasUnarchivedChildren(Vec<TimelineId>),
777 :
778 : #[error("Timeline archival is already in progress")]
779 : AlreadyInProgress,
780 :
781 : #[error(transparent)]
782 : Other(anyhow::Error),
783 : }
784 :
785 : #[derive(thiserror::Error, Debug)]
786 : pub(crate) enum TenantManifestError {
787 : #[error("Remote storage error: {0}")]
788 : RemoteStorage(anyhow::Error),
789 :
790 : #[error("Cancelled")]
791 : Cancelled,
792 : }
793 :
794 : impl From<TenantManifestError> for TimelineArchivalError {
795 0 : fn from(e: TenantManifestError) -> Self {
796 0 : match e {
797 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
798 0 : TenantManifestError::Cancelled => Self::Cancelled,
799 : }
800 0 : }
801 : }
802 :
803 : impl Debug for TimelineArchivalError {
804 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
805 0 : match self {
806 0 : Self::NotFound => write!(f, "NotFound"),
807 0 : Self::Timeout => write!(f, "Timeout"),
808 0 : Self::Cancelled => write!(f, "Cancelled"),
809 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
810 0 : Self::HasUnarchivedChildren(c) => {
811 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
812 : }
813 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
814 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
815 : }
816 0 : }
817 : }
818 :
819 : pub enum SetStoppingError {
820 : AlreadyStopping(completion::Barrier),
821 : Broken,
822 : }
823 :
824 : impl Debug for SetStoppingError {
825 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
826 0 : match self {
827 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
828 0 : Self::Broken => write!(f, "Broken"),
829 : }
830 0 : }
831 : }
832 :
833 : /// Arguments to [`Tenant::create_timeline`].
834 : ///
835 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
836 : /// is `None`, the result of the timeline create call is not deterministic.
837 : ///
838 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
839 : #[derive(Debug)]
840 : pub(crate) enum CreateTimelineParams {
841 : Bootstrap(CreateTimelineParamsBootstrap),
842 : Branch(CreateTimelineParamsBranch),
843 : ImportPgdata(CreateTimelineParamsImportPgdata),
844 : }
845 :
846 : #[derive(Debug)]
847 : pub(crate) struct CreateTimelineParamsBootstrap {
848 : pub(crate) new_timeline_id: TimelineId,
849 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
850 : pub(crate) pg_version: u32,
851 : }
852 :
853 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
854 : #[derive(Debug)]
855 : pub(crate) struct CreateTimelineParamsBranch {
856 : pub(crate) new_timeline_id: TimelineId,
857 : pub(crate) ancestor_timeline_id: TimelineId,
858 : pub(crate) ancestor_start_lsn: Option<Lsn>,
859 : }
860 :
861 : #[derive(Debug)]
862 : pub(crate) struct CreateTimelineParamsImportPgdata {
863 : pub(crate) new_timeline_id: TimelineId,
864 : pub(crate) location: import_pgdata::index_part_format::Location,
865 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
866 : }
867 :
868 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`] in [`Tenant::start_creating_timeline`].
869 : ///
870 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
871 : ///
872 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
873 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
874 : ///
875 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
876 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
877 : ///
878 : /// Notes:
879 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
880 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
881 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
882 : ///
883 : #[derive(Debug, Clone, PartialEq, Eq)]
884 : pub(crate) enum CreateTimelineIdempotency {
885 : /// NB: special treatment, see comment in [`Self`].
886 : FailWithConflict,
887 : Bootstrap {
888 : pg_version: u32,
889 : },
890 : /// NB: branches always have the same `pg_version` as their ancestor.
891 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
892 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
893 : /// determining the child branch pg_version.
894 : Branch {
895 : ancestor_timeline_id: TimelineId,
896 : ancestor_start_lsn: Lsn,
897 : },
898 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
899 : }
900 :
901 : #[derive(Debug, Clone, PartialEq, Eq)]
902 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
903 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
904 : }
905 :
906 : /// What is returned by [`Tenant::start_creating_timeline`].
907 : #[must_use]
908 : enum StartCreatingTimelineResult {
909 : CreateGuard(TimelineCreateGuard),
910 : Idempotent(Arc<Timeline>),
911 : }
912 :
913 : enum TimelineInitAndSyncResult {
914 : ReadyToActivate(Arc<Timeline>),
915 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
916 : }
917 :
918 : impl TimelineInitAndSyncResult {
919 0 : fn ready_to_activate(self) -> Option<Arc<Timeline>> {
920 0 : match self {
921 0 : Self::ReadyToActivate(timeline) => Some(timeline),
922 0 : _ => None,
923 : }
924 0 : }
925 : }
926 :
927 : #[must_use]
928 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
929 : timeline: Arc<Timeline>,
930 : import_pgdata: import_pgdata::index_part_format::Root,
931 : guard: TimelineCreateGuard,
932 : }
933 :
934 : /// What is returned by [`Tenant::create_timeline`].
935 : enum CreateTimelineResult {
936 : Created(Arc<Timeline>),
937 : Idempotent(Arc<Timeline>),
938 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`Tenant::timelines`] when
939 : /// we return this result, nor will this concrete object ever be added there.
940 : /// Cf method comment on [`Tenant::create_timeline_import_pgdata`].
941 : ImportSpawned(Arc<Timeline>),
942 : }
943 :
944 : impl CreateTimelineResult {
945 0 : fn discriminant(&self) -> &'static str {
946 0 : match self {
947 0 : Self::Created(_) => "Created",
948 0 : Self::Idempotent(_) => "Idempotent",
949 0 : Self::ImportSpawned(_) => "ImportSpawned",
950 : }
951 0 : }
952 0 : fn timeline(&self) -> &Arc<Timeline> {
953 0 : match self {
954 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
955 0 : }
956 0 : }
957 : /// Unit test timelines aren't activated, test has to do it if it needs to.
958 : #[cfg(test)]
959 460 : fn into_timeline_for_test(self) -> Arc<Timeline> {
960 460 : match self {
961 460 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
962 460 : }
963 460 : }
964 : }
965 :
966 : #[derive(thiserror::Error, Debug)]
967 : pub enum CreateTimelineError {
968 : #[error("creation of timeline with the given ID is in progress")]
969 : AlreadyCreating,
970 : #[error("timeline already exists with different parameters")]
971 : Conflict,
972 : #[error(transparent)]
973 : AncestorLsn(anyhow::Error),
974 : #[error("ancestor timeline is not active")]
975 : AncestorNotActive,
976 : #[error("ancestor timeline is archived")]
977 : AncestorArchived,
978 : #[error("tenant shutting down")]
979 : ShuttingDown,
980 : #[error(transparent)]
981 : Other(#[from] anyhow::Error),
982 : }
983 :
984 : #[derive(thiserror::Error, Debug)]
985 : pub enum InitdbError {
986 : #[error("Operation was cancelled")]
987 : Cancelled,
988 : #[error(transparent)]
989 : Other(anyhow::Error),
990 : #[error(transparent)]
991 : Inner(postgres_initdb::Error),
992 : }
993 :
994 : enum CreateTimelineCause {
995 : Load,
996 : Delete,
997 : }
998 :
999 : enum LoadTimelineCause {
1000 : Attach,
1001 : Unoffload,
1002 : ImportPgdata {
1003 : create_guard: TimelineCreateGuard,
1004 : activate: ActivateTimelineArgs,
1005 : },
1006 : }
1007 :
1008 : #[derive(thiserror::Error, Debug)]
1009 : pub(crate) enum GcError {
1010 : // The tenant is shutting down
1011 : #[error("tenant shutting down")]
1012 : TenantCancelled,
1013 :
1014 : // The tenant is shutting down
1015 : #[error("timeline shutting down")]
1016 : TimelineCancelled,
1017 :
1018 : // The tenant is in a state inelegible to run GC
1019 : #[error("not active")]
1020 : NotActive,
1021 :
1022 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1023 : #[error("not active")]
1024 : BadLsn { why: String },
1025 :
1026 : // A remote storage error while scheduling updates after compaction
1027 : #[error(transparent)]
1028 : Remote(anyhow::Error),
1029 :
1030 : // An error reading while calculating GC cutoffs
1031 : #[error(transparent)]
1032 : GcCutoffs(PageReconstructError),
1033 :
1034 : // If GC was invoked for a particular timeline, this error means it didn't exist
1035 : #[error("timeline not found")]
1036 : TimelineNotFound,
1037 : }
1038 :
1039 : impl From<PageReconstructError> for GcError {
1040 0 : fn from(value: PageReconstructError) -> Self {
1041 0 : match value {
1042 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1043 0 : other => Self::GcCutoffs(other),
1044 : }
1045 0 : }
1046 : }
1047 :
1048 : impl From<NotInitialized> for GcError {
1049 0 : fn from(value: NotInitialized) -> Self {
1050 0 : match value {
1051 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1052 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1053 : }
1054 0 : }
1055 : }
1056 :
1057 : impl From<timeline::layer_manager::Shutdown> for GcError {
1058 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1059 0 : GcError::TimelineCancelled
1060 0 : }
1061 : }
1062 :
1063 : #[derive(thiserror::Error, Debug)]
1064 : pub(crate) enum LoadConfigError {
1065 : #[error("TOML deserialization error: '{0}'")]
1066 : DeserializeToml(#[from] toml_edit::de::Error),
1067 :
1068 : #[error("Config not found at {0}")]
1069 : NotFound(Utf8PathBuf),
1070 : }
1071 :
1072 : impl Tenant {
1073 : /// Yet another helper for timeline initialization.
1074 : ///
1075 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1076 : /// - Scans the local timeline directory for layer files and builds the layer map
1077 : /// - Downloads remote index file and adds remote files to the layer map
1078 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1079 : ///
1080 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1081 : /// it is marked as Active.
1082 : #[allow(clippy::too_many_arguments)]
1083 12 : async fn timeline_init_and_sync(
1084 12 : self: &Arc<Self>,
1085 12 : timeline_id: TimelineId,
1086 12 : resources: TimelineResources,
1087 12 : mut index_part: IndexPart,
1088 12 : metadata: TimelineMetadata,
1089 12 : previous_heatmap: Option<PreviousHeatmap>,
1090 12 : ancestor: Option<Arc<Timeline>>,
1091 12 : cause: LoadTimelineCause,
1092 12 : ctx: &RequestContext,
1093 12 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1094 12 : let tenant_id = self.tenant_shard_id;
1095 12 :
1096 12 : let import_pgdata = index_part.import_pgdata.take();
1097 12 : let idempotency = match &import_pgdata {
1098 0 : Some(import_pgdata) => {
1099 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1100 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1101 0 : })
1102 : }
1103 : None => {
1104 12 : if metadata.ancestor_timeline().is_none() {
1105 8 : CreateTimelineIdempotency::Bootstrap {
1106 8 : pg_version: metadata.pg_version(),
1107 8 : }
1108 : } else {
1109 4 : CreateTimelineIdempotency::Branch {
1110 4 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1111 4 : ancestor_start_lsn: metadata.ancestor_lsn(),
1112 4 : }
1113 : }
1114 : }
1115 : };
1116 :
1117 12 : let timeline = self.create_timeline_struct(
1118 12 : timeline_id,
1119 12 : &metadata,
1120 12 : previous_heatmap,
1121 12 : ancestor.clone(),
1122 12 : resources,
1123 12 : CreateTimelineCause::Load,
1124 12 : idempotency.clone(),
1125 12 : index_part.gc_compaction.clone(),
1126 12 : )?;
1127 12 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1128 12 : anyhow::ensure!(
1129 12 : disk_consistent_lsn.is_valid(),
1130 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1131 : );
1132 12 : assert_eq!(
1133 12 : disk_consistent_lsn,
1134 12 : metadata.disk_consistent_lsn(),
1135 0 : "these are used interchangeably"
1136 : );
1137 :
1138 12 : timeline.remote_client.init_upload_queue(&index_part)?;
1139 :
1140 12 : timeline
1141 12 : .load_layer_map(disk_consistent_lsn, index_part)
1142 12 : .await
1143 12 : .with_context(|| {
1144 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1145 12 : })?;
1146 :
1147 : // When unarchiving, we've mostly likely lost the heatmap generated prior
1148 : // to the archival operation. To allow warming this timeline up, generate
1149 : // a previous heatmap which contains all visible layers in the layer map.
1150 : // This previous heatmap will be used whenever a fresh heatmap is generated
1151 : // for the timeline.
1152 12 : if matches!(cause, LoadTimelineCause::Unoffload) {
1153 0 : let mut tline_ending_at = Some((&timeline, timeline.get_last_record_lsn()));
1154 0 : while let Some((tline, end_lsn)) = tline_ending_at {
1155 0 : let unarchival_heatmap = tline.generate_unarchival_heatmap(end_lsn).await;
1156 : // Another unearchived timeline might have generated a heatmap for this ancestor.
1157 : // If the current branch point greater than the previous one use the the heatmap
1158 : // we just generated - it should include more layers.
1159 0 : if !tline.should_keep_previous_heatmap(end_lsn) {
1160 0 : tline
1161 0 : .previous_heatmap
1162 0 : .store(Some(Arc::new(unarchival_heatmap)));
1163 0 : } else {
1164 0 : tracing::info!("Previous heatmap preferred. Dropping unarchival heatmap.")
1165 : }
1166 :
1167 0 : match tline.ancestor_timeline() {
1168 0 : Some(ancestor) => {
1169 0 : if ancestor.update_layer_visibility().await.is_err() {
1170 : // Ancestor timeline is shutting down.
1171 0 : break;
1172 0 : }
1173 0 :
1174 0 : tline_ending_at = Some((ancestor, tline.get_ancestor_lsn()));
1175 : }
1176 0 : None => {
1177 0 : tline_ending_at = None;
1178 0 : }
1179 : }
1180 : }
1181 12 : }
1182 :
1183 0 : match import_pgdata {
1184 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1185 0 : match cause {
1186 0 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1187 : LoadTimelineCause::ImportPgdata { .. } => {
1188 0 : unreachable!(
1189 0 : "ImportPgdata should not be reloading timeline import is done and persisted as such in s3"
1190 0 : )
1191 : }
1192 : }
1193 0 : let mut guard = self.timelines_creating.lock().unwrap();
1194 0 : if !guard.insert(timeline_id) {
1195 : // We should never try and load the same timeline twice during startup
1196 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1197 0 : }
1198 0 : let timeline_create_guard = TimelineCreateGuard {
1199 0 : _tenant_gate_guard: self.gate.enter()?,
1200 0 : owning_tenant: self.clone(),
1201 0 : timeline_id,
1202 0 : idempotency,
1203 0 : // The users of this specific return value don't need the timline_path in there.
1204 0 : timeline_path: timeline
1205 0 : .conf
1206 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1207 0 : };
1208 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1209 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1210 0 : timeline,
1211 0 : import_pgdata,
1212 0 : guard: timeline_create_guard,
1213 0 : },
1214 0 : ))
1215 : }
1216 : Some(_) | None => {
1217 : {
1218 12 : let mut timelines_accessor = self.timelines.lock().unwrap();
1219 12 : match timelines_accessor.entry(timeline_id) {
1220 : // We should never try and load the same timeline twice during startup
1221 : Entry::Occupied(_) => {
1222 0 : unreachable!(
1223 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1224 0 : );
1225 : }
1226 12 : Entry::Vacant(v) => {
1227 12 : v.insert(Arc::clone(&timeline));
1228 12 : timeline.maybe_spawn_flush_loop();
1229 12 : }
1230 : }
1231 : }
1232 :
1233 : // Sanity check: a timeline should have some content.
1234 12 : anyhow::ensure!(
1235 12 : ancestor.is_some()
1236 8 : || timeline
1237 8 : .layers
1238 8 : .read()
1239 8 : .await
1240 8 : .layer_map()
1241 8 : .expect("currently loading, layer manager cannot be shutdown already")
1242 8 : .iter_historic_layers()
1243 8 : .next()
1244 8 : .is_some(),
1245 0 : "Timeline has no ancestor and no layer files"
1246 : );
1247 :
1248 12 : match cause {
1249 12 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1250 : LoadTimelineCause::ImportPgdata {
1251 0 : create_guard,
1252 0 : activate,
1253 0 : } => {
1254 0 : // TODO: see the comment in the task code above how I'm not so certain
1255 0 : // it is safe to activate here because of concurrent shutdowns.
1256 0 : match activate {
1257 0 : ActivateTimelineArgs::Yes { broker_client } => {
1258 0 : info!("activating timeline after reload from pgdata import task");
1259 0 : timeline.activate(self.clone(), broker_client, None, ctx);
1260 : }
1261 0 : ActivateTimelineArgs::No => (),
1262 : }
1263 0 : drop(create_guard);
1264 : }
1265 : }
1266 :
1267 12 : Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
1268 : }
1269 : }
1270 12 : }
1271 :
1272 : /// Attach a tenant that's available in cloud storage.
1273 : ///
1274 : /// This returns quickly, after just creating the in-memory object
1275 : /// Tenant struct and launching a background task to download
1276 : /// the remote index files. On return, the tenant is most likely still in
1277 : /// Attaching state, and it will become Active once the background task
1278 : /// finishes. You can use wait_until_active() to wait for the task to
1279 : /// complete.
1280 : ///
1281 : #[allow(clippy::too_many_arguments)]
1282 0 : pub(crate) fn spawn(
1283 0 : conf: &'static PageServerConf,
1284 0 : tenant_shard_id: TenantShardId,
1285 0 : resources: TenantSharedResources,
1286 0 : attached_conf: AttachedTenantConf,
1287 0 : shard_identity: ShardIdentity,
1288 0 : init_order: Option<InitializationOrder>,
1289 0 : mode: SpawnMode,
1290 0 : ctx: &RequestContext,
1291 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1292 0 : let wal_redo_manager =
1293 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1294 :
1295 : let TenantSharedResources {
1296 0 : broker_client,
1297 0 : remote_storage,
1298 0 : deletion_queue_client,
1299 0 : l0_flush_global_state,
1300 0 : } = resources;
1301 0 :
1302 0 : let attach_mode = attached_conf.location.attach_mode;
1303 0 : let generation = attached_conf.location.generation;
1304 0 :
1305 0 : let tenant = Arc::new(Tenant::new(
1306 0 : TenantState::Attaching,
1307 0 : conf,
1308 0 : attached_conf,
1309 0 : shard_identity,
1310 0 : Some(wal_redo_manager),
1311 0 : tenant_shard_id,
1312 0 : remote_storage.clone(),
1313 0 : deletion_queue_client,
1314 0 : l0_flush_global_state,
1315 0 : ));
1316 0 :
1317 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1318 0 : // we shut down while attaching.
1319 0 : let attach_gate_guard = tenant
1320 0 : .gate
1321 0 : .enter()
1322 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1323 0 :
1324 0 : // Do all the hard work in the background
1325 0 : let tenant_clone = Arc::clone(&tenant);
1326 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1327 0 : task_mgr::spawn(
1328 0 : &tokio::runtime::Handle::current(),
1329 0 : TaskKind::Attach,
1330 0 : tenant_shard_id,
1331 0 : None,
1332 0 : "attach tenant",
1333 0 : async move {
1334 0 :
1335 0 : info!(
1336 : ?attach_mode,
1337 0 : "Attaching tenant"
1338 : );
1339 :
1340 0 : let _gate_guard = attach_gate_guard;
1341 0 :
1342 0 : // Is this tenant being spawned as part of process startup?
1343 0 : let starting_up = init_order.is_some();
1344 0 : scopeguard::defer! {
1345 0 : if starting_up {
1346 0 : TENANT.startup_complete.inc();
1347 0 : }
1348 0 : }
1349 :
1350 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1351 : enum BrokenVerbosity {
1352 : Error,
1353 : Info
1354 : }
1355 0 : let make_broken =
1356 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1357 0 : match verbosity {
1358 : BrokenVerbosity::Info => {
1359 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1360 : },
1361 : BrokenVerbosity::Error => {
1362 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1363 : }
1364 : }
1365 0 : t.state.send_modify(|state| {
1366 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1367 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1368 0 : assert!(
1369 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1370 0 : "the attach task owns the tenant state until activation is complete"
1371 : );
1372 :
1373 0 : *state = TenantState::broken_from_reason(err.to_string());
1374 0 : });
1375 0 : };
1376 :
1377 : // TODO: should also be rejecting tenant conf changes that violate this check.
1378 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1379 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1380 0 : return Ok(());
1381 0 : }
1382 0 :
1383 0 : let mut init_order = init_order;
1384 0 : // take the completion because initial tenant loading will complete when all of
1385 0 : // these tasks complete.
1386 0 : let _completion = init_order
1387 0 : .as_mut()
1388 0 : .and_then(|x| x.initial_tenant_load.take());
1389 0 : let remote_load_completion = init_order
1390 0 : .as_mut()
1391 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1392 :
1393 : enum AttachType<'a> {
1394 : /// We are attaching this tenant lazily in the background.
1395 : Warmup {
1396 : _permit: tokio::sync::SemaphorePermit<'a>,
1397 : during_startup: bool
1398 : },
1399 : /// We are attaching this tenant as soon as we can, because for example an
1400 : /// endpoint tried to access it.
1401 : OnDemand,
1402 : /// During normal operations after startup, we are attaching a tenant, and
1403 : /// eager attach was requested.
1404 : Normal,
1405 : }
1406 :
1407 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1408 : // Before doing any I/O, wait for at least one of:
1409 : // - A client attempting to access to this tenant (on-demand loading)
1410 : // - A permit becoming available in the warmup semaphore (background warmup)
1411 :
1412 0 : tokio::select!(
1413 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1414 0 : let _ = permit.expect("activate_now_sem is never closed");
1415 0 : tracing::info!("Activating tenant (on-demand)");
1416 0 : AttachType::OnDemand
1417 : },
1418 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1419 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1420 0 : tracing::info!("Activating tenant (warmup)");
1421 0 : AttachType::Warmup {
1422 0 : _permit,
1423 0 : during_startup: init_order.is_some()
1424 0 : }
1425 : }
1426 0 : _ = tenant_clone.cancel.cancelled() => {
1427 : // This is safe, but should be pretty rare: it is interesting if a tenant
1428 : // stayed in Activating for such a long time that shutdown found it in
1429 : // that state.
1430 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1431 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1432 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1433 : // just shutting down), but ensures progress.
1434 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1435 0 : return Ok(());
1436 : },
1437 : )
1438 : } else {
1439 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1440 : // concurrent_tenant_warmup queue
1441 0 : AttachType::Normal
1442 : };
1443 :
1444 0 : let preload = match &mode {
1445 : SpawnMode::Eager | SpawnMode::Lazy => {
1446 0 : let _preload_timer = TENANT.preload.start_timer();
1447 0 : let res = tenant_clone
1448 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1449 0 : .await;
1450 0 : match res {
1451 0 : Ok(p) => Some(p),
1452 0 : Err(e) => {
1453 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1454 0 : return Ok(());
1455 : }
1456 : }
1457 : }
1458 :
1459 : };
1460 :
1461 : // Remote preload is complete.
1462 0 : drop(remote_load_completion);
1463 0 :
1464 0 :
1465 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1466 0 : let attach_start = std::time::Instant::now();
1467 0 : let attached = {
1468 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1469 0 : tenant_clone.attach(preload, &ctx).await
1470 : };
1471 0 : let attach_duration = attach_start.elapsed();
1472 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1473 0 :
1474 0 : match attached {
1475 : Ok(()) => {
1476 0 : info!("attach finished, activating");
1477 0 : tenant_clone.activate(broker_client, None, &ctx);
1478 : }
1479 0 : Err(e) => {
1480 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1481 0 : }
1482 : }
1483 :
1484 : // If we are doing an opportunistic warmup attachment at startup, initialize
1485 : // logical size at the same time. This is better than starting a bunch of idle tenants
1486 : // with cold caches and then coming back later to initialize their logical sizes.
1487 : //
1488 : // It also prevents the warmup proccess competing with the concurrency limit on
1489 : // logical size calculations: if logical size calculation semaphore is saturated,
1490 : // then warmup will wait for that before proceeding to the next tenant.
1491 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1492 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1493 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1494 0 : while futs.next().await.is_some() {}
1495 0 : tracing::info!("Warm-up complete");
1496 0 : }
1497 :
1498 0 : Ok(())
1499 0 : }
1500 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1501 : );
1502 0 : Ok(tenant)
1503 0 : }
1504 :
1505 : #[instrument(skip_all)]
1506 : pub(crate) async fn preload(
1507 : self: &Arc<Self>,
1508 : remote_storage: &GenericRemoteStorage,
1509 : cancel: CancellationToken,
1510 : ) -> anyhow::Result<TenantPreload> {
1511 : span::debug_assert_current_span_has_tenant_id();
1512 : // Get list of remote timelines
1513 : // download index files for every tenant timeline
1514 : info!("listing remote timelines");
1515 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1516 : remote_storage,
1517 : self.tenant_shard_id,
1518 : cancel.clone(),
1519 : )
1520 : .await?;
1521 : let (offloaded_add, tenant_manifest) =
1522 : match remote_timeline_client::download_tenant_manifest(
1523 : remote_storage,
1524 : &self.tenant_shard_id,
1525 : self.generation,
1526 : &cancel,
1527 : )
1528 : .await
1529 : {
1530 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1531 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1532 : tenant_manifest,
1533 : ),
1534 : Err(DownloadError::NotFound) => {
1535 : ("no manifest".to_string(), TenantManifest::empty())
1536 : }
1537 : Err(e) => Err(e)?,
1538 : };
1539 :
1540 : info!(
1541 : "found {} timelines, and {offloaded_add}",
1542 : remote_timeline_ids.len()
1543 : );
1544 :
1545 : for k in other_keys {
1546 : warn!("Unexpected non timeline key {k}");
1547 : }
1548 :
1549 : // Avoid downloading IndexPart of offloaded timelines.
1550 : let mut offloaded_with_prefix = HashSet::new();
1551 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1552 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1553 : offloaded_with_prefix.insert(offloaded.timeline_id);
1554 : } else {
1555 : // We'll take care later of timelines in the manifest without a prefix
1556 : }
1557 : }
1558 :
1559 : // TODO(vlad): Could go to S3 if the secondary is freezing cold and hasn't even
1560 : // pulled the first heatmap. Not entirely necessary since the storage controller
1561 : // will kick the secondary in any case and cause a download.
1562 : let maybe_heatmap_at = self.read_on_disk_heatmap().await;
1563 :
1564 : let timelines = self
1565 : .load_timelines_metadata(
1566 : remote_timeline_ids,
1567 : remote_storage,
1568 : maybe_heatmap_at,
1569 : cancel,
1570 : )
1571 : .await?;
1572 :
1573 : Ok(TenantPreload {
1574 : tenant_manifest,
1575 : timelines: timelines
1576 : .into_iter()
1577 12 : .map(|(id, tl)| (id, Some(tl)))
1578 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1579 : .collect(),
1580 : })
1581 : }
1582 :
1583 444 : async fn read_on_disk_heatmap(&self) -> Option<(HeatMapTenant, std::time::Instant)> {
1584 444 : let on_disk_heatmap_path = self.conf.tenant_heatmap_path(&self.tenant_shard_id);
1585 444 : match tokio::fs::read_to_string(on_disk_heatmap_path).await {
1586 0 : Ok(heatmap) => match serde_json::from_str::<HeatMapTenant>(&heatmap) {
1587 0 : Ok(heatmap) => Some((heatmap, std::time::Instant::now())),
1588 0 : Err(err) => {
1589 0 : error!("Failed to deserialize old heatmap: {err}");
1590 0 : None
1591 : }
1592 : },
1593 444 : Err(err) => match err.kind() {
1594 444 : std::io::ErrorKind::NotFound => None,
1595 : _ => {
1596 0 : error!("Unexpected IO error reading old heatmap: {err}");
1597 0 : None
1598 : }
1599 : },
1600 : }
1601 444 : }
1602 :
1603 : ///
1604 : /// Background task that downloads all data for a tenant and brings it to Active state.
1605 : ///
1606 : /// No background tasks are started as part of this routine.
1607 : ///
1608 444 : async fn attach(
1609 444 : self: &Arc<Tenant>,
1610 444 : preload: Option<TenantPreload>,
1611 444 : ctx: &RequestContext,
1612 444 : ) -> anyhow::Result<()> {
1613 444 : span::debug_assert_current_span_has_tenant_id();
1614 444 :
1615 444 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1616 :
1617 444 : let Some(preload) = preload else {
1618 0 : anyhow::bail!(
1619 0 : "local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624"
1620 0 : );
1621 : };
1622 :
1623 444 : let mut offloaded_timeline_ids = HashSet::new();
1624 444 : let mut offloaded_timelines_list = Vec::new();
1625 444 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1626 0 : let timeline_id = timeline_manifest.timeline_id;
1627 0 : let offloaded_timeline =
1628 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1629 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1630 0 : offloaded_timeline_ids.insert(timeline_id);
1631 0 : }
1632 : // Complete deletions for offloaded timeline id's from manifest.
1633 : // The manifest will be uploaded later in this function.
1634 444 : offloaded_timelines_list
1635 444 : .retain(|(offloaded_id, offloaded)| {
1636 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1637 0 : // If there is dangling references in another location, they need to be cleaned up.
1638 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1639 0 : if delete {
1640 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1641 0 : offloaded.defuse_for_tenant_drop();
1642 0 : }
1643 0 : !delete
1644 444 : });
1645 444 :
1646 444 : let mut timelines_to_resume_deletions = vec![];
1647 444 :
1648 444 : let mut remote_index_and_client = HashMap::new();
1649 444 : let mut timeline_ancestors = HashMap::new();
1650 444 : let mut existent_timelines = HashSet::new();
1651 456 : for (timeline_id, preload) in preload.timelines {
1652 12 : let Some(preload) = preload else { continue };
1653 : // This is an invariant of the `preload` function's API
1654 12 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1655 12 : let index_part = match preload.index_part {
1656 12 : Ok(i) => {
1657 12 : debug!("remote index part exists for timeline {timeline_id}");
1658 : // We found index_part on the remote, this is the standard case.
1659 12 : existent_timelines.insert(timeline_id);
1660 12 : i
1661 : }
1662 : Err(DownloadError::NotFound) => {
1663 : // There is no index_part on the remote. We only get here
1664 : // if there is some prefix for the timeline in the remote storage.
1665 : // This can e.g. be the initdb.tar.zst archive, maybe a
1666 : // remnant from a prior incomplete creation or deletion attempt.
1667 : // Delete the local directory as the deciding criterion for a
1668 : // timeline's existence is presence of index_part.
1669 0 : info!(%timeline_id, "index_part not found on remote");
1670 0 : continue;
1671 : }
1672 0 : Err(DownloadError::Fatal(why)) => {
1673 0 : // If, while loading one remote timeline, we saw an indication that our generation
1674 0 : // number is likely invalid, then we should not load the whole tenant.
1675 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1676 0 : anyhow::bail!(why.to_string());
1677 : }
1678 0 : Err(e) => {
1679 0 : // Some (possibly ephemeral) error happened during index_part download.
1680 0 : // Pretend the timeline exists to not delete the timeline directory,
1681 0 : // as it might be a temporary issue and we don't want to re-download
1682 0 : // everything after it resolves.
1683 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1684 :
1685 0 : existent_timelines.insert(timeline_id);
1686 0 : continue;
1687 : }
1688 : };
1689 12 : match index_part {
1690 12 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1691 12 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1692 12 : remote_index_and_client.insert(
1693 12 : timeline_id,
1694 12 : (index_part, preload.client, preload.previous_heatmap),
1695 12 : );
1696 12 : }
1697 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1698 0 : info!(
1699 0 : "timeline {} is deleted, picking to resume deletion",
1700 : timeline_id
1701 : );
1702 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1703 : }
1704 : }
1705 : }
1706 :
1707 444 : let mut gc_blocks = HashMap::new();
1708 :
1709 : // For every timeline, download the metadata file, scan the local directory,
1710 : // and build a layer map that contains an entry for each remote and local
1711 : // layer file.
1712 444 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1713 456 : for (timeline_id, remote_metadata) in sorted_timelines {
1714 12 : let (index_part, remote_client, previous_heatmap) = remote_index_and_client
1715 12 : .remove(&timeline_id)
1716 12 : .expect("just put it in above");
1717 :
1718 12 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1719 : // could just filter these away, but it helps while testing
1720 0 : anyhow::ensure!(
1721 0 : !blocking.reasons.is_empty(),
1722 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1723 : );
1724 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1725 0 : assert!(prev.is_none());
1726 12 : }
1727 :
1728 : // TODO again handle early failure
1729 12 : let effect = self
1730 12 : .load_remote_timeline(
1731 12 : timeline_id,
1732 12 : index_part,
1733 12 : remote_metadata,
1734 12 : previous_heatmap,
1735 12 : self.get_timeline_resources_for(remote_client),
1736 12 : LoadTimelineCause::Attach,
1737 12 : ctx,
1738 12 : )
1739 12 : .await
1740 12 : .with_context(|| {
1741 0 : format!(
1742 0 : "failed to load remote timeline {} for tenant {}",
1743 0 : timeline_id, self.tenant_shard_id
1744 0 : )
1745 12 : })?;
1746 :
1747 12 : match effect {
1748 12 : TimelineInitAndSyncResult::ReadyToActivate(_) => {
1749 12 : // activation happens later, on Tenant::activate
1750 12 : }
1751 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1752 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1753 0 : timeline,
1754 0 : import_pgdata,
1755 0 : guard,
1756 0 : },
1757 0 : ) => {
1758 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1759 0 : timeline,
1760 0 : import_pgdata,
1761 0 : ActivateTimelineArgs::No,
1762 0 : guard,
1763 0 : ));
1764 0 : }
1765 : }
1766 : }
1767 :
1768 : // Walk through deleted timelines, resume deletion
1769 444 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1770 0 : remote_timeline_client
1771 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1772 0 : .context("init queue stopped")
1773 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1774 :
1775 0 : DeleteTimelineFlow::resume_deletion(
1776 0 : Arc::clone(self),
1777 0 : timeline_id,
1778 0 : &index_part.metadata,
1779 0 : remote_timeline_client,
1780 0 : )
1781 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1782 0 : .await
1783 0 : .context("resume_deletion")
1784 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1785 : }
1786 444 : let needs_manifest_upload =
1787 444 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1788 444 : {
1789 444 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1790 444 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1791 444 : }
1792 444 : if needs_manifest_upload {
1793 0 : self.store_tenant_manifest().await?;
1794 444 : }
1795 :
1796 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1797 : // IndexPart is the source of truth.
1798 444 : self.clean_up_timelines(&existent_timelines)?;
1799 :
1800 444 : self.gc_block.set_scanned(gc_blocks);
1801 444 :
1802 444 : fail::fail_point!("attach-before-activate", |_| {
1803 0 : anyhow::bail!("attach-before-activate");
1804 444 : });
1805 444 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1806 :
1807 444 : info!("Done");
1808 :
1809 444 : Ok(())
1810 444 : }
1811 :
1812 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1813 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1814 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1815 444 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1816 444 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1817 :
1818 444 : let entries = match timelines_dir.read_dir_utf8() {
1819 444 : Ok(d) => d,
1820 0 : Err(e) => {
1821 0 : if e.kind() == std::io::ErrorKind::NotFound {
1822 0 : return Ok(());
1823 : } else {
1824 0 : return Err(e).context("list timelines directory for tenant");
1825 : }
1826 : }
1827 : };
1828 :
1829 460 : for entry in entries {
1830 16 : let entry = entry.context("read timeline dir entry")?;
1831 16 : let entry_path = entry.path();
1832 :
1833 16 : let purge = if crate::is_temporary(entry_path) {
1834 0 : true
1835 : } else {
1836 16 : match TimelineId::try_from(entry_path.file_name()) {
1837 16 : Ok(i) => {
1838 16 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1839 16 : !existent_timelines.contains(&i)
1840 : }
1841 0 : Err(e) => {
1842 0 : tracing::warn!(
1843 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1844 : );
1845 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1846 0 : false
1847 : }
1848 : }
1849 : };
1850 :
1851 16 : if purge {
1852 4 : tracing::info!("Purging stale timeline dentry {entry_path}");
1853 4 : if let Err(e) = match entry.file_type() {
1854 4 : Ok(t) => if t.is_dir() {
1855 4 : std::fs::remove_dir_all(entry_path)
1856 : } else {
1857 0 : std::fs::remove_file(entry_path)
1858 : }
1859 4 : .or_else(fs_ext::ignore_not_found),
1860 0 : Err(e) => Err(e),
1861 : } {
1862 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1863 4 : }
1864 12 : }
1865 : }
1866 :
1867 444 : Ok(())
1868 444 : }
1869 :
1870 : /// Get sum of all remote timelines sizes
1871 : ///
1872 : /// This function relies on the index_part instead of listing the remote storage
1873 0 : pub fn remote_size(&self) -> u64 {
1874 0 : let mut size = 0;
1875 :
1876 0 : for timeline in self.list_timelines() {
1877 0 : size += timeline.remote_client.get_remote_physical_size();
1878 0 : }
1879 :
1880 0 : size
1881 0 : }
1882 :
1883 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1884 : #[allow(clippy::too_many_arguments)]
1885 : async fn load_remote_timeline(
1886 : self: &Arc<Self>,
1887 : timeline_id: TimelineId,
1888 : index_part: IndexPart,
1889 : remote_metadata: TimelineMetadata,
1890 : previous_heatmap: Option<PreviousHeatmap>,
1891 : resources: TimelineResources,
1892 : cause: LoadTimelineCause,
1893 : ctx: &RequestContext,
1894 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1895 : span::debug_assert_current_span_has_tenant_id();
1896 :
1897 : info!("downloading index file for timeline {}", timeline_id);
1898 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1899 : .await
1900 : .context("Failed to create new timeline directory")?;
1901 :
1902 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1903 : let timelines = self.timelines.lock().unwrap();
1904 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1905 0 : || {
1906 0 : anyhow::anyhow!(
1907 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1908 0 : )
1909 0 : },
1910 : )?))
1911 : } else {
1912 : None
1913 : };
1914 :
1915 : self.timeline_init_and_sync(
1916 : timeline_id,
1917 : resources,
1918 : index_part,
1919 : remote_metadata,
1920 : previous_heatmap,
1921 : ancestor,
1922 : cause,
1923 : ctx,
1924 : )
1925 : .await
1926 : }
1927 :
1928 444 : async fn load_timelines_metadata(
1929 444 : self: &Arc<Tenant>,
1930 444 : timeline_ids: HashSet<TimelineId>,
1931 444 : remote_storage: &GenericRemoteStorage,
1932 444 : heatmap: Option<(HeatMapTenant, std::time::Instant)>,
1933 444 : cancel: CancellationToken,
1934 444 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1935 444 : let mut timeline_heatmaps = heatmap.map(|h| (h.0.into_timelines_index(), h.1));
1936 444 :
1937 444 : let mut part_downloads = JoinSet::new();
1938 456 : for timeline_id in timeline_ids {
1939 12 : let cancel_clone = cancel.clone();
1940 12 :
1941 12 : let previous_timeline_heatmap = timeline_heatmaps.as_mut().and_then(|hs| {
1942 0 : hs.0.remove(&timeline_id).map(|h| PreviousHeatmap::Active {
1943 0 : heatmap: h,
1944 0 : read_at: hs.1,
1945 0 : end_lsn: None,
1946 0 : })
1947 12 : });
1948 12 : part_downloads.spawn(
1949 12 : self.load_timeline_metadata(
1950 12 : timeline_id,
1951 12 : remote_storage.clone(),
1952 12 : previous_timeline_heatmap,
1953 12 : cancel_clone,
1954 12 : )
1955 12 : .instrument(info_span!("download_index_part", %timeline_id)),
1956 : );
1957 : }
1958 :
1959 444 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1960 :
1961 : loop {
1962 456 : tokio::select!(
1963 456 : next = part_downloads.join_next() => {
1964 456 : match next {
1965 12 : Some(result) => {
1966 12 : let preload = result.context("join preload task")?;
1967 12 : timeline_preloads.insert(preload.timeline_id, preload);
1968 : },
1969 : None => {
1970 444 : break;
1971 : }
1972 : }
1973 : },
1974 456 : _ = cancel.cancelled() => {
1975 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1976 : }
1977 : )
1978 : }
1979 :
1980 444 : Ok(timeline_preloads)
1981 444 : }
1982 :
1983 12 : fn build_timeline_client(
1984 12 : &self,
1985 12 : timeline_id: TimelineId,
1986 12 : remote_storage: GenericRemoteStorage,
1987 12 : ) -> RemoteTimelineClient {
1988 12 : RemoteTimelineClient::new(
1989 12 : remote_storage.clone(),
1990 12 : self.deletion_queue_client.clone(),
1991 12 : self.conf,
1992 12 : self.tenant_shard_id,
1993 12 : timeline_id,
1994 12 : self.generation,
1995 12 : &self.tenant_conf.load().location,
1996 12 : )
1997 12 : }
1998 :
1999 12 : fn load_timeline_metadata(
2000 12 : self: &Arc<Tenant>,
2001 12 : timeline_id: TimelineId,
2002 12 : remote_storage: GenericRemoteStorage,
2003 12 : previous_heatmap: Option<PreviousHeatmap>,
2004 12 : cancel: CancellationToken,
2005 12 : ) -> impl Future<Output = TimelinePreload> + use<> {
2006 12 : let client = self.build_timeline_client(timeline_id, remote_storage);
2007 12 : async move {
2008 12 : debug_assert_current_span_has_tenant_and_timeline_id();
2009 12 : debug!("starting index part download");
2010 :
2011 12 : let index_part = client.download_index_file(&cancel).await;
2012 :
2013 12 : debug!("finished index part download");
2014 :
2015 12 : TimelinePreload {
2016 12 : client,
2017 12 : timeline_id,
2018 12 : index_part,
2019 12 : previous_heatmap,
2020 12 : }
2021 12 : }
2022 12 : }
2023 :
2024 0 : fn check_to_be_archived_has_no_unarchived_children(
2025 0 : timeline_id: TimelineId,
2026 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2027 0 : ) -> Result<(), TimelineArchivalError> {
2028 0 : let children: Vec<TimelineId> = timelines
2029 0 : .iter()
2030 0 : .filter_map(|(id, entry)| {
2031 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
2032 0 : return None;
2033 0 : }
2034 0 : if entry.is_archived() == Some(true) {
2035 0 : return None;
2036 0 : }
2037 0 : Some(*id)
2038 0 : })
2039 0 : .collect();
2040 0 :
2041 0 : if !children.is_empty() {
2042 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
2043 0 : }
2044 0 : Ok(())
2045 0 : }
2046 :
2047 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
2048 0 : ancestor_timeline_id: TimelineId,
2049 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2050 0 : offloaded_timelines: &std::sync::MutexGuard<
2051 0 : '_,
2052 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
2053 0 : >,
2054 0 : ) -> Result<(), TimelineArchivalError> {
2055 0 : let has_archived_parent =
2056 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2057 0 : ancestor_timeline.is_archived() == Some(true)
2058 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2059 0 : true
2060 : } else {
2061 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2062 0 : if cfg!(debug_assertions) {
2063 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2064 0 : }
2065 0 : return Err(TimelineArchivalError::NotFound);
2066 : };
2067 0 : if has_archived_parent {
2068 0 : return Err(TimelineArchivalError::HasArchivedParent(
2069 0 : ancestor_timeline_id,
2070 0 : ));
2071 0 : }
2072 0 : Ok(())
2073 0 : }
2074 :
2075 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2076 0 : timeline: &Arc<Timeline>,
2077 0 : ) -> Result<(), TimelineArchivalError> {
2078 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2079 0 : if ancestor_timeline.is_archived() == Some(true) {
2080 0 : return Err(TimelineArchivalError::HasArchivedParent(
2081 0 : ancestor_timeline.timeline_id,
2082 0 : ));
2083 0 : }
2084 0 : }
2085 0 : Ok(())
2086 0 : }
2087 :
2088 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2089 : ///
2090 : /// Counterpart to [`offload_timeline`].
2091 0 : async fn unoffload_timeline(
2092 0 : self: &Arc<Self>,
2093 0 : timeline_id: TimelineId,
2094 0 : broker_client: storage_broker::BrokerClientChannel,
2095 0 : ctx: RequestContext,
2096 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2097 0 : info!("unoffloading timeline");
2098 :
2099 : // We activate the timeline below manually, so this must be called on an active tenant.
2100 : // We expect callers of this function to ensure this.
2101 0 : match self.current_state() {
2102 : TenantState::Activating { .. }
2103 : | TenantState::Attaching
2104 : | TenantState::Broken { .. } => {
2105 0 : panic!("Timeline expected to be active")
2106 : }
2107 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2108 0 : TenantState::Active => {}
2109 0 : }
2110 0 : let cancel = self.cancel.clone();
2111 0 :
2112 0 : // Protect against concurrent attempts to use this TimelineId
2113 0 : // We don't care much about idempotency, as it's ensured a layer above.
2114 0 : let allow_offloaded = true;
2115 0 : let _create_guard = self
2116 0 : .create_timeline_create_guard(
2117 0 : timeline_id,
2118 0 : CreateTimelineIdempotency::FailWithConflict,
2119 0 : allow_offloaded,
2120 0 : )
2121 0 : .map_err(|err| match err {
2122 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2123 : TimelineExclusionError::AlreadyExists { .. } => {
2124 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2125 : }
2126 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2127 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2128 0 : })?;
2129 :
2130 0 : let timeline_preload = self
2131 0 : .load_timeline_metadata(
2132 0 : timeline_id,
2133 0 : self.remote_storage.clone(),
2134 0 : None,
2135 0 : cancel.clone(),
2136 0 : )
2137 0 : .await;
2138 :
2139 0 : let index_part = match timeline_preload.index_part {
2140 0 : Ok(index_part) => {
2141 0 : debug!("remote index part exists for timeline {timeline_id}");
2142 0 : index_part
2143 : }
2144 : Err(DownloadError::NotFound) => {
2145 0 : error!(%timeline_id, "index_part not found on remote");
2146 0 : return Err(TimelineArchivalError::NotFound);
2147 : }
2148 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2149 0 : Err(e) => {
2150 0 : // Some (possibly ephemeral) error happened during index_part download.
2151 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2152 0 : return Err(TimelineArchivalError::Other(
2153 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2154 0 : ));
2155 : }
2156 : };
2157 0 : let index_part = match index_part {
2158 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2159 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2160 0 : info!("timeline is deleted according to index_part.json");
2161 0 : return Err(TimelineArchivalError::NotFound);
2162 : }
2163 : };
2164 0 : let remote_metadata = index_part.metadata.clone();
2165 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2166 0 : self.load_remote_timeline(
2167 0 : timeline_id,
2168 0 : index_part,
2169 0 : remote_metadata,
2170 0 : None,
2171 0 : timeline_resources,
2172 0 : LoadTimelineCause::Unoffload,
2173 0 : &ctx,
2174 0 : )
2175 0 : .await
2176 0 : .with_context(|| {
2177 0 : format!(
2178 0 : "failed to load remote timeline {} for tenant {}",
2179 0 : timeline_id, self.tenant_shard_id
2180 0 : )
2181 0 : })
2182 0 : .map_err(TimelineArchivalError::Other)?;
2183 :
2184 0 : let timeline = {
2185 0 : let timelines = self.timelines.lock().unwrap();
2186 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2187 0 : warn!("timeline not available directly after attach");
2188 : // This is not a panic because no locks are held between `load_remote_timeline`
2189 : // which puts the timeline into timelines, and our look into the timeline map.
2190 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2191 0 : "timeline not available directly after attach"
2192 0 : )));
2193 : };
2194 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2195 0 : match offloaded_timelines.remove(&timeline_id) {
2196 0 : Some(offloaded) => {
2197 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2198 0 : }
2199 0 : None => warn!("timeline already removed from offloaded timelines"),
2200 : }
2201 :
2202 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2203 0 :
2204 0 : Arc::clone(timeline)
2205 0 : };
2206 0 :
2207 0 : // Upload new list of offloaded timelines to S3
2208 0 : self.store_tenant_manifest().await?;
2209 :
2210 : // Activate the timeline (if it makes sense)
2211 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2212 0 : let background_jobs_can_start = None;
2213 0 : timeline.activate(
2214 0 : self.clone(),
2215 0 : broker_client.clone(),
2216 0 : background_jobs_can_start,
2217 0 : &ctx,
2218 0 : );
2219 0 : }
2220 :
2221 0 : info!("timeline unoffloading complete");
2222 0 : Ok(timeline)
2223 0 : }
2224 :
2225 0 : pub(crate) async fn apply_timeline_archival_config(
2226 0 : self: &Arc<Self>,
2227 0 : timeline_id: TimelineId,
2228 0 : new_state: TimelineArchivalState,
2229 0 : broker_client: storage_broker::BrokerClientChannel,
2230 0 : ctx: RequestContext,
2231 0 : ) -> Result<(), TimelineArchivalError> {
2232 0 : info!("setting timeline archival config");
2233 : // First part: figure out what is needed to do, and do validation
2234 0 : let timeline_or_unarchive_offloaded = 'outer: {
2235 0 : let timelines = self.timelines.lock().unwrap();
2236 :
2237 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2238 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2239 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2240 0 : return Err(TimelineArchivalError::NotFound);
2241 : };
2242 0 : if new_state == TimelineArchivalState::Archived {
2243 : // It's offloaded already, so nothing to do
2244 0 : return Ok(());
2245 0 : }
2246 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2247 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2248 0 : ancestor_timeline_id,
2249 0 : &timelines,
2250 0 : &offloaded_timelines,
2251 0 : )?;
2252 0 : }
2253 0 : break 'outer None;
2254 : };
2255 :
2256 : // Do some validation. We release the timelines lock below, so there is potential
2257 : // for race conditions: these checks are more present to prevent misunderstandings of
2258 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2259 0 : match new_state {
2260 : TimelineArchivalState::Unarchived => {
2261 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2262 : }
2263 : TimelineArchivalState::Archived => {
2264 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2265 : }
2266 : }
2267 0 : Some(Arc::clone(timeline))
2268 : };
2269 :
2270 : // Second part: unoffload timeline (if needed)
2271 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2272 0 : timeline
2273 : } else {
2274 : // Turn offloaded timeline into a non-offloaded one
2275 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2276 0 : .await?
2277 : };
2278 :
2279 : // Third part: upload new timeline archival state and block until it is present in S3
2280 0 : let upload_needed = match timeline
2281 0 : .remote_client
2282 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2283 : {
2284 0 : Ok(upload_needed) => upload_needed,
2285 0 : Err(e) => {
2286 0 : if timeline.cancel.is_cancelled() {
2287 0 : return Err(TimelineArchivalError::Cancelled);
2288 : } else {
2289 0 : return Err(TimelineArchivalError::Other(e));
2290 : }
2291 : }
2292 : };
2293 :
2294 0 : if upload_needed {
2295 0 : info!("Uploading new state");
2296 : const MAX_WAIT: Duration = Duration::from_secs(10);
2297 0 : let Ok(v) =
2298 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2299 : else {
2300 0 : tracing::warn!("reached timeout for waiting on upload queue");
2301 0 : return Err(TimelineArchivalError::Timeout);
2302 : };
2303 0 : v.map_err(|e| match e {
2304 0 : WaitCompletionError::NotInitialized(e) => {
2305 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2306 : }
2307 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2308 0 : TimelineArchivalError::Cancelled
2309 : }
2310 0 : })?;
2311 0 : }
2312 0 : Ok(())
2313 0 : }
2314 :
2315 4 : pub fn get_offloaded_timeline(
2316 4 : &self,
2317 4 : timeline_id: TimelineId,
2318 4 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2319 4 : self.timelines_offloaded
2320 4 : .lock()
2321 4 : .unwrap()
2322 4 : .get(&timeline_id)
2323 4 : .map(Arc::clone)
2324 4 : .ok_or(GetTimelineError::NotFound {
2325 4 : tenant_id: self.tenant_shard_id,
2326 4 : timeline_id,
2327 4 : })
2328 4 : }
2329 :
2330 8 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2331 8 : self.tenant_shard_id
2332 8 : }
2333 :
2334 : /// Get Timeline handle for given Neon timeline ID.
2335 : /// This function is idempotent. It doesn't change internal state in any way.
2336 444 : pub fn get_timeline(
2337 444 : &self,
2338 444 : timeline_id: TimelineId,
2339 444 : active_only: bool,
2340 444 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2341 444 : let timelines_accessor = self.timelines.lock().unwrap();
2342 444 : let timeline = timelines_accessor
2343 444 : .get(&timeline_id)
2344 444 : .ok_or(GetTimelineError::NotFound {
2345 444 : tenant_id: self.tenant_shard_id,
2346 444 : timeline_id,
2347 444 : })?;
2348 :
2349 440 : if active_only && !timeline.is_active() {
2350 0 : Err(GetTimelineError::NotActive {
2351 0 : tenant_id: self.tenant_shard_id,
2352 0 : timeline_id,
2353 0 : state: timeline.current_state(),
2354 0 : })
2355 : } else {
2356 440 : Ok(Arc::clone(timeline))
2357 : }
2358 444 : }
2359 :
2360 : /// Lists timelines the tenant contains.
2361 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2362 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2363 0 : self.timelines
2364 0 : .lock()
2365 0 : .unwrap()
2366 0 : .values()
2367 0 : .map(Arc::clone)
2368 0 : .collect()
2369 0 : }
2370 :
2371 : /// Lists timelines the tenant manages, including offloaded ones.
2372 : ///
2373 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2374 0 : pub fn list_timelines_and_offloaded(
2375 0 : &self,
2376 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2377 0 : let timelines = self
2378 0 : .timelines
2379 0 : .lock()
2380 0 : .unwrap()
2381 0 : .values()
2382 0 : .map(Arc::clone)
2383 0 : .collect();
2384 0 : let offloaded = self
2385 0 : .timelines_offloaded
2386 0 : .lock()
2387 0 : .unwrap()
2388 0 : .values()
2389 0 : .map(Arc::clone)
2390 0 : .collect();
2391 0 : (timelines, offloaded)
2392 0 : }
2393 :
2394 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2395 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2396 0 : }
2397 :
2398 : /// This is used by tests & import-from-basebackup.
2399 : ///
2400 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2401 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2402 : ///
2403 : /// The caller is responsible for getting the timeline into a state that will be accepted
2404 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2405 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2406 : /// to the [`Tenant::timelines`].
2407 : ///
2408 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2409 428 : pub(crate) async fn create_empty_timeline(
2410 428 : self: &Arc<Self>,
2411 428 : new_timeline_id: TimelineId,
2412 428 : initdb_lsn: Lsn,
2413 428 : pg_version: u32,
2414 428 : _ctx: &RequestContext,
2415 428 : ) -> anyhow::Result<UninitializedTimeline> {
2416 428 : anyhow::ensure!(
2417 428 : self.is_active(),
2418 0 : "Cannot create empty timelines on inactive tenant"
2419 : );
2420 :
2421 : // Protect against concurrent attempts to use this TimelineId
2422 428 : let create_guard = match self
2423 428 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2424 428 : .await?
2425 : {
2426 424 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2427 : StartCreatingTimelineResult::Idempotent(_) => {
2428 0 : unreachable!("FailWithConflict implies we get an error instead")
2429 : }
2430 : };
2431 :
2432 424 : let new_metadata = TimelineMetadata::new(
2433 424 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2434 424 : // make it valid, before calling finish_creation()
2435 424 : Lsn(0),
2436 424 : None,
2437 424 : None,
2438 424 : Lsn(0),
2439 424 : initdb_lsn,
2440 424 : initdb_lsn,
2441 424 : pg_version,
2442 424 : );
2443 424 : self.prepare_new_timeline(
2444 424 : new_timeline_id,
2445 424 : &new_metadata,
2446 424 : create_guard,
2447 424 : initdb_lsn,
2448 424 : None,
2449 424 : )
2450 424 : .await
2451 428 : }
2452 :
2453 : /// Helper for unit tests to create an empty timeline.
2454 : ///
2455 : /// The timeline is has state value `Active` but its background loops are not running.
2456 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2457 : // Our current tests don't need the background loops.
2458 : #[cfg(test)]
2459 408 : pub async fn create_test_timeline(
2460 408 : self: &Arc<Self>,
2461 408 : new_timeline_id: TimelineId,
2462 408 : initdb_lsn: Lsn,
2463 408 : pg_version: u32,
2464 408 : ctx: &RequestContext,
2465 408 : ) -> anyhow::Result<Arc<Timeline>> {
2466 408 : let uninit_tl = self
2467 408 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2468 408 : .await?;
2469 408 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2470 408 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2471 :
2472 : // Setup minimum keys required for the timeline to be usable.
2473 408 : let mut modification = tline.begin_modification(initdb_lsn);
2474 408 : modification
2475 408 : .init_empty_test_timeline()
2476 408 : .context("init_empty_test_timeline")?;
2477 408 : modification
2478 408 : .commit(ctx)
2479 408 : .await
2480 408 : .context("commit init_empty_test_timeline modification")?;
2481 :
2482 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2483 408 : tline.maybe_spawn_flush_loop();
2484 408 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2485 :
2486 : // Make sure the freeze_and_flush reaches remote storage.
2487 408 : tline.remote_client.wait_completion().await.unwrap();
2488 :
2489 408 : let tl = uninit_tl.finish_creation().await?;
2490 : // The non-test code would call tl.activate() here.
2491 408 : tl.set_state(TimelineState::Active);
2492 408 : Ok(tl)
2493 408 : }
2494 :
2495 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2496 : #[cfg(test)]
2497 : #[allow(clippy::too_many_arguments)]
2498 80 : pub async fn create_test_timeline_with_layers(
2499 80 : self: &Arc<Self>,
2500 80 : new_timeline_id: TimelineId,
2501 80 : initdb_lsn: Lsn,
2502 80 : pg_version: u32,
2503 80 : ctx: &RequestContext,
2504 80 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2505 80 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2506 80 : end_lsn: Lsn,
2507 80 : ) -> anyhow::Result<Arc<Timeline>> {
2508 : use checks::check_valid_layermap;
2509 : use itertools::Itertools;
2510 :
2511 80 : let tline = self
2512 80 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2513 80 : .await?;
2514 80 : tline.force_advance_lsn(end_lsn);
2515 252 : for deltas in delta_layer_desc {
2516 172 : tline
2517 172 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2518 172 : .await?;
2519 : }
2520 192 : for (lsn, images) in image_layer_desc {
2521 112 : tline
2522 112 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2523 112 : .await?;
2524 : }
2525 80 : let layer_names = tline
2526 80 : .layers
2527 80 : .read()
2528 80 : .await
2529 80 : .layer_map()
2530 80 : .unwrap()
2531 80 : .iter_historic_layers()
2532 364 : .map(|layer| layer.layer_name())
2533 80 : .collect_vec();
2534 80 : if let Some(err) = check_valid_layermap(&layer_names) {
2535 0 : bail!("invalid layermap: {err}");
2536 80 : }
2537 80 : Ok(tline)
2538 80 : }
2539 :
2540 : /// Create a new timeline.
2541 : ///
2542 : /// Returns the new timeline ID and reference to its Timeline object.
2543 : ///
2544 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2545 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2546 : #[allow(clippy::too_many_arguments)]
2547 0 : pub(crate) async fn create_timeline(
2548 0 : self: &Arc<Tenant>,
2549 0 : params: CreateTimelineParams,
2550 0 : broker_client: storage_broker::BrokerClientChannel,
2551 0 : ctx: &RequestContext,
2552 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2553 0 : if !self.is_active() {
2554 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2555 0 : return Err(CreateTimelineError::ShuttingDown);
2556 : } else {
2557 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2558 0 : "Cannot create timelines on inactive tenant"
2559 0 : )));
2560 : }
2561 0 : }
2562 :
2563 0 : let _gate = self
2564 0 : .gate
2565 0 : .enter()
2566 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2567 :
2568 0 : let result: CreateTimelineResult = match params {
2569 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2570 0 : new_timeline_id,
2571 0 : existing_initdb_timeline_id,
2572 0 : pg_version,
2573 0 : }) => {
2574 0 : self.bootstrap_timeline(
2575 0 : new_timeline_id,
2576 0 : pg_version,
2577 0 : existing_initdb_timeline_id,
2578 0 : ctx,
2579 0 : )
2580 0 : .await?
2581 : }
2582 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2583 0 : new_timeline_id,
2584 0 : ancestor_timeline_id,
2585 0 : mut ancestor_start_lsn,
2586 : }) => {
2587 0 : let ancestor_timeline = self
2588 0 : .get_timeline(ancestor_timeline_id, false)
2589 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2590 :
2591 : // instead of waiting around, just deny the request because ancestor is not yet
2592 : // ready for other purposes either.
2593 0 : if !ancestor_timeline.is_active() {
2594 0 : return Err(CreateTimelineError::AncestorNotActive);
2595 0 : }
2596 0 :
2597 0 : if ancestor_timeline.is_archived() == Some(true) {
2598 0 : info!("tried to branch archived timeline");
2599 0 : return Err(CreateTimelineError::AncestorArchived);
2600 0 : }
2601 :
2602 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2603 0 : *lsn = lsn.align();
2604 0 :
2605 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2606 0 : if ancestor_ancestor_lsn > *lsn {
2607 : // can we safely just branch from the ancestor instead?
2608 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2609 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2610 0 : lsn,
2611 0 : ancestor_timeline_id,
2612 0 : ancestor_ancestor_lsn,
2613 0 : )));
2614 0 : }
2615 0 :
2616 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2617 0 : // to the requested branch point. The repository code itself doesn't
2618 0 : // require it, but if we start to receive WAL on the new timeline,
2619 0 : // decoding the new WAL might need to look up previous pages, relation
2620 0 : // sizes etc. and that would get confused if the previous page versions
2621 0 : // are not in the repository yet.
2622 0 : ancestor_timeline
2623 0 : .wait_lsn(
2624 0 : *lsn,
2625 0 : timeline::WaitLsnWaiter::Tenant,
2626 0 : timeline::WaitLsnTimeout::Default,
2627 0 : ctx,
2628 0 : )
2629 0 : .await
2630 0 : .map_err(|e| match e {
2631 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2632 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2633 : }
2634 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2635 0 : })?;
2636 0 : }
2637 :
2638 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2639 0 : .await?
2640 : }
2641 0 : CreateTimelineParams::ImportPgdata(params) => {
2642 0 : self.create_timeline_import_pgdata(
2643 0 : params,
2644 0 : ActivateTimelineArgs::Yes {
2645 0 : broker_client: broker_client.clone(),
2646 0 : },
2647 0 : ctx,
2648 0 : )
2649 0 : .await?
2650 : }
2651 : };
2652 :
2653 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2654 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2655 : // not send a success to the caller until it is. The same applies to idempotent retries.
2656 : //
2657 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2658 : // assume that, because they can see the timeline via API, that the creation is done and
2659 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2660 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2661 : // interacts with UninitializedTimeline and is generally a bit tricky.
2662 : //
2663 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2664 : // creation API until it returns success. Only then is durability guaranteed.
2665 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2666 0 : result
2667 0 : .timeline()
2668 0 : .remote_client
2669 0 : .wait_completion()
2670 0 : .await
2671 0 : .map_err(|e| match e {
2672 : WaitCompletionError::NotInitialized(
2673 0 : e, // If the queue is already stopped, it's a shutdown error.
2674 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2675 : WaitCompletionError::NotInitialized(_) => {
2676 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2677 0 : debug_assert!(false);
2678 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2679 : }
2680 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2681 0 : CreateTimelineError::ShuttingDown
2682 : }
2683 0 : })?;
2684 :
2685 : // The creating task is responsible for activating the timeline.
2686 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2687 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2688 0 : let activated_timeline = match result {
2689 0 : CreateTimelineResult::Created(timeline) => {
2690 0 : timeline.activate(self.clone(), broker_client, None, ctx);
2691 0 : timeline
2692 : }
2693 0 : CreateTimelineResult::Idempotent(timeline) => {
2694 0 : info!(
2695 0 : "request was deemed idempotent, activation will be done by the creating task"
2696 : );
2697 0 : timeline
2698 : }
2699 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2700 0 : info!(
2701 0 : "import task spawned, timeline will become visible and activated once the import is done"
2702 : );
2703 0 : timeline
2704 : }
2705 : };
2706 :
2707 0 : Ok(activated_timeline)
2708 0 : }
2709 :
2710 : /// The returned [`Arc<Timeline>`] is NOT in the [`Tenant::timelines`] map until the import
2711 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2712 : /// [`Tenant::timelines`] map when the import completes.
2713 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2714 : /// for the response.
2715 0 : async fn create_timeline_import_pgdata(
2716 0 : self: &Arc<Tenant>,
2717 0 : params: CreateTimelineParamsImportPgdata,
2718 0 : activate: ActivateTimelineArgs,
2719 0 : ctx: &RequestContext,
2720 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2721 0 : let CreateTimelineParamsImportPgdata {
2722 0 : new_timeline_id,
2723 0 : location,
2724 0 : idempotency_key,
2725 0 : } = params;
2726 0 :
2727 0 : let started_at = chrono::Utc::now().naive_utc();
2728 :
2729 : //
2730 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2731 : // is the canonical way we do it.
2732 : // - create an empty timeline in-memory
2733 : // - use its remote_timeline_client to do the upload
2734 : // - dispose of the uninit timeline
2735 : // - keep the creation guard alive
2736 :
2737 0 : let timeline_create_guard = match self
2738 0 : .start_creating_timeline(
2739 0 : new_timeline_id,
2740 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2741 0 : idempotency_key: idempotency_key.clone(),
2742 0 : }),
2743 0 : )
2744 0 : .await?
2745 : {
2746 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2747 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2748 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
2749 : }
2750 : };
2751 :
2752 0 : let mut uninit_timeline = {
2753 0 : let this = &self;
2754 0 : let initdb_lsn = Lsn(0);
2755 0 : let _ctx = ctx;
2756 0 : async move {
2757 0 : let new_metadata = TimelineMetadata::new(
2758 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2759 0 : // make it valid, before calling finish_creation()
2760 0 : Lsn(0),
2761 0 : None,
2762 0 : None,
2763 0 : Lsn(0),
2764 0 : initdb_lsn,
2765 0 : initdb_lsn,
2766 0 : 15,
2767 0 : );
2768 0 : this.prepare_new_timeline(
2769 0 : new_timeline_id,
2770 0 : &new_metadata,
2771 0 : timeline_create_guard,
2772 0 : initdb_lsn,
2773 0 : None,
2774 0 : )
2775 0 : .await
2776 0 : }
2777 0 : }
2778 0 : .await?;
2779 :
2780 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2781 0 : idempotency_key,
2782 0 : location,
2783 0 : started_at,
2784 0 : };
2785 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2786 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2787 0 : );
2788 0 : uninit_timeline
2789 0 : .raw_timeline()
2790 0 : .unwrap()
2791 0 : .remote_client
2792 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2793 :
2794 : // wait_completion happens in caller
2795 :
2796 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2797 0 :
2798 0 : tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2799 0 : timeline.clone(),
2800 0 : index_part,
2801 0 : activate,
2802 0 : timeline_create_guard,
2803 0 : ));
2804 0 :
2805 0 : // NB: the timeline doesn't exist in self.timelines at this point
2806 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2807 0 : }
2808 :
2809 : #[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))]
2810 : async fn create_timeline_import_pgdata_task(
2811 : self: Arc<Tenant>,
2812 : timeline: Arc<Timeline>,
2813 : index_part: import_pgdata::index_part_format::Root,
2814 : activate: ActivateTimelineArgs,
2815 : timeline_create_guard: TimelineCreateGuard,
2816 : ) {
2817 : debug_assert_current_span_has_tenant_and_timeline_id();
2818 : info!("starting");
2819 : scopeguard::defer! {info!("exiting")};
2820 :
2821 : let res = self
2822 : .create_timeline_import_pgdata_task_impl(
2823 : timeline,
2824 : index_part,
2825 : activate,
2826 : timeline_create_guard,
2827 : )
2828 : .await;
2829 : if let Err(err) = &res {
2830 : error!(?err, "task failed");
2831 : // TODO sleep & retry, sensitive to tenant shutdown
2832 : // TODO: allow timeline deletion requests => should cancel the task
2833 : }
2834 : }
2835 :
2836 0 : async fn create_timeline_import_pgdata_task_impl(
2837 0 : self: Arc<Tenant>,
2838 0 : timeline: Arc<Timeline>,
2839 0 : index_part: import_pgdata::index_part_format::Root,
2840 0 : activate: ActivateTimelineArgs,
2841 0 : timeline_create_guard: TimelineCreateGuard,
2842 0 : ) -> Result<(), anyhow::Error> {
2843 0 : let ctx = RequestContext::new(TaskKind::ImportPgdata, DownloadBehavior::Warn);
2844 0 :
2845 0 : info!("importing pgdata");
2846 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2847 0 : .await
2848 0 : .context("import")?;
2849 0 : info!("import done");
2850 :
2851 : //
2852 : // Reload timeline from remote.
2853 : // This proves that the remote state is attachable, and it reuses the code.
2854 : //
2855 : // TODO: think about whether this is safe to do with concurrent Tenant::shutdown.
2856 : // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
2857 : // But our activate() call might launch new background tasks after Tenant::shutdown
2858 : // already went past shutting down the Tenant::timelines, which this timeline here is no part of.
2859 : // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
2860 : // down while bootstrapping/branching + activating), but, the race condition is much more likely
2861 : // to manifest because of the long runtime of this import task.
2862 :
2863 : // in theory this shouldn't even .await anything except for coop yield
2864 0 : info!("shutting down timeline");
2865 0 : timeline.shutdown(ShutdownMode::Hard).await;
2866 0 : info!("timeline shut down, reloading from remote");
2867 : // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
2868 : // let Some(timeline) = Arc::into_inner(timeline) else {
2869 : // anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
2870 : // };
2871 0 : let timeline_id = timeline.timeline_id;
2872 0 :
2873 0 : // load from object storage like Tenant::attach does
2874 0 : let resources = self.build_timeline_resources(timeline_id);
2875 0 : let index_part = resources
2876 0 : .remote_client
2877 0 : .download_index_file(&self.cancel)
2878 0 : .await?;
2879 0 : let index_part = match index_part {
2880 : MaybeDeletedIndexPart::Deleted(_) => {
2881 : // likely concurrent delete call, cplane should prevent this
2882 0 : anyhow::bail!(
2883 0 : "index part says deleted but we are not done creating yet, this should not happen but"
2884 0 : )
2885 : }
2886 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2887 0 : };
2888 0 : let metadata = index_part.metadata.clone();
2889 0 : self
2890 0 : .load_remote_timeline(timeline_id, index_part, metadata, None, resources, LoadTimelineCause::ImportPgdata{
2891 0 : create_guard: timeline_create_guard, activate, }, &ctx)
2892 0 : .await?
2893 0 : .ready_to_activate()
2894 0 : .context("implementation error: reloaded timeline still needs import after import reported success")?;
2895 :
2896 0 : anyhow::Ok(())
2897 0 : }
2898 :
2899 0 : pub(crate) async fn delete_timeline(
2900 0 : self: Arc<Self>,
2901 0 : timeline_id: TimelineId,
2902 0 : ) -> Result<(), DeleteTimelineError> {
2903 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2904 :
2905 0 : Ok(())
2906 0 : }
2907 :
2908 : /// perform one garbage collection iteration, removing old data files from disk.
2909 : /// this function is periodically called by gc task.
2910 : /// also it can be explicitly requested through page server api 'do_gc' command.
2911 : ///
2912 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2913 : ///
2914 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2915 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2916 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2917 : /// `pitr` specifies the same as a time difference from the current time. The effective
2918 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2919 : /// requires more history to be retained.
2920 : //
2921 1508 : pub(crate) async fn gc_iteration(
2922 1508 : &self,
2923 1508 : target_timeline_id: Option<TimelineId>,
2924 1508 : horizon: u64,
2925 1508 : pitr: Duration,
2926 1508 : cancel: &CancellationToken,
2927 1508 : ctx: &RequestContext,
2928 1508 : ) -> Result<GcResult, GcError> {
2929 1508 : // Don't start doing work during shutdown
2930 1508 : if let TenantState::Stopping { .. } = self.current_state() {
2931 0 : return Ok(GcResult::default());
2932 1508 : }
2933 1508 :
2934 1508 : // there is a global allowed_error for this
2935 1508 : if !self.is_active() {
2936 0 : return Err(GcError::NotActive);
2937 1508 : }
2938 1508 :
2939 1508 : {
2940 1508 : let conf = self.tenant_conf.load();
2941 1508 :
2942 1508 : // If we may not delete layers, then simply skip GC. Even though a tenant
2943 1508 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2944 1508 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2945 1508 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2946 1508 : if !conf.location.may_delete_layers_hint() {
2947 0 : info!("Skipping GC in location state {:?}", conf.location);
2948 0 : return Ok(GcResult::default());
2949 1508 : }
2950 1508 :
2951 1508 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2952 1500 : info!("Skipping GC because lsn lease deadline is not reached");
2953 1500 : return Ok(GcResult::default());
2954 8 : }
2955 : }
2956 :
2957 8 : let _guard = match self.gc_block.start().await {
2958 8 : Ok(guard) => guard,
2959 0 : Err(reasons) => {
2960 0 : info!("Skipping GC: {reasons}");
2961 0 : return Ok(GcResult::default());
2962 : }
2963 : };
2964 :
2965 8 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2966 8 : .await
2967 1508 : }
2968 :
2969 : /// Performs one compaction iteration. Called periodically from the compaction loop. Returns
2970 : /// whether another compaction is needed, if we still have pending work or if we yield for
2971 : /// immediate L0 compaction.
2972 : ///
2973 : /// Compaction can also be explicitly requested for a timeline via the HTTP API.
2974 0 : async fn compaction_iteration(
2975 0 : self: &Arc<Self>,
2976 0 : cancel: &CancellationToken,
2977 0 : ctx: &RequestContext,
2978 0 : ) -> Result<CompactionOutcome, CompactionError> {
2979 0 : // Don't compact inactive tenants.
2980 0 : if !self.is_active() {
2981 0 : return Ok(CompactionOutcome::Skipped);
2982 0 : }
2983 0 :
2984 0 : // Don't compact tenants that can't upload layers. We don't check `may_delete_layers_hint`,
2985 0 : // since we need to compact L0 even in AttachedMulti to bound read amplification.
2986 0 : let location = self.tenant_conf.load().location;
2987 0 : if !location.may_upload_layers_hint() {
2988 0 : info!("skipping compaction in location state {location:?}");
2989 0 : return Ok(CompactionOutcome::Skipped);
2990 0 : }
2991 0 :
2992 0 : // Don't compact if the circuit breaker is tripped.
2993 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
2994 0 : info!("skipping compaction due to previous failures");
2995 0 : return Ok(CompactionOutcome::Skipped);
2996 0 : }
2997 0 :
2998 0 : // Collect all timelines to compact, along with offload instructions and L0 counts.
2999 0 : let mut compact: Vec<Arc<Timeline>> = Vec::new();
3000 0 : let mut offload: HashSet<TimelineId> = HashSet::new();
3001 0 : let mut l0_counts: HashMap<TimelineId, usize> = HashMap::new();
3002 0 :
3003 0 : {
3004 0 : let offload_enabled = self.get_timeline_offloading_enabled();
3005 0 : let timelines = self.timelines.lock().unwrap();
3006 0 : for (&timeline_id, timeline) in timelines.iter() {
3007 : // Skip inactive timelines.
3008 0 : if !timeline.is_active() {
3009 0 : continue;
3010 0 : }
3011 0 :
3012 0 : // Schedule the timeline for compaction.
3013 0 : compact.push(timeline.clone());
3014 :
3015 : // Schedule the timeline for offloading if eligible.
3016 0 : let can_offload = offload_enabled
3017 0 : && timeline.can_offload().0
3018 0 : && !timelines
3019 0 : .iter()
3020 0 : .any(|(_, tli)| tli.get_ancestor_timeline_id() == Some(timeline_id));
3021 0 : if can_offload {
3022 0 : offload.insert(timeline_id);
3023 0 : }
3024 : }
3025 : } // release timelines lock
3026 :
3027 0 : for timeline in &compact {
3028 : // Collect L0 counts. Can't await while holding lock above.
3029 0 : if let Ok(lm) = timeline.layers.read().await.layer_map() {
3030 0 : l0_counts.insert(timeline.timeline_id, lm.level0_deltas().len());
3031 0 : }
3032 : }
3033 :
3034 : // Pass 1: L0 compaction across all timelines, in order of L0 count. We prioritize this to
3035 : // bound read amplification.
3036 : //
3037 : // TODO: this may spin on one or more ingest-heavy timelines, starving out image/GC
3038 : // compaction and offloading. We leave that as a potential problem to solve later. Consider
3039 : // splitting L0 and image/GC compaction to separate background jobs.
3040 0 : if self.get_compaction_l0_first() {
3041 0 : let compaction_threshold = self.get_compaction_threshold();
3042 0 : let compact_l0 = compact
3043 0 : .iter()
3044 0 : .map(|tli| (tli, l0_counts.get(&tli.timeline_id).copied().unwrap_or(0)))
3045 0 : .filter(|&(_, l0)| l0 >= compaction_threshold)
3046 0 : .sorted_by_key(|&(_, l0)| l0)
3047 0 : .rev()
3048 0 : .map(|(tli, _)| tli.clone())
3049 0 : .collect_vec();
3050 0 :
3051 0 : let mut has_pending_l0 = false;
3052 0 : for timeline in compact_l0 {
3053 0 : let outcome = timeline
3054 0 : .compact(cancel, CompactFlags::OnlyL0Compaction.into(), ctx)
3055 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3056 0 : .await
3057 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3058 0 : match outcome {
3059 0 : CompactionOutcome::Done => {}
3060 0 : CompactionOutcome::Skipped => {}
3061 0 : CompactionOutcome::Pending => has_pending_l0 = true,
3062 0 : CompactionOutcome::YieldForL0 => has_pending_l0 = true,
3063 : }
3064 : }
3065 0 : if has_pending_l0 {
3066 0 : return Ok(CompactionOutcome::YieldForL0); // do another pass
3067 0 : }
3068 0 : }
3069 :
3070 : // Pass 2: image compaction and timeline offloading. If any timelines have accumulated
3071 : // more L0 layers, they may also be compacted here.
3072 : //
3073 : // NB: image compaction may yield if there is pending L0 compaction.
3074 : //
3075 : // TODO: it will only yield if there is pending L0 compaction on the same timeline. If a
3076 : // different timeline needs compaction, it won't. It should check `l0_compaction_trigger`.
3077 : // We leave this for a later PR.
3078 : //
3079 : // TODO: consider ordering timelines by some priority, e.g. time since last full compaction,
3080 : // amount of L1 delta debt or garbage, offload-eligible timelines first, etc.
3081 0 : let mut has_pending = false;
3082 0 : for timeline in compact {
3083 0 : if !timeline.is_active() {
3084 0 : continue;
3085 0 : }
3086 :
3087 0 : let mut outcome = timeline
3088 0 : .compact(cancel, EnumSet::default(), ctx)
3089 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3090 0 : .await
3091 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3092 :
3093 : // If we're done compacting, check the scheduled GC compaction queue for more work.
3094 0 : if outcome == CompactionOutcome::Done {
3095 0 : let queue = {
3096 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3097 0 : guard
3098 0 : .entry(timeline.timeline_id)
3099 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()))
3100 0 : .clone()
3101 0 : };
3102 0 : outcome = queue
3103 0 : .iteration(cancel, ctx, &self.gc_block, &timeline)
3104 0 : .instrument(
3105 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id),
3106 : )
3107 0 : .await?;
3108 0 : }
3109 :
3110 : // If we're done compacting, offload the timeline if requested.
3111 0 : if outcome == CompactionOutcome::Done && offload.contains(&timeline.timeline_id) {
3112 0 : pausable_failpoint!("before-timeline-auto-offload");
3113 0 : offload_timeline(self, &timeline)
3114 0 : .instrument(info_span!("offload_timeline", timeline_id = %timeline.timeline_id))
3115 0 : .await
3116 0 : .or_else(|err| match err {
3117 : // Ignore this, we likely raced with unarchival.
3118 0 : OffloadError::NotArchived => Ok(()),
3119 0 : err => Err(err),
3120 0 : })?;
3121 0 : }
3122 :
3123 0 : match outcome {
3124 0 : CompactionOutcome::Done => {}
3125 0 : CompactionOutcome::Skipped => {}
3126 0 : CompactionOutcome::Pending => has_pending = true,
3127 : // This mostly makes sense when the L0-only pass above is enabled, since there's
3128 : // otherwise no guarantee that we'll start with the timeline that has high L0.
3129 0 : CompactionOutcome::YieldForL0 => return Ok(CompactionOutcome::YieldForL0),
3130 : }
3131 : }
3132 :
3133 : // Success! Untrip the breaker if necessary.
3134 0 : self.compaction_circuit_breaker
3135 0 : .lock()
3136 0 : .unwrap()
3137 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3138 0 :
3139 0 : match has_pending {
3140 0 : true => Ok(CompactionOutcome::Pending),
3141 0 : false => Ok(CompactionOutcome::Done),
3142 : }
3143 0 : }
3144 :
3145 : /// Trips the compaction circuit breaker if appropriate.
3146 0 : pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
3147 0 : match err {
3148 0 : err if err.is_cancel() => {}
3149 0 : CompactionError::ShuttingDown => (),
3150 : // Offload failures don't trip the circuit breaker, since they're cheap to retry and
3151 : // shouldn't block compaction.
3152 0 : CompactionError::Offload(_) => {}
3153 0 : CompactionError::CollectKeySpaceError(err) => {
3154 0 : // CollectKeySpaceError::Cancelled and PageRead::Cancelled are handled in `err.is_cancel` branch.
3155 0 : self.compaction_circuit_breaker
3156 0 : .lock()
3157 0 : .unwrap()
3158 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3159 0 : }
3160 0 : CompactionError::Other(err) => {
3161 0 : self.compaction_circuit_breaker
3162 0 : .lock()
3163 0 : .unwrap()
3164 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3165 0 : }
3166 0 : CompactionError::AlreadyRunning(_) => {}
3167 : }
3168 0 : }
3169 :
3170 : /// Cancel scheduled compaction tasks
3171 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3172 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3173 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3174 0 : q.cancel_scheduled();
3175 0 : }
3176 0 : }
3177 :
3178 0 : pub(crate) fn get_scheduled_compaction_tasks(
3179 0 : &self,
3180 0 : timeline_id: TimelineId,
3181 0 : ) -> Vec<CompactInfoResponse> {
3182 0 : let res = {
3183 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3184 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3185 : };
3186 0 : let Some((running, remaining)) = res else {
3187 0 : return Vec::new();
3188 : };
3189 0 : let mut result = Vec::new();
3190 0 : if let Some((id, running)) = running {
3191 0 : result.extend(running.into_compact_info_resp(id, true));
3192 0 : }
3193 0 : for (id, job) in remaining {
3194 0 : result.extend(job.into_compact_info_resp(id, false));
3195 0 : }
3196 0 : result
3197 0 : }
3198 :
3199 : /// Schedule a compaction task for a timeline.
3200 0 : pub(crate) async fn schedule_compaction(
3201 0 : &self,
3202 0 : timeline_id: TimelineId,
3203 0 : options: CompactOptions,
3204 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3205 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3206 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3207 0 : let q = guard
3208 0 : .entry(timeline_id)
3209 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3210 0 : q.schedule_manual_compaction(options, Some(tx));
3211 0 : Ok(rx)
3212 0 : }
3213 :
3214 : /// Performs periodic housekeeping, via the tenant housekeeping background task.
3215 0 : async fn housekeeping(&self) {
3216 0 : // Call through to all timelines to freeze ephemeral layers as needed. This usually happens
3217 0 : // during ingest, but we don't want idle timelines to hold open layers for too long.
3218 0 : let timelines = self
3219 0 : .timelines
3220 0 : .lock()
3221 0 : .unwrap()
3222 0 : .values()
3223 0 : .filter(|tli| tli.is_active())
3224 0 : .cloned()
3225 0 : .collect_vec();
3226 :
3227 0 : for timeline in timelines {
3228 0 : timeline.maybe_freeze_ephemeral_layer().await;
3229 : }
3230 :
3231 : // Shut down walredo if idle.
3232 : const WALREDO_IDLE_TIMEOUT: Duration = Duration::from_secs(180);
3233 0 : if let Some(ref walredo_mgr) = self.walredo_mgr {
3234 0 : walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
3235 0 : }
3236 0 : }
3237 :
3238 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3239 0 : let timelines = self.timelines.lock().unwrap();
3240 0 : !timelines
3241 0 : .iter()
3242 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3243 0 : }
3244 :
3245 3468 : pub fn current_state(&self) -> TenantState {
3246 3468 : self.state.borrow().clone()
3247 3468 : }
3248 :
3249 1944 : pub fn is_active(&self) -> bool {
3250 1944 : self.current_state() == TenantState::Active
3251 1944 : }
3252 :
3253 0 : pub fn generation(&self) -> Generation {
3254 0 : self.generation
3255 0 : }
3256 :
3257 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3258 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3259 0 : }
3260 :
3261 : /// Changes tenant status to active, unless shutdown was already requested.
3262 : ///
3263 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3264 : /// to delay background jobs. Background jobs can be started right away when None is given.
3265 0 : fn activate(
3266 0 : self: &Arc<Self>,
3267 0 : broker_client: BrokerClientChannel,
3268 0 : background_jobs_can_start: Option<&completion::Barrier>,
3269 0 : ctx: &RequestContext,
3270 0 : ) {
3271 0 : span::debug_assert_current_span_has_tenant_id();
3272 0 :
3273 0 : let mut activating = false;
3274 0 : self.state.send_modify(|current_state| {
3275 : use pageserver_api::models::ActivatingFrom;
3276 0 : match &*current_state {
3277 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3278 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3279 : }
3280 0 : TenantState::Attaching => {
3281 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3282 0 : }
3283 0 : }
3284 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3285 0 : activating = true;
3286 0 : // Continue outside the closure. We need to grab timelines.lock()
3287 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3288 0 : });
3289 0 :
3290 0 : if activating {
3291 0 : let timelines_accessor = self.timelines.lock().unwrap();
3292 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3293 0 : let timelines_to_activate = timelines_accessor
3294 0 : .values()
3295 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3296 0 :
3297 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3298 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3299 0 :
3300 0 : // Spawn gc and compaction loops. The loops will shut themselves
3301 0 : // down when they notice that the tenant is inactive.
3302 0 : tasks::start_background_loops(self, background_jobs_can_start);
3303 0 :
3304 0 : let mut activated_timelines = 0;
3305 :
3306 0 : for timeline in timelines_to_activate {
3307 0 : timeline.activate(
3308 0 : self.clone(),
3309 0 : broker_client.clone(),
3310 0 : background_jobs_can_start,
3311 0 : ctx,
3312 0 : );
3313 0 : activated_timelines += 1;
3314 0 : }
3315 :
3316 0 : self.state.send_modify(move |current_state| {
3317 0 : assert!(
3318 0 : matches!(current_state, TenantState::Activating(_)),
3319 0 : "set_stopping and set_broken wait for us to leave Activating state",
3320 : );
3321 0 : *current_state = TenantState::Active;
3322 0 :
3323 0 : let elapsed = self.constructed_at.elapsed();
3324 0 : let total_timelines = timelines_accessor.len();
3325 0 :
3326 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3327 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3328 0 : info!(
3329 0 : since_creation_millis = elapsed.as_millis(),
3330 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3331 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3332 0 : activated_timelines,
3333 0 : total_timelines,
3334 0 : post_state = <&'static str>::from(&*current_state),
3335 0 : "activation attempt finished"
3336 : );
3337 :
3338 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3339 0 : });
3340 0 : }
3341 0 : }
3342 :
3343 : /// Shutdown the tenant and join all of the spawned tasks.
3344 : ///
3345 : /// The method caters for all use-cases:
3346 : /// - pageserver shutdown (freeze_and_flush == true)
3347 : /// - detach + ignore (freeze_and_flush == false)
3348 : ///
3349 : /// This will attempt to shutdown even if tenant is broken.
3350 : ///
3351 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3352 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3353 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3354 : /// the ongoing shutdown.
3355 12 : async fn shutdown(
3356 12 : &self,
3357 12 : shutdown_progress: completion::Barrier,
3358 12 : shutdown_mode: timeline::ShutdownMode,
3359 12 : ) -> Result<(), completion::Barrier> {
3360 12 : span::debug_assert_current_span_has_tenant_id();
3361 :
3362 : // Set tenant (and its timlines) to Stoppping state.
3363 : //
3364 : // Since we can only transition into Stopping state after activation is complete,
3365 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3366 : //
3367 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3368 : // 1. Lock out any new requests to the tenants.
3369 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3370 : // 3. Signal cancellation for other tenant background loops.
3371 : // 4. ???
3372 : //
3373 : // The waiting for the cancellation is not done uniformly.
3374 : // We certainly wait for WAL receivers to shut down.
3375 : // That is necessary so that no new data comes in before the freeze_and_flush.
3376 : // But the tenant background loops are joined-on in our caller.
3377 : // It's mesed up.
3378 : // we just ignore the failure to stop
3379 :
3380 : // If we're still attaching, fire the cancellation token early to drop out: this
3381 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3382 : // is very slow.
3383 12 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3384 0 : self.cancel.cancel();
3385 0 :
3386 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3387 0 : // are children of ours, so their flush loops will have shut down already
3388 0 : timeline::ShutdownMode::Hard
3389 : } else {
3390 12 : shutdown_mode
3391 : };
3392 :
3393 12 : match self.set_stopping(shutdown_progress, false, false).await {
3394 12 : Ok(()) => {}
3395 0 : Err(SetStoppingError::Broken) => {
3396 0 : // assume that this is acceptable
3397 0 : }
3398 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3399 0 : // give caller the option to wait for this this shutdown
3400 0 : info!("Tenant::shutdown: AlreadyStopping");
3401 0 : return Err(other);
3402 : }
3403 : };
3404 :
3405 12 : let mut js = tokio::task::JoinSet::new();
3406 12 : {
3407 12 : let timelines = self.timelines.lock().unwrap();
3408 12 : timelines.values().for_each(|timeline| {
3409 12 : let timeline = Arc::clone(timeline);
3410 12 : let timeline_id = timeline.timeline_id;
3411 12 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3412 12 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3413 12 : });
3414 12 : }
3415 12 : {
3416 12 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3417 12 : timelines_offloaded.values().for_each(|timeline| {
3418 0 : timeline.defuse_for_tenant_drop();
3419 12 : });
3420 12 : }
3421 12 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3422 12 : tracing::info!("Waiting for timelines...");
3423 24 : while let Some(res) = js.join_next().await {
3424 0 : match res {
3425 12 : Ok(()) => {}
3426 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3427 0 : Err(je) if je.is_panic() => { /* logged already */ }
3428 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3429 : }
3430 : }
3431 :
3432 12 : if let ShutdownMode::Reload = shutdown_mode {
3433 0 : tracing::info!("Flushing deletion queue");
3434 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3435 0 : match e {
3436 0 : DeletionQueueError::ShuttingDown => {
3437 0 : // This is the only error we expect for now. In the future, if more error
3438 0 : // variants are added, we should handle them here.
3439 0 : }
3440 : }
3441 0 : }
3442 12 : }
3443 :
3444 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3445 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3446 12 : tracing::debug!("Cancelling CancellationToken");
3447 12 : self.cancel.cancel();
3448 12 :
3449 12 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3450 12 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3451 12 : //
3452 12 : // this will additionally shutdown and await all timeline tasks.
3453 12 : tracing::debug!("Waiting for tasks...");
3454 12 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3455 :
3456 12 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3457 12 : walredo_mgr.shutdown().await;
3458 0 : }
3459 :
3460 : // Wait for any in-flight operations to complete
3461 12 : self.gate.close().await;
3462 :
3463 12 : remove_tenant_metrics(&self.tenant_shard_id);
3464 12 :
3465 12 : Ok(())
3466 12 : }
3467 :
3468 : /// Change tenant status to Stopping, to mark that it is being shut down.
3469 : ///
3470 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3471 : ///
3472 : /// This function is not cancel-safe!
3473 : ///
3474 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
3475 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
3476 12 : async fn set_stopping(
3477 12 : &self,
3478 12 : progress: completion::Barrier,
3479 12 : _allow_transition_from_loading: bool,
3480 12 : allow_transition_from_attaching: bool,
3481 12 : ) -> Result<(), SetStoppingError> {
3482 12 : let mut rx = self.state.subscribe();
3483 12 :
3484 12 : // cannot stop before we're done activating, so wait out until we're done activating
3485 12 : rx.wait_for(|state| match state {
3486 0 : TenantState::Attaching if allow_transition_from_attaching => true,
3487 : TenantState::Activating(_) | TenantState::Attaching => {
3488 0 : info!(
3489 0 : "waiting for {} to turn Active|Broken|Stopping",
3490 0 : <&'static str>::from(state)
3491 : );
3492 0 : false
3493 : }
3494 12 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3495 12 : })
3496 12 : .await
3497 12 : .expect("cannot drop self.state while on a &self method");
3498 12 :
3499 12 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3500 12 : let mut err = None;
3501 12 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3502 : TenantState::Activating(_) => {
3503 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
3504 : }
3505 : TenantState::Attaching => {
3506 0 : if !allow_transition_from_attaching {
3507 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
3508 0 : };
3509 0 : *current_state = TenantState::Stopping { progress };
3510 0 : true
3511 : }
3512 : TenantState::Active => {
3513 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3514 : // are created after the transition to Stopping. That's harmless, as the Timelines
3515 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3516 12 : *current_state = TenantState::Stopping { progress };
3517 12 : // Continue stopping outside the closure. We need to grab timelines.lock()
3518 12 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3519 12 : true
3520 : }
3521 0 : TenantState::Broken { reason, .. } => {
3522 0 : info!(
3523 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3524 : );
3525 0 : err = Some(SetStoppingError::Broken);
3526 0 : false
3527 : }
3528 0 : TenantState::Stopping { progress } => {
3529 0 : info!("Tenant is already in Stopping state");
3530 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3531 0 : false
3532 : }
3533 12 : });
3534 12 : match (stopping, err) {
3535 12 : (true, None) => {} // continue
3536 0 : (false, Some(err)) => return Err(err),
3537 0 : (true, Some(_)) => unreachable!(
3538 0 : "send_if_modified closure must error out if not transitioning to Stopping"
3539 0 : ),
3540 0 : (false, None) => unreachable!(
3541 0 : "send_if_modified closure must return true if transitioning to Stopping"
3542 0 : ),
3543 : }
3544 :
3545 12 : let timelines_accessor = self.timelines.lock().unwrap();
3546 12 : let not_broken_timelines = timelines_accessor
3547 12 : .values()
3548 12 : .filter(|timeline| !timeline.is_broken());
3549 24 : for timeline in not_broken_timelines {
3550 12 : timeline.set_state(TimelineState::Stopping);
3551 12 : }
3552 12 : Ok(())
3553 12 : }
3554 :
3555 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3556 : /// `remove_tenant_from_memory`
3557 : ///
3558 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3559 : ///
3560 : /// In tests, we also use this to set tenants to Broken state on purpose.
3561 0 : pub(crate) async fn set_broken(&self, reason: String) {
3562 0 : let mut rx = self.state.subscribe();
3563 0 :
3564 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3565 0 : // So, wait until it's done.
3566 0 : rx.wait_for(|state| match state {
3567 : TenantState::Activating(_) | TenantState::Attaching => {
3568 0 : info!(
3569 0 : "waiting for {} to turn Active|Broken|Stopping",
3570 0 : <&'static str>::from(state)
3571 : );
3572 0 : false
3573 : }
3574 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3575 0 : })
3576 0 : .await
3577 0 : .expect("cannot drop self.state while on a &self method");
3578 0 :
3579 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3580 0 : self.set_broken_no_wait(reason)
3581 0 : }
3582 :
3583 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3584 0 : let reason = reason.to_string();
3585 0 : self.state.send_modify(|current_state| {
3586 0 : match *current_state {
3587 : TenantState::Activating(_) | TenantState::Attaching => {
3588 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3589 : }
3590 : TenantState::Active => {
3591 0 : if cfg!(feature = "testing") {
3592 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3593 0 : *current_state = TenantState::broken_from_reason(reason);
3594 : } else {
3595 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3596 : }
3597 : }
3598 : TenantState::Broken { .. } => {
3599 0 : warn!("Tenant is already in Broken state");
3600 : }
3601 : // This is the only "expected" path, any other path is a bug.
3602 : TenantState::Stopping { .. } => {
3603 0 : warn!(
3604 0 : "Marking Stopping tenant as Broken state, reason: {}",
3605 : reason
3606 : );
3607 0 : *current_state = TenantState::broken_from_reason(reason);
3608 : }
3609 : }
3610 0 : });
3611 0 : }
3612 :
3613 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3614 0 : self.state.subscribe()
3615 0 : }
3616 :
3617 : /// The activate_now semaphore is initialized with zero units. As soon as
3618 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3619 0 : pub(crate) fn activate_now(&self) {
3620 0 : self.activate_now_sem.add_permits(1);
3621 0 : }
3622 :
3623 0 : pub(crate) async fn wait_to_become_active(
3624 0 : &self,
3625 0 : timeout: Duration,
3626 0 : ) -> Result<(), GetActiveTenantError> {
3627 0 : let mut receiver = self.state.subscribe();
3628 : loop {
3629 0 : let current_state = receiver.borrow_and_update().clone();
3630 0 : match current_state {
3631 : TenantState::Attaching | TenantState::Activating(_) => {
3632 : // in these states, there's a chance that we can reach ::Active
3633 0 : self.activate_now();
3634 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3635 0 : Ok(r) => {
3636 0 : r.map_err(
3637 0 : |_e: tokio::sync::watch::error::RecvError|
3638 : // Tenant existed but was dropped: report it as non-existent
3639 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3640 0 : )?
3641 : }
3642 : Err(TimeoutCancellableError::Cancelled) => {
3643 0 : return Err(GetActiveTenantError::Cancelled);
3644 : }
3645 : Err(TimeoutCancellableError::Timeout) => {
3646 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3647 0 : latest_state: Some(self.current_state()),
3648 0 : wait_time: timeout,
3649 0 : });
3650 : }
3651 : }
3652 : }
3653 : TenantState::Active { .. } => {
3654 0 : return Ok(());
3655 : }
3656 0 : TenantState::Broken { reason, .. } => {
3657 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3658 0 : // it's logically a 500 to external API users (broken is always a bug).
3659 0 : return Err(GetActiveTenantError::Broken(reason));
3660 : }
3661 : TenantState::Stopping { .. } => {
3662 : // There's no chance the tenant can transition back into ::Active
3663 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3664 : }
3665 : }
3666 : }
3667 0 : }
3668 :
3669 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3670 0 : self.tenant_conf.load().location.attach_mode
3671 0 : }
3672 :
3673 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3674 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3675 : /// rare external API calls, like a reconciliation at startup.
3676 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3677 0 : let conf = self.tenant_conf.load();
3678 :
3679 0 : let location_config_mode = match conf.location.attach_mode {
3680 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3681 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3682 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3683 : };
3684 :
3685 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3686 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3687 0 :
3688 0 : models::LocationConfig {
3689 0 : mode: location_config_mode,
3690 0 : generation: self.generation.into(),
3691 0 : secondary_conf: None,
3692 0 : shard_number: self.shard_identity.number.0,
3693 0 : shard_count: self.shard_identity.count.literal(),
3694 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3695 0 : tenant_conf: tenant_config,
3696 0 : }
3697 0 : }
3698 :
3699 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3700 0 : &self.tenant_shard_id
3701 0 : }
3702 :
3703 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3704 0 : self.shard_identity.stripe_size
3705 0 : }
3706 :
3707 0 : pub(crate) fn get_generation(&self) -> Generation {
3708 0 : self.generation
3709 0 : }
3710 :
3711 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3712 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3713 : /// resetting this tenant to a valid state if we fail.
3714 0 : pub(crate) async fn split_prepare(
3715 0 : &self,
3716 0 : child_shards: &Vec<TenantShardId>,
3717 0 : ) -> anyhow::Result<()> {
3718 0 : let (timelines, offloaded) = {
3719 0 : let timelines = self.timelines.lock().unwrap();
3720 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3721 0 : (timelines.clone(), offloaded.clone())
3722 0 : };
3723 0 : let timelines_iter = timelines
3724 0 : .values()
3725 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3726 0 : .chain(
3727 0 : offloaded
3728 0 : .values()
3729 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3730 0 : );
3731 0 : for timeline in timelines_iter {
3732 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3733 : // to ensure that they do not start a split if currently in the process of doing these.
3734 :
3735 0 : let timeline_id = timeline.timeline_id();
3736 :
3737 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3738 : // Upload an index from the parent: this is partly to provide freshness for the
3739 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3740 : // always be a parent shard index in the same generation as we wrote the child shard index.
3741 0 : tracing::info!(%timeline_id, "Uploading index");
3742 0 : timeline
3743 0 : .remote_client
3744 0 : .schedule_index_upload_for_file_changes()?;
3745 0 : timeline.remote_client.wait_completion().await?;
3746 0 : }
3747 :
3748 0 : let remote_client = match timeline {
3749 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3750 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3751 0 : let remote_client = self
3752 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3753 0 : Arc::new(remote_client)
3754 : }
3755 : };
3756 :
3757 : // Shut down the timeline's remote client: this means that the indices we write
3758 : // for child shards will not be invalidated by the parent shard deleting layers.
3759 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3760 0 : remote_client.shutdown().await;
3761 :
3762 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3763 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3764 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3765 : // we use here really is the remotely persistent one).
3766 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3767 0 : let result = remote_client
3768 0 : .download_index_file(&self.cancel)
3769 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))
3770 0 : .await?;
3771 0 : let index_part = match result {
3772 : MaybeDeletedIndexPart::Deleted(_) => {
3773 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3774 : }
3775 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3776 : };
3777 :
3778 0 : for child_shard in child_shards {
3779 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3780 0 : upload_index_part(
3781 0 : &self.remote_storage,
3782 0 : child_shard,
3783 0 : &timeline_id,
3784 0 : self.generation,
3785 0 : &index_part,
3786 0 : &self.cancel,
3787 0 : )
3788 0 : .await?;
3789 : }
3790 : }
3791 :
3792 0 : let tenant_manifest = self.build_tenant_manifest();
3793 0 : for child_shard in child_shards {
3794 0 : tracing::info!(
3795 0 : "Uploading tenant manifest for child {}",
3796 0 : child_shard.to_index()
3797 : );
3798 0 : upload_tenant_manifest(
3799 0 : &self.remote_storage,
3800 0 : child_shard,
3801 0 : self.generation,
3802 0 : &tenant_manifest,
3803 0 : &self.cancel,
3804 0 : )
3805 0 : .await?;
3806 : }
3807 :
3808 0 : Ok(())
3809 0 : }
3810 :
3811 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3812 0 : let mut result = TopTenantShardItem {
3813 0 : id: self.tenant_shard_id,
3814 0 : resident_size: 0,
3815 0 : physical_size: 0,
3816 0 : max_logical_size: 0,
3817 0 : };
3818 :
3819 0 : for timeline in self.timelines.lock().unwrap().values() {
3820 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3821 0 :
3822 0 : result.physical_size += timeline
3823 0 : .remote_client
3824 0 : .metrics
3825 0 : .remote_physical_size_gauge
3826 0 : .get();
3827 0 : result.max_logical_size = std::cmp::max(
3828 0 : result.max_logical_size,
3829 0 : timeline.metrics.current_logical_size_gauge.get(),
3830 0 : );
3831 0 : }
3832 :
3833 0 : result
3834 0 : }
3835 : }
3836 :
3837 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3838 : /// perform a topological sort, so that the parent of each timeline comes
3839 : /// before the children.
3840 : /// E extracts the ancestor from T
3841 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3842 444 : fn tree_sort_timelines<T, E>(
3843 444 : timelines: HashMap<TimelineId, T>,
3844 444 : extractor: E,
3845 444 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3846 444 : where
3847 444 : E: Fn(&T) -> Option<TimelineId>,
3848 444 : {
3849 444 : let mut result = Vec::with_capacity(timelines.len());
3850 444 :
3851 444 : let mut now = Vec::with_capacity(timelines.len());
3852 444 : // (ancestor, children)
3853 444 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3854 444 : HashMap::with_capacity(timelines.len());
3855 :
3856 456 : for (timeline_id, value) in timelines {
3857 12 : if let Some(ancestor_id) = extractor(&value) {
3858 4 : let children = later.entry(ancestor_id).or_default();
3859 4 : children.push((timeline_id, value));
3860 8 : } else {
3861 8 : now.push((timeline_id, value));
3862 8 : }
3863 : }
3864 :
3865 456 : while let Some((timeline_id, metadata)) = now.pop() {
3866 12 : result.push((timeline_id, metadata));
3867 : // All children of this can be loaded now
3868 12 : if let Some(mut children) = later.remove(&timeline_id) {
3869 4 : now.append(&mut children);
3870 8 : }
3871 : }
3872 :
3873 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3874 444 : if !later.is_empty() {
3875 0 : for (missing_id, orphan_ids) in later {
3876 0 : for (orphan_id, _) in orphan_ids {
3877 0 : error!(
3878 0 : "could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded"
3879 : );
3880 : }
3881 : }
3882 0 : bail!("could not load tenant because some timelines are missing ancestors");
3883 444 : }
3884 444 :
3885 444 : Ok(result)
3886 444 : }
3887 :
3888 : enum ActivateTimelineArgs {
3889 : Yes {
3890 : broker_client: storage_broker::BrokerClientChannel,
3891 : },
3892 : No,
3893 : }
3894 :
3895 : impl Tenant {
3896 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3897 0 : self.tenant_conf.load().tenant_conf.clone()
3898 0 : }
3899 :
3900 0 : pub fn effective_config(&self) -> TenantConf {
3901 0 : self.tenant_specific_overrides()
3902 0 : .merge(self.conf.default_tenant_conf.clone())
3903 0 : }
3904 :
3905 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3906 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3907 0 : tenant_conf
3908 0 : .checkpoint_distance
3909 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3910 0 : }
3911 :
3912 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3913 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3914 0 : tenant_conf
3915 0 : .checkpoint_timeout
3916 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3917 0 : }
3918 :
3919 0 : pub fn get_compaction_target_size(&self) -> u64 {
3920 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3921 0 : tenant_conf
3922 0 : .compaction_target_size
3923 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3924 0 : }
3925 :
3926 0 : pub fn get_compaction_period(&self) -> Duration {
3927 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3928 0 : tenant_conf
3929 0 : .compaction_period
3930 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3931 0 : }
3932 :
3933 0 : pub fn get_compaction_threshold(&self) -> usize {
3934 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3935 0 : tenant_conf
3936 0 : .compaction_threshold
3937 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3938 0 : }
3939 :
3940 0 : pub fn get_rel_size_v2_enabled(&self) -> bool {
3941 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3942 0 : tenant_conf
3943 0 : .rel_size_v2_enabled
3944 0 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
3945 0 : }
3946 :
3947 0 : pub fn get_compaction_upper_limit(&self) -> usize {
3948 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3949 0 : tenant_conf
3950 0 : .compaction_upper_limit
3951 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
3952 0 : }
3953 :
3954 0 : pub fn get_compaction_l0_first(&self) -> bool {
3955 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3956 0 : tenant_conf
3957 0 : .compaction_l0_first
3958 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
3959 0 : }
3960 :
3961 0 : pub fn get_gc_horizon(&self) -> u64 {
3962 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3963 0 : tenant_conf
3964 0 : .gc_horizon
3965 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3966 0 : }
3967 :
3968 0 : pub fn get_gc_period(&self) -> Duration {
3969 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3970 0 : tenant_conf
3971 0 : .gc_period
3972 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
3973 0 : }
3974 :
3975 0 : pub fn get_image_creation_threshold(&self) -> usize {
3976 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3977 0 : tenant_conf
3978 0 : .image_creation_threshold
3979 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3980 0 : }
3981 :
3982 0 : pub fn get_pitr_interval(&self) -> Duration {
3983 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3984 0 : tenant_conf
3985 0 : .pitr_interval
3986 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
3987 0 : }
3988 :
3989 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
3990 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3991 0 : tenant_conf
3992 0 : .min_resident_size_override
3993 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
3994 0 : }
3995 :
3996 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
3997 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3998 0 : let heatmap_period = tenant_conf
3999 0 : .heatmap_period
4000 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
4001 0 : if heatmap_period.is_zero() {
4002 0 : None
4003 : } else {
4004 0 : Some(heatmap_period)
4005 : }
4006 0 : }
4007 :
4008 8 : pub fn get_lsn_lease_length(&self) -> Duration {
4009 8 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4010 8 : tenant_conf
4011 8 : .lsn_lease_length
4012 8 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
4013 8 : }
4014 :
4015 0 : pub fn get_timeline_offloading_enabled(&self) -> bool {
4016 0 : if self.conf.timeline_offloading {
4017 0 : return true;
4018 0 : }
4019 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4020 0 : tenant_conf
4021 0 : .timeline_offloading
4022 0 : .unwrap_or(self.conf.default_tenant_conf.timeline_offloading)
4023 0 : }
4024 :
4025 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
4026 4 : fn build_tenant_manifest(&self) -> TenantManifest {
4027 4 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
4028 4 :
4029 4 : let mut timeline_manifests = timelines_offloaded
4030 4 : .iter()
4031 4 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
4032 4 : .collect::<Vec<_>>();
4033 4 : // Sort the manifests so that our output is deterministic
4034 4 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
4035 4 :
4036 4 : TenantManifest {
4037 4 : version: LATEST_TENANT_MANIFEST_VERSION,
4038 4 : offloaded_timelines: timeline_manifests,
4039 4 : }
4040 4 : }
4041 :
4042 0 : pub fn update_tenant_config<F: Fn(TenantConfOpt) -> anyhow::Result<TenantConfOpt>>(
4043 0 : &self,
4044 0 : update: F,
4045 0 : ) -> anyhow::Result<TenantConfOpt> {
4046 0 : // Use read-copy-update in order to avoid overwriting the location config
4047 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
4048 0 : // this race is not possible if both request types come from the storage
4049 0 : // controller (as they should!) because an exclusive op lock is required
4050 0 : // on the storage controller side.
4051 0 :
4052 0 : self.tenant_conf
4053 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
4054 0 : Ok(Arc::new(AttachedTenantConf {
4055 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
4056 0 : location: attached_conf.location,
4057 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
4058 : }))
4059 0 : })?;
4060 :
4061 0 : let updated = self.tenant_conf.load();
4062 0 :
4063 0 : self.tenant_conf_updated(&updated.tenant_conf);
4064 0 : // Don't hold self.timelines.lock() during the notifies.
4065 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4066 0 : // mutexes in struct Timeline in the future.
4067 0 : let timelines = self.list_timelines();
4068 0 : for timeline in timelines {
4069 0 : timeline.tenant_conf_updated(&updated);
4070 0 : }
4071 :
4072 0 : Ok(updated.tenant_conf.clone())
4073 0 : }
4074 :
4075 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4076 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4077 0 :
4078 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4079 0 :
4080 0 : self.tenant_conf_updated(&new_tenant_conf);
4081 0 : // Don't hold self.timelines.lock() during the notifies.
4082 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4083 0 : // mutexes in struct Timeline in the future.
4084 0 : let timelines = self.list_timelines();
4085 0 : for timeline in timelines {
4086 0 : timeline.tenant_conf_updated(&new_conf);
4087 0 : }
4088 0 : }
4089 :
4090 444 : fn get_pagestream_throttle_config(
4091 444 : psconf: &'static PageServerConf,
4092 444 : overrides: &TenantConfOpt,
4093 444 : ) -> throttle::Config {
4094 444 : overrides
4095 444 : .timeline_get_throttle
4096 444 : .clone()
4097 444 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4098 444 : }
4099 :
4100 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
4101 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4102 0 : self.pagestream_throttle.reconfigure(conf)
4103 0 : }
4104 :
4105 : /// Helper function to create a new Timeline struct.
4106 : ///
4107 : /// The returned Timeline is in Loading state. The caller is responsible for
4108 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4109 : /// map.
4110 : ///
4111 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4112 : /// and we might not have the ancestor present anymore which is fine for to be
4113 : /// deleted timelines.
4114 : #[allow(clippy::too_many_arguments)]
4115 896 : fn create_timeline_struct(
4116 896 : &self,
4117 896 : new_timeline_id: TimelineId,
4118 896 : new_metadata: &TimelineMetadata,
4119 896 : previous_heatmap: Option<PreviousHeatmap>,
4120 896 : ancestor: Option<Arc<Timeline>>,
4121 896 : resources: TimelineResources,
4122 896 : cause: CreateTimelineCause,
4123 896 : create_idempotency: CreateTimelineIdempotency,
4124 896 : gc_compaction_state: Option<GcCompactionState>,
4125 896 : ) -> anyhow::Result<Arc<Timeline>> {
4126 896 : let state = match cause {
4127 : CreateTimelineCause::Load => {
4128 896 : let ancestor_id = new_metadata.ancestor_timeline();
4129 896 : anyhow::ensure!(
4130 896 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4131 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4132 : );
4133 896 : TimelineState::Loading
4134 : }
4135 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4136 : };
4137 :
4138 896 : let pg_version = new_metadata.pg_version();
4139 896 :
4140 896 : let timeline = Timeline::new(
4141 896 : self.conf,
4142 896 : Arc::clone(&self.tenant_conf),
4143 896 : new_metadata,
4144 896 : previous_heatmap,
4145 896 : ancestor,
4146 896 : new_timeline_id,
4147 896 : self.tenant_shard_id,
4148 896 : self.generation,
4149 896 : self.shard_identity,
4150 896 : self.walredo_mgr.clone(),
4151 896 : resources,
4152 896 : pg_version,
4153 896 : state,
4154 896 : self.attach_wal_lag_cooldown.clone(),
4155 896 : create_idempotency,
4156 896 : gc_compaction_state,
4157 896 : self.cancel.child_token(),
4158 896 : );
4159 896 :
4160 896 : Ok(timeline)
4161 896 : }
4162 :
4163 : /// [`Tenant::shutdown`] must be called before dropping the returned [`Tenant`] object
4164 : /// to ensure proper cleanup of background tasks and metrics.
4165 : //
4166 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4167 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4168 : #[allow(clippy::too_many_arguments)]
4169 444 : fn new(
4170 444 : state: TenantState,
4171 444 : conf: &'static PageServerConf,
4172 444 : attached_conf: AttachedTenantConf,
4173 444 : shard_identity: ShardIdentity,
4174 444 : walredo_mgr: Option<Arc<WalRedoManager>>,
4175 444 : tenant_shard_id: TenantShardId,
4176 444 : remote_storage: GenericRemoteStorage,
4177 444 : deletion_queue_client: DeletionQueueClient,
4178 444 : l0_flush_global_state: L0FlushGlobalState,
4179 444 : ) -> Tenant {
4180 444 : debug_assert!(
4181 444 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
4182 : );
4183 :
4184 444 : let (state, mut rx) = watch::channel(state);
4185 444 :
4186 444 : tokio::spawn(async move {
4187 444 : // reflect tenant state in metrics:
4188 444 : // - global per tenant state: TENANT_STATE_METRIC
4189 444 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4190 444 : //
4191 444 : // set of broken tenants should not have zero counts so that it remains accessible for
4192 444 : // alerting.
4193 444 :
4194 444 : let tid = tenant_shard_id.to_string();
4195 444 : let shard_id = tenant_shard_id.shard_slug().to_string();
4196 444 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4197 :
4198 888 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4199 888 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4200 888 : }
4201 :
4202 444 : let mut tuple = inspect_state(&rx.borrow_and_update());
4203 444 :
4204 444 : let is_broken = tuple.1;
4205 444 : let mut counted_broken = if is_broken {
4206 : // add the id to the set right away, there should not be any updates on the channel
4207 : // after before tenant is removed, if ever
4208 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4209 0 : true
4210 : } else {
4211 444 : false
4212 : };
4213 :
4214 : loop {
4215 888 : let labels = &tuple.0;
4216 888 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4217 888 : current.inc();
4218 888 :
4219 888 : if rx.changed().await.is_err() {
4220 : // tenant has been dropped
4221 28 : current.dec();
4222 28 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4223 28 : break;
4224 444 : }
4225 444 :
4226 444 : current.dec();
4227 444 : tuple = inspect_state(&rx.borrow_and_update());
4228 444 :
4229 444 : let is_broken = tuple.1;
4230 444 : if is_broken && !counted_broken {
4231 0 : counted_broken = true;
4232 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4233 0 : // access
4234 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4235 444 : }
4236 : }
4237 444 : });
4238 444 :
4239 444 : Tenant {
4240 444 : tenant_shard_id,
4241 444 : shard_identity,
4242 444 : generation: attached_conf.location.generation,
4243 444 : conf,
4244 444 : // using now here is good enough approximation to catch tenants with really long
4245 444 : // activation times.
4246 444 : constructed_at: Instant::now(),
4247 444 : timelines: Mutex::new(HashMap::new()),
4248 444 : timelines_creating: Mutex::new(HashSet::new()),
4249 444 : timelines_offloaded: Mutex::new(HashMap::new()),
4250 444 : tenant_manifest_upload: Default::default(),
4251 444 : gc_cs: tokio::sync::Mutex::new(()),
4252 444 : walredo_mgr,
4253 444 : remote_storage,
4254 444 : deletion_queue_client,
4255 444 : state,
4256 444 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4257 444 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4258 444 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4259 444 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4260 444 : format!("compaction-{tenant_shard_id}"),
4261 444 : 5,
4262 444 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4263 444 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4264 444 : // use an extremely long backoff.
4265 444 : Some(Duration::from_secs(3600 * 24)),
4266 444 : )),
4267 444 : l0_compaction_trigger: Arc::new(Notify::new()),
4268 444 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4269 444 : activate_now_sem: tokio::sync::Semaphore::new(0),
4270 444 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4271 444 : cancel: CancellationToken::default(),
4272 444 : gate: Gate::default(),
4273 444 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4274 444 : Tenant::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4275 444 : )),
4276 444 : pagestream_throttle_metrics: Arc::new(
4277 444 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4278 444 : ),
4279 444 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4280 444 : ongoing_timeline_detach: std::sync::Mutex::default(),
4281 444 : gc_block: Default::default(),
4282 444 : l0_flush_global_state,
4283 444 : }
4284 444 : }
4285 :
4286 : /// Locate and load config
4287 0 : pub(super) fn load_tenant_config(
4288 0 : conf: &'static PageServerConf,
4289 0 : tenant_shard_id: &TenantShardId,
4290 0 : ) -> Result<LocationConf, LoadConfigError> {
4291 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4292 0 :
4293 0 : info!("loading tenant configuration from {config_path}");
4294 :
4295 : // load and parse file
4296 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4297 0 : match e.kind() {
4298 : std::io::ErrorKind::NotFound => {
4299 : // The config should almost always exist for a tenant directory:
4300 : // - When attaching a tenant, the config is the first thing we write
4301 : // - When detaching a tenant, we atomically move the directory to a tmp location
4302 : // before deleting contents.
4303 : //
4304 : // The very rare edge case that can result in a missing config is if we crash during attach
4305 : // between creating directory and writing config. Callers should handle that as if the
4306 : // directory didn't exist.
4307 :
4308 0 : LoadConfigError::NotFound(config_path)
4309 : }
4310 : _ => {
4311 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4312 : // that we cannot cleanly recover
4313 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4314 : }
4315 : }
4316 0 : })?;
4317 :
4318 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4319 0 : }
4320 :
4321 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4322 : pub(super) async fn persist_tenant_config(
4323 : conf: &'static PageServerConf,
4324 : tenant_shard_id: &TenantShardId,
4325 : location_conf: &LocationConf,
4326 : ) -> std::io::Result<()> {
4327 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4328 :
4329 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4330 : }
4331 :
4332 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4333 : pub(super) async fn persist_tenant_config_at(
4334 : tenant_shard_id: &TenantShardId,
4335 : config_path: &Utf8Path,
4336 : location_conf: &LocationConf,
4337 : ) -> std::io::Result<()> {
4338 : debug!("persisting tenantconf to {config_path}");
4339 :
4340 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4341 : # It is read in case of pageserver restart.
4342 : "#
4343 : .to_string();
4344 :
4345 0 : fail::fail_point!("tenant-config-before-write", |_| {
4346 0 : Err(std::io::Error::new(
4347 0 : std::io::ErrorKind::Other,
4348 0 : "tenant-config-before-write",
4349 0 : ))
4350 0 : });
4351 :
4352 : // Convert the config to a toml file.
4353 : conf_content +=
4354 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4355 :
4356 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4357 :
4358 : let conf_content = conf_content.into_bytes();
4359 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4360 : }
4361 :
4362 : //
4363 : // How garbage collection works:
4364 : //
4365 : // +--bar------------->
4366 : // /
4367 : // +----+-----foo---------------->
4368 : // /
4369 : // ----main--+-------------------------->
4370 : // \
4371 : // +-----baz-------->
4372 : //
4373 : //
4374 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4375 : // `gc_infos` are being refreshed
4376 : // 2. Scan collected timelines, and on each timeline, make note of the
4377 : // all the points where other timelines have been branched off.
4378 : // We will refrain from removing page versions at those LSNs.
4379 : // 3. For each timeline, scan all layer files on the timeline.
4380 : // Remove all files for which a newer file exists and which
4381 : // don't cover any branch point LSNs.
4382 : //
4383 : // TODO:
4384 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4385 : // don't need to keep that in the parent anymore. But currently
4386 : // we do.
4387 8 : async fn gc_iteration_internal(
4388 8 : &self,
4389 8 : target_timeline_id: Option<TimelineId>,
4390 8 : horizon: u64,
4391 8 : pitr: Duration,
4392 8 : cancel: &CancellationToken,
4393 8 : ctx: &RequestContext,
4394 8 : ) -> Result<GcResult, GcError> {
4395 8 : let mut totals: GcResult = Default::default();
4396 8 : let now = Instant::now();
4397 :
4398 8 : let gc_timelines = self
4399 8 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4400 8 : .await?;
4401 :
4402 8 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4403 :
4404 : // If there is nothing to GC, we don't want any messages in the INFO log.
4405 8 : if !gc_timelines.is_empty() {
4406 8 : info!("{} timelines need GC", gc_timelines.len());
4407 : } else {
4408 0 : debug!("{} timelines need GC", gc_timelines.len());
4409 : }
4410 :
4411 : // Perform GC for each timeline.
4412 : //
4413 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
4414 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4415 : // with branch creation.
4416 : //
4417 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
4418 : // creation task can run concurrently with timeline's GC iteration.
4419 16 : for timeline in gc_timelines {
4420 8 : if cancel.is_cancelled() {
4421 : // We were requested to shut down. Stop and return with the progress we
4422 : // made.
4423 0 : break;
4424 8 : }
4425 8 : let result = match timeline.gc().await {
4426 : Err(GcError::TimelineCancelled) => {
4427 0 : if target_timeline_id.is_some() {
4428 : // If we were targetting this specific timeline, surface cancellation to caller
4429 0 : return Err(GcError::TimelineCancelled);
4430 : } else {
4431 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4432 : // skip past this and proceed to try GC on other timelines.
4433 0 : continue;
4434 : }
4435 : }
4436 8 : r => r?,
4437 : };
4438 8 : totals += result;
4439 : }
4440 :
4441 8 : totals.elapsed = now.elapsed();
4442 8 : Ok(totals)
4443 8 : }
4444 :
4445 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4446 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4447 : /// [`Tenant::get_gc_horizon`].
4448 : ///
4449 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4450 0 : pub(crate) async fn refresh_gc_info(
4451 0 : &self,
4452 0 : cancel: &CancellationToken,
4453 0 : ctx: &RequestContext,
4454 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4455 0 : // since this method can now be called at different rates than the configured gc loop, it
4456 0 : // might be that these configuration values get applied faster than what it was previously,
4457 0 : // since these were only read from the gc task.
4458 0 : let horizon = self.get_gc_horizon();
4459 0 : let pitr = self.get_pitr_interval();
4460 0 :
4461 0 : // refresh all timelines
4462 0 : let target_timeline_id = None;
4463 0 :
4464 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4465 0 : .await
4466 0 : }
4467 :
4468 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4469 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4470 : ///
4471 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4472 0 : fn initialize_gc_info(
4473 0 : &self,
4474 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4475 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4476 0 : restrict_to_timeline: Option<TimelineId>,
4477 0 : ) {
4478 0 : if restrict_to_timeline.is_none() {
4479 : // This function must be called before activation: after activation timeline create/delete operations
4480 : // might happen, and this function is not safe to run concurrently with those.
4481 0 : assert!(!self.is_active());
4482 0 : }
4483 :
4484 : // Scan all timelines. For each timeline, remember the timeline ID and
4485 : // the branch point where it was created.
4486 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4487 0 : BTreeMap::new();
4488 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4489 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4490 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4491 0 : ancestor_children.push((
4492 0 : timeline_entry.get_ancestor_lsn(),
4493 0 : *timeline_id,
4494 0 : MaybeOffloaded::No,
4495 0 : ));
4496 0 : }
4497 0 : });
4498 0 : timelines_offloaded
4499 0 : .iter()
4500 0 : .for_each(|(timeline_id, timeline_entry)| {
4501 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4502 0 : return;
4503 : };
4504 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4505 0 : return;
4506 : };
4507 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4508 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4509 0 : });
4510 0 :
4511 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4512 0 : let horizon = self.get_gc_horizon();
4513 :
4514 : // Populate each timeline's GcInfo with information about its child branches
4515 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4516 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4517 : } else {
4518 0 : itertools::Either::Right(timelines.values())
4519 : };
4520 0 : for timeline in timelines_to_write {
4521 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4522 0 : .remove(&timeline.timeline_id)
4523 0 : .unwrap_or_default();
4524 0 :
4525 0 : branchpoints.sort_by_key(|b| b.0);
4526 0 :
4527 0 : let mut target = timeline.gc_info.write().unwrap();
4528 0 :
4529 0 : target.retain_lsns = branchpoints;
4530 0 :
4531 0 : let space_cutoff = timeline
4532 0 : .get_last_record_lsn()
4533 0 : .checked_sub(horizon)
4534 0 : .unwrap_or(Lsn(0));
4535 0 :
4536 0 : target.cutoffs = GcCutoffs {
4537 0 : space: space_cutoff,
4538 0 : time: Lsn::INVALID,
4539 0 : };
4540 0 : }
4541 0 : }
4542 :
4543 8 : async fn refresh_gc_info_internal(
4544 8 : &self,
4545 8 : target_timeline_id: Option<TimelineId>,
4546 8 : horizon: u64,
4547 8 : pitr: Duration,
4548 8 : cancel: &CancellationToken,
4549 8 : ctx: &RequestContext,
4550 8 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4551 8 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4552 8 : // currently visible timelines.
4553 8 : let timelines = self
4554 8 : .timelines
4555 8 : .lock()
4556 8 : .unwrap()
4557 8 : .values()
4558 8 : .filter(|tl| match target_timeline_id.as_ref() {
4559 8 : Some(target) => &tl.timeline_id == target,
4560 0 : None => true,
4561 8 : })
4562 8 : .cloned()
4563 8 : .collect::<Vec<_>>();
4564 8 :
4565 8 : if target_timeline_id.is_some() && timelines.is_empty() {
4566 : // We were to act on a particular timeline and it wasn't found
4567 0 : return Err(GcError::TimelineNotFound);
4568 8 : }
4569 8 :
4570 8 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4571 8 : HashMap::with_capacity(timelines.len());
4572 8 :
4573 8 : // Ensures all timelines use the same start time when computing the time cutoff.
4574 8 : let now_ts_for_pitr_calc = SystemTime::now();
4575 8 : for timeline in timelines.iter() {
4576 8 : let cutoff = timeline
4577 8 : .get_last_record_lsn()
4578 8 : .checked_sub(horizon)
4579 8 : .unwrap_or(Lsn(0));
4580 :
4581 8 : let cutoffs = timeline
4582 8 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4583 8 : .await?;
4584 8 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4585 8 : assert!(old.is_none());
4586 : }
4587 :
4588 8 : if !self.is_active() || self.cancel.is_cancelled() {
4589 0 : return Err(GcError::TenantCancelled);
4590 8 : }
4591 :
4592 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4593 : // because that will stall branch creation.
4594 8 : let gc_cs = self.gc_cs.lock().await;
4595 :
4596 : // Ok, we now know all the branch points.
4597 : // Update the GC information for each timeline.
4598 8 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4599 16 : for timeline in timelines {
4600 : // We filtered the timeline list above
4601 8 : if let Some(target_timeline_id) = target_timeline_id {
4602 8 : assert_eq!(target_timeline_id, timeline.timeline_id);
4603 0 : }
4604 :
4605 : {
4606 8 : let mut target = timeline.gc_info.write().unwrap();
4607 8 :
4608 8 : // Cull any expired leases
4609 8 : let now = SystemTime::now();
4610 12 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4611 8 :
4612 8 : timeline
4613 8 : .metrics
4614 8 : .valid_lsn_lease_count_gauge
4615 8 : .set(target.leases.len() as u64);
4616 :
4617 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4618 8 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4619 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4620 0 : target.within_ancestor_pitr =
4621 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4622 0 : }
4623 8 : }
4624 :
4625 : // Update metrics that depend on GC state
4626 8 : timeline
4627 8 : .metrics
4628 8 : .archival_size
4629 8 : .set(if target.within_ancestor_pitr {
4630 0 : timeline.metrics.current_logical_size_gauge.get()
4631 : } else {
4632 8 : 0
4633 : });
4634 8 : timeline.metrics.pitr_history_size.set(
4635 8 : timeline
4636 8 : .get_last_record_lsn()
4637 8 : .checked_sub(target.cutoffs.time)
4638 8 : .unwrap_or(Lsn(0))
4639 8 : .0,
4640 8 : );
4641 :
4642 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4643 : // - this timeline was created while we were finding cutoffs
4644 : // - lsn for timestamp search fails for this timeline repeatedly
4645 8 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4646 8 : let original_cutoffs = target.cutoffs.clone();
4647 8 : // GC cutoffs should never go back
4648 8 : target.cutoffs = GcCutoffs {
4649 8 : space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
4650 8 : time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
4651 8 : }
4652 0 : }
4653 : }
4654 :
4655 8 : gc_timelines.push(timeline);
4656 : }
4657 8 : drop(gc_cs);
4658 8 : Ok(gc_timelines)
4659 8 : }
4660 :
4661 : /// A substitute for `branch_timeline` for use in unit tests.
4662 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4663 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4664 : /// timeline background tasks are launched, except the flush loop.
4665 : #[cfg(test)]
4666 464 : async fn branch_timeline_test(
4667 464 : self: &Arc<Self>,
4668 464 : src_timeline: &Arc<Timeline>,
4669 464 : dst_id: TimelineId,
4670 464 : ancestor_lsn: Option<Lsn>,
4671 464 : ctx: &RequestContext,
4672 464 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4673 464 : let tl = self
4674 464 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4675 464 : .await?
4676 456 : .into_timeline_for_test();
4677 456 : tl.set_state(TimelineState::Active);
4678 456 : Ok(tl)
4679 464 : }
4680 :
4681 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4682 : #[cfg(test)]
4683 : #[allow(clippy::too_many_arguments)]
4684 12 : pub async fn branch_timeline_test_with_layers(
4685 12 : self: &Arc<Self>,
4686 12 : src_timeline: &Arc<Timeline>,
4687 12 : dst_id: TimelineId,
4688 12 : ancestor_lsn: Option<Lsn>,
4689 12 : ctx: &RequestContext,
4690 12 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4691 12 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4692 12 : end_lsn: Lsn,
4693 12 : ) -> anyhow::Result<Arc<Timeline>> {
4694 : use checks::check_valid_layermap;
4695 : use itertools::Itertools;
4696 :
4697 12 : let tline = self
4698 12 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4699 12 : .await?;
4700 12 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4701 12 : ancestor_lsn
4702 : } else {
4703 0 : tline.get_last_record_lsn()
4704 : };
4705 12 : assert!(end_lsn >= ancestor_lsn);
4706 12 : tline.force_advance_lsn(end_lsn);
4707 24 : for deltas in delta_layer_desc {
4708 12 : tline
4709 12 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4710 12 : .await?;
4711 : }
4712 20 : for (lsn, images) in image_layer_desc {
4713 8 : tline
4714 8 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4715 8 : .await?;
4716 : }
4717 12 : let layer_names = tline
4718 12 : .layers
4719 12 : .read()
4720 12 : .await
4721 12 : .layer_map()
4722 12 : .unwrap()
4723 12 : .iter_historic_layers()
4724 20 : .map(|layer| layer.layer_name())
4725 12 : .collect_vec();
4726 12 : if let Some(err) = check_valid_layermap(&layer_names) {
4727 0 : bail!("invalid layermap: {err}");
4728 12 : }
4729 12 : Ok(tline)
4730 12 : }
4731 :
4732 : /// Branch an existing timeline.
4733 0 : async fn branch_timeline(
4734 0 : self: &Arc<Self>,
4735 0 : src_timeline: &Arc<Timeline>,
4736 0 : dst_id: TimelineId,
4737 0 : start_lsn: Option<Lsn>,
4738 0 : ctx: &RequestContext,
4739 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4740 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4741 0 : .await
4742 0 : }
4743 :
4744 464 : async fn branch_timeline_impl(
4745 464 : self: &Arc<Self>,
4746 464 : src_timeline: &Arc<Timeline>,
4747 464 : dst_id: TimelineId,
4748 464 : start_lsn: Option<Lsn>,
4749 464 : _ctx: &RequestContext,
4750 464 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4751 464 : let src_id = src_timeline.timeline_id;
4752 :
4753 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4754 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4755 : // valid while we are creating the branch.
4756 464 : let _gc_cs = self.gc_cs.lock().await;
4757 :
4758 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4759 464 : let start_lsn = start_lsn.unwrap_or_else(|| {
4760 4 : let lsn = src_timeline.get_last_record_lsn();
4761 4 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4762 4 : lsn
4763 464 : });
4764 :
4765 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4766 464 : let timeline_create_guard = match self
4767 464 : .start_creating_timeline(
4768 464 : dst_id,
4769 464 : CreateTimelineIdempotency::Branch {
4770 464 : ancestor_timeline_id: src_timeline.timeline_id,
4771 464 : ancestor_start_lsn: start_lsn,
4772 464 : },
4773 464 : )
4774 464 : .await?
4775 : {
4776 464 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4777 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4778 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4779 : }
4780 : };
4781 :
4782 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4783 : // horizon on the source timeline
4784 : //
4785 : // We check it against both the planned GC cutoff stored in 'gc_info',
4786 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4787 : // planned GC cutoff in 'gc_info' is normally larger than
4788 : // 'applied_gc_cutoff_lsn', but beware of corner cases like if you just
4789 : // changed the GC settings for the tenant to make the PITR window
4790 : // larger, but some of the data was already removed by an earlier GC
4791 : // iteration.
4792 :
4793 : // check against last actual 'latest_gc_cutoff' first
4794 464 : let applied_gc_cutoff_lsn = src_timeline.get_applied_gc_cutoff_lsn();
4795 464 : {
4796 464 : let gc_info = src_timeline.gc_info.read().unwrap();
4797 464 : let planned_cutoff = gc_info.min_cutoff();
4798 464 : if gc_info.lsn_covered_by_lease(start_lsn) {
4799 0 : tracing::info!(
4800 0 : "skipping comparison of {start_lsn} with gc cutoff {} and planned gc cutoff {planned_cutoff} due to lsn lease",
4801 0 : *applied_gc_cutoff_lsn
4802 : );
4803 : } else {
4804 464 : src_timeline
4805 464 : .check_lsn_is_in_scope(start_lsn, &applied_gc_cutoff_lsn)
4806 464 : .context(format!(
4807 464 : "invalid branch start lsn: less than latest GC cutoff {}",
4808 464 : *applied_gc_cutoff_lsn,
4809 464 : ))
4810 464 : .map_err(CreateTimelineError::AncestorLsn)?;
4811 :
4812 : // and then the planned GC cutoff
4813 456 : if start_lsn < planned_cutoff {
4814 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4815 0 : "invalid branch start lsn: less than planned GC cutoff {planned_cutoff}"
4816 0 : )));
4817 456 : }
4818 : }
4819 : }
4820 :
4821 : //
4822 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4823 : // so that GC cannot advance the GC cutoff until we are finished.
4824 : // Proceed with the branch creation.
4825 : //
4826 :
4827 : // Determine prev-LSN for the new timeline. We can only determine it if
4828 : // the timeline was branched at the current end of the source timeline.
4829 : let RecordLsn {
4830 456 : last: src_last,
4831 456 : prev: src_prev,
4832 456 : } = src_timeline.get_last_record_rlsn();
4833 456 : let dst_prev = if src_last == start_lsn {
4834 432 : Some(src_prev)
4835 : } else {
4836 24 : None
4837 : };
4838 :
4839 : // Create the metadata file, noting the ancestor of the new timeline.
4840 : // There is initially no data in it, but all the read-calls know to look
4841 : // into the ancestor.
4842 456 : let metadata = TimelineMetadata::new(
4843 456 : start_lsn,
4844 456 : dst_prev,
4845 456 : Some(src_id),
4846 456 : start_lsn,
4847 456 : *src_timeline.applied_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4848 456 : src_timeline.initdb_lsn,
4849 456 : src_timeline.pg_version,
4850 456 : );
4851 :
4852 456 : let uninitialized_timeline = self
4853 456 : .prepare_new_timeline(
4854 456 : dst_id,
4855 456 : &metadata,
4856 456 : timeline_create_guard,
4857 456 : start_lsn + 1,
4858 456 : Some(Arc::clone(src_timeline)),
4859 456 : )
4860 456 : .await?;
4861 :
4862 456 : let new_timeline = uninitialized_timeline.finish_creation().await?;
4863 :
4864 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4865 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4866 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4867 : // could get incorrect information and remove more layers, than needed.
4868 : // See also https://github.com/neondatabase/neon/issues/3865
4869 456 : new_timeline
4870 456 : .remote_client
4871 456 : .schedule_index_upload_for_full_metadata_update(&metadata)
4872 456 : .context("branch initial metadata upload")?;
4873 :
4874 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4875 :
4876 456 : Ok(CreateTimelineResult::Created(new_timeline))
4877 464 : }
4878 :
4879 : /// For unit tests, make this visible so that other modules can directly create timelines
4880 : #[cfg(test)]
4881 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4882 : pub(crate) async fn bootstrap_timeline_test(
4883 : self: &Arc<Self>,
4884 : timeline_id: TimelineId,
4885 : pg_version: u32,
4886 : load_existing_initdb: Option<TimelineId>,
4887 : ctx: &RequestContext,
4888 : ) -> anyhow::Result<Arc<Timeline>> {
4889 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4890 : .await
4891 : .map_err(anyhow::Error::new)
4892 4 : .map(|r| r.into_timeline_for_test())
4893 : }
4894 :
4895 : /// Get exclusive access to the timeline ID for creation.
4896 : ///
4897 : /// Timeline-creating code paths must use this function before making changes
4898 : /// to in-memory or persistent state.
4899 : ///
4900 : /// The `state` parameter is a description of the timeline creation operation
4901 : /// we intend to perform.
4902 : /// If the timeline was already created in the meantime, we check whether this
4903 : /// request conflicts or is idempotent , based on `state`.
4904 896 : async fn start_creating_timeline(
4905 896 : self: &Arc<Self>,
4906 896 : new_timeline_id: TimelineId,
4907 896 : idempotency: CreateTimelineIdempotency,
4908 896 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4909 896 : let allow_offloaded = false;
4910 896 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4911 892 : Ok(create_guard) => {
4912 892 : pausable_failpoint!("timeline-creation-after-uninit");
4913 892 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4914 : }
4915 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4916 : Err(TimelineExclusionError::AlreadyCreating) => {
4917 : // Creation is in progress, we cannot create it again, and we cannot
4918 : // check if this request matches the existing one, so caller must try
4919 : // again later.
4920 0 : Err(CreateTimelineError::AlreadyCreating)
4921 : }
4922 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4923 : Err(TimelineExclusionError::AlreadyExists {
4924 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4925 0 : ..
4926 0 : }) => {
4927 0 : info!("timeline already exists but is offloaded");
4928 0 : Err(CreateTimelineError::Conflict)
4929 : }
4930 : Err(TimelineExclusionError::AlreadyExists {
4931 4 : existing: TimelineOrOffloaded::Timeline(existing),
4932 4 : arg,
4933 4 : }) => {
4934 4 : {
4935 4 : let existing = &existing.create_idempotency;
4936 4 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4937 4 : debug!("timeline already exists");
4938 :
4939 4 : match (existing, &arg) {
4940 : // FailWithConflict => no idempotency check
4941 : (CreateTimelineIdempotency::FailWithConflict, _)
4942 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4943 4 : warn!("timeline already exists, failing request");
4944 4 : return Err(CreateTimelineError::Conflict);
4945 : }
4946 : // Idempotent <=> CreateTimelineIdempotency is identical
4947 0 : (x, y) if x == y => {
4948 0 : info!(
4949 0 : "timeline already exists and idempotency matches, succeeding request"
4950 : );
4951 : // fallthrough
4952 : }
4953 : (_, _) => {
4954 0 : warn!("idempotency conflict, failing request");
4955 0 : return Err(CreateTimelineError::Conflict);
4956 : }
4957 : }
4958 : }
4959 :
4960 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
4961 : }
4962 : }
4963 896 : }
4964 :
4965 0 : async fn upload_initdb(
4966 0 : &self,
4967 0 : timelines_path: &Utf8PathBuf,
4968 0 : pgdata_path: &Utf8PathBuf,
4969 0 : timeline_id: &TimelineId,
4970 0 : ) -> anyhow::Result<()> {
4971 0 : let temp_path = timelines_path.join(format!(
4972 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
4973 0 : ));
4974 0 :
4975 0 : scopeguard::defer! {
4976 0 : if let Err(e) = fs::remove_file(&temp_path) {
4977 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
4978 0 : }
4979 0 : }
4980 :
4981 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
4982 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
4983 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
4984 0 : warn!(
4985 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
4986 : );
4987 0 : }
4988 :
4989 0 : pausable_failpoint!("before-initdb-upload");
4990 :
4991 0 : backoff::retry(
4992 0 : || async {
4993 0 : self::remote_timeline_client::upload_initdb_dir(
4994 0 : &self.remote_storage,
4995 0 : &self.tenant_shard_id.tenant_id,
4996 0 : timeline_id,
4997 0 : pgdata_zstd.try_clone().await?,
4998 0 : tar_zst_size,
4999 0 : &self.cancel,
5000 0 : )
5001 0 : .await
5002 0 : },
5003 0 : |_| false,
5004 0 : 3,
5005 0 : u32::MAX,
5006 0 : "persist_initdb_tar_zst",
5007 0 : &self.cancel,
5008 0 : )
5009 0 : .await
5010 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
5011 0 : .and_then(|x| x)
5012 0 : }
5013 :
5014 : /// - run initdb to init temporary instance and get bootstrap data
5015 : /// - after initialization completes, tar up the temp dir and upload it to S3.
5016 4 : async fn bootstrap_timeline(
5017 4 : self: &Arc<Self>,
5018 4 : timeline_id: TimelineId,
5019 4 : pg_version: u32,
5020 4 : load_existing_initdb: Option<TimelineId>,
5021 4 : ctx: &RequestContext,
5022 4 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
5023 4 : let timeline_create_guard = match self
5024 4 : .start_creating_timeline(
5025 4 : timeline_id,
5026 4 : CreateTimelineIdempotency::Bootstrap { pg_version },
5027 4 : )
5028 4 : .await?
5029 : {
5030 4 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5031 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5032 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5033 : }
5034 : };
5035 :
5036 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
5037 : // temporary directory for basebackup files for the given timeline.
5038 :
5039 4 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
5040 4 : let pgdata_path = path_with_suffix_extension(
5041 4 : timelines_path.join(format!("basebackup-{timeline_id}")),
5042 4 : TEMP_FILE_SUFFIX,
5043 4 : );
5044 4 :
5045 4 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
5046 4 : // we won't race with other creations or existent timelines with the same path.
5047 4 : if pgdata_path.exists() {
5048 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
5049 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
5050 0 : })?;
5051 4 : }
5052 :
5053 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
5054 4 : let pgdata_path_deferred = pgdata_path.clone();
5055 4 : scopeguard::defer! {
5056 4 : if let Err(e) = fs::remove_dir_all(&pgdata_path_deferred) {
5057 4 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
5058 4 : error!("Failed to remove temporary initdb directory '{pgdata_path_deferred}': {e}");
5059 4 : }
5060 4 : }
5061 4 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
5062 4 : if existing_initdb_timeline_id != timeline_id {
5063 0 : let source_path = &remote_initdb_archive_path(
5064 0 : &self.tenant_shard_id.tenant_id,
5065 0 : &existing_initdb_timeline_id,
5066 0 : );
5067 0 : let dest_path =
5068 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
5069 0 :
5070 0 : // if this fails, it will get retried by retried control plane requests
5071 0 : self.remote_storage
5072 0 : .copy_object(source_path, dest_path, &self.cancel)
5073 0 : .await
5074 0 : .context("copy initdb tar")?;
5075 4 : }
5076 4 : let (initdb_tar_zst_path, initdb_tar_zst) =
5077 4 : self::remote_timeline_client::download_initdb_tar_zst(
5078 4 : self.conf,
5079 4 : &self.remote_storage,
5080 4 : &self.tenant_shard_id,
5081 4 : &existing_initdb_timeline_id,
5082 4 : &self.cancel,
5083 4 : )
5084 4 : .await
5085 4 : .context("download initdb tar")?;
5086 :
5087 4 : scopeguard::defer! {
5088 4 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
5089 4 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
5090 4 : }
5091 4 : }
5092 4 :
5093 4 : let buf_read =
5094 4 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5095 4 : extract_zst_tarball(&pgdata_path, buf_read)
5096 4 : .await
5097 4 : .context("extract initdb tar")?;
5098 : } else {
5099 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5100 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5101 0 : .await
5102 0 : .context("run initdb")?;
5103 :
5104 : // Upload the created data dir to S3
5105 0 : if self.tenant_shard_id().is_shard_zero() {
5106 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5107 0 : .await?;
5108 0 : }
5109 : }
5110 4 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5111 4 :
5112 4 : // Import the contents of the data directory at the initial checkpoint
5113 4 : // LSN, and any WAL after that.
5114 4 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5115 4 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5116 4 : let new_metadata = TimelineMetadata::new(
5117 4 : Lsn(0),
5118 4 : None,
5119 4 : None,
5120 4 : Lsn(0),
5121 4 : pgdata_lsn,
5122 4 : pgdata_lsn,
5123 4 : pg_version,
5124 4 : );
5125 4 : let mut raw_timeline = self
5126 4 : .prepare_new_timeline(
5127 4 : timeline_id,
5128 4 : &new_metadata,
5129 4 : timeline_create_guard,
5130 4 : pgdata_lsn,
5131 4 : None,
5132 4 : )
5133 4 : .await?;
5134 :
5135 4 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5136 4 : raw_timeline
5137 4 : .write(|unfinished_timeline| async move {
5138 4 : import_datadir::import_timeline_from_postgres_datadir(
5139 4 : &unfinished_timeline,
5140 4 : &pgdata_path,
5141 4 : pgdata_lsn,
5142 4 : ctx,
5143 4 : )
5144 4 : .await
5145 4 : .with_context(|| {
5146 0 : format!(
5147 0 : "Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}"
5148 0 : )
5149 4 : })?;
5150 :
5151 4 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5152 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5153 0 : "failpoint before-checkpoint-new-timeline"
5154 0 : )))
5155 4 : });
5156 :
5157 4 : Ok(())
5158 8 : })
5159 4 : .await?;
5160 :
5161 : // All done!
5162 4 : let timeline = raw_timeline.finish_creation().await?;
5163 :
5164 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5165 :
5166 4 : Ok(CreateTimelineResult::Created(timeline))
5167 4 : }
5168 :
5169 884 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5170 884 : RemoteTimelineClient::new(
5171 884 : self.remote_storage.clone(),
5172 884 : self.deletion_queue_client.clone(),
5173 884 : self.conf,
5174 884 : self.tenant_shard_id,
5175 884 : timeline_id,
5176 884 : self.generation,
5177 884 : &self.tenant_conf.load().location,
5178 884 : )
5179 884 : }
5180 :
5181 : /// Builds required resources for a new timeline.
5182 884 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5183 884 : let remote_client = self.build_timeline_remote_client(timeline_id);
5184 884 : self.get_timeline_resources_for(remote_client)
5185 884 : }
5186 :
5187 : /// Builds timeline resources for the given remote client.
5188 896 : fn get_timeline_resources_for(&self, remote_client: RemoteTimelineClient) -> TimelineResources {
5189 896 : TimelineResources {
5190 896 : remote_client,
5191 896 : pagestream_throttle: self.pagestream_throttle.clone(),
5192 896 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5193 896 : l0_compaction_trigger: self.l0_compaction_trigger.clone(),
5194 896 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5195 896 : }
5196 896 : }
5197 :
5198 : /// Creates intermediate timeline structure and its files.
5199 : ///
5200 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5201 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5202 : /// `finish_creation` to insert the Timeline into the timelines map.
5203 884 : async fn prepare_new_timeline<'a>(
5204 884 : &'a self,
5205 884 : new_timeline_id: TimelineId,
5206 884 : new_metadata: &TimelineMetadata,
5207 884 : create_guard: TimelineCreateGuard,
5208 884 : start_lsn: Lsn,
5209 884 : ancestor: Option<Arc<Timeline>>,
5210 884 : ) -> anyhow::Result<UninitializedTimeline<'a>> {
5211 884 : let tenant_shard_id = self.tenant_shard_id;
5212 884 :
5213 884 : let resources = self.build_timeline_resources(new_timeline_id);
5214 884 : resources
5215 884 : .remote_client
5216 884 : .init_upload_queue_for_empty_remote(new_metadata)?;
5217 :
5218 884 : let timeline_struct = self
5219 884 : .create_timeline_struct(
5220 884 : new_timeline_id,
5221 884 : new_metadata,
5222 884 : None,
5223 884 : ancestor,
5224 884 : resources,
5225 884 : CreateTimelineCause::Load,
5226 884 : create_guard.idempotency.clone(),
5227 884 : None,
5228 884 : )
5229 884 : .context("Failed to create timeline data structure")?;
5230 :
5231 884 : timeline_struct.init_empty_layer_map(start_lsn);
5232 :
5233 884 : if let Err(e) = self
5234 884 : .create_timeline_files(&create_guard.timeline_path)
5235 884 : .await
5236 : {
5237 0 : error!(
5238 0 : "Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}"
5239 : );
5240 0 : cleanup_timeline_directory(create_guard);
5241 0 : return Err(e);
5242 884 : }
5243 884 :
5244 884 : debug!(
5245 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5246 : );
5247 :
5248 884 : Ok(UninitializedTimeline::new(
5249 884 : self,
5250 884 : new_timeline_id,
5251 884 : Some((timeline_struct, create_guard)),
5252 884 : ))
5253 884 : }
5254 :
5255 884 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5256 884 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5257 :
5258 884 : fail::fail_point!("after-timeline-dir-creation", |_| {
5259 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5260 884 : });
5261 :
5262 884 : Ok(())
5263 884 : }
5264 :
5265 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5266 : /// concurrent attempts to create the same timeline.
5267 : ///
5268 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5269 : /// offloaded timelines or not.
5270 896 : fn create_timeline_create_guard(
5271 896 : self: &Arc<Self>,
5272 896 : timeline_id: TimelineId,
5273 896 : idempotency: CreateTimelineIdempotency,
5274 896 : allow_offloaded: bool,
5275 896 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5276 896 : let tenant_shard_id = self.tenant_shard_id;
5277 896 :
5278 896 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5279 :
5280 896 : let create_guard = TimelineCreateGuard::new(
5281 896 : self,
5282 896 : timeline_id,
5283 896 : timeline_path.clone(),
5284 896 : idempotency,
5285 896 : allow_offloaded,
5286 896 : )?;
5287 :
5288 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5289 : // for creation.
5290 : // A timeline directory should never exist on disk already:
5291 : // - a previous failed creation would have cleaned up after itself
5292 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5293 : //
5294 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5295 : // this error may indicate a bug in cleanup on failed creations.
5296 892 : if timeline_path.exists() {
5297 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5298 0 : "Timeline directory already exists! This is a bug."
5299 0 : )));
5300 892 : }
5301 892 :
5302 892 : Ok(create_guard)
5303 896 : }
5304 :
5305 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5306 : ///
5307 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5308 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5309 : pub async fn gather_size_inputs(
5310 : &self,
5311 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5312 : // (only if it is shorter than the real cutoff).
5313 : max_retention_period: Option<u64>,
5314 : cause: LogicalSizeCalculationCause,
5315 : cancel: &CancellationToken,
5316 : ctx: &RequestContext,
5317 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5318 : let logical_sizes_at_once = self
5319 : .conf
5320 : .concurrent_tenant_size_logical_size_queries
5321 : .inner();
5322 :
5323 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5324 : //
5325 : // But the only case where we need to run multiple of these at once is when we
5326 : // request a size for a tenant manually via API, while another background calculation
5327 : // is in progress (which is not a common case).
5328 : //
5329 : // See more for on the issue #2748 condenced out of the initial PR review.
5330 : let mut shared_cache = tokio::select! {
5331 : locked = self.cached_logical_sizes.lock() => locked,
5332 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5333 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5334 : };
5335 :
5336 : size::gather_inputs(
5337 : self,
5338 : logical_sizes_at_once,
5339 : max_retention_period,
5340 : &mut shared_cache,
5341 : cause,
5342 : cancel,
5343 : ctx,
5344 : )
5345 : .await
5346 : }
5347 :
5348 : /// Calculate synthetic tenant size and cache the result.
5349 : /// This is periodically called by background worker.
5350 : /// result is cached in tenant struct
5351 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5352 : pub async fn calculate_synthetic_size(
5353 : &self,
5354 : cause: LogicalSizeCalculationCause,
5355 : cancel: &CancellationToken,
5356 : ctx: &RequestContext,
5357 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5358 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5359 :
5360 : let size = inputs.calculate();
5361 :
5362 : self.set_cached_synthetic_size(size);
5363 :
5364 : Ok(size)
5365 : }
5366 :
5367 : /// Cache given synthetic size and update the metric value
5368 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5369 0 : self.cached_synthetic_tenant_size
5370 0 : .store(size, Ordering::Relaxed);
5371 0 :
5372 0 : // Only shard zero should be calculating synthetic sizes
5373 0 : debug_assert!(self.shard_identity.is_shard_zero());
5374 :
5375 0 : TENANT_SYNTHETIC_SIZE_METRIC
5376 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5377 0 : .unwrap()
5378 0 : .set(size);
5379 0 : }
5380 :
5381 0 : pub fn cached_synthetic_size(&self) -> u64 {
5382 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5383 0 : }
5384 :
5385 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5386 : ///
5387 : /// This function can take a long time: callers should wrap it in a timeout if calling
5388 : /// from an external API handler.
5389 : ///
5390 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5391 : /// still bounded by tenant/timeline shutdown.
5392 : #[tracing::instrument(skip_all)]
5393 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5394 : let timelines = self.timelines.lock().unwrap().clone();
5395 :
5396 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5397 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5398 0 : timeline.freeze_and_flush().await?;
5399 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5400 0 : timeline.remote_client.wait_completion().await?;
5401 :
5402 0 : Ok(())
5403 0 : }
5404 :
5405 : // We do not use a JoinSet for these tasks, because we don't want them to be
5406 : // aborted when this function's future is cancelled: they should stay alive
5407 : // holding their GateGuard until they complete, to ensure their I/Os complete
5408 : // before Timeline shutdown completes.
5409 : let mut results = FuturesUnordered::new();
5410 :
5411 : for (_timeline_id, timeline) in timelines {
5412 : // Run each timeline's flush in a task holding the timeline's gate: this
5413 : // means that if this function's future is cancelled, the Timeline shutdown
5414 : // will still wait for any I/O in here to complete.
5415 : let Ok(gate) = timeline.gate.enter() else {
5416 : continue;
5417 : };
5418 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5419 : results.push(jh);
5420 : }
5421 :
5422 : while let Some(r) = results.next().await {
5423 : if let Err(e) = r {
5424 : if !e.is_cancelled() && !e.is_panic() {
5425 : tracing::error!("unexpected join error: {e:?}");
5426 : }
5427 : }
5428 : }
5429 :
5430 : // The flushes we did above were just writes, but the Tenant might have had
5431 : // pending deletions as well from recent compaction/gc: we want to flush those
5432 : // as well. This requires flushing the global delete queue. This is cheap
5433 : // because it's typically a no-op.
5434 : match self.deletion_queue_client.flush_execute().await {
5435 : Ok(_) => {}
5436 : Err(DeletionQueueError::ShuttingDown) => {}
5437 : }
5438 :
5439 : Ok(())
5440 : }
5441 :
5442 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
5443 0 : self.tenant_conf.load().tenant_conf.clone()
5444 0 : }
5445 :
5446 : /// How much local storage would this tenant like to have? It can cope with
5447 : /// less than this (via eviction and on-demand downloads), but this function enables
5448 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
5449 : /// by keeping important things on local disk.
5450 : ///
5451 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5452 : /// than they report here, due to layer eviction. Tenants with many active branches may
5453 : /// actually use more than they report here.
5454 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5455 0 : let timelines = self.timelines.lock().unwrap();
5456 0 :
5457 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5458 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5459 0 : // of them is used actively enough to occupy space on disk.
5460 0 : timelines
5461 0 : .values()
5462 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5463 0 : .max()
5464 0 : .unwrap_or(0)
5465 0 : }
5466 :
5467 : /// Serialize and write the latest TenantManifest to remote storage.
5468 4 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5469 : // Only one manifest write may be done at at time, and the contents of the manifest
5470 : // must be loaded while holding this lock. This makes it safe to call this function
5471 : // from anywhere without worrying about colliding updates.
5472 4 : let mut guard = tokio::select! {
5473 4 : g = self.tenant_manifest_upload.lock() => {
5474 4 : g
5475 : },
5476 4 : _ = self.cancel.cancelled() => {
5477 0 : return Err(TenantManifestError::Cancelled);
5478 : }
5479 : };
5480 :
5481 4 : let manifest = self.build_tenant_manifest();
5482 4 : if Some(&manifest) == (*guard).as_ref() {
5483 : // Optimisation: skip uploads that don't change anything.
5484 0 : return Ok(());
5485 4 : }
5486 4 :
5487 4 : // Remote storage does no retries internally, so wrap it
5488 4 : match backoff::retry(
5489 4 : || async {
5490 4 : upload_tenant_manifest(
5491 4 : &self.remote_storage,
5492 4 : &self.tenant_shard_id,
5493 4 : self.generation,
5494 4 : &manifest,
5495 4 : &self.cancel,
5496 4 : )
5497 4 : .await
5498 8 : },
5499 4 : |_e| self.cancel.is_cancelled(),
5500 4 : FAILED_UPLOAD_WARN_THRESHOLD,
5501 4 : FAILED_REMOTE_OP_RETRIES,
5502 4 : "uploading tenant manifest",
5503 4 : &self.cancel,
5504 4 : )
5505 4 : .await
5506 : {
5507 0 : None => Err(TenantManifestError::Cancelled),
5508 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5509 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5510 : Some(Ok(_)) => {
5511 : // Store the successfully uploaded manifest, so that future callers can avoid
5512 : // re-uploading the same thing.
5513 4 : *guard = Some(manifest);
5514 4 :
5515 4 : Ok(())
5516 : }
5517 : }
5518 4 : }
5519 : }
5520 :
5521 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5522 : /// to get bootstrap data for timeline initialization.
5523 0 : async fn run_initdb(
5524 0 : conf: &'static PageServerConf,
5525 0 : initdb_target_dir: &Utf8Path,
5526 0 : pg_version: u32,
5527 0 : cancel: &CancellationToken,
5528 0 : ) -> Result<(), InitdbError> {
5529 0 : let initdb_bin_path = conf
5530 0 : .pg_bin_dir(pg_version)
5531 0 : .map_err(InitdbError::Other)?
5532 0 : .join("initdb");
5533 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5534 0 : info!(
5535 0 : "running {} in {}, libdir: {}",
5536 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5537 : );
5538 :
5539 0 : let _permit = {
5540 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5541 0 : INIT_DB_SEMAPHORE.acquire().await
5542 : };
5543 :
5544 0 : CONCURRENT_INITDBS.inc();
5545 0 : scopeguard::defer! {
5546 0 : CONCURRENT_INITDBS.dec();
5547 0 : }
5548 0 :
5549 0 : let _timer = INITDB_RUN_TIME.start_timer();
5550 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5551 0 : superuser: &conf.superuser,
5552 0 : locale: &conf.locale,
5553 0 : initdb_bin: &initdb_bin_path,
5554 0 : pg_version,
5555 0 : library_search_path: &initdb_lib_dir,
5556 0 : pgdata: initdb_target_dir,
5557 0 : })
5558 0 : .await
5559 0 : .map_err(InitdbError::Inner);
5560 0 :
5561 0 : // This isn't true cancellation support, see above. Still return an error to
5562 0 : // excercise the cancellation code path.
5563 0 : if cancel.is_cancelled() {
5564 0 : return Err(InitdbError::Cancelled);
5565 0 : }
5566 0 :
5567 0 : res
5568 0 : }
5569 :
5570 : /// Dump contents of a layer file to stdout.
5571 0 : pub async fn dump_layerfile_from_path(
5572 0 : path: &Utf8Path,
5573 0 : verbose: bool,
5574 0 : ctx: &RequestContext,
5575 0 : ) -> anyhow::Result<()> {
5576 : use std::os::unix::fs::FileExt;
5577 :
5578 : // All layer files start with a two-byte "magic" value, to identify the kind of
5579 : // file.
5580 0 : let file = File::open(path)?;
5581 0 : let mut header_buf = [0u8; 2];
5582 0 : file.read_exact_at(&mut header_buf, 0)?;
5583 :
5584 0 : match u16::from_be_bytes(header_buf) {
5585 : crate::IMAGE_FILE_MAGIC => {
5586 0 : ImageLayer::new_for_path(path, file)?
5587 0 : .dump(verbose, ctx)
5588 0 : .await?
5589 : }
5590 : crate::DELTA_FILE_MAGIC => {
5591 0 : DeltaLayer::new_for_path(path, file)?
5592 0 : .dump(verbose, ctx)
5593 0 : .await?
5594 : }
5595 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5596 : }
5597 :
5598 0 : Ok(())
5599 0 : }
5600 :
5601 : #[cfg(test)]
5602 : pub(crate) mod harness {
5603 : use bytes::{Bytes, BytesMut};
5604 : use hex_literal::hex;
5605 : use once_cell::sync::OnceCell;
5606 : use pageserver_api::key::Key;
5607 : use pageserver_api::models::ShardParameters;
5608 : use pageserver_api::record::NeonWalRecord;
5609 : use pageserver_api::shard::ShardIndex;
5610 : use utils::id::TenantId;
5611 : use utils::logging;
5612 :
5613 : use super::*;
5614 : use crate::deletion_queue::mock::MockDeletionQueue;
5615 : use crate::l0_flush::L0FlushConfig;
5616 : use crate::walredo::apply_neon;
5617 :
5618 : pub const TIMELINE_ID: TimelineId =
5619 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5620 : pub const NEW_TIMELINE_ID: TimelineId =
5621 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5622 :
5623 : /// Convenience function to create a page image with given string as the only content
5624 10057572 : pub fn test_img(s: &str) -> Bytes {
5625 10057572 : let mut buf = BytesMut::new();
5626 10057572 : buf.extend_from_slice(s.as_bytes());
5627 10057572 : buf.resize(64, 0);
5628 10057572 :
5629 10057572 : buf.freeze()
5630 10057572 : }
5631 :
5632 : impl From<TenantConf> for TenantConfOpt {
5633 444 : fn from(tenant_conf: TenantConf) -> Self {
5634 444 : Self {
5635 444 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5636 444 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5637 444 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5638 444 : compaction_period: Some(tenant_conf.compaction_period),
5639 444 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5640 444 : compaction_upper_limit: Some(tenant_conf.compaction_upper_limit),
5641 444 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5642 444 : compaction_l0_first: Some(tenant_conf.compaction_l0_first),
5643 444 : compaction_l0_semaphore: Some(tenant_conf.compaction_l0_semaphore),
5644 444 : l0_flush_delay_threshold: tenant_conf.l0_flush_delay_threshold,
5645 444 : l0_flush_stall_threshold: tenant_conf.l0_flush_stall_threshold,
5646 444 : l0_flush_wait_upload: Some(tenant_conf.l0_flush_wait_upload),
5647 444 : gc_horizon: Some(tenant_conf.gc_horizon),
5648 444 : gc_period: Some(tenant_conf.gc_period),
5649 444 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5650 444 : pitr_interval: Some(tenant_conf.pitr_interval),
5651 444 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5652 444 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5653 444 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5654 444 : eviction_policy: Some(tenant_conf.eviction_policy),
5655 444 : min_resident_size_override: tenant_conf.min_resident_size_override,
5656 444 : evictions_low_residence_duration_metric_threshold: Some(
5657 444 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5658 444 : ),
5659 444 : heatmap_period: Some(tenant_conf.heatmap_period),
5660 444 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5661 444 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5662 444 : image_layer_creation_check_threshold: Some(
5663 444 : tenant_conf.image_layer_creation_check_threshold,
5664 444 : ),
5665 444 : image_creation_preempt_threshold: Some(
5666 444 : tenant_conf.image_creation_preempt_threshold,
5667 444 : ),
5668 444 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5669 444 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5670 444 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5671 444 : wal_receiver_protocol_override: tenant_conf.wal_receiver_protocol_override,
5672 444 : rel_size_v2_enabled: Some(tenant_conf.rel_size_v2_enabled),
5673 444 : gc_compaction_enabled: Some(tenant_conf.gc_compaction_enabled),
5674 444 : gc_compaction_initial_threshold_kb: Some(
5675 444 : tenant_conf.gc_compaction_initial_threshold_kb,
5676 444 : ),
5677 444 : gc_compaction_ratio_percent: Some(tenant_conf.gc_compaction_ratio_percent),
5678 444 : }
5679 444 : }
5680 : }
5681 :
5682 : pub struct TenantHarness {
5683 : pub conf: &'static PageServerConf,
5684 : pub tenant_conf: TenantConf,
5685 : pub tenant_shard_id: TenantShardId,
5686 : pub generation: Generation,
5687 : pub shard: ShardIndex,
5688 : pub remote_storage: GenericRemoteStorage,
5689 : pub remote_fs_dir: Utf8PathBuf,
5690 : pub deletion_queue: MockDeletionQueue,
5691 : }
5692 :
5693 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5694 :
5695 492 : pub(crate) fn setup_logging() {
5696 492 : LOG_HANDLE.get_or_init(|| {
5697 468 : logging::init(
5698 468 : logging::LogFormat::Test,
5699 468 : // enable it in case the tests exercise code paths that use
5700 468 : // debug_assert_current_span_has_tenant_and_timeline_id
5701 468 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5702 468 : logging::Output::Stdout,
5703 468 : )
5704 468 : .expect("Failed to init test logging")
5705 492 : });
5706 492 : }
5707 :
5708 : impl TenantHarness {
5709 444 : pub async fn create_custom(
5710 444 : test_name: &'static str,
5711 444 : tenant_conf: TenantConf,
5712 444 : tenant_id: TenantId,
5713 444 : shard_identity: ShardIdentity,
5714 444 : generation: Generation,
5715 444 : ) -> anyhow::Result<Self> {
5716 444 : setup_logging();
5717 444 :
5718 444 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5719 444 : let _ = fs::remove_dir_all(&repo_dir);
5720 444 : fs::create_dir_all(&repo_dir)?;
5721 :
5722 444 : let conf = PageServerConf::dummy_conf(repo_dir);
5723 444 : // Make a static copy of the config. This can never be free'd, but that's
5724 444 : // OK in a test.
5725 444 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5726 444 :
5727 444 : let shard = shard_identity.shard_index();
5728 444 : let tenant_shard_id = TenantShardId {
5729 444 : tenant_id,
5730 444 : shard_number: shard.shard_number,
5731 444 : shard_count: shard.shard_count,
5732 444 : };
5733 444 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5734 444 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5735 :
5736 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5737 444 : let remote_fs_dir = conf.workdir.join("localfs");
5738 444 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5739 444 : let config = RemoteStorageConfig {
5740 444 : storage: RemoteStorageKind::LocalFs {
5741 444 : local_path: remote_fs_dir.clone(),
5742 444 : },
5743 444 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5744 444 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5745 444 : };
5746 444 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5747 444 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5748 444 :
5749 444 : Ok(Self {
5750 444 : conf,
5751 444 : tenant_conf,
5752 444 : tenant_shard_id,
5753 444 : generation,
5754 444 : shard,
5755 444 : remote_storage,
5756 444 : remote_fs_dir,
5757 444 : deletion_queue,
5758 444 : })
5759 444 : }
5760 :
5761 420 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5762 420 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5763 420 : // The tests perform them manually if needed.
5764 420 : let tenant_conf = TenantConf {
5765 420 : gc_period: Duration::ZERO,
5766 420 : compaction_period: Duration::ZERO,
5767 420 : ..TenantConf::default()
5768 420 : };
5769 420 : let tenant_id = TenantId::generate();
5770 420 : let shard = ShardIdentity::unsharded();
5771 420 : Self::create_custom(
5772 420 : test_name,
5773 420 : tenant_conf,
5774 420 : tenant_id,
5775 420 : shard,
5776 420 : Generation::new(0xdeadbeef),
5777 420 : )
5778 420 : .await
5779 420 : }
5780 :
5781 40 : pub fn span(&self) -> tracing::Span {
5782 40 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5783 40 : }
5784 :
5785 444 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5786 444 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
5787 444 : (
5788 444 : self.do_try_load(&ctx)
5789 444 : .await
5790 444 : .expect("failed to load test tenant"),
5791 444 : ctx,
5792 444 : )
5793 444 : }
5794 :
5795 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5796 : pub(crate) async fn do_try_load(
5797 : &self,
5798 : ctx: &RequestContext,
5799 : ) -> anyhow::Result<Arc<Tenant>> {
5800 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5801 :
5802 : let tenant = Arc::new(Tenant::new(
5803 : TenantState::Attaching,
5804 : self.conf,
5805 : AttachedTenantConf::try_from(LocationConf::attached_single(
5806 : TenantConfOpt::from(self.tenant_conf.clone()),
5807 : self.generation,
5808 : &ShardParameters::default(),
5809 : ))
5810 : .unwrap(),
5811 : // This is a legacy/test code path: sharding isn't supported here.
5812 : ShardIdentity::unsharded(),
5813 : Some(walredo_mgr),
5814 : self.tenant_shard_id,
5815 : self.remote_storage.clone(),
5816 : self.deletion_queue.new_client(),
5817 : // TODO: ideally we should run all unit tests with both configs
5818 : L0FlushGlobalState::new(L0FlushConfig::default()),
5819 : ));
5820 :
5821 : let preload = tenant
5822 : .preload(&self.remote_storage, CancellationToken::new())
5823 : .await?;
5824 : tenant.attach(Some(preload), ctx).await?;
5825 :
5826 : tenant.state.send_replace(TenantState::Active);
5827 : for timeline in tenant.timelines.lock().unwrap().values() {
5828 : timeline.set_state(TimelineState::Active);
5829 : }
5830 : Ok(tenant)
5831 : }
5832 :
5833 4 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5834 4 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5835 4 : }
5836 : }
5837 :
5838 : // Mock WAL redo manager that doesn't do much
5839 : pub(crate) struct TestRedoManager;
5840 :
5841 : impl TestRedoManager {
5842 : /// # Cancel-Safety
5843 : ///
5844 : /// This method is cancellation-safe.
5845 1636 : pub async fn request_redo(
5846 1636 : &self,
5847 1636 : key: Key,
5848 1636 : lsn: Lsn,
5849 1636 : base_img: Option<(Lsn, Bytes)>,
5850 1636 : records: Vec<(Lsn, NeonWalRecord)>,
5851 1636 : _pg_version: u32,
5852 1636 : ) -> Result<Bytes, walredo::Error> {
5853 2392 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5854 1636 : if records_neon {
5855 : // For Neon wal records, we can decode without spawning postgres, so do so.
5856 1636 : let mut page = match (base_img, records.first()) {
5857 1504 : (Some((_lsn, img)), _) => {
5858 1504 : let mut page = BytesMut::new();
5859 1504 : page.extend_from_slice(&img);
5860 1504 : page
5861 : }
5862 132 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5863 : _ => {
5864 0 : panic!("Neon WAL redo requires base image or will init record");
5865 : }
5866 : };
5867 :
5868 4028 : for (record_lsn, record) in records {
5869 2392 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5870 : }
5871 1636 : Ok(page.freeze())
5872 : } else {
5873 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5874 0 : let s = format!(
5875 0 : "redo for {} to get to {}, with {} and {} records",
5876 0 : key,
5877 0 : lsn,
5878 0 : if base_img.is_some() {
5879 0 : "base image"
5880 : } else {
5881 0 : "no base image"
5882 : },
5883 0 : records.len()
5884 0 : );
5885 0 : println!("{s}");
5886 0 :
5887 0 : Ok(test_img(&s))
5888 : }
5889 1636 : }
5890 : }
5891 : }
5892 :
5893 : #[cfg(test)]
5894 : mod tests {
5895 : use std::collections::{BTreeMap, BTreeSet};
5896 :
5897 : use bytes::{Bytes, BytesMut};
5898 : use hex_literal::hex;
5899 : use itertools::Itertools;
5900 : #[cfg(feature = "testing")]
5901 : use models::CompactLsnRange;
5902 : use pageserver_api::key::{AUX_KEY_PREFIX, Key, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX};
5903 : use pageserver_api::keyspace::KeySpace;
5904 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5905 : #[cfg(feature = "testing")]
5906 : use pageserver_api::record::NeonWalRecord;
5907 : use pageserver_api::value::Value;
5908 : use pageserver_compaction::helpers::overlaps_with;
5909 : use rand::{Rng, thread_rng};
5910 : use storage_layer::{IoConcurrency, PersistentLayerKey};
5911 : use tests::storage_layer::ValuesReconstructState;
5912 : use tests::timeline::{GetVectoredError, ShutdownMode};
5913 : #[cfg(feature = "testing")]
5914 : use timeline::GcInfo;
5915 : #[cfg(feature = "testing")]
5916 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5917 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5918 : use utils::id::TenantId;
5919 :
5920 : use super::*;
5921 : use crate::DEFAULT_PG_VERSION;
5922 : use crate::keyspace::KeySpaceAccum;
5923 : use crate::tenant::harness::*;
5924 : use crate::tenant::timeline::CompactFlags;
5925 :
5926 : static TEST_KEY: Lazy<Key> =
5927 36 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5928 :
5929 : #[tokio::test]
5930 4 : async fn test_basic() -> anyhow::Result<()> {
5931 4 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5932 4 : let tline = tenant
5933 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5934 4 : .await?;
5935 4 :
5936 4 : let mut writer = tline.writer().await;
5937 4 : writer
5938 4 : .put(
5939 4 : *TEST_KEY,
5940 4 : Lsn(0x10),
5941 4 : &Value::Image(test_img("foo at 0x10")),
5942 4 : &ctx,
5943 4 : )
5944 4 : .await?;
5945 4 : writer.finish_write(Lsn(0x10));
5946 4 : drop(writer);
5947 4 :
5948 4 : let mut writer = tline.writer().await;
5949 4 : writer
5950 4 : .put(
5951 4 : *TEST_KEY,
5952 4 : Lsn(0x20),
5953 4 : &Value::Image(test_img("foo at 0x20")),
5954 4 : &ctx,
5955 4 : )
5956 4 : .await?;
5957 4 : writer.finish_write(Lsn(0x20));
5958 4 : drop(writer);
5959 4 :
5960 4 : assert_eq!(
5961 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
5962 4 : test_img("foo at 0x10")
5963 4 : );
5964 4 : assert_eq!(
5965 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
5966 4 : test_img("foo at 0x10")
5967 4 : );
5968 4 : assert_eq!(
5969 4 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
5970 4 : test_img("foo at 0x20")
5971 4 : );
5972 4 :
5973 4 : Ok(())
5974 4 : }
5975 :
5976 : #[tokio::test]
5977 4 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
5978 4 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
5979 4 : .await?
5980 4 : .load()
5981 4 : .await;
5982 4 : let _ = tenant
5983 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5984 4 : .await?;
5985 4 :
5986 4 : match tenant
5987 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5988 4 : .await
5989 4 : {
5990 4 : Ok(_) => panic!("duplicate timeline creation should fail"),
5991 4 : Err(e) => assert_eq!(
5992 4 : e.to_string(),
5993 4 : "timeline already exists with different parameters".to_string()
5994 4 : ),
5995 4 : }
5996 4 :
5997 4 : Ok(())
5998 4 : }
5999 :
6000 : /// Convenience function to create a page image with given string as the only content
6001 20 : pub fn test_value(s: &str) -> Value {
6002 20 : let mut buf = BytesMut::new();
6003 20 : buf.extend_from_slice(s.as_bytes());
6004 20 : Value::Image(buf.freeze())
6005 20 : }
6006 :
6007 : ///
6008 : /// Test branch creation
6009 : ///
6010 : #[tokio::test]
6011 4 : async fn test_branch() -> anyhow::Result<()> {
6012 4 : use std::str::from_utf8;
6013 4 :
6014 4 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
6015 4 : let tline = tenant
6016 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6017 4 : .await?;
6018 4 : let mut writer = tline.writer().await;
6019 4 :
6020 4 : #[allow(non_snake_case)]
6021 4 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
6022 4 : #[allow(non_snake_case)]
6023 4 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
6024 4 :
6025 4 : // Insert a value on the timeline
6026 4 : writer
6027 4 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
6028 4 : .await?;
6029 4 : writer
6030 4 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
6031 4 : .await?;
6032 4 : writer.finish_write(Lsn(0x20));
6033 4 :
6034 4 : writer
6035 4 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
6036 4 : .await?;
6037 4 : writer.finish_write(Lsn(0x30));
6038 4 : writer
6039 4 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
6040 4 : .await?;
6041 4 : writer.finish_write(Lsn(0x40));
6042 4 :
6043 4 : //assert_current_logical_size(&tline, Lsn(0x40));
6044 4 :
6045 4 : // Branch the history, modify relation differently on the new timeline
6046 4 : tenant
6047 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
6048 4 : .await?;
6049 4 : let newtline = tenant
6050 4 : .get_timeline(NEW_TIMELINE_ID, true)
6051 4 : .expect("Should have a local timeline");
6052 4 : let mut new_writer = newtline.writer().await;
6053 4 : new_writer
6054 4 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
6055 4 : .await?;
6056 4 : new_writer.finish_write(Lsn(0x40));
6057 4 :
6058 4 : // Check page contents on both branches
6059 4 : assert_eq!(
6060 4 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6061 4 : "foo at 0x40"
6062 4 : );
6063 4 : assert_eq!(
6064 4 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6065 4 : "bar at 0x40"
6066 4 : );
6067 4 : assert_eq!(
6068 4 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
6069 4 : "foobar at 0x20"
6070 4 : );
6071 4 :
6072 4 : //assert_current_logical_size(&tline, Lsn(0x40));
6073 4 :
6074 4 : Ok(())
6075 4 : }
6076 :
6077 40 : async fn make_some_layers(
6078 40 : tline: &Timeline,
6079 40 : start_lsn: Lsn,
6080 40 : ctx: &RequestContext,
6081 40 : ) -> anyhow::Result<()> {
6082 40 : let mut lsn = start_lsn;
6083 : {
6084 40 : let mut writer = tline.writer().await;
6085 : // Create a relation on the timeline
6086 40 : writer
6087 40 : .put(
6088 40 : *TEST_KEY,
6089 40 : lsn,
6090 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6091 40 : ctx,
6092 40 : )
6093 40 : .await?;
6094 40 : writer.finish_write(lsn);
6095 40 : lsn += 0x10;
6096 40 : writer
6097 40 : .put(
6098 40 : *TEST_KEY,
6099 40 : lsn,
6100 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6101 40 : ctx,
6102 40 : )
6103 40 : .await?;
6104 40 : writer.finish_write(lsn);
6105 40 : lsn += 0x10;
6106 40 : }
6107 40 : tline.freeze_and_flush().await?;
6108 : {
6109 40 : let mut writer = tline.writer().await;
6110 40 : writer
6111 40 : .put(
6112 40 : *TEST_KEY,
6113 40 : lsn,
6114 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6115 40 : ctx,
6116 40 : )
6117 40 : .await?;
6118 40 : writer.finish_write(lsn);
6119 40 : lsn += 0x10;
6120 40 : writer
6121 40 : .put(
6122 40 : *TEST_KEY,
6123 40 : lsn,
6124 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6125 40 : ctx,
6126 40 : )
6127 40 : .await?;
6128 40 : writer.finish_write(lsn);
6129 40 : }
6130 40 : tline.freeze_and_flush().await.map_err(|e| e.into())
6131 40 : }
6132 :
6133 : #[tokio::test(start_paused = true)]
6134 4 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6135 4 : let (tenant, ctx) =
6136 4 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6137 4 : .await?
6138 4 : .load()
6139 4 : .await;
6140 4 : // Advance to the lsn lease deadline so that GC is not blocked by
6141 4 : // initial transition into AttachedSingle.
6142 4 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
6143 4 : tokio::time::resume();
6144 4 : let tline = tenant
6145 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6146 4 : .await?;
6147 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6148 4 :
6149 4 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6150 4 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6151 4 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6152 4 : // below should fail.
6153 4 : tenant
6154 4 : .gc_iteration(
6155 4 : Some(TIMELINE_ID),
6156 4 : 0x10,
6157 4 : Duration::ZERO,
6158 4 : &CancellationToken::new(),
6159 4 : &ctx,
6160 4 : )
6161 4 : .await?;
6162 4 :
6163 4 : // try to branch at lsn 25, should fail because we already garbage collected the data
6164 4 : match tenant
6165 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6166 4 : .await
6167 4 : {
6168 4 : Ok(_) => panic!("branching should have failed"),
6169 4 : Err(err) => {
6170 4 : let CreateTimelineError::AncestorLsn(err) = err else {
6171 4 : panic!("wrong error type")
6172 4 : };
6173 4 : assert!(err.to_string().contains("invalid branch start lsn"));
6174 4 : assert!(
6175 4 : err.source()
6176 4 : .unwrap()
6177 4 : .to_string()
6178 4 : .contains("we might've already garbage collected needed data")
6179 4 : )
6180 4 : }
6181 4 : }
6182 4 :
6183 4 : Ok(())
6184 4 : }
6185 :
6186 : #[tokio::test]
6187 4 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6188 4 : let (tenant, ctx) =
6189 4 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6190 4 : .await?
6191 4 : .load()
6192 4 : .await;
6193 4 :
6194 4 : let tline = tenant
6195 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6196 4 : .await?;
6197 4 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6198 4 : match tenant
6199 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6200 4 : .await
6201 4 : {
6202 4 : Ok(_) => panic!("branching should have failed"),
6203 4 : Err(err) => {
6204 4 : let CreateTimelineError::AncestorLsn(err) = err else {
6205 4 : panic!("wrong error type");
6206 4 : };
6207 4 : assert!(&err.to_string().contains("invalid branch start lsn"));
6208 4 : assert!(
6209 4 : &err.source()
6210 4 : .unwrap()
6211 4 : .to_string()
6212 4 : .contains("is earlier than latest GC cutoff")
6213 4 : );
6214 4 : }
6215 4 : }
6216 4 :
6217 4 : Ok(())
6218 4 : }
6219 :
6220 : /*
6221 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6222 : // remove the old value, we'd need to work a little harder
6223 : #[tokio::test]
6224 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6225 : let repo =
6226 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6227 : .load();
6228 :
6229 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6230 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6231 :
6232 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6233 : let applied_gc_cutoff_lsn = tline.get_applied_gc_cutoff_lsn();
6234 : assert!(*applied_gc_cutoff_lsn > Lsn(0x25));
6235 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6236 : Ok(_) => panic!("request for page should have failed"),
6237 : Err(err) => assert!(err.to_string().contains("not found at")),
6238 : }
6239 : Ok(())
6240 : }
6241 : */
6242 :
6243 : #[tokio::test]
6244 4 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6245 4 : let (tenant, ctx) =
6246 4 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6247 4 : .await?
6248 4 : .load()
6249 4 : .await;
6250 4 : let tline = tenant
6251 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6252 4 : .await?;
6253 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6254 4 :
6255 4 : tenant
6256 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6257 4 : .await?;
6258 4 : let newtline = tenant
6259 4 : .get_timeline(NEW_TIMELINE_ID, true)
6260 4 : .expect("Should have a local timeline");
6261 4 :
6262 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6263 4 :
6264 4 : tline.set_broken("test".to_owned());
6265 4 :
6266 4 : tenant
6267 4 : .gc_iteration(
6268 4 : Some(TIMELINE_ID),
6269 4 : 0x10,
6270 4 : Duration::ZERO,
6271 4 : &CancellationToken::new(),
6272 4 : &ctx,
6273 4 : )
6274 4 : .await?;
6275 4 :
6276 4 : // The branchpoints should contain all timelines, even ones marked
6277 4 : // as Broken.
6278 4 : {
6279 4 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6280 4 : assert_eq!(branchpoints.len(), 1);
6281 4 : assert_eq!(
6282 4 : branchpoints[0],
6283 4 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6284 4 : );
6285 4 : }
6286 4 :
6287 4 : // You can read the key from the child branch even though the parent is
6288 4 : // Broken, as long as you don't need to access data from the parent.
6289 4 : assert_eq!(
6290 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6291 4 : test_img(&format!("foo at {}", Lsn(0x70)))
6292 4 : );
6293 4 :
6294 4 : // This needs to traverse to the parent, and fails.
6295 4 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6296 4 : assert!(
6297 4 : err.to_string().starts_with(&format!(
6298 4 : "bad state on timeline {}: Broken",
6299 4 : tline.timeline_id
6300 4 : )),
6301 4 : "{err}"
6302 4 : );
6303 4 :
6304 4 : Ok(())
6305 4 : }
6306 :
6307 : #[tokio::test]
6308 4 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6309 4 : let (tenant, ctx) =
6310 4 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6311 4 : .await?
6312 4 : .load()
6313 4 : .await;
6314 4 : let tline = tenant
6315 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6316 4 : .await?;
6317 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6318 4 :
6319 4 : tenant
6320 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6321 4 : .await?;
6322 4 : let newtline = tenant
6323 4 : .get_timeline(NEW_TIMELINE_ID, true)
6324 4 : .expect("Should have a local timeline");
6325 4 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6326 4 : tenant
6327 4 : .gc_iteration(
6328 4 : Some(TIMELINE_ID),
6329 4 : 0x10,
6330 4 : Duration::ZERO,
6331 4 : &CancellationToken::new(),
6332 4 : &ctx,
6333 4 : )
6334 4 : .await?;
6335 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6336 4 :
6337 4 : Ok(())
6338 4 : }
6339 : #[tokio::test]
6340 4 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6341 4 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6342 4 : .await?
6343 4 : .load()
6344 4 : .await;
6345 4 : let tline = tenant
6346 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6347 4 : .await?;
6348 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6349 4 :
6350 4 : tenant
6351 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6352 4 : .await?;
6353 4 : let newtline = tenant
6354 4 : .get_timeline(NEW_TIMELINE_ID, true)
6355 4 : .expect("Should have a local timeline");
6356 4 :
6357 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6358 4 :
6359 4 : // run gc on parent
6360 4 : tenant
6361 4 : .gc_iteration(
6362 4 : Some(TIMELINE_ID),
6363 4 : 0x10,
6364 4 : Duration::ZERO,
6365 4 : &CancellationToken::new(),
6366 4 : &ctx,
6367 4 : )
6368 4 : .await?;
6369 4 :
6370 4 : // Check that the data is still accessible on the branch.
6371 4 : assert_eq!(
6372 4 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6373 4 : test_img(&format!("foo at {}", Lsn(0x40)))
6374 4 : );
6375 4 :
6376 4 : Ok(())
6377 4 : }
6378 :
6379 : #[tokio::test]
6380 4 : async fn timeline_load() -> anyhow::Result<()> {
6381 4 : const TEST_NAME: &str = "timeline_load";
6382 4 : let harness = TenantHarness::create(TEST_NAME).await?;
6383 4 : {
6384 4 : let (tenant, ctx) = harness.load().await;
6385 4 : let tline = tenant
6386 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6387 4 : .await?;
6388 4 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6389 4 : // so that all uploads finish & we can call harness.load() below again
6390 4 : tenant
6391 4 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6392 4 : .instrument(harness.span())
6393 4 : .await
6394 4 : .ok()
6395 4 : .unwrap();
6396 4 : }
6397 4 :
6398 4 : let (tenant, _ctx) = harness.load().await;
6399 4 : tenant
6400 4 : .get_timeline(TIMELINE_ID, true)
6401 4 : .expect("cannot load timeline");
6402 4 :
6403 4 : Ok(())
6404 4 : }
6405 :
6406 : #[tokio::test]
6407 4 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6408 4 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6409 4 : let harness = TenantHarness::create(TEST_NAME).await?;
6410 4 : // create two timelines
6411 4 : {
6412 4 : let (tenant, ctx) = harness.load().await;
6413 4 : let tline = tenant
6414 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6415 4 : .await?;
6416 4 :
6417 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6418 4 :
6419 4 : let child_tline = tenant
6420 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6421 4 : .await?;
6422 4 : child_tline.set_state(TimelineState::Active);
6423 4 :
6424 4 : let newtline = tenant
6425 4 : .get_timeline(NEW_TIMELINE_ID, true)
6426 4 : .expect("Should have a local timeline");
6427 4 :
6428 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6429 4 :
6430 4 : // so that all uploads finish & we can call harness.load() below again
6431 4 : tenant
6432 4 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6433 4 : .instrument(harness.span())
6434 4 : .await
6435 4 : .ok()
6436 4 : .unwrap();
6437 4 : }
6438 4 :
6439 4 : // check that both of them are initially unloaded
6440 4 : let (tenant, _ctx) = harness.load().await;
6441 4 :
6442 4 : // check that both, child and ancestor are loaded
6443 4 : let _child_tline = tenant
6444 4 : .get_timeline(NEW_TIMELINE_ID, true)
6445 4 : .expect("cannot get child timeline loaded");
6446 4 :
6447 4 : let _ancestor_tline = tenant
6448 4 : .get_timeline(TIMELINE_ID, true)
6449 4 : .expect("cannot get ancestor timeline loaded");
6450 4 :
6451 4 : Ok(())
6452 4 : }
6453 :
6454 : #[tokio::test]
6455 4 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6456 4 : use storage_layer::AsLayerDesc;
6457 4 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6458 4 : .await?
6459 4 : .load()
6460 4 : .await;
6461 4 : let tline = tenant
6462 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6463 4 : .await?;
6464 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6465 4 :
6466 4 : let layer_map = tline.layers.read().await;
6467 4 : let level0_deltas = layer_map
6468 4 : .layer_map()?
6469 4 : .level0_deltas()
6470 4 : .iter()
6471 8 : .map(|desc| layer_map.get_from_desc(desc))
6472 4 : .collect::<Vec<_>>();
6473 4 :
6474 4 : assert!(!level0_deltas.is_empty());
6475 4 :
6476 12 : for delta in level0_deltas {
6477 4 : // Ensure we are dumping a delta layer here
6478 8 : assert!(delta.layer_desc().is_delta);
6479 8 : delta.dump(true, &ctx).await.unwrap();
6480 4 : }
6481 4 :
6482 4 : Ok(())
6483 4 : }
6484 :
6485 : #[tokio::test]
6486 4 : async fn test_images() -> anyhow::Result<()> {
6487 4 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6488 4 : let tline = tenant
6489 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6490 4 : .await?;
6491 4 :
6492 4 : let mut writer = tline.writer().await;
6493 4 : writer
6494 4 : .put(
6495 4 : *TEST_KEY,
6496 4 : Lsn(0x10),
6497 4 : &Value::Image(test_img("foo at 0x10")),
6498 4 : &ctx,
6499 4 : )
6500 4 : .await?;
6501 4 : writer.finish_write(Lsn(0x10));
6502 4 : drop(writer);
6503 4 :
6504 4 : tline.freeze_and_flush().await?;
6505 4 : tline
6506 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6507 4 : .await?;
6508 4 :
6509 4 : let mut writer = tline.writer().await;
6510 4 : writer
6511 4 : .put(
6512 4 : *TEST_KEY,
6513 4 : Lsn(0x20),
6514 4 : &Value::Image(test_img("foo at 0x20")),
6515 4 : &ctx,
6516 4 : )
6517 4 : .await?;
6518 4 : writer.finish_write(Lsn(0x20));
6519 4 : drop(writer);
6520 4 :
6521 4 : tline.freeze_and_flush().await?;
6522 4 : tline
6523 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6524 4 : .await?;
6525 4 :
6526 4 : let mut writer = tline.writer().await;
6527 4 : writer
6528 4 : .put(
6529 4 : *TEST_KEY,
6530 4 : Lsn(0x30),
6531 4 : &Value::Image(test_img("foo at 0x30")),
6532 4 : &ctx,
6533 4 : )
6534 4 : .await?;
6535 4 : writer.finish_write(Lsn(0x30));
6536 4 : drop(writer);
6537 4 :
6538 4 : tline.freeze_and_flush().await?;
6539 4 : tline
6540 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6541 4 : .await?;
6542 4 :
6543 4 : let mut writer = tline.writer().await;
6544 4 : writer
6545 4 : .put(
6546 4 : *TEST_KEY,
6547 4 : Lsn(0x40),
6548 4 : &Value::Image(test_img("foo at 0x40")),
6549 4 : &ctx,
6550 4 : )
6551 4 : .await?;
6552 4 : writer.finish_write(Lsn(0x40));
6553 4 : drop(writer);
6554 4 :
6555 4 : tline.freeze_and_flush().await?;
6556 4 : tline
6557 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6558 4 : .await?;
6559 4 :
6560 4 : assert_eq!(
6561 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6562 4 : test_img("foo at 0x10")
6563 4 : );
6564 4 : assert_eq!(
6565 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6566 4 : test_img("foo at 0x10")
6567 4 : );
6568 4 : assert_eq!(
6569 4 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6570 4 : test_img("foo at 0x20")
6571 4 : );
6572 4 : assert_eq!(
6573 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6574 4 : test_img("foo at 0x30")
6575 4 : );
6576 4 : assert_eq!(
6577 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6578 4 : test_img("foo at 0x40")
6579 4 : );
6580 4 :
6581 4 : Ok(())
6582 4 : }
6583 :
6584 8 : async fn bulk_insert_compact_gc(
6585 8 : tenant: &Tenant,
6586 8 : timeline: &Arc<Timeline>,
6587 8 : ctx: &RequestContext,
6588 8 : lsn: Lsn,
6589 8 : repeat: usize,
6590 8 : key_count: usize,
6591 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6592 8 : let compact = true;
6593 8 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6594 8 : }
6595 :
6596 16 : async fn bulk_insert_maybe_compact_gc(
6597 16 : tenant: &Tenant,
6598 16 : timeline: &Arc<Timeline>,
6599 16 : ctx: &RequestContext,
6600 16 : mut lsn: Lsn,
6601 16 : repeat: usize,
6602 16 : key_count: usize,
6603 16 : compact: bool,
6604 16 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6605 16 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6606 16 :
6607 16 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6608 16 : let mut blknum = 0;
6609 16 :
6610 16 : // Enforce that key range is monotonously increasing
6611 16 : let mut keyspace = KeySpaceAccum::new();
6612 16 :
6613 16 : let cancel = CancellationToken::new();
6614 16 :
6615 16 : for _ in 0..repeat {
6616 800 : for _ in 0..key_count {
6617 8000000 : test_key.field6 = blknum;
6618 8000000 : let mut writer = timeline.writer().await;
6619 8000000 : writer
6620 8000000 : .put(
6621 8000000 : test_key,
6622 8000000 : lsn,
6623 8000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6624 8000000 : ctx,
6625 8000000 : )
6626 8000000 : .await?;
6627 8000000 : inserted.entry(test_key).or_default().insert(lsn);
6628 8000000 : writer.finish_write(lsn);
6629 8000000 : drop(writer);
6630 8000000 :
6631 8000000 : keyspace.add_key(test_key);
6632 8000000 :
6633 8000000 : lsn = Lsn(lsn.0 + 0x10);
6634 8000000 : blknum += 1;
6635 : }
6636 :
6637 800 : timeline.freeze_and_flush().await?;
6638 800 : if compact {
6639 : // this requires timeline to be &Arc<Timeline>
6640 400 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
6641 400 : }
6642 :
6643 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6644 : // originally was.
6645 800 : let res = tenant
6646 800 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
6647 800 : .await?;
6648 :
6649 800 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6650 : }
6651 :
6652 16 : Ok(inserted)
6653 16 : }
6654 :
6655 : //
6656 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6657 : // Repeat 50 times.
6658 : //
6659 : #[tokio::test]
6660 4 : async fn test_bulk_insert() -> anyhow::Result<()> {
6661 4 : let harness = TenantHarness::create("test_bulk_insert").await?;
6662 4 : let (tenant, ctx) = harness.load().await;
6663 4 : let tline = tenant
6664 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6665 4 : .await?;
6666 4 :
6667 4 : let lsn = Lsn(0x10);
6668 4 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6669 4 :
6670 4 : Ok(())
6671 4 : }
6672 :
6673 : // Test the vectored get real implementation against a simple sequential implementation.
6674 : //
6675 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6676 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6677 : // grow to the right on the X axis.
6678 : // [Delta]
6679 : // [Delta]
6680 : // [Delta]
6681 : // [Delta]
6682 : // ------------ Image ---------------
6683 : //
6684 : // After layer generation we pick the ranges to query as follows:
6685 : // 1. The beginning of each delta layer
6686 : // 2. At the seam between two adjacent delta layers
6687 : //
6688 : // There's one major downside to this test: delta layers only contains images,
6689 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6690 : #[tokio::test]
6691 4 : async fn test_get_vectored() -> anyhow::Result<()> {
6692 4 : let harness = TenantHarness::create("test_get_vectored").await?;
6693 4 : let (tenant, ctx) = harness.load().await;
6694 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6695 4 : let tline = tenant
6696 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6697 4 : .await?;
6698 4 :
6699 4 : let lsn = Lsn(0x10);
6700 4 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6701 4 :
6702 4 : let guard = tline.layers.read().await;
6703 4 : let lm = guard.layer_map()?;
6704 4 :
6705 4 : lm.dump(true, &ctx).await?;
6706 4 :
6707 4 : let mut reads = Vec::new();
6708 4 : let mut prev = None;
6709 24 : lm.iter_historic_layers().for_each(|desc| {
6710 24 : if !desc.is_delta() {
6711 4 : prev = Some(desc.clone());
6712 4 : return;
6713 20 : }
6714 20 :
6715 20 : let start = desc.key_range.start;
6716 20 : let end = desc
6717 20 : .key_range
6718 20 : .start
6719 20 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6720 20 : reads.push(KeySpace {
6721 20 : ranges: vec![start..end],
6722 20 : });
6723 4 :
6724 20 : if let Some(prev) = &prev {
6725 20 : if !prev.is_delta() {
6726 20 : return;
6727 4 : }
6728 0 :
6729 0 : let first_range = Key {
6730 0 : field6: prev.key_range.end.field6 - 4,
6731 0 : ..prev.key_range.end
6732 0 : }..prev.key_range.end;
6733 0 :
6734 0 : let second_range = desc.key_range.start..Key {
6735 0 : field6: desc.key_range.start.field6 + 4,
6736 0 : ..desc.key_range.start
6737 0 : };
6738 0 :
6739 0 : reads.push(KeySpace {
6740 0 : ranges: vec![first_range, second_range],
6741 0 : });
6742 4 : };
6743 4 :
6744 4 : prev = Some(desc.clone());
6745 24 : });
6746 4 :
6747 4 : drop(guard);
6748 4 :
6749 4 : // Pick a big LSN such that we query over all the changes.
6750 4 : let reads_lsn = Lsn(u64::MAX - 1);
6751 4 :
6752 24 : for read in reads {
6753 20 : info!("Doing vectored read on {:?}", read);
6754 4 :
6755 20 : let vectored_res = tline
6756 20 : .get_vectored_impl(
6757 20 : read.clone(),
6758 20 : reads_lsn,
6759 20 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6760 20 : &ctx,
6761 20 : )
6762 20 : .await;
6763 4 :
6764 20 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6765 20 : let mut expect_missing = false;
6766 20 : let mut key = read.start().unwrap();
6767 660 : while key != read.end().unwrap() {
6768 640 : if let Some(lsns) = inserted.get(&key) {
6769 640 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6770 640 : match expected_lsn {
6771 640 : Some(lsn) => {
6772 640 : expected_lsns.insert(key, *lsn);
6773 640 : }
6774 4 : None => {
6775 4 : expect_missing = true;
6776 0 : break;
6777 4 : }
6778 4 : }
6779 4 : } else {
6780 4 : expect_missing = true;
6781 0 : break;
6782 4 : }
6783 4 :
6784 640 : key = key.next();
6785 4 : }
6786 4 :
6787 20 : if expect_missing {
6788 4 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6789 4 : } else {
6790 640 : for (key, image) in vectored_res? {
6791 640 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6792 640 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6793 640 : assert_eq!(image?, expected_image);
6794 4 : }
6795 4 : }
6796 4 : }
6797 4 :
6798 4 : Ok(())
6799 4 : }
6800 :
6801 : #[tokio::test]
6802 4 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6803 4 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6804 4 :
6805 4 : let (tenant, ctx) = harness.load().await;
6806 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6807 4 : let tline = tenant
6808 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6809 4 : .await?;
6810 4 : let tline = tline.raw_timeline().unwrap();
6811 4 :
6812 4 : let mut modification = tline.begin_modification(Lsn(0x1000));
6813 4 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6814 4 : modification.set_lsn(Lsn(0x1008))?;
6815 4 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6816 4 : modification.commit(&ctx).await?;
6817 4 :
6818 4 : let child_timeline_id = TimelineId::generate();
6819 4 : tenant
6820 4 : .branch_timeline_test(
6821 4 : tline,
6822 4 : child_timeline_id,
6823 4 : Some(tline.get_last_record_lsn()),
6824 4 : &ctx,
6825 4 : )
6826 4 : .await?;
6827 4 :
6828 4 : let child_timeline = tenant
6829 4 : .get_timeline(child_timeline_id, true)
6830 4 : .expect("Should have the branched timeline");
6831 4 :
6832 4 : let aux_keyspace = KeySpace {
6833 4 : ranges: vec![NON_INHERITED_RANGE],
6834 4 : };
6835 4 : let read_lsn = child_timeline.get_last_record_lsn();
6836 4 :
6837 4 : let vectored_res = child_timeline
6838 4 : .get_vectored_impl(
6839 4 : aux_keyspace.clone(),
6840 4 : read_lsn,
6841 4 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6842 4 : &ctx,
6843 4 : )
6844 4 : .await;
6845 4 :
6846 4 : let images = vectored_res?;
6847 4 : assert!(images.is_empty());
6848 4 : Ok(())
6849 4 : }
6850 :
6851 : // Test that vectored get handles layer gaps correctly
6852 : // by advancing into the next ancestor timeline if required.
6853 : //
6854 : // The test generates timelines that look like the diagram below.
6855 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6856 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6857 : //
6858 : // ```
6859 : //-------------------------------+
6860 : // ... |
6861 : // [ L1 ] |
6862 : // [ / L1 ] | Child Timeline
6863 : // ... |
6864 : // ------------------------------+
6865 : // [ X L1 ] | Parent Timeline
6866 : // ------------------------------+
6867 : // ```
6868 : #[tokio::test]
6869 4 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6870 4 : let tenant_conf = TenantConf {
6871 4 : // Make compaction deterministic
6872 4 : gc_period: Duration::ZERO,
6873 4 : compaction_period: Duration::ZERO,
6874 4 : // Encourage creation of L1 layers
6875 4 : checkpoint_distance: 16 * 1024,
6876 4 : compaction_target_size: 8 * 1024,
6877 4 : ..TenantConf::default()
6878 4 : };
6879 4 :
6880 4 : let harness = TenantHarness::create_custom(
6881 4 : "test_get_vectored_key_gap",
6882 4 : tenant_conf,
6883 4 : TenantId::generate(),
6884 4 : ShardIdentity::unsharded(),
6885 4 : Generation::new(0xdeadbeef),
6886 4 : )
6887 4 : .await?;
6888 4 : let (tenant, ctx) = harness.load().await;
6889 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6890 4 :
6891 4 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6892 4 : let gap_at_key = current_key.add(100);
6893 4 : let mut current_lsn = Lsn(0x10);
6894 4 :
6895 4 : const KEY_COUNT: usize = 10_000;
6896 4 :
6897 4 : let timeline_id = TimelineId::generate();
6898 4 : let current_timeline = tenant
6899 4 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6900 4 : .await?;
6901 4 :
6902 4 : current_lsn += 0x100;
6903 4 :
6904 4 : let mut writer = current_timeline.writer().await;
6905 4 : writer
6906 4 : .put(
6907 4 : gap_at_key,
6908 4 : current_lsn,
6909 4 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6910 4 : &ctx,
6911 4 : )
6912 4 : .await?;
6913 4 : writer.finish_write(current_lsn);
6914 4 : drop(writer);
6915 4 :
6916 4 : let mut latest_lsns = HashMap::new();
6917 4 : latest_lsns.insert(gap_at_key, current_lsn);
6918 4 :
6919 4 : current_timeline.freeze_and_flush().await?;
6920 4 :
6921 4 : let child_timeline_id = TimelineId::generate();
6922 4 :
6923 4 : tenant
6924 4 : .branch_timeline_test(
6925 4 : ¤t_timeline,
6926 4 : child_timeline_id,
6927 4 : Some(current_lsn),
6928 4 : &ctx,
6929 4 : )
6930 4 : .await?;
6931 4 : let child_timeline = tenant
6932 4 : .get_timeline(child_timeline_id, true)
6933 4 : .expect("Should have the branched timeline");
6934 4 :
6935 40004 : for i in 0..KEY_COUNT {
6936 40000 : if current_key == gap_at_key {
6937 4 : current_key = current_key.next();
6938 4 : continue;
6939 39996 : }
6940 39996 :
6941 39996 : current_lsn += 0x10;
6942 4 :
6943 39996 : let mut writer = child_timeline.writer().await;
6944 39996 : writer
6945 39996 : .put(
6946 39996 : current_key,
6947 39996 : current_lsn,
6948 39996 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
6949 39996 : &ctx,
6950 39996 : )
6951 39996 : .await?;
6952 39996 : writer.finish_write(current_lsn);
6953 39996 : drop(writer);
6954 39996 :
6955 39996 : latest_lsns.insert(current_key, current_lsn);
6956 39996 : current_key = current_key.next();
6957 39996 :
6958 39996 : // Flush every now and then to encourage layer file creation.
6959 39996 : if i % 500 == 0 {
6960 80 : child_timeline.freeze_and_flush().await?;
6961 39916 : }
6962 4 : }
6963 4 :
6964 4 : child_timeline.freeze_and_flush().await?;
6965 4 : let mut flags = EnumSet::new();
6966 4 : flags.insert(CompactFlags::ForceRepartition);
6967 4 : child_timeline
6968 4 : .compact(&CancellationToken::new(), flags, &ctx)
6969 4 : .await?;
6970 4 :
6971 4 : let key_near_end = {
6972 4 : let mut tmp = current_key;
6973 4 : tmp.field6 -= 10;
6974 4 : tmp
6975 4 : };
6976 4 :
6977 4 : let key_near_gap = {
6978 4 : let mut tmp = gap_at_key;
6979 4 : tmp.field6 -= 10;
6980 4 : tmp
6981 4 : };
6982 4 :
6983 4 : let read = KeySpace {
6984 4 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
6985 4 : };
6986 4 : let results = child_timeline
6987 4 : .get_vectored_impl(
6988 4 : read.clone(),
6989 4 : current_lsn,
6990 4 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6991 4 : &ctx,
6992 4 : )
6993 4 : .await?;
6994 4 :
6995 88 : for (key, img_res) in results {
6996 84 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
6997 84 : assert_eq!(img_res?, expected);
6998 4 : }
6999 4 :
7000 4 : Ok(())
7001 4 : }
7002 :
7003 : // Test that vectored get descends into ancestor timelines correctly and
7004 : // does not return an image that's newer than requested.
7005 : //
7006 : // The diagram below ilustrates an interesting case. We have a parent timeline
7007 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
7008 : // from the child timeline, so the parent timeline must be visited. When advacing into
7009 : // the child timeline, the read path needs to remember what the requested Lsn was in
7010 : // order to avoid returning an image that's too new. The test below constructs such
7011 : // a timeline setup and does a few queries around the Lsn of each page image.
7012 : // ```
7013 : // LSN
7014 : // ^
7015 : // |
7016 : // |
7017 : // 500 | --------------------------------------> branch point
7018 : // 400 | X
7019 : // 300 | X
7020 : // 200 | --------------------------------------> requested lsn
7021 : // 100 | X
7022 : // |---------------------------------------> Key
7023 : // |
7024 : // ------> requested key
7025 : //
7026 : // Legend:
7027 : // * X - page images
7028 : // ```
7029 : #[tokio::test]
7030 4 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
7031 4 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
7032 4 : let (tenant, ctx) = harness.load().await;
7033 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7034 4 :
7035 4 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7036 4 : let end_key = start_key.add(1000);
7037 4 : let child_gap_at_key = start_key.add(500);
7038 4 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
7039 4 :
7040 4 : let mut current_lsn = Lsn(0x10);
7041 4 :
7042 4 : let timeline_id = TimelineId::generate();
7043 4 : let parent_timeline = tenant
7044 4 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7045 4 : .await?;
7046 4 :
7047 4 : current_lsn += 0x100;
7048 4 :
7049 16 : for _ in 0..3 {
7050 12 : let mut key = start_key;
7051 12012 : while key < end_key {
7052 12000 : current_lsn += 0x10;
7053 12000 :
7054 12000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
7055 4 :
7056 12000 : let mut writer = parent_timeline.writer().await;
7057 12000 : writer
7058 12000 : .put(
7059 12000 : key,
7060 12000 : current_lsn,
7061 12000 : &Value::Image(test_img(&image_value)),
7062 12000 : &ctx,
7063 12000 : )
7064 12000 : .await?;
7065 12000 : writer.finish_write(current_lsn);
7066 12000 :
7067 12000 : if key == child_gap_at_key {
7068 12 : parent_gap_lsns.insert(current_lsn, image_value);
7069 11988 : }
7070 4 :
7071 12000 : key = key.next();
7072 4 : }
7073 4 :
7074 12 : parent_timeline.freeze_and_flush().await?;
7075 4 : }
7076 4 :
7077 4 : let child_timeline_id = TimelineId::generate();
7078 4 :
7079 4 : let child_timeline = tenant
7080 4 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
7081 4 : .await?;
7082 4 :
7083 4 : let mut key = start_key;
7084 4004 : while key < end_key {
7085 4000 : if key == child_gap_at_key {
7086 4 : key = key.next();
7087 4 : continue;
7088 3996 : }
7089 3996 :
7090 3996 : current_lsn += 0x10;
7091 4 :
7092 3996 : let mut writer = child_timeline.writer().await;
7093 3996 : writer
7094 3996 : .put(
7095 3996 : key,
7096 3996 : current_lsn,
7097 3996 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
7098 3996 : &ctx,
7099 3996 : )
7100 3996 : .await?;
7101 3996 : writer.finish_write(current_lsn);
7102 3996 :
7103 3996 : key = key.next();
7104 4 : }
7105 4 :
7106 4 : child_timeline.freeze_and_flush().await?;
7107 4 :
7108 4 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
7109 4 : let mut query_lsns = Vec::new();
7110 12 : for image_lsn in parent_gap_lsns.keys().rev() {
7111 72 : for offset in lsn_offsets {
7112 60 : query_lsns.push(Lsn(image_lsn
7113 60 : .0
7114 60 : .checked_add_signed(offset)
7115 60 : .expect("Shouldn't overflow")));
7116 60 : }
7117 4 : }
7118 4 :
7119 64 : for query_lsn in query_lsns {
7120 60 : let results = child_timeline
7121 60 : .get_vectored_impl(
7122 60 : KeySpace {
7123 60 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
7124 60 : },
7125 60 : query_lsn,
7126 60 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7127 60 : &ctx,
7128 60 : )
7129 60 : .await;
7130 4 :
7131 60 : let expected_item = parent_gap_lsns
7132 60 : .iter()
7133 60 : .rev()
7134 136 : .find(|(lsn, _)| **lsn <= query_lsn);
7135 60 :
7136 60 : info!(
7137 4 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7138 4 : query_lsn, expected_item
7139 4 : );
7140 4 :
7141 60 : match expected_item {
7142 52 : Some((_, img_value)) => {
7143 52 : let key_results = results.expect("No vectored get error expected");
7144 52 : let key_result = &key_results[&child_gap_at_key];
7145 52 : let returned_img = key_result
7146 52 : .as_ref()
7147 52 : .expect("No page reconstruct error expected");
7148 52 :
7149 52 : info!(
7150 4 : "Vectored read at LSN {} returned image {}",
7151 0 : query_lsn,
7152 0 : std::str::from_utf8(returned_img)?
7153 4 : );
7154 52 : assert_eq!(*returned_img, test_img(img_value));
7155 4 : }
7156 4 : None => {
7157 8 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7158 4 : }
7159 4 : }
7160 4 : }
7161 4 :
7162 4 : Ok(())
7163 4 : }
7164 :
7165 : #[tokio::test]
7166 4 : async fn test_random_updates() -> anyhow::Result<()> {
7167 4 : let names_algorithms = [
7168 4 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7169 4 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7170 4 : ];
7171 12 : for (name, algorithm) in names_algorithms {
7172 8 : test_random_updates_algorithm(name, algorithm).await?;
7173 4 : }
7174 4 : Ok(())
7175 4 : }
7176 :
7177 8 : async fn test_random_updates_algorithm(
7178 8 : name: &'static str,
7179 8 : compaction_algorithm: CompactionAlgorithm,
7180 8 : ) -> anyhow::Result<()> {
7181 8 : let mut harness = TenantHarness::create(name).await?;
7182 8 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7183 8 : kind: compaction_algorithm,
7184 8 : };
7185 8 : let (tenant, ctx) = harness.load().await;
7186 8 : let tline = tenant
7187 8 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7188 8 : .await?;
7189 :
7190 : const NUM_KEYS: usize = 1000;
7191 8 : let cancel = CancellationToken::new();
7192 8 :
7193 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7194 8 : let mut test_key_end = test_key;
7195 8 : test_key_end.field6 = NUM_KEYS as u32;
7196 8 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7197 8 :
7198 8 : let mut keyspace = KeySpaceAccum::new();
7199 8 :
7200 8 : // Track when each page was last modified. Used to assert that
7201 8 : // a read sees the latest page version.
7202 8 : let mut updated = [Lsn(0); NUM_KEYS];
7203 8 :
7204 8 : let mut lsn = Lsn(0x10);
7205 : #[allow(clippy::needless_range_loop)]
7206 8008 : for blknum in 0..NUM_KEYS {
7207 8000 : lsn = Lsn(lsn.0 + 0x10);
7208 8000 : test_key.field6 = blknum as u32;
7209 8000 : let mut writer = tline.writer().await;
7210 8000 : writer
7211 8000 : .put(
7212 8000 : test_key,
7213 8000 : lsn,
7214 8000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7215 8000 : &ctx,
7216 8000 : )
7217 8000 : .await?;
7218 8000 : writer.finish_write(lsn);
7219 8000 : updated[blknum] = lsn;
7220 8000 : drop(writer);
7221 8000 :
7222 8000 : keyspace.add_key(test_key);
7223 : }
7224 :
7225 408 : for _ in 0..50 {
7226 400400 : for _ in 0..NUM_KEYS {
7227 400000 : lsn = Lsn(lsn.0 + 0x10);
7228 400000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7229 400000 : test_key.field6 = blknum as u32;
7230 400000 : let mut writer = tline.writer().await;
7231 400000 : writer
7232 400000 : .put(
7233 400000 : test_key,
7234 400000 : lsn,
7235 400000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7236 400000 : &ctx,
7237 400000 : )
7238 400000 : .await?;
7239 400000 : writer.finish_write(lsn);
7240 400000 : drop(writer);
7241 400000 : updated[blknum] = lsn;
7242 : }
7243 :
7244 : // Read all the blocks
7245 400000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7246 400000 : test_key.field6 = blknum as u32;
7247 400000 : assert_eq!(
7248 400000 : tline.get(test_key, lsn, &ctx).await?,
7249 400000 : test_img(&format!("{} at {}", blknum, last_lsn))
7250 : );
7251 : }
7252 :
7253 : // Perform a cycle of flush, and GC
7254 400 : tline.freeze_and_flush().await?;
7255 400 : tenant
7256 400 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7257 400 : .await?;
7258 : }
7259 :
7260 8 : Ok(())
7261 8 : }
7262 :
7263 : #[tokio::test]
7264 4 : async fn test_traverse_branches() -> anyhow::Result<()> {
7265 4 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7266 4 : .await?
7267 4 : .load()
7268 4 : .await;
7269 4 : let mut tline = tenant
7270 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7271 4 : .await?;
7272 4 :
7273 4 : const NUM_KEYS: usize = 1000;
7274 4 :
7275 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7276 4 :
7277 4 : let mut keyspace = KeySpaceAccum::new();
7278 4 :
7279 4 : let cancel = CancellationToken::new();
7280 4 :
7281 4 : // Track when each page was last modified. Used to assert that
7282 4 : // a read sees the latest page version.
7283 4 : let mut updated = [Lsn(0); NUM_KEYS];
7284 4 :
7285 4 : let mut lsn = Lsn(0x10);
7286 4 : #[allow(clippy::needless_range_loop)]
7287 4004 : for blknum in 0..NUM_KEYS {
7288 4000 : lsn = Lsn(lsn.0 + 0x10);
7289 4000 : test_key.field6 = blknum as u32;
7290 4000 : let mut writer = tline.writer().await;
7291 4000 : writer
7292 4000 : .put(
7293 4000 : test_key,
7294 4000 : lsn,
7295 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7296 4000 : &ctx,
7297 4000 : )
7298 4000 : .await?;
7299 4000 : writer.finish_write(lsn);
7300 4000 : updated[blknum] = lsn;
7301 4000 : drop(writer);
7302 4000 :
7303 4000 : keyspace.add_key(test_key);
7304 4 : }
7305 4 :
7306 204 : for _ in 0..50 {
7307 200 : let new_tline_id = TimelineId::generate();
7308 200 : tenant
7309 200 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7310 200 : .await?;
7311 200 : tline = tenant
7312 200 : .get_timeline(new_tline_id, true)
7313 200 : .expect("Should have the branched timeline");
7314 4 :
7315 200200 : for _ in 0..NUM_KEYS {
7316 200000 : lsn = Lsn(lsn.0 + 0x10);
7317 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7318 200000 : test_key.field6 = blknum as u32;
7319 200000 : let mut writer = tline.writer().await;
7320 200000 : writer
7321 200000 : .put(
7322 200000 : test_key,
7323 200000 : lsn,
7324 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7325 200000 : &ctx,
7326 200000 : )
7327 200000 : .await?;
7328 200000 : println!("updating {} at {}", blknum, lsn);
7329 200000 : writer.finish_write(lsn);
7330 200000 : drop(writer);
7331 200000 : updated[blknum] = lsn;
7332 4 : }
7333 4 :
7334 4 : // Read all the blocks
7335 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7336 200000 : test_key.field6 = blknum as u32;
7337 200000 : assert_eq!(
7338 200000 : tline.get(test_key, lsn, &ctx).await?,
7339 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
7340 4 : );
7341 4 : }
7342 4 :
7343 4 : // Perform a cycle of flush, compact, and GC
7344 200 : tline.freeze_and_flush().await?;
7345 200 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7346 200 : tenant
7347 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7348 200 : .await?;
7349 4 : }
7350 4 :
7351 4 : Ok(())
7352 4 : }
7353 :
7354 : #[tokio::test]
7355 4 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7356 4 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7357 4 : .await?
7358 4 : .load()
7359 4 : .await;
7360 4 : let mut tline = tenant
7361 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7362 4 : .await?;
7363 4 :
7364 4 : const NUM_KEYS: usize = 100;
7365 4 : const NUM_TLINES: usize = 50;
7366 4 :
7367 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7368 4 : // Track page mutation lsns across different timelines.
7369 4 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7370 4 :
7371 4 : let mut lsn = Lsn(0x10);
7372 4 :
7373 4 : #[allow(clippy::needless_range_loop)]
7374 204 : for idx in 0..NUM_TLINES {
7375 200 : let new_tline_id = TimelineId::generate();
7376 200 : tenant
7377 200 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7378 200 : .await?;
7379 200 : tline = tenant
7380 200 : .get_timeline(new_tline_id, true)
7381 200 : .expect("Should have the branched timeline");
7382 4 :
7383 20200 : for _ in 0..NUM_KEYS {
7384 20000 : lsn = Lsn(lsn.0 + 0x10);
7385 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7386 20000 : test_key.field6 = blknum as u32;
7387 20000 : let mut writer = tline.writer().await;
7388 20000 : writer
7389 20000 : .put(
7390 20000 : test_key,
7391 20000 : lsn,
7392 20000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7393 20000 : &ctx,
7394 20000 : )
7395 20000 : .await?;
7396 20000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7397 20000 : writer.finish_write(lsn);
7398 20000 : drop(writer);
7399 20000 : updated[idx][blknum] = lsn;
7400 4 : }
7401 4 : }
7402 4 :
7403 4 : // Read pages from leaf timeline across all ancestors.
7404 200 : for (idx, lsns) in updated.iter().enumerate() {
7405 20000 : for (blknum, lsn) in lsns.iter().enumerate() {
7406 4 : // Skip empty mutations.
7407 20000 : if lsn.0 == 0 {
7408 7316 : continue;
7409 12684 : }
7410 12684 : println!("checking [{idx}][{blknum}] at {lsn}");
7411 12684 : test_key.field6 = blknum as u32;
7412 12684 : assert_eq!(
7413 12684 : tline.get(test_key, *lsn, &ctx).await?,
7414 12684 : test_img(&format!("{idx} {blknum} at {lsn}"))
7415 4 : );
7416 4 : }
7417 4 : }
7418 4 : Ok(())
7419 4 : }
7420 :
7421 : #[tokio::test]
7422 4 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7423 4 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7424 4 : .await?
7425 4 : .load()
7426 4 : .await;
7427 4 :
7428 4 : let initdb_lsn = Lsn(0x20);
7429 4 : let utline = tenant
7430 4 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7431 4 : .await?;
7432 4 : let tline = utline.raw_timeline().unwrap();
7433 4 :
7434 4 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7435 4 : tline.maybe_spawn_flush_loop();
7436 4 :
7437 4 : // Make sure the timeline has the minimum set of required keys for operation.
7438 4 : // The only operation you can always do on an empty timeline is to `put` new data.
7439 4 : // Except if you `put` at `initdb_lsn`.
7440 4 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7441 4 : // It uses `repartition()`, which assumes some keys to be present.
7442 4 : // Let's make sure the test timeline can handle that case.
7443 4 : {
7444 4 : let mut state = tline.flush_loop_state.lock().unwrap();
7445 4 : assert_eq!(
7446 4 : timeline::FlushLoopState::Running {
7447 4 : expect_initdb_optimization: false,
7448 4 : initdb_optimization_count: 0,
7449 4 : },
7450 4 : *state
7451 4 : );
7452 4 : *state = timeline::FlushLoopState::Running {
7453 4 : expect_initdb_optimization: true,
7454 4 : initdb_optimization_count: 0,
7455 4 : };
7456 4 : }
7457 4 :
7458 4 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7459 4 : // As explained above, the optimization requires some keys to be present.
7460 4 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7461 4 : // This is what `create_test_timeline` does, by the way.
7462 4 : let mut modification = tline.begin_modification(initdb_lsn);
7463 4 : modification
7464 4 : .init_empty_test_timeline()
7465 4 : .context("init_empty_test_timeline")?;
7466 4 : modification
7467 4 : .commit(&ctx)
7468 4 : .await
7469 4 : .context("commit init_empty_test_timeline modification")?;
7470 4 :
7471 4 : // Do the flush. The flush code will check the expectations that we set above.
7472 4 : tline.freeze_and_flush().await?;
7473 4 :
7474 4 : // assert freeze_and_flush exercised the initdb optimization
7475 4 : {
7476 4 : let state = tline.flush_loop_state.lock().unwrap();
7477 4 : let timeline::FlushLoopState::Running {
7478 4 : expect_initdb_optimization,
7479 4 : initdb_optimization_count,
7480 4 : } = *state
7481 4 : else {
7482 4 : panic!("unexpected state: {:?}", *state);
7483 4 : };
7484 4 : assert!(expect_initdb_optimization);
7485 4 : assert!(initdb_optimization_count > 0);
7486 4 : }
7487 4 : Ok(())
7488 4 : }
7489 :
7490 : #[tokio::test]
7491 4 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7492 4 : let name = "test_create_guard_crash";
7493 4 : let harness = TenantHarness::create(name).await?;
7494 4 : {
7495 4 : let (tenant, ctx) = harness.load().await;
7496 4 : let tline = tenant
7497 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7498 4 : .await?;
7499 4 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
7500 4 : let raw_tline = tline.raw_timeline().unwrap();
7501 4 : raw_tline
7502 4 : .shutdown(super::timeline::ShutdownMode::Hard)
7503 4 : .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))
7504 4 : .await;
7505 4 : std::mem::forget(tline);
7506 4 : }
7507 4 :
7508 4 : let (tenant, _) = harness.load().await;
7509 4 : match tenant.get_timeline(TIMELINE_ID, false) {
7510 4 : Ok(_) => panic!("timeline should've been removed during load"),
7511 4 : Err(e) => {
7512 4 : assert_eq!(
7513 4 : e,
7514 4 : GetTimelineError::NotFound {
7515 4 : tenant_id: tenant.tenant_shard_id,
7516 4 : timeline_id: TIMELINE_ID,
7517 4 : }
7518 4 : )
7519 4 : }
7520 4 : }
7521 4 :
7522 4 : assert!(
7523 4 : !harness
7524 4 : .conf
7525 4 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7526 4 : .exists()
7527 4 : );
7528 4 :
7529 4 : Ok(())
7530 4 : }
7531 :
7532 : #[tokio::test]
7533 4 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7534 4 : let names_algorithms = [
7535 4 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7536 4 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7537 4 : ];
7538 12 : for (name, algorithm) in names_algorithms {
7539 8 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7540 4 : }
7541 4 : Ok(())
7542 4 : }
7543 :
7544 8 : async fn test_read_at_max_lsn_algorithm(
7545 8 : name: &'static str,
7546 8 : compaction_algorithm: CompactionAlgorithm,
7547 8 : ) -> anyhow::Result<()> {
7548 8 : let mut harness = TenantHarness::create(name).await?;
7549 8 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7550 8 : kind: compaction_algorithm,
7551 8 : };
7552 8 : let (tenant, ctx) = harness.load().await;
7553 8 : let tline = tenant
7554 8 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7555 8 : .await?;
7556 :
7557 8 : let lsn = Lsn(0x10);
7558 8 : let compact = false;
7559 8 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7560 :
7561 8 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7562 8 : let read_lsn = Lsn(u64::MAX - 1);
7563 :
7564 8 : let result = tline.get(test_key, read_lsn, &ctx).await;
7565 8 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7566 :
7567 8 : Ok(())
7568 8 : }
7569 :
7570 : #[tokio::test]
7571 4 : async fn test_metadata_scan() -> anyhow::Result<()> {
7572 4 : let harness = TenantHarness::create("test_metadata_scan").await?;
7573 4 : let (tenant, ctx) = harness.load().await;
7574 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7575 4 : let tline = tenant
7576 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7577 4 : .await?;
7578 4 :
7579 4 : const NUM_KEYS: usize = 1000;
7580 4 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7581 4 :
7582 4 : let cancel = CancellationToken::new();
7583 4 :
7584 4 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7585 4 : base_key.field1 = AUX_KEY_PREFIX;
7586 4 : let mut test_key = base_key;
7587 4 :
7588 4 : // Track when each page was last modified. Used to assert that
7589 4 : // a read sees the latest page version.
7590 4 : let mut updated = [Lsn(0); NUM_KEYS];
7591 4 :
7592 4 : let mut lsn = Lsn(0x10);
7593 4 : #[allow(clippy::needless_range_loop)]
7594 4004 : for blknum in 0..NUM_KEYS {
7595 4000 : lsn = Lsn(lsn.0 + 0x10);
7596 4000 : test_key.field6 = (blknum * STEP) as u32;
7597 4000 : let mut writer = tline.writer().await;
7598 4000 : writer
7599 4000 : .put(
7600 4000 : test_key,
7601 4000 : lsn,
7602 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7603 4000 : &ctx,
7604 4000 : )
7605 4000 : .await?;
7606 4000 : writer.finish_write(lsn);
7607 4000 : updated[blknum] = lsn;
7608 4000 : drop(writer);
7609 4 : }
7610 4 :
7611 4 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7612 4 :
7613 48 : for iter in 0..=10 {
7614 4 : // Read all the blocks
7615 44000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7616 44000 : test_key.field6 = (blknum * STEP) as u32;
7617 44000 : assert_eq!(
7618 44000 : tline.get(test_key, lsn, &ctx).await?,
7619 44000 : test_img(&format!("{} at {}", blknum, last_lsn))
7620 4 : );
7621 4 : }
7622 4 :
7623 44 : let mut cnt = 0;
7624 44000 : for (key, value) in tline
7625 44 : .get_vectored_impl(
7626 44 : keyspace.clone(),
7627 44 : lsn,
7628 44 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7629 44 : &ctx,
7630 44 : )
7631 44 : .await?
7632 4 : {
7633 44000 : let blknum = key.field6 as usize;
7634 44000 : let value = value?;
7635 44000 : assert!(blknum % STEP == 0);
7636 44000 : let blknum = blknum / STEP;
7637 44000 : assert_eq!(
7638 44000 : value,
7639 44000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
7640 44000 : );
7641 44000 : cnt += 1;
7642 4 : }
7643 4 :
7644 44 : assert_eq!(cnt, NUM_KEYS);
7645 4 :
7646 44044 : for _ in 0..NUM_KEYS {
7647 44000 : lsn = Lsn(lsn.0 + 0x10);
7648 44000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7649 44000 : test_key.field6 = (blknum * STEP) as u32;
7650 44000 : let mut writer = tline.writer().await;
7651 44000 : writer
7652 44000 : .put(
7653 44000 : test_key,
7654 44000 : lsn,
7655 44000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7656 44000 : &ctx,
7657 44000 : )
7658 44000 : .await?;
7659 44000 : writer.finish_write(lsn);
7660 44000 : drop(writer);
7661 44000 : updated[blknum] = lsn;
7662 4 : }
7663 4 :
7664 4 : // Perform two cycles of flush, compact, and GC
7665 132 : for round in 0..2 {
7666 88 : tline.freeze_and_flush().await?;
7667 88 : tline
7668 88 : .compact(
7669 88 : &cancel,
7670 88 : if iter % 5 == 0 && round == 0 {
7671 12 : let mut flags = EnumSet::new();
7672 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
7673 12 : flags.insert(CompactFlags::ForceRepartition);
7674 12 : flags
7675 4 : } else {
7676 76 : EnumSet::empty()
7677 4 : },
7678 88 : &ctx,
7679 88 : )
7680 88 : .await?;
7681 88 : tenant
7682 88 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7683 88 : .await?;
7684 4 : }
7685 4 : }
7686 4 :
7687 4 : Ok(())
7688 4 : }
7689 :
7690 : #[tokio::test]
7691 4 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7692 4 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7693 4 : let (tenant, ctx) = harness.load().await;
7694 4 : let tline = tenant
7695 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7696 4 : .await?;
7697 4 :
7698 4 : let cancel = CancellationToken::new();
7699 4 :
7700 4 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7701 4 : base_key.field1 = AUX_KEY_PREFIX;
7702 4 : let test_key = base_key;
7703 4 : let mut lsn = Lsn(0x10);
7704 4 :
7705 84 : for _ in 0..20 {
7706 80 : lsn = Lsn(lsn.0 + 0x10);
7707 80 : let mut writer = tline.writer().await;
7708 80 : writer
7709 80 : .put(
7710 80 : test_key,
7711 80 : lsn,
7712 80 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7713 80 : &ctx,
7714 80 : )
7715 80 : .await?;
7716 80 : writer.finish_write(lsn);
7717 80 : drop(writer);
7718 80 : tline.freeze_and_flush().await?; // force create a delta layer
7719 4 : }
7720 4 :
7721 4 : let before_num_l0_delta_files =
7722 4 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7723 4 :
7724 4 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7725 4 :
7726 4 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7727 4 :
7728 4 : assert!(
7729 4 : after_num_l0_delta_files < before_num_l0_delta_files,
7730 4 : "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}"
7731 4 : );
7732 4 :
7733 4 : assert_eq!(
7734 4 : tline.get(test_key, lsn, &ctx).await?,
7735 4 : test_img(&format!("{} at {}", 0, lsn))
7736 4 : );
7737 4 :
7738 4 : Ok(())
7739 4 : }
7740 :
7741 : #[tokio::test]
7742 4 : async fn test_aux_file_e2e() {
7743 4 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7744 4 :
7745 4 : let (tenant, ctx) = harness.load().await;
7746 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7747 4 :
7748 4 : let mut lsn = Lsn(0x08);
7749 4 :
7750 4 : let tline: Arc<Timeline> = tenant
7751 4 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7752 4 : .await
7753 4 : .unwrap();
7754 4 :
7755 4 : {
7756 4 : lsn += 8;
7757 4 : let mut modification = tline.begin_modification(lsn);
7758 4 : modification
7759 4 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7760 4 : .await
7761 4 : .unwrap();
7762 4 : modification.commit(&ctx).await.unwrap();
7763 4 : }
7764 4 :
7765 4 : // we can read everything from the storage
7766 4 : let files = tline
7767 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7768 4 : .await
7769 4 : .unwrap();
7770 4 : assert_eq!(
7771 4 : files.get("pg_logical/mappings/test1"),
7772 4 : Some(&bytes::Bytes::from_static(b"first"))
7773 4 : );
7774 4 :
7775 4 : {
7776 4 : lsn += 8;
7777 4 : let mut modification = tline.begin_modification(lsn);
7778 4 : modification
7779 4 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7780 4 : .await
7781 4 : .unwrap();
7782 4 : modification.commit(&ctx).await.unwrap();
7783 4 : }
7784 4 :
7785 4 : let files = tline
7786 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7787 4 : .await
7788 4 : .unwrap();
7789 4 : assert_eq!(
7790 4 : files.get("pg_logical/mappings/test2"),
7791 4 : Some(&bytes::Bytes::from_static(b"second"))
7792 4 : );
7793 4 :
7794 4 : let child = tenant
7795 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7796 4 : .await
7797 4 : .unwrap();
7798 4 :
7799 4 : let files = child
7800 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7801 4 : .await
7802 4 : .unwrap();
7803 4 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7804 4 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7805 4 : }
7806 :
7807 : #[tokio::test]
7808 4 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7809 4 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7810 4 : let (tenant, ctx) = harness.load().await;
7811 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7812 4 : let tline = tenant
7813 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7814 4 : .await?;
7815 4 :
7816 4 : const NUM_KEYS: usize = 1000;
7817 4 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7818 4 :
7819 4 : let cancel = CancellationToken::new();
7820 4 :
7821 4 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7822 4 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7823 4 : let mut test_key = base_key;
7824 4 : let mut lsn = Lsn(0x10);
7825 4 :
7826 16 : async fn scan_with_statistics(
7827 16 : tline: &Timeline,
7828 16 : keyspace: &KeySpace,
7829 16 : lsn: Lsn,
7830 16 : ctx: &RequestContext,
7831 16 : io_concurrency: IoConcurrency,
7832 16 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7833 16 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
7834 16 : let res = tline
7835 16 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7836 16 : .await?;
7837 16 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7838 16 : }
7839 4 :
7840 4 : #[allow(clippy::needless_range_loop)]
7841 4004 : for blknum in 0..NUM_KEYS {
7842 4000 : lsn = Lsn(lsn.0 + 0x10);
7843 4000 : test_key.field6 = (blknum * STEP) as u32;
7844 4000 : let mut writer = tline.writer().await;
7845 4000 : writer
7846 4000 : .put(
7847 4000 : test_key,
7848 4000 : lsn,
7849 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7850 4000 : &ctx,
7851 4000 : )
7852 4000 : .await?;
7853 4000 : writer.finish_write(lsn);
7854 4000 : drop(writer);
7855 4 : }
7856 4 :
7857 4 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7858 4 :
7859 44 : for iter in 1..=10 {
7860 40040 : for _ in 0..NUM_KEYS {
7861 40000 : lsn = Lsn(lsn.0 + 0x10);
7862 40000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7863 40000 : test_key.field6 = (blknum * STEP) as u32;
7864 40000 : let mut writer = tline.writer().await;
7865 40000 : writer
7866 40000 : .put(
7867 40000 : test_key,
7868 40000 : lsn,
7869 40000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7870 40000 : &ctx,
7871 40000 : )
7872 40000 : .await?;
7873 40000 : writer.finish_write(lsn);
7874 40000 : drop(writer);
7875 4 : }
7876 4 :
7877 40 : tline.freeze_and_flush().await?;
7878 4 :
7879 40 : if iter % 5 == 0 {
7880 8 : let (_, before_delta_file_accessed) =
7881 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
7882 8 : .await?;
7883 8 : tline
7884 8 : .compact(
7885 8 : &cancel,
7886 8 : {
7887 8 : let mut flags = EnumSet::new();
7888 8 : flags.insert(CompactFlags::ForceImageLayerCreation);
7889 8 : flags.insert(CompactFlags::ForceRepartition);
7890 8 : flags
7891 8 : },
7892 8 : &ctx,
7893 8 : )
7894 8 : .await?;
7895 8 : let (_, after_delta_file_accessed) =
7896 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
7897 8 : .await?;
7898 8 : assert!(
7899 8 : after_delta_file_accessed < before_delta_file_accessed,
7900 4 : "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}"
7901 4 : );
7902 4 : // 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.
7903 8 : assert!(
7904 8 : after_delta_file_accessed <= 2,
7905 4 : "after_delta_file_accessed={after_delta_file_accessed}"
7906 4 : );
7907 32 : }
7908 4 : }
7909 4 :
7910 4 : Ok(())
7911 4 : }
7912 :
7913 : #[tokio::test]
7914 4 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7915 4 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7916 4 : let (tenant, ctx) = harness.load().await;
7917 4 :
7918 4 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7919 4 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7920 4 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7921 4 :
7922 4 : let tline = tenant
7923 4 : .create_test_timeline_with_layers(
7924 4 : TIMELINE_ID,
7925 4 : Lsn(0x10),
7926 4 : DEFAULT_PG_VERSION,
7927 4 : &ctx,
7928 4 : Vec::new(), // delta layers
7929 4 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7930 4 : 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
7931 4 : )
7932 4 : .await?;
7933 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7934 4 :
7935 4 : let child = tenant
7936 4 : .branch_timeline_test_with_layers(
7937 4 : &tline,
7938 4 : NEW_TIMELINE_ID,
7939 4 : Some(Lsn(0x20)),
7940 4 : &ctx,
7941 4 : Vec::new(), // delta layers
7942 4 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
7943 4 : Lsn(0x30),
7944 4 : )
7945 4 : .await
7946 4 : .unwrap();
7947 4 :
7948 4 : let lsn = Lsn(0x30);
7949 4 :
7950 4 : // test vectored get on parent timeline
7951 4 : assert_eq!(
7952 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
7953 4 : Some(test_img("data key 1"))
7954 4 : );
7955 4 : assert!(
7956 4 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
7957 4 : .await
7958 4 : .unwrap_err()
7959 4 : .is_missing_key_error()
7960 4 : );
7961 4 : assert!(
7962 4 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
7963 4 : .await
7964 4 : .unwrap_err()
7965 4 : .is_missing_key_error()
7966 4 : );
7967 4 :
7968 4 : // test vectored get on child timeline
7969 4 : assert_eq!(
7970 4 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
7971 4 : Some(test_img("data key 1"))
7972 4 : );
7973 4 : assert_eq!(
7974 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
7975 4 : Some(test_img("data key 2"))
7976 4 : );
7977 4 : assert!(
7978 4 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
7979 4 : .await
7980 4 : .unwrap_err()
7981 4 : .is_missing_key_error()
7982 4 : );
7983 4 :
7984 4 : Ok(())
7985 4 : }
7986 :
7987 : #[tokio::test]
7988 4 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
7989 4 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
7990 4 : let (tenant, ctx) = harness.load().await;
7991 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7992 4 :
7993 4 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7994 4 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
7995 4 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
7996 4 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
7997 4 :
7998 4 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
7999 4 : let base_inherited_key_child =
8000 4 : Key::from_hex("610000000033333333444444445500000001").unwrap();
8001 4 : let base_inherited_key_nonexist =
8002 4 : Key::from_hex("610000000033333333444444445500000002").unwrap();
8003 4 : let base_inherited_key_overwrite =
8004 4 : Key::from_hex("610000000033333333444444445500000003").unwrap();
8005 4 :
8006 4 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8007 4 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
8008 4 :
8009 4 : let tline = tenant
8010 4 : .create_test_timeline_with_layers(
8011 4 : TIMELINE_ID,
8012 4 : Lsn(0x10),
8013 4 : DEFAULT_PG_VERSION,
8014 4 : &ctx,
8015 4 : Vec::new(), // delta layers
8016 4 : vec![(
8017 4 : Lsn(0x20),
8018 4 : vec![
8019 4 : (base_inherited_key, test_img("metadata inherited key 1")),
8020 4 : (
8021 4 : base_inherited_key_overwrite,
8022 4 : test_img("metadata key overwrite 1a"),
8023 4 : ),
8024 4 : (base_key, test_img("metadata key 1")),
8025 4 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8026 4 : ],
8027 4 : )], // image layers
8028 4 : 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
8029 4 : )
8030 4 : .await?;
8031 4 :
8032 4 : let child = tenant
8033 4 : .branch_timeline_test_with_layers(
8034 4 : &tline,
8035 4 : NEW_TIMELINE_ID,
8036 4 : Some(Lsn(0x20)),
8037 4 : &ctx,
8038 4 : Vec::new(), // delta layers
8039 4 : vec![(
8040 4 : Lsn(0x30),
8041 4 : vec![
8042 4 : (
8043 4 : base_inherited_key_child,
8044 4 : test_img("metadata inherited key 2"),
8045 4 : ),
8046 4 : (
8047 4 : base_inherited_key_overwrite,
8048 4 : test_img("metadata key overwrite 2a"),
8049 4 : ),
8050 4 : (base_key_child, test_img("metadata key 2")),
8051 4 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8052 4 : ],
8053 4 : )], // image layers
8054 4 : Lsn(0x30),
8055 4 : )
8056 4 : .await
8057 4 : .unwrap();
8058 4 :
8059 4 : let lsn = Lsn(0x30);
8060 4 :
8061 4 : // test vectored get on parent timeline
8062 4 : assert_eq!(
8063 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8064 4 : Some(test_img("metadata key 1"))
8065 4 : );
8066 4 : assert_eq!(
8067 4 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
8068 4 : None
8069 4 : );
8070 4 : assert_eq!(
8071 4 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
8072 4 : None
8073 4 : );
8074 4 : assert_eq!(
8075 4 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
8076 4 : Some(test_img("metadata key overwrite 1b"))
8077 4 : );
8078 4 : assert_eq!(
8079 4 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
8080 4 : Some(test_img("metadata inherited key 1"))
8081 4 : );
8082 4 : assert_eq!(
8083 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
8084 4 : None
8085 4 : );
8086 4 : assert_eq!(
8087 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
8088 4 : None
8089 4 : );
8090 4 : assert_eq!(
8091 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
8092 4 : Some(test_img("metadata key overwrite 1a"))
8093 4 : );
8094 4 :
8095 4 : // test vectored get on child timeline
8096 4 : assert_eq!(
8097 4 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8098 4 : None
8099 4 : );
8100 4 : assert_eq!(
8101 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8102 4 : Some(test_img("metadata key 2"))
8103 4 : );
8104 4 : assert_eq!(
8105 4 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
8106 4 : None
8107 4 : );
8108 4 : assert_eq!(
8109 4 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
8110 4 : Some(test_img("metadata inherited key 1"))
8111 4 : );
8112 4 : assert_eq!(
8113 4 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
8114 4 : Some(test_img("metadata inherited key 2"))
8115 4 : );
8116 4 : assert_eq!(
8117 4 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
8118 4 : None
8119 4 : );
8120 4 : assert_eq!(
8121 4 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
8122 4 : Some(test_img("metadata key overwrite 2b"))
8123 4 : );
8124 4 : assert_eq!(
8125 4 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
8126 4 : Some(test_img("metadata key overwrite 2a"))
8127 4 : );
8128 4 :
8129 4 : // test vectored scan on parent timeline
8130 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8131 4 : let res = tline
8132 4 : .get_vectored_impl(
8133 4 : KeySpace::single(Key::metadata_key_range()),
8134 4 : lsn,
8135 4 : &mut reconstruct_state,
8136 4 : &ctx,
8137 4 : )
8138 4 : .await?;
8139 4 :
8140 4 : assert_eq!(
8141 4 : res.into_iter()
8142 16 : .map(|(k, v)| (k, v.unwrap()))
8143 4 : .collect::<Vec<_>>(),
8144 4 : vec![
8145 4 : (base_inherited_key, test_img("metadata inherited key 1")),
8146 4 : (
8147 4 : base_inherited_key_overwrite,
8148 4 : test_img("metadata key overwrite 1a")
8149 4 : ),
8150 4 : (base_key, test_img("metadata key 1")),
8151 4 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8152 4 : ]
8153 4 : );
8154 4 :
8155 4 : // test vectored scan on child timeline
8156 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8157 4 : let res = child
8158 4 : .get_vectored_impl(
8159 4 : KeySpace::single(Key::metadata_key_range()),
8160 4 : lsn,
8161 4 : &mut reconstruct_state,
8162 4 : &ctx,
8163 4 : )
8164 4 : .await?;
8165 4 :
8166 4 : assert_eq!(
8167 4 : res.into_iter()
8168 20 : .map(|(k, v)| (k, v.unwrap()))
8169 4 : .collect::<Vec<_>>(),
8170 4 : vec![
8171 4 : (base_inherited_key, test_img("metadata inherited key 1")),
8172 4 : (
8173 4 : base_inherited_key_child,
8174 4 : test_img("metadata inherited key 2")
8175 4 : ),
8176 4 : (
8177 4 : base_inherited_key_overwrite,
8178 4 : test_img("metadata key overwrite 2a")
8179 4 : ),
8180 4 : (base_key_child, test_img("metadata key 2")),
8181 4 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8182 4 : ]
8183 4 : );
8184 4 :
8185 4 : Ok(())
8186 4 : }
8187 :
8188 112 : async fn get_vectored_impl_wrapper(
8189 112 : tline: &Arc<Timeline>,
8190 112 : key: Key,
8191 112 : lsn: Lsn,
8192 112 : ctx: &RequestContext,
8193 112 : ) -> Result<Option<Bytes>, GetVectoredError> {
8194 112 : let io_concurrency =
8195 112 : IoConcurrency::spawn_from_conf(tline.conf, tline.gate.enter().unwrap());
8196 112 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8197 112 : let mut res = tline
8198 112 : .get_vectored_impl(
8199 112 : KeySpace::single(key..key.next()),
8200 112 : lsn,
8201 112 : &mut reconstruct_state,
8202 112 : ctx,
8203 112 : )
8204 112 : .await?;
8205 100 : Ok(res.pop_last().map(|(k, v)| {
8206 64 : assert_eq!(k, key);
8207 64 : v.unwrap()
8208 100 : }))
8209 112 : }
8210 :
8211 : #[tokio::test]
8212 4 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8213 4 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8214 4 : let (tenant, ctx) = harness.load().await;
8215 4 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8216 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8217 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8218 4 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8219 4 :
8220 4 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8221 4 : // Lsn 0x30 key0, key3, no key1+key2
8222 4 : // Lsn 0x20 key1+key2 tomestones
8223 4 : // Lsn 0x10 key1 in image, key2 in delta
8224 4 : let tline = tenant
8225 4 : .create_test_timeline_with_layers(
8226 4 : TIMELINE_ID,
8227 4 : Lsn(0x10),
8228 4 : DEFAULT_PG_VERSION,
8229 4 : &ctx,
8230 4 : // delta layers
8231 4 : vec![
8232 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8233 4 : Lsn(0x10)..Lsn(0x20),
8234 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8235 4 : ),
8236 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8237 4 : Lsn(0x20)..Lsn(0x30),
8238 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8239 4 : ),
8240 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8241 4 : Lsn(0x20)..Lsn(0x30),
8242 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8243 4 : ),
8244 4 : ],
8245 4 : // image layers
8246 4 : vec![
8247 4 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8248 4 : (
8249 4 : Lsn(0x30),
8250 4 : vec![
8251 4 : (key0, test_img("metadata key 0")),
8252 4 : (key3, test_img("metadata key 3")),
8253 4 : ],
8254 4 : ),
8255 4 : ],
8256 4 : Lsn(0x30),
8257 4 : )
8258 4 : .await?;
8259 4 :
8260 4 : let lsn = Lsn(0x30);
8261 4 : let old_lsn = Lsn(0x20);
8262 4 :
8263 4 : assert_eq!(
8264 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8265 4 : Some(test_img("metadata key 0"))
8266 4 : );
8267 4 : assert_eq!(
8268 4 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8269 4 : None,
8270 4 : );
8271 4 : assert_eq!(
8272 4 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8273 4 : None,
8274 4 : );
8275 4 : assert_eq!(
8276 4 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8277 4 : Some(Bytes::new()),
8278 4 : );
8279 4 : assert_eq!(
8280 4 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8281 4 : Some(Bytes::new()),
8282 4 : );
8283 4 : assert_eq!(
8284 4 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8285 4 : Some(test_img("metadata key 3"))
8286 4 : );
8287 4 :
8288 4 : Ok(())
8289 4 : }
8290 :
8291 : #[tokio::test]
8292 4 : async fn test_metadata_tombstone_image_creation() {
8293 4 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8294 4 : .await
8295 4 : .unwrap();
8296 4 : let (tenant, ctx) = harness.load().await;
8297 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8298 4 :
8299 4 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8300 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8301 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8302 4 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8303 4 :
8304 4 : let tline = tenant
8305 4 : .create_test_timeline_with_layers(
8306 4 : TIMELINE_ID,
8307 4 : Lsn(0x10),
8308 4 : DEFAULT_PG_VERSION,
8309 4 : &ctx,
8310 4 : // delta layers
8311 4 : vec![
8312 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8313 4 : Lsn(0x10)..Lsn(0x20),
8314 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8315 4 : ),
8316 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8317 4 : Lsn(0x20)..Lsn(0x30),
8318 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8319 4 : ),
8320 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8321 4 : Lsn(0x20)..Lsn(0x30),
8322 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8323 4 : ),
8324 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8325 4 : Lsn(0x30)..Lsn(0x40),
8326 4 : vec![
8327 4 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8328 4 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8329 4 : ],
8330 4 : ),
8331 4 : ],
8332 4 : // image layers
8333 4 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8334 4 : Lsn(0x40),
8335 4 : )
8336 4 : .await
8337 4 : .unwrap();
8338 4 :
8339 4 : let cancel = CancellationToken::new();
8340 4 :
8341 4 : tline
8342 4 : .compact(
8343 4 : &cancel,
8344 4 : {
8345 4 : let mut flags = EnumSet::new();
8346 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
8347 4 : flags.insert(CompactFlags::ForceRepartition);
8348 4 : flags
8349 4 : },
8350 4 : &ctx,
8351 4 : )
8352 4 : .await
8353 4 : .unwrap();
8354 4 :
8355 4 : // Image layers are created at last_record_lsn
8356 4 : let images = tline
8357 4 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
8358 4 : .await
8359 4 : .unwrap()
8360 4 : .into_iter()
8361 36 : .filter(|(k, _)| k.is_metadata_key())
8362 4 : .collect::<Vec<_>>();
8363 4 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8364 4 : }
8365 :
8366 : #[tokio::test]
8367 4 : async fn test_metadata_tombstone_empty_image_creation() {
8368 4 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8369 4 : .await
8370 4 : .unwrap();
8371 4 : let (tenant, ctx) = harness.load().await;
8372 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8373 4 :
8374 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8375 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8376 4 :
8377 4 : let tline = tenant
8378 4 : .create_test_timeline_with_layers(
8379 4 : TIMELINE_ID,
8380 4 : Lsn(0x10),
8381 4 : DEFAULT_PG_VERSION,
8382 4 : &ctx,
8383 4 : // delta layers
8384 4 : vec![
8385 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8386 4 : Lsn(0x10)..Lsn(0x20),
8387 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8388 4 : ),
8389 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8390 4 : Lsn(0x20)..Lsn(0x30),
8391 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8392 4 : ),
8393 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8394 4 : Lsn(0x20)..Lsn(0x30),
8395 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8396 4 : ),
8397 4 : ],
8398 4 : // image layers
8399 4 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8400 4 : Lsn(0x30),
8401 4 : )
8402 4 : .await
8403 4 : .unwrap();
8404 4 :
8405 4 : let cancel = CancellationToken::new();
8406 4 :
8407 4 : tline
8408 4 : .compact(
8409 4 : &cancel,
8410 4 : {
8411 4 : let mut flags = EnumSet::new();
8412 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
8413 4 : flags.insert(CompactFlags::ForceRepartition);
8414 4 : flags
8415 4 : },
8416 4 : &ctx,
8417 4 : )
8418 4 : .await
8419 4 : .unwrap();
8420 4 :
8421 4 : // Image layers are created at last_record_lsn
8422 4 : let images = tline
8423 4 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8424 4 : .await
8425 4 : .unwrap()
8426 4 : .into_iter()
8427 28 : .filter(|(k, _)| k.is_metadata_key())
8428 4 : .collect::<Vec<_>>();
8429 4 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8430 4 : }
8431 :
8432 : #[tokio::test]
8433 4 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8434 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8435 4 : let (tenant, ctx) = harness.load().await;
8436 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8437 4 :
8438 204 : fn get_key(id: u32) -> Key {
8439 204 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8440 204 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8441 204 : key.field6 = id;
8442 204 : key
8443 204 : }
8444 4 :
8445 4 : // We create
8446 4 : // - one bottom-most image layer,
8447 4 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8448 4 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8449 4 : // - a delta layer D3 above the horizon.
8450 4 : //
8451 4 : // | D3 |
8452 4 : // | D1 |
8453 4 : // -| |-- gc horizon -----------------
8454 4 : // | | | D2 |
8455 4 : // --------- img layer ------------------
8456 4 : //
8457 4 : // What we should expact from this compaction is:
8458 4 : // | D3 |
8459 4 : // | Part of D1 |
8460 4 : // --------- img layer with D1+D2 at GC horizon------------------
8461 4 :
8462 4 : // img layer at 0x10
8463 4 : let img_layer = (0..10)
8464 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8465 4 : .collect_vec();
8466 4 :
8467 4 : let delta1 = vec![
8468 4 : (
8469 4 : get_key(1),
8470 4 : Lsn(0x20),
8471 4 : Value::Image(Bytes::from("value 1@0x20")),
8472 4 : ),
8473 4 : (
8474 4 : get_key(2),
8475 4 : Lsn(0x30),
8476 4 : Value::Image(Bytes::from("value 2@0x30")),
8477 4 : ),
8478 4 : (
8479 4 : get_key(3),
8480 4 : Lsn(0x40),
8481 4 : Value::Image(Bytes::from("value 3@0x40")),
8482 4 : ),
8483 4 : ];
8484 4 : let delta2 = vec![
8485 4 : (
8486 4 : get_key(5),
8487 4 : Lsn(0x20),
8488 4 : Value::Image(Bytes::from("value 5@0x20")),
8489 4 : ),
8490 4 : (
8491 4 : get_key(6),
8492 4 : Lsn(0x20),
8493 4 : Value::Image(Bytes::from("value 6@0x20")),
8494 4 : ),
8495 4 : ];
8496 4 : let delta3 = vec![
8497 4 : (
8498 4 : get_key(8),
8499 4 : Lsn(0x48),
8500 4 : Value::Image(Bytes::from("value 8@0x48")),
8501 4 : ),
8502 4 : (
8503 4 : get_key(9),
8504 4 : Lsn(0x48),
8505 4 : Value::Image(Bytes::from("value 9@0x48")),
8506 4 : ),
8507 4 : ];
8508 4 :
8509 4 : let tline = tenant
8510 4 : .create_test_timeline_with_layers(
8511 4 : TIMELINE_ID,
8512 4 : Lsn(0x10),
8513 4 : DEFAULT_PG_VERSION,
8514 4 : &ctx,
8515 4 : vec![
8516 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8517 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8518 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8519 4 : ], // delta layers
8520 4 : vec![(Lsn(0x10), img_layer)], // image layers
8521 4 : Lsn(0x50),
8522 4 : )
8523 4 : .await?;
8524 4 : {
8525 4 : tline
8526 4 : .applied_gc_cutoff_lsn
8527 4 : .lock_for_write()
8528 4 : .store_and_unlock(Lsn(0x30))
8529 4 : .wait()
8530 4 : .await;
8531 4 : // Update GC info
8532 4 : let mut guard = tline.gc_info.write().unwrap();
8533 4 : guard.cutoffs.time = Lsn(0x30);
8534 4 : guard.cutoffs.space = Lsn(0x30);
8535 4 : }
8536 4 :
8537 4 : let expected_result = [
8538 4 : Bytes::from_static(b"value 0@0x10"),
8539 4 : Bytes::from_static(b"value 1@0x20"),
8540 4 : Bytes::from_static(b"value 2@0x30"),
8541 4 : Bytes::from_static(b"value 3@0x40"),
8542 4 : Bytes::from_static(b"value 4@0x10"),
8543 4 : Bytes::from_static(b"value 5@0x20"),
8544 4 : Bytes::from_static(b"value 6@0x20"),
8545 4 : Bytes::from_static(b"value 7@0x10"),
8546 4 : Bytes::from_static(b"value 8@0x48"),
8547 4 : Bytes::from_static(b"value 9@0x48"),
8548 4 : ];
8549 4 :
8550 40 : for (idx, expected) in expected_result.iter().enumerate() {
8551 40 : assert_eq!(
8552 40 : tline
8553 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8554 40 : .await
8555 40 : .unwrap(),
8556 4 : expected
8557 4 : );
8558 4 : }
8559 4 :
8560 4 : let cancel = CancellationToken::new();
8561 4 : tline
8562 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8563 4 : .await
8564 4 : .unwrap();
8565 4 :
8566 40 : for (idx, expected) in expected_result.iter().enumerate() {
8567 40 : assert_eq!(
8568 40 : tline
8569 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8570 40 : .await
8571 40 : .unwrap(),
8572 4 : expected
8573 4 : );
8574 4 : }
8575 4 :
8576 4 : // Check if the image layer at the GC horizon contains exactly what we want
8577 4 : let image_at_gc_horizon = tline
8578 4 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8579 4 : .await
8580 4 : .unwrap()
8581 4 : .into_iter()
8582 68 : .filter(|(k, _)| k.is_metadata_key())
8583 4 : .collect::<Vec<_>>();
8584 4 :
8585 4 : assert_eq!(image_at_gc_horizon.len(), 10);
8586 4 : let expected_result = [
8587 4 : Bytes::from_static(b"value 0@0x10"),
8588 4 : Bytes::from_static(b"value 1@0x20"),
8589 4 : Bytes::from_static(b"value 2@0x30"),
8590 4 : Bytes::from_static(b"value 3@0x10"),
8591 4 : Bytes::from_static(b"value 4@0x10"),
8592 4 : Bytes::from_static(b"value 5@0x20"),
8593 4 : Bytes::from_static(b"value 6@0x20"),
8594 4 : Bytes::from_static(b"value 7@0x10"),
8595 4 : Bytes::from_static(b"value 8@0x10"),
8596 4 : Bytes::from_static(b"value 9@0x10"),
8597 4 : ];
8598 44 : for idx in 0..10 {
8599 40 : assert_eq!(
8600 40 : image_at_gc_horizon[idx],
8601 40 : (get_key(idx as u32), expected_result[idx].clone())
8602 40 : );
8603 4 : }
8604 4 :
8605 4 : // Check if old layers are removed / new layers have the expected LSN
8606 4 : let all_layers = inspect_and_sort(&tline, None).await;
8607 4 : assert_eq!(
8608 4 : all_layers,
8609 4 : vec![
8610 4 : // Image layer at GC horizon
8611 4 : PersistentLayerKey {
8612 4 : key_range: Key::MIN..Key::MAX,
8613 4 : lsn_range: Lsn(0x30)..Lsn(0x31),
8614 4 : is_delta: false
8615 4 : },
8616 4 : // The delta layer below the horizon
8617 4 : PersistentLayerKey {
8618 4 : key_range: get_key(3)..get_key(4),
8619 4 : lsn_range: Lsn(0x30)..Lsn(0x48),
8620 4 : is_delta: true
8621 4 : },
8622 4 : // The delta3 layer that should not be picked for the compaction
8623 4 : PersistentLayerKey {
8624 4 : key_range: get_key(8)..get_key(10),
8625 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
8626 4 : is_delta: true
8627 4 : }
8628 4 : ]
8629 4 : );
8630 4 :
8631 4 : // increase GC horizon and compact again
8632 4 : {
8633 4 : tline
8634 4 : .applied_gc_cutoff_lsn
8635 4 : .lock_for_write()
8636 4 : .store_and_unlock(Lsn(0x40))
8637 4 : .wait()
8638 4 : .await;
8639 4 : // Update GC info
8640 4 : let mut guard = tline.gc_info.write().unwrap();
8641 4 : guard.cutoffs.time = Lsn(0x40);
8642 4 : guard.cutoffs.space = Lsn(0x40);
8643 4 : }
8644 4 : tline
8645 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8646 4 : .await
8647 4 : .unwrap();
8648 4 :
8649 4 : Ok(())
8650 4 : }
8651 :
8652 : #[cfg(feature = "testing")]
8653 : #[tokio::test]
8654 4 : async fn test_neon_test_record() -> anyhow::Result<()> {
8655 4 : let harness = TenantHarness::create("test_neon_test_record").await?;
8656 4 : let (tenant, ctx) = harness.load().await;
8657 4 :
8658 48 : fn get_key(id: u32) -> Key {
8659 48 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8660 48 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8661 48 : key.field6 = id;
8662 48 : key
8663 48 : }
8664 4 :
8665 4 : let delta1 = vec![
8666 4 : (
8667 4 : get_key(1),
8668 4 : Lsn(0x20),
8669 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8670 4 : ),
8671 4 : (
8672 4 : get_key(1),
8673 4 : Lsn(0x30),
8674 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8675 4 : ),
8676 4 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
8677 4 : (
8678 4 : get_key(2),
8679 4 : Lsn(0x20),
8680 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8681 4 : ),
8682 4 : (
8683 4 : get_key(2),
8684 4 : Lsn(0x30),
8685 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8686 4 : ),
8687 4 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
8688 4 : (
8689 4 : get_key(3),
8690 4 : Lsn(0x20),
8691 4 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
8692 4 : ),
8693 4 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
8694 4 : (
8695 4 : get_key(4),
8696 4 : Lsn(0x20),
8697 4 : Value::WalRecord(NeonWalRecord::wal_init("i")),
8698 4 : ),
8699 4 : ];
8700 4 : let image1 = vec![(get_key(1), "0x10".into())];
8701 4 :
8702 4 : let tline = tenant
8703 4 : .create_test_timeline_with_layers(
8704 4 : TIMELINE_ID,
8705 4 : Lsn(0x10),
8706 4 : DEFAULT_PG_VERSION,
8707 4 : &ctx,
8708 4 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8709 4 : Lsn(0x10)..Lsn(0x40),
8710 4 : delta1,
8711 4 : )], // delta layers
8712 4 : vec![(Lsn(0x10), image1)], // image layers
8713 4 : Lsn(0x50),
8714 4 : )
8715 4 : .await?;
8716 4 :
8717 4 : assert_eq!(
8718 4 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
8719 4 : Bytes::from_static(b"0x10,0x20,0x30")
8720 4 : );
8721 4 : assert_eq!(
8722 4 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
8723 4 : Bytes::from_static(b"0x10,0x20,0x30")
8724 4 : );
8725 4 :
8726 4 : // Need to remove the limit of "Neon WAL redo requires base image".
8727 4 :
8728 4 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
8729 4 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
8730 4 :
8731 4 : Ok(())
8732 4 : }
8733 :
8734 : #[tokio::test(start_paused = true)]
8735 4 : async fn test_lsn_lease() -> anyhow::Result<()> {
8736 4 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
8737 4 : .await
8738 4 : .unwrap()
8739 4 : .load()
8740 4 : .await;
8741 4 : // Advance to the lsn lease deadline so that GC is not blocked by
8742 4 : // initial transition into AttachedSingle.
8743 4 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
8744 4 : tokio::time::resume();
8745 4 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8746 4 :
8747 4 : let end_lsn = Lsn(0x100);
8748 4 : let image_layers = (0x20..=0x90)
8749 4 : .step_by(0x10)
8750 32 : .map(|n| {
8751 32 : (
8752 32 : Lsn(n),
8753 32 : vec![(key, test_img(&format!("data key at {:x}", n)))],
8754 32 : )
8755 32 : })
8756 4 : .collect();
8757 4 :
8758 4 : let timeline = tenant
8759 4 : .create_test_timeline_with_layers(
8760 4 : TIMELINE_ID,
8761 4 : Lsn(0x10),
8762 4 : DEFAULT_PG_VERSION,
8763 4 : &ctx,
8764 4 : Vec::new(),
8765 4 : image_layers,
8766 4 : end_lsn,
8767 4 : )
8768 4 : .await?;
8769 4 :
8770 4 : let leased_lsns = [0x30, 0x50, 0x70];
8771 4 : let mut leases = Vec::new();
8772 12 : leased_lsns.iter().for_each(|n| {
8773 12 : leases.push(
8774 12 : timeline
8775 12 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
8776 12 : .expect("lease request should succeed"),
8777 12 : );
8778 12 : });
8779 4 :
8780 4 : let updated_lease_0 = timeline
8781 4 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
8782 4 : .expect("lease renewal should succeed");
8783 4 : assert_eq!(
8784 4 : updated_lease_0.valid_until, leases[0].valid_until,
8785 4 : " Renewing with shorter lease should not change the lease."
8786 4 : );
8787 4 :
8788 4 : let updated_lease_1 = timeline
8789 4 : .renew_lsn_lease(
8790 4 : Lsn(leased_lsns[1]),
8791 4 : timeline.get_lsn_lease_length() * 2,
8792 4 : &ctx,
8793 4 : )
8794 4 : .expect("lease renewal should succeed");
8795 4 : assert!(
8796 4 : updated_lease_1.valid_until > leases[1].valid_until,
8797 4 : "Renewing with a long lease should renew lease with later expiration time."
8798 4 : );
8799 4 :
8800 4 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
8801 4 : info!(
8802 4 : "applied_gc_cutoff_lsn: {}",
8803 0 : *timeline.get_applied_gc_cutoff_lsn()
8804 4 : );
8805 4 : timeline.force_set_disk_consistent_lsn(end_lsn);
8806 4 :
8807 4 : let res = tenant
8808 4 : .gc_iteration(
8809 4 : Some(TIMELINE_ID),
8810 4 : 0,
8811 4 : Duration::ZERO,
8812 4 : &CancellationToken::new(),
8813 4 : &ctx,
8814 4 : )
8815 4 : .await
8816 4 : .unwrap();
8817 4 :
8818 4 : // Keeping everything <= Lsn(0x80) b/c leases:
8819 4 : // 0/10: initdb layer
8820 4 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
8821 4 : assert_eq!(res.layers_needed_by_leases, 7);
8822 4 : // Keeping 0/90 b/c it is the latest layer.
8823 4 : assert_eq!(res.layers_not_updated, 1);
8824 4 : // Removed 0/80.
8825 4 : assert_eq!(res.layers_removed, 1);
8826 4 :
8827 4 : // Make lease on a already GC-ed LSN.
8828 4 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8829 4 : assert!(Lsn(0x80) < *timeline.get_applied_gc_cutoff_lsn());
8830 4 : timeline
8831 4 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8832 4 : .expect_err("lease request on GC-ed LSN should fail");
8833 4 :
8834 4 : // Should still be able to renew a currently valid lease
8835 4 : // Assumption: original lease to is still valid for 0/50.
8836 4 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8837 4 : timeline
8838 4 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8839 4 : .expect("lease renewal with validation should succeed");
8840 4 :
8841 4 : Ok(())
8842 4 : }
8843 :
8844 : #[cfg(feature = "testing")]
8845 : #[tokio::test]
8846 4 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8847 4 : test_simple_bottom_most_compaction_deltas_helper(
8848 4 : "test_simple_bottom_most_compaction_deltas_1",
8849 4 : false,
8850 4 : )
8851 4 : .await
8852 4 : }
8853 :
8854 : #[cfg(feature = "testing")]
8855 : #[tokio::test]
8856 4 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8857 4 : test_simple_bottom_most_compaction_deltas_helper(
8858 4 : "test_simple_bottom_most_compaction_deltas_2",
8859 4 : true,
8860 4 : )
8861 4 : .await
8862 4 : }
8863 :
8864 : #[cfg(feature = "testing")]
8865 8 : async fn test_simple_bottom_most_compaction_deltas_helper(
8866 8 : test_name: &'static str,
8867 8 : use_delta_bottom_layer: bool,
8868 8 : ) -> anyhow::Result<()> {
8869 8 : let harness = TenantHarness::create(test_name).await?;
8870 8 : let (tenant, ctx) = harness.load().await;
8871 :
8872 552 : fn get_key(id: u32) -> Key {
8873 552 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8874 552 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8875 552 : key.field6 = id;
8876 552 : key
8877 552 : }
8878 :
8879 : // We create
8880 : // - one bottom-most image layer,
8881 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8882 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8883 : // - a delta layer D3 above the horizon.
8884 : //
8885 : // | D3 |
8886 : // | D1 |
8887 : // -| |-- gc horizon -----------------
8888 : // | | | D2 |
8889 : // --------- img layer ------------------
8890 : //
8891 : // What we should expact from this compaction is:
8892 : // | D3 |
8893 : // | Part of D1 |
8894 : // --------- img layer with D1+D2 at GC horizon------------------
8895 :
8896 : // img layer at 0x10
8897 8 : let img_layer = (0..10)
8898 80 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8899 8 : .collect_vec();
8900 8 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8901 8 : let delta4 = (0..10)
8902 80 : .map(|id| {
8903 80 : (
8904 80 : get_key(id),
8905 80 : Lsn(0x08),
8906 80 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8907 80 : )
8908 80 : })
8909 8 : .collect_vec();
8910 8 :
8911 8 : let delta1 = vec![
8912 8 : (
8913 8 : get_key(1),
8914 8 : Lsn(0x20),
8915 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8916 8 : ),
8917 8 : (
8918 8 : get_key(2),
8919 8 : Lsn(0x30),
8920 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8921 8 : ),
8922 8 : (
8923 8 : get_key(3),
8924 8 : Lsn(0x28),
8925 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8926 8 : ),
8927 8 : (
8928 8 : get_key(3),
8929 8 : Lsn(0x30),
8930 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8931 8 : ),
8932 8 : (
8933 8 : get_key(3),
8934 8 : Lsn(0x40),
8935 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
8936 8 : ),
8937 8 : ];
8938 8 : let delta2 = vec![
8939 8 : (
8940 8 : get_key(5),
8941 8 : Lsn(0x20),
8942 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8943 8 : ),
8944 8 : (
8945 8 : get_key(6),
8946 8 : Lsn(0x20),
8947 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8948 8 : ),
8949 8 : ];
8950 8 : let delta3 = vec![
8951 8 : (
8952 8 : get_key(8),
8953 8 : Lsn(0x48),
8954 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8955 8 : ),
8956 8 : (
8957 8 : get_key(9),
8958 8 : Lsn(0x48),
8959 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
8960 8 : ),
8961 8 : ];
8962 :
8963 8 : let tline = if use_delta_bottom_layer {
8964 4 : tenant
8965 4 : .create_test_timeline_with_layers(
8966 4 : TIMELINE_ID,
8967 4 : Lsn(0x08),
8968 4 : DEFAULT_PG_VERSION,
8969 4 : &ctx,
8970 4 : vec![
8971 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8972 4 : Lsn(0x08)..Lsn(0x10),
8973 4 : delta4,
8974 4 : ),
8975 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8976 4 : Lsn(0x20)..Lsn(0x48),
8977 4 : delta1,
8978 4 : ),
8979 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8980 4 : Lsn(0x20)..Lsn(0x48),
8981 4 : delta2,
8982 4 : ),
8983 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8984 4 : Lsn(0x48)..Lsn(0x50),
8985 4 : delta3,
8986 4 : ),
8987 4 : ], // delta layers
8988 4 : vec![], // image layers
8989 4 : Lsn(0x50),
8990 4 : )
8991 4 : .await?
8992 : } else {
8993 4 : tenant
8994 4 : .create_test_timeline_with_layers(
8995 4 : TIMELINE_ID,
8996 4 : Lsn(0x10),
8997 4 : DEFAULT_PG_VERSION,
8998 4 : &ctx,
8999 4 : vec![
9000 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
9001 4 : Lsn(0x10)..Lsn(0x48),
9002 4 : delta1,
9003 4 : ),
9004 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
9005 4 : Lsn(0x10)..Lsn(0x48),
9006 4 : delta2,
9007 4 : ),
9008 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
9009 4 : Lsn(0x48)..Lsn(0x50),
9010 4 : delta3,
9011 4 : ),
9012 4 : ], // delta layers
9013 4 : vec![(Lsn(0x10), img_layer)], // image layers
9014 4 : Lsn(0x50),
9015 4 : )
9016 4 : .await?
9017 : };
9018 : {
9019 8 : tline
9020 8 : .applied_gc_cutoff_lsn
9021 8 : .lock_for_write()
9022 8 : .store_and_unlock(Lsn(0x30))
9023 8 : .wait()
9024 8 : .await;
9025 : // Update GC info
9026 8 : let mut guard = tline.gc_info.write().unwrap();
9027 8 : *guard = GcInfo {
9028 8 : retain_lsns: vec![],
9029 8 : cutoffs: GcCutoffs {
9030 8 : time: Lsn(0x30),
9031 8 : space: Lsn(0x30),
9032 8 : },
9033 8 : leases: Default::default(),
9034 8 : within_ancestor_pitr: false,
9035 8 : };
9036 8 : }
9037 8 :
9038 8 : let expected_result = [
9039 8 : Bytes::from_static(b"value 0@0x10"),
9040 8 : Bytes::from_static(b"value 1@0x10@0x20"),
9041 8 : Bytes::from_static(b"value 2@0x10@0x30"),
9042 8 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9043 8 : Bytes::from_static(b"value 4@0x10"),
9044 8 : Bytes::from_static(b"value 5@0x10@0x20"),
9045 8 : Bytes::from_static(b"value 6@0x10@0x20"),
9046 8 : Bytes::from_static(b"value 7@0x10"),
9047 8 : Bytes::from_static(b"value 8@0x10@0x48"),
9048 8 : Bytes::from_static(b"value 9@0x10@0x48"),
9049 8 : ];
9050 8 :
9051 8 : let expected_result_at_gc_horizon = [
9052 8 : Bytes::from_static(b"value 0@0x10"),
9053 8 : Bytes::from_static(b"value 1@0x10@0x20"),
9054 8 : Bytes::from_static(b"value 2@0x10@0x30"),
9055 8 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9056 8 : Bytes::from_static(b"value 4@0x10"),
9057 8 : Bytes::from_static(b"value 5@0x10@0x20"),
9058 8 : Bytes::from_static(b"value 6@0x10@0x20"),
9059 8 : Bytes::from_static(b"value 7@0x10"),
9060 8 : Bytes::from_static(b"value 8@0x10"),
9061 8 : Bytes::from_static(b"value 9@0x10"),
9062 8 : ];
9063 :
9064 88 : for idx in 0..10 {
9065 80 : assert_eq!(
9066 80 : tline
9067 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9068 80 : .await
9069 80 : .unwrap(),
9070 80 : &expected_result[idx]
9071 : );
9072 80 : assert_eq!(
9073 80 : tline
9074 80 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9075 80 : .await
9076 80 : .unwrap(),
9077 80 : &expected_result_at_gc_horizon[idx]
9078 : );
9079 : }
9080 :
9081 8 : let cancel = CancellationToken::new();
9082 8 : tline
9083 8 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9084 8 : .await
9085 8 : .unwrap();
9086 :
9087 88 : for idx in 0..10 {
9088 80 : assert_eq!(
9089 80 : tline
9090 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9091 80 : .await
9092 80 : .unwrap(),
9093 80 : &expected_result[idx]
9094 : );
9095 80 : assert_eq!(
9096 80 : tline
9097 80 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9098 80 : .await
9099 80 : .unwrap(),
9100 80 : &expected_result_at_gc_horizon[idx]
9101 : );
9102 : }
9103 :
9104 : // increase GC horizon and compact again
9105 : {
9106 8 : tline
9107 8 : .applied_gc_cutoff_lsn
9108 8 : .lock_for_write()
9109 8 : .store_and_unlock(Lsn(0x40))
9110 8 : .wait()
9111 8 : .await;
9112 : // Update GC info
9113 8 : let mut guard = tline.gc_info.write().unwrap();
9114 8 : guard.cutoffs.time = Lsn(0x40);
9115 8 : guard.cutoffs.space = Lsn(0x40);
9116 8 : }
9117 8 : tline
9118 8 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9119 8 : .await
9120 8 : .unwrap();
9121 8 :
9122 8 : Ok(())
9123 8 : }
9124 :
9125 : #[cfg(feature = "testing")]
9126 : #[tokio::test]
9127 4 : async fn test_generate_key_retention() -> anyhow::Result<()> {
9128 4 : let harness = TenantHarness::create("test_generate_key_retention").await?;
9129 4 : let (tenant, ctx) = harness.load().await;
9130 4 : let tline = tenant
9131 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9132 4 : .await?;
9133 4 : tline.force_advance_lsn(Lsn(0x70));
9134 4 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9135 4 : let history = vec![
9136 4 : (
9137 4 : key,
9138 4 : Lsn(0x10),
9139 4 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
9140 4 : ),
9141 4 : (
9142 4 : key,
9143 4 : Lsn(0x20),
9144 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9145 4 : ),
9146 4 : (
9147 4 : key,
9148 4 : Lsn(0x30),
9149 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9150 4 : ),
9151 4 : (
9152 4 : key,
9153 4 : Lsn(0x40),
9154 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9155 4 : ),
9156 4 : (
9157 4 : key,
9158 4 : Lsn(0x50),
9159 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9160 4 : ),
9161 4 : (
9162 4 : key,
9163 4 : Lsn(0x60),
9164 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9165 4 : ),
9166 4 : (
9167 4 : key,
9168 4 : Lsn(0x70),
9169 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9170 4 : ),
9171 4 : (
9172 4 : key,
9173 4 : Lsn(0x80),
9174 4 : Value::Image(Bytes::copy_from_slice(
9175 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9176 4 : )),
9177 4 : ),
9178 4 : (
9179 4 : key,
9180 4 : Lsn(0x90),
9181 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9182 4 : ),
9183 4 : ];
9184 4 : let res = tline
9185 4 : .generate_key_retention(
9186 4 : key,
9187 4 : &history,
9188 4 : Lsn(0x60),
9189 4 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9190 4 : 3,
9191 4 : None,
9192 4 : )
9193 4 : .await
9194 4 : .unwrap();
9195 4 : let expected_res = KeyHistoryRetention {
9196 4 : below_horizon: vec![
9197 4 : (
9198 4 : Lsn(0x20),
9199 4 : KeyLogAtLsn(vec![(
9200 4 : Lsn(0x20),
9201 4 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9202 4 : )]),
9203 4 : ),
9204 4 : (
9205 4 : Lsn(0x40),
9206 4 : KeyLogAtLsn(vec![
9207 4 : (
9208 4 : Lsn(0x30),
9209 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9210 4 : ),
9211 4 : (
9212 4 : Lsn(0x40),
9213 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9214 4 : ),
9215 4 : ]),
9216 4 : ),
9217 4 : (
9218 4 : Lsn(0x50),
9219 4 : KeyLogAtLsn(vec![(
9220 4 : Lsn(0x50),
9221 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9222 4 : )]),
9223 4 : ),
9224 4 : (
9225 4 : Lsn(0x60),
9226 4 : KeyLogAtLsn(vec![(
9227 4 : Lsn(0x60),
9228 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9229 4 : )]),
9230 4 : ),
9231 4 : ],
9232 4 : above_horizon: KeyLogAtLsn(vec![
9233 4 : (
9234 4 : Lsn(0x70),
9235 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9236 4 : ),
9237 4 : (
9238 4 : Lsn(0x80),
9239 4 : Value::Image(Bytes::copy_from_slice(
9240 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9241 4 : )),
9242 4 : ),
9243 4 : (
9244 4 : Lsn(0x90),
9245 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9246 4 : ),
9247 4 : ]),
9248 4 : };
9249 4 : assert_eq!(res, expected_res);
9250 4 :
9251 4 : // We expect GC-compaction to run with the original GC. This would create a situation that
9252 4 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9253 4 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9254 4 : // For example, we have
9255 4 : // ```plain
9256 4 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9257 4 : // ```
9258 4 : // Now the GC horizon moves up, and we have
9259 4 : // ```plain
9260 4 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9261 4 : // ```
9262 4 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9263 4 : // We will end up with
9264 4 : // ```plain
9265 4 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9266 4 : // ```
9267 4 : // Now we run the GC-compaction, and this key does not have a full history.
9268 4 : // We should be able to handle this partial history and drop everything before the
9269 4 : // gc_horizon image.
9270 4 :
9271 4 : let history = vec![
9272 4 : (
9273 4 : key,
9274 4 : Lsn(0x20),
9275 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9276 4 : ),
9277 4 : (
9278 4 : key,
9279 4 : Lsn(0x30),
9280 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9281 4 : ),
9282 4 : (
9283 4 : key,
9284 4 : Lsn(0x40),
9285 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9286 4 : ),
9287 4 : (
9288 4 : key,
9289 4 : Lsn(0x50),
9290 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9291 4 : ),
9292 4 : (
9293 4 : key,
9294 4 : Lsn(0x60),
9295 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9296 4 : ),
9297 4 : (
9298 4 : key,
9299 4 : Lsn(0x70),
9300 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9301 4 : ),
9302 4 : (
9303 4 : key,
9304 4 : Lsn(0x80),
9305 4 : Value::Image(Bytes::copy_from_slice(
9306 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9307 4 : )),
9308 4 : ),
9309 4 : (
9310 4 : key,
9311 4 : Lsn(0x90),
9312 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9313 4 : ),
9314 4 : ];
9315 4 : let res = tline
9316 4 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
9317 4 : .await
9318 4 : .unwrap();
9319 4 : let expected_res = KeyHistoryRetention {
9320 4 : below_horizon: vec![
9321 4 : (
9322 4 : Lsn(0x40),
9323 4 : KeyLogAtLsn(vec![(
9324 4 : Lsn(0x40),
9325 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9326 4 : )]),
9327 4 : ),
9328 4 : (
9329 4 : Lsn(0x50),
9330 4 : KeyLogAtLsn(vec![(
9331 4 : Lsn(0x50),
9332 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9333 4 : )]),
9334 4 : ),
9335 4 : (
9336 4 : Lsn(0x60),
9337 4 : KeyLogAtLsn(vec![(
9338 4 : Lsn(0x60),
9339 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9340 4 : )]),
9341 4 : ),
9342 4 : ],
9343 4 : above_horizon: KeyLogAtLsn(vec![
9344 4 : (
9345 4 : Lsn(0x70),
9346 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9347 4 : ),
9348 4 : (
9349 4 : Lsn(0x80),
9350 4 : Value::Image(Bytes::copy_from_slice(
9351 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9352 4 : )),
9353 4 : ),
9354 4 : (
9355 4 : Lsn(0x90),
9356 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9357 4 : ),
9358 4 : ]),
9359 4 : };
9360 4 : assert_eq!(res, expected_res);
9361 4 :
9362 4 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9363 4 : // the ancestor image in the test case.
9364 4 :
9365 4 : let history = vec![
9366 4 : (
9367 4 : key,
9368 4 : Lsn(0x20),
9369 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9370 4 : ),
9371 4 : (
9372 4 : key,
9373 4 : Lsn(0x30),
9374 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9375 4 : ),
9376 4 : (
9377 4 : key,
9378 4 : Lsn(0x40),
9379 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9380 4 : ),
9381 4 : (
9382 4 : key,
9383 4 : Lsn(0x70),
9384 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9385 4 : ),
9386 4 : ];
9387 4 : let res = tline
9388 4 : .generate_key_retention(
9389 4 : key,
9390 4 : &history,
9391 4 : Lsn(0x60),
9392 4 : &[],
9393 4 : 3,
9394 4 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9395 4 : )
9396 4 : .await
9397 4 : .unwrap();
9398 4 : let expected_res = KeyHistoryRetention {
9399 4 : below_horizon: vec![(
9400 4 : Lsn(0x60),
9401 4 : KeyLogAtLsn(vec![(
9402 4 : Lsn(0x60),
9403 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9404 4 : )]),
9405 4 : )],
9406 4 : above_horizon: KeyLogAtLsn(vec![(
9407 4 : Lsn(0x70),
9408 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9409 4 : )]),
9410 4 : };
9411 4 : assert_eq!(res, expected_res);
9412 4 :
9413 4 : let history = vec![
9414 4 : (
9415 4 : key,
9416 4 : Lsn(0x20),
9417 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9418 4 : ),
9419 4 : (
9420 4 : key,
9421 4 : Lsn(0x40),
9422 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9423 4 : ),
9424 4 : (
9425 4 : key,
9426 4 : Lsn(0x60),
9427 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9428 4 : ),
9429 4 : (
9430 4 : key,
9431 4 : Lsn(0x70),
9432 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9433 4 : ),
9434 4 : ];
9435 4 : let res = tline
9436 4 : .generate_key_retention(
9437 4 : key,
9438 4 : &history,
9439 4 : Lsn(0x60),
9440 4 : &[Lsn(0x30)],
9441 4 : 3,
9442 4 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9443 4 : )
9444 4 : .await
9445 4 : .unwrap();
9446 4 : let expected_res = KeyHistoryRetention {
9447 4 : below_horizon: vec![
9448 4 : (
9449 4 : Lsn(0x30),
9450 4 : KeyLogAtLsn(vec![(
9451 4 : Lsn(0x20),
9452 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9453 4 : )]),
9454 4 : ),
9455 4 : (
9456 4 : Lsn(0x60),
9457 4 : KeyLogAtLsn(vec![(
9458 4 : Lsn(0x60),
9459 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9460 4 : )]),
9461 4 : ),
9462 4 : ],
9463 4 : above_horizon: KeyLogAtLsn(vec![(
9464 4 : Lsn(0x70),
9465 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9466 4 : )]),
9467 4 : };
9468 4 : assert_eq!(res, expected_res);
9469 4 :
9470 4 : Ok(())
9471 4 : }
9472 :
9473 : #[cfg(feature = "testing")]
9474 : #[tokio::test]
9475 4 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9476 4 : let harness =
9477 4 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9478 4 : let (tenant, ctx) = harness.load().await;
9479 4 :
9480 1036 : fn get_key(id: u32) -> Key {
9481 1036 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9482 1036 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9483 1036 : key.field6 = id;
9484 1036 : key
9485 1036 : }
9486 4 :
9487 4 : let img_layer = (0..10)
9488 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9489 4 : .collect_vec();
9490 4 :
9491 4 : let delta1 = vec![
9492 4 : (
9493 4 : get_key(1),
9494 4 : Lsn(0x20),
9495 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9496 4 : ),
9497 4 : (
9498 4 : get_key(2),
9499 4 : Lsn(0x30),
9500 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9501 4 : ),
9502 4 : (
9503 4 : get_key(3),
9504 4 : Lsn(0x28),
9505 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9506 4 : ),
9507 4 : (
9508 4 : get_key(3),
9509 4 : Lsn(0x30),
9510 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9511 4 : ),
9512 4 : (
9513 4 : get_key(3),
9514 4 : Lsn(0x40),
9515 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9516 4 : ),
9517 4 : ];
9518 4 : let delta2 = vec![
9519 4 : (
9520 4 : get_key(5),
9521 4 : Lsn(0x20),
9522 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9523 4 : ),
9524 4 : (
9525 4 : get_key(6),
9526 4 : Lsn(0x20),
9527 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9528 4 : ),
9529 4 : ];
9530 4 : let delta3 = vec![
9531 4 : (
9532 4 : get_key(8),
9533 4 : Lsn(0x48),
9534 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9535 4 : ),
9536 4 : (
9537 4 : get_key(9),
9538 4 : Lsn(0x48),
9539 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9540 4 : ),
9541 4 : ];
9542 4 :
9543 4 : let tline = tenant
9544 4 : .create_test_timeline_with_layers(
9545 4 : TIMELINE_ID,
9546 4 : Lsn(0x10),
9547 4 : DEFAULT_PG_VERSION,
9548 4 : &ctx,
9549 4 : vec![
9550 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
9551 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
9552 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9553 4 : ], // delta layers
9554 4 : vec![(Lsn(0x10), img_layer)], // image layers
9555 4 : Lsn(0x50),
9556 4 : )
9557 4 : .await?;
9558 4 : {
9559 4 : tline
9560 4 : .applied_gc_cutoff_lsn
9561 4 : .lock_for_write()
9562 4 : .store_and_unlock(Lsn(0x30))
9563 4 : .wait()
9564 4 : .await;
9565 4 : // Update GC info
9566 4 : let mut guard = tline.gc_info.write().unwrap();
9567 4 : *guard = GcInfo {
9568 4 : retain_lsns: vec![
9569 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9570 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9571 4 : ],
9572 4 : cutoffs: GcCutoffs {
9573 4 : time: Lsn(0x30),
9574 4 : space: Lsn(0x30),
9575 4 : },
9576 4 : leases: Default::default(),
9577 4 : within_ancestor_pitr: false,
9578 4 : };
9579 4 : }
9580 4 :
9581 4 : let expected_result = [
9582 4 : Bytes::from_static(b"value 0@0x10"),
9583 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9584 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9585 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9586 4 : Bytes::from_static(b"value 4@0x10"),
9587 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9588 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9589 4 : Bytes::from_static(b"value 7@0x10"),
9590 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9591 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9592 4 : ];
9593 4 :
9594 4 : let expected_result_at_gc_horizon = [
9595 4 : Bytes::from_static(b"value 0@0x10"),
9596 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9597 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9598 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9599 4 : Bytes::from_static(b"value 4@0x10"),
9600 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9601 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9602 4 : Bytes::from_static(b"value 7@0x10"),
9603 4 : Bytes::from_static(b"value 8@0x10"),
9604 4 : Bytes::from_static(b"value 9@0x10"),
9605 4 : ];
9606 4 :
9607 4 : let expected_result_at_lsn_20 = [
9608 4 : Bytes::from_static(b"value 0@0x10"),
9609 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9610 4 : Bytes::from_static(b"value 2@0x10"),
9611 4 : Bytes::from_static(b"value 3@0x10"),
9612 4 : Bytes::from_static(b"value 4@0x10"),
9613 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9614 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9615 4 : Bytes::from_static(b"value 7@0x10"),
9616 4 : Bytes::from_static(b"value 8@0x10"),
9617 4 : Bytes::from_static(b"value 9@0x10"),
9618 4 : ];
9619 4 :
9620 4 : let expected_result_at_lsn_10 = [
9621 4 : Bytes::from_static(b"value 0@0x10"),
9622 4 : Bytes::from_static(b"value 1@0x10"),
9623 4 : Bytes::from_static(b"value 2@0x10"),
9624 4 : Bytes::from_static(b"value 3@0x10"),
9625 4 : Bytes::from_static(b"value 4@0x10"),
9626 4 : Bytes::from_static(b"value 5@0x10"),
9627 4 : Bytes::from_static(b"value 6@0x10"),
9628 4 : Bytes::from_static(b"value 7@0x10"),
9629 4 : Bytes::from_static(b"value 8@0x10"),
9630 4 : Bytes::from_static(b"value 9@0x10"),
9631 4 : ];
9632 4 :
9633 24 : let verify_result = || async {
9634 24 : let gc_horizon = {
9635 24 : let gc_info = tline.gc_info.read().unwrap();
9636 24 : gc_info.cutoffs.time
9637 4 : };
9638 264 : for idx in 0..10 {
9639 240 : assert_eq!(
9640 240 : tline
9641 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9642 240 : .await
9643 240 : .unwrap(),
9644 240 : &expected_result[idx]
9645 4 : );
9646 240 : assert_eq!(
9647 240 : tline
9648 240 : .get(get_key(idx as u32), gc_horizon, &ctx)
9649 240 : .await
9650 240 : .unwrap(),
9651 240 : &expected_result_at_gc_horizon[idx]
9652 4 : );
9653 240 : assert_eq!(
9654 240 : tline
9655 240 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9656 240 : .await
9657 240 : .unwrap(),
9658 240 : &expected_result_at_lsn_20[idx]
9659 4 : );
9660 240 : assert_eq!(
9661 240 : tline
9662 240 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9663 240 : .await
9664 240 : .unwrap(),
9665 240 : &expected_result_at_lsn_10[idx]
9666 4 : );
9667 4 : }
9668 48 : };
9669 4 :
9670 4 : verify_result().await;
9671 4 :
9672 4 : let cancel = CancellationToken::new();
9673 4 : let mut dryrun_flags = EnumSet::new();
9674 4 : dryrun_flags.insert(CompactFlags::DryRun);
9675 4 :
9676 4 : tline
9677 4 : .compact_with_gc(
9678 4 : &cancel,
9679 4 : CompactOptions {
9680 4 : flags: dryrun_flags,
9681 4 : ..Default::default()
9682 4 : },
9683 4 : &ctx,
9684 4 : )
9685 4 : .await
9686 4 : .unwrap();
9687 4 : // 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
9688 4 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9689 4 : verify_result().await;
9690 4 :
9691 4 : tline
9692 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9693 4 : .await
9694 4 : .unwrap();
9695 4 : verify_result().await;
9696 4 :
9697 4 : // compact again
9698 4 : tline
9699 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9700 4 : .await
9701 4 : .unwrap();
9702 4 : verify_result().await;
9703 4 :
9704 4 : // increase GC horizon and compact again
9705 4 : {
9706 4 : tline
9707 4 : .applied_gc_cutoff_lsn
9708 4 : .lock_for_write()
9709 4 : .store_and_unlock(Lsn(0x38))
9710 4 : .wait()
9711 4 : .await;
9712 4 : // Update GC info
9713 4 : let mut guard = tline.gc_info.write().unwrap();
9714 4 : guard.cutoffs.time = Lsn(0x38);
9715 4 : guard.cutoffs.space = Lsn(0x38);
9716 4 : }
9717 4 : tline
9718 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9719 4 : .await
9720 4 : .unwrap();
9721 4 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
9722 4 :
9723 4 : // not increasing the GC horizon and compact again
9724 4 : tline
9725 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9726 4 : .await
9727 4 : .unwrap();
9728 4 : verify_result().await;
9729 4 :
9730 4 : Ok(())
9731 4 : }
9732 :
9733 : #[cfg(feature = "testing")]
9734 : #[tokio::test]
9735 4 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
9736 4 : {
9737 4 : let harness =
9738 4 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
9739 4 : .await?;
9740 4 : let (tenant, ctx) = harness.load().await;
9741 4 :
9742 704 : fn get_key(id: u32) -> Key {
9743 704 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9744 704 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9745 704 : key.field6 = id;
9746 704 : key
9747 704 : }
9748 4 :
9749 4 : let img_layer = (0..10)
9750 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9751 4 : .collect_vec();
9752 4 :
9753 4 : let delta1 = vec![
9754 4 : (
9755 4 : get_key(1),
9756 4 : Lsn(0x20),
9757 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9758 4 : ),
9759 4 : (
9760 4 : get_key(1),
9761 4 : Lsn(0x28),
9762 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9763 4 : ),
9764 4 : ];
9765 4 : let delta2 = vec![
9766 4 : (
9767 4 : get_key(1),
9768 4 : Lsn(0x30),
9769 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9770 4 : ),
9771 4 : (
9772 4 : get_key(1),
9773 4 : Lsn(0x38),
9774 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
9775 4 : ),
9776 4 : ];
9777 4 : let delta3 = vec![
9778 4 : (
9779 4 : get_key(8),
9780 4 : Lsn(0x48),
9781 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9782 4 : ),
9783 4 : (
9784 4 : get_key(9),
9785 4 : Lsn(0x48),
9786 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9787 4 : ),
9788 4 : ];
9789 4 :
9790 4 : let tline = tenant
9791 4 : .create_test_timeline_with_layers(
9792 4 : TIMELINE_ID,
9793 4 : Lsn(0x10),
9794 4 : DEFAULT_PG_VERSION,
9795 4 : &ctx,
9796 4 : vec![
9797 4 : // delta1 and delta 2 only contain a single key but multiple updates
9798 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
9799 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
9800 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
9801 4 : ], // delta layers
9802 4 : vec![(Lsn(0x10), img_layer)], // image layers
9803 4 : Lsn(0x50),
9804 4 : )
9805 4 : .await?;
9806 4 : {
9807 4 : tline
9808 4 : .applied_gc_cutoff_lsn
9809 4 : .lock_for_write()
9810 4 : .store_and_unlock(Lsn(0x30))
9811 4 : .wait()
9812 4 : .await;
9813 4 : // Update GC info
9814 4 : let mut guard = tline.gc_info.write().unwrap();
9815 4 : *guard = GcInfo {
9816 4 : retain_lsns: vec![
9817 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9818 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9819 4 : ],
9820 4 : cutoffs: GcCutoffs {
9821 4 : time: Lsn(0x30),
9822 4 : space: Lsn(0x30),
9823 4 : },
9824 4 : leases: Default::default(),
9825 4 : within_ancestor_pitr: false,
9826 4 : };
9827 4 : }
9828 4 :
9829 4 : let expected_result = [
9830 4 : Bytes::from_static(b"value 0@0x10"),
9831 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
9832 4 : Bytes::from_static(b"value 2@0x10"),
9833 4 : Bytes::from_static(b"value 3@0x10"),
9834 4 : Bytes::from_static(b"value 4@0x10"),
9835 4 : Bytes::from_static(b"value 5@0x10"),
9836 4 : Bytes::from_static(b"value 6@0x10"),
9837 4 : Bytes::from_static(b"value 7@0x10"),
9838 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9839 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9840 4 : ];
9841 4 :
9842 4 : let expected_result_at_gc_horizon = [
9843 4 : Bytes::from_static(b"value 0@0x10"),
9844 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
9845 4 : Bytes::from_static(b"value 2@0x10"),
9846 4 : Bytes::from_static(b"value 3@0x10"),
9847 4 : Bytes::from_static(b"value 4@0x10"),
9848 4 : Bytes::from_static(b"value 5@0x10"),
9849 4 : Bytes::from_static(b"value 6@0x10"),
9850 4 : Bytes::from_static(b"value 7@0x10"),
9851 4 : Bytes::from_static(b"value 8@0x10"),
9852 4 : Bytes::from_static(b"value 9@0x10"),
9853 4 : ];
9854 4 :
9855 4 : let expected_result_at_lsn_20 = [
9856 4 : Bytes::from_static(b"value 0@0x10"),
9857 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9858 4 : Bytes::from_static(b"value 2@0x10"),
9859 4 : Bytes::from_static(b"value 3@0x10"),
9860 4 : Bytes::from_static(b"value 4@0x10"),
9861 4 : Bytes::from_static(b"value 5@0x10"),
9862 4 : Bytes::from_static(b"value 6@0x10"),
9863 4 : Bytes::from_static(b"value 7@0x10"),
9864 4 : Bytes::from_static(b"value 8@0x10"),
9865 4 : Bytes::from_static(b"value 9@0x10"),
9866 4 : ];
9867 4 :
9868 4 : let expected_result_at_lsn_10 = [
9869 4 : Bytes::from_static(b"value 0@0x10"),
9870 4 : Bytes::from_static(b"value 1@0x10"),
9871 4 : Bytes::from_static(b"value 2@0x10"),
9872 4 : Bytes::from_static(b"value 3@0x10"),
9873 4 : Bytes::from_static(b"value 4@0x10"),
9874 4 : Bytes::from_static(b"value 5@0x10"),
9875 4 : Bytes::from_static(b"value 6@0x10"),
9876 4 : Bytes::from_static(b"value 7@0x10"),
9877 4 : Bytes::from_static(b"value 8@0x10"),
9878 4 : Bytes::from_static(b"value 9@0x10"),
9879 4 : ];
9880 4 :
9881 16 : let verify_result = || async {
9882 16 : let gc_horizon = {
9883 16 : let gc_info = tline.gc_info.read().unwrap();
9884 16 : gc_info.cutoffs.time
9885 4 : };
9886 176 : for idx in 0..10 {
9887 160 : assert_eq!(
9888 160 : tline
9889 160 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9890 160 : .await
9891 160 : .unwrap(),
9892 160 : &expected_result[idx]
9893 4 : );
9894 160 : assert_eq!(
9895 160 : tline
9896 160 : .get(get_key(idx as u32), gc_horizon, &ctx)
9897 160 : .await
9898 160 : .unwrap(),
9899 160 : &expected_result_at_gc_horizon[idx]
9900 4 : );
9901 160 : assert_eq!(
9902 160 : tline
9903 160 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9904 160 : .await
9905 160 : .unwrap(),
9906 160 : &expected_result_at_lsn_20[idx]
9907 4 : );
9908 160 : assert_eq!(
9909 160 : tline
9910 160 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9911 160 : .await
9912 160 : .unwrap(),
9913 160 : &expected_result_at_lsn_10[idx]
9914 4 : );
9915 4 : }
9916 32 : };
9917 4 :
9918 4 : verify_result().await;
9919 4 :
9920 4 : let cancel = CancellationToken::new();
9921 4 : let mut dryrun_flags = EnumSet::new();
9922 4 : dryrun_flags.insert(CompactFlags::DryRun);
9923 4 :
9924 4 : tline
9925 4 : .compact_with_gc(
9926 4 : &cancel,
9927 4 : CompactOptions {
9928 4 : flags: dryrun_flags,
9929 4 : ..Default::default()
9930 4 : },
9931 4 : &ctx,
9932 4 : )
9933 4 : .await
9934 4 : .unwrap();
9935 4 : // 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
9936 4 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9937 4 : verify_result().await;
9938 4 :
9939 4 : tline
9940 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9941 4 : .await
9942 4 : .unwrap();
9943 4 : verify_result().await;
9944 4 :
9945 4 : // compact again
9946 4 : tline
9947 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9948 4 : .await
9949 4 : .unwrap();
9950 4 : verify_result().await;
9951 4 :
9952 4 : Ok(())
9953 4 : }
9954 :
9955 : #[cfg(feature = "testing")]
9956 : #[tokio::test]
9957 4 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
9958 4 : use models::CompactLsnRange;
9959 4 :
9960 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
9961 4 : let (tenant, ctx) = harness.load().await;
9962 4 :
9963 332 : fn get_key(id: u32) -> Key {
9964 332 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
9965 332 : key.field6 = id;
9966 332 : key
9967 332 : }
9968 4 :
9969 4 : let img_layer = (0..10)
9970 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9971 4 : .collect_vec();
9972 4 :
9973 4 : let delta1 = vec![
9974 4 : (
9975 4 : get_key(1),
9976 4 : Lsn(0x20),
9977 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9978 4 : ),
9979 4 : (
9980 4 : get_key(2),
9981 4 : Lsn(0x30),
9982 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9983 4 : ),
9984 4 : (
9985 4 : get_key(3),
9986 4 : Lsn(0x28),
9987 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9988 4 : ),
9989 4 : (
9990 4 : get_key(3),
9991 4 : Lsn(0x30),
9992 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9993 4 : ),
9994 4 : (
9995 4 : get_key(3),
9996 4 : Lsn(0x40),
9997 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9998 4 : ),
9999 4 : ];
10000 4 : let delta2 = vec![
10001 4 : (
10002 4 : get_key(5),
10003 4 : Lsn(0x20),
10004 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10005 4 : ),
10006 4 : (
10007 4 : get_key(6),
10008 4 : Lsn(0x20),
10009 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10010 4 : ),
10011 4 : ];
10012 4 : let delta3 = vec![
10013 4 : (
10014 4 : get_key(8),
10015 4 : Lsn(0x48),
10016 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10017 4 : ),
10018 4 : (
10019 4 : get_key(9),
10020 4 : Lsn(0x48),
10021 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10022 4 : ),
10023 4 : ];
10024 4 :
10025 4 : let parent_tline = tenant
10026 4 : .create_test_timeline_with_layers(
10027 4 : TIMELINE_ID,
10028 4 : Lsn(0x10),
10029 4 : DEFAULT_PG_VERSION,
10030 4 : &ctx,
10031 4 : vec![], // delta layers
10032 4 : vec![(Lsn(0x18), img_layer)], // image layers
10033 4 : Lsn(0x18),
10034 4 : )
10035 4 : .await?;
10036 4 :
10037 4 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10038 4 :
10039 4 : let branch_tline = tenant
10040 4 : .branch_timeline_test_with_layers(
10041 4 : &parent_tline,
10042 4 : NEW_TIMELINE_ID,
10043 4 : Some(Lsn(0x18)),
10044 4 : &ctx,
10045 4 : vec![
10046 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10047 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10048 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10049 4 : ], // delta layers
10050 4 : vec![], // image layers
10051 4 : Lsn(0x50),
10052 4 : )
10053 4 : .await?;
10054 4 :
10055 4 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10056 4 :
10057 4 : {
10058 4 : parent_tline
10059 4 : .applied_gc_cutoff_lsn
10060 4 : .lock_for_write()
10061 4 : .store_and_unlock(Lsn(0x10))
10062 4 : .wait()
10063 4 : .await;
10064 4 : // Update GC info
10065 4 : let mut guard = parent_tline.gc_info.write().unwrap();
10066 4 : *guard = GcInfo {
10067 4 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
10068 4 : cutoffs: GcCutoffs {
10069 4 : time: Lsn(0x10),
10070 4 : space: Lsn(0x10),
10071 4 : },
10072 4 : leases: Default::default(),
10073 4 : within_ancestor_pitr: false,
10074 4 : };
10075 4 : }
10076 4 :
10077 4 : {
10078 4 : branch_tline
10079 4 : .applied_gc_cutoff_lsn
10080 4 : .lock_for_write()
10081 4 : .store_and_unlock(Lsn(0x50))
10082 4 : .wait()
10083 4 : .await;
10084 4 : // Update GC info
10085 4 : let mut guard = branch_tline.gc_info.write().unwrap();
10086 4 : *guard = GcInfo {
10087 4 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
10088 4 : cutoffs: GcCutoffs {
10089 4 : time: Lsn(0x50),
10090 4 : space: Lsn(0x50),
10091 4 : },
10092 4 : leases: Default::default(),
10093 4 : within_ancestor_pitr: false,
10094 4 : };
10095 4 : }
10096 4 :
10097 4 : let expected_result_at_gc_horizon = [
10098 4 : Bytes::from_static(b"value 0@0x10"),
10099 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10100 4 : Bytes::from_static(b"value 2@0x10@0x30"),
10101 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10102 4 : Bytes::from_static(b"value 4@0x10"),
10103 4 : Bytes::from_static(b"value 5@0x10@0x20"),
10104 4 : Bytes::from_static(b"value 6@0x10@0x20"),
10105 4 : Bytes::from_static(b"value 7@0x10"),
10106 4 : Bytes::from_static(b"value 8@0x10@0x48"),
10107 4 : Bytes::from_static(b"value 9@0x10@0x48"),
10108 4 : ];
10109 4 :
10110 4 : let expected_result_at_lsn_40 = [
10111 4 : Bytes::from_static(b"value 0@0x10"),
10112 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10113 4 : Bytes::from_static(b"value 2@0x10@0x30"),
10114 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10115 4 : Bytes::from_static(b"value 4@0x10"),
10116 4 : Bytes::from_static(b"value 5@0x10@0x20"),
10117 4 : Bytes::from_static(b"value 6@0x10@0x20"),
10118 4 : Bytes::from_static(b"value 7@0x10"),
10119 4 : Bytes::from_static(b"value 8@0x10"),
10120 4 : Bytes::from_static(b"value 9@0x10"),
10121 4 : ];
10122 4 :
10123 12 : let verify_result = || async {
10124 132 : for idx in 0..10 {
10125 120 : assert_eq!(
10126 120 : branch_tline
10127 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10128 120 : .await
10129 120 : .unwrap(),
10130 120 : &expected_result_at_gc_horizon[idx]
10131 4 : );
10132 120 : assert_eq!(
10133 120 : branch_tline
10134 120 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
10135 120 : .await
10136 120 : .unwrap(),
10137 120 : &expected_result_at_lsn_40[idx]
10138 4 : );
10139 4 : }
10140 24 : };
10141 4 :
10142 4 : verify_result().await;
10143 4 :
10144 4 : let cancel = CancellationToken::new();
10145 4 : branch_tline
10146 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10147 4 : .await
10148 4 : .unwrap();
10149 4 :
10150 4 : verify_result().await;
10151 4 :
10152 4 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
10153 4 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
10154 4 : branch_tline
10155 4 : .compact_with_gc(
10156 4 : &cancel,
10157 4 : CompactOptions {
10158 4 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
10159 4 : ..Default::default()
10160 4 : },
10161 4 : &ctx,
10162 4 : )
10163 4 : .await
10164 4 : .unwrap();
10165 4 :
10166 4 : verify_result().await;
10167 4 :
10168 4 : Ok(())
10169 4 : }
10170 :
10171 : // Regression test for https://github.com/neondatabase/neon/issues/9012
10172 : // Create an image arrangement where we have to read at different LSN ranges
10173 : // from a delta layer. This is achieved by overlapping an image layer on top of
10174 : // a delta layer. Like so:
10175 : //
10176 : // A B
10177 : // +----------------+ -> delta_layer
10178 : // | | ^ lsn
10179 : // | =========|-> nested_image_layer |
10180 : // | C | |
10181 : // +----------------+ |
10182 : // ======== -> baseline_image_layer +-------> key
10183 : //
10184 : //
10185 : // When querying the key range [A, B) we need to read at different LSN ranges
10186 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
10187 : #[cfg(feature = "testing")]
10188 : #[tokio::test]
10189 4 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
10190 4 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
10191 4 : let (tenant, ctx) = harness.load().await;
10192 4 :
10193 4 : let will_init_keys = [2, 6];
10194 88 : fn get_key(id: u32) -> Key {
10195 88 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10196 88 : key.field6 = id;
10197 88 : key
10198 88 : }
10199 4 :
10200 4 : let mut expected_key_values = HashMap::new();
10201 4 :
10202 4 : let baseline_image_layer_lsn = Lsn(0x10);
10203 4 : let mut baseline_img_layer = Vec::new();
10204 24 : for i in 0..5 {
10205 20 : let key = get_key(i);
10206 20 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10207 20 :
10208 20 : let removed = expected_key_values.insert(key, value.clone());
10209 20 : assert!(removed.is_none());
10210 4 :
10211 20 : baseline_img_layer.push((key, Bytes::from(value)));
10212 4 : }
10213 4 :
10214 4 : let nested_image_layer_lsn = Lsn(0x50);
10215 4 : let mut nested_img_layer = Vec::new();
10216 24 : for i in 5..10 {
10217 20 : let key = get_key(i);
10218 20 : let value = format!("value {i}@{nested_image_layer_lsn}");
10219 20 :
10220 20 : let removed = expected_key_values.insert(key, value.clone());
10221 20 : assert!(removed.is_none());
10222 4 :
10223 20 : nested_img_layer.push((key, Bytes::from(value)));
10224 4 : }
10225 4 :
10226 4 : let mut delta_layer_spec = Vec::default();
10227 4 : let delta_layer_start_lsn = Lsn(0x20);
10228 4 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10229 4 :
10230 44 : for i in 0..10 {
10231 40 : let key = get_key(i);
10232 40 : let key_in_nested = nested_img_layer
10233 40 : .iter()
10234 160 : .any(|(key_with_img, _)| *key_with_img == key);
10235 40 : let lsn = {
10236 40 : if key_in_nested {
10237 20 : Lsn(nested_image_layer_lsn.0 + 0x10)
10238 4 : } else {
10239 20 : delta_layer_start_lsn
10240 4 : }
10241 4 : };
10242 4 :
10243 40 : let will_init = will_init_keys.contains(&i);
10244 40 : if will_init {
10245 8 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10246 8 :
10247 8 : expected_key_values.insert(key, "".to_string());
10248 32 : } else {
10249 32 : let delta = format!("@{lsn}");
10250 32 : delta_layer_spec.push((
10251 32 : key,
10252 32 : lsn,
10253 32 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10254 32 : ));
10255 32 :
10256 32 : expected_key_values
10257 32 : .get_mut(&key)
10258 32 : .expect("An image exists for each key")
10259 32 : .push_str(delta.as_str());
10260 32 : }
10261 40 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
10262 4 : }
10263 4 :
10264 4 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
10265 4 :
10266 4 : assert!(
10267 4 : nested_image_layer_lsn > delta_layer_start_lsn
10268 4 : && nested_image_layer_lsn < delta_layer_end_lsn
10269 4 : );
10270 4 :
10271 4 : let tline = tenant
10272 4 : .create_test_timeline_with_layers(
10273 4 : TIMELINE_ID,
10274 4 : baseline_image_layer_lsn,
10275 4 : DEFAULT_PG_VERSION,
10276 4 : &ctx,
10277 4 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
10278 4 : delta_layer_start_lsn..delta_layer_end_lsn,
10279 4 : delta_layer_spec,
10280 4 : )], // delta layers
10281 4 : vec![
10282 4 : (baseline_image_layer_lsn, baseline_img_layer),
10283 4 : (nested_image_layer_lsn, nested_img_layer),
10284 4 : ], // image layers
10285 4 : delta_layer_end_lsn,
10286 4 : )
10287 4 : .await?;
10288 4 :
10289 4 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
10290 4 : let results = tline
10291 4 : .get_vectored(
10292 4 : keyspace,
10293 4 : delta_layer_end_lsn,
10294 4 : IoConcurrency::sequential(),
10295 4 : &ctx,
10296 4 : )
10297 4 : .await
10298 4 : .expect("No vectored errors");
10299 44 : for (key, res) in results {
10300 40 : let value = res.expect("No key errors");
10301 40 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10302 40 : assert_eq!(value, Bytes::from(expected_value));
10303 4 : }
10304 4 :
10305 4 : Ok(())
10306 4 : }
10307 :
10308 428 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
10309 428 : (
10310 428 : k1.is_delta,
10311 428 : k1.key_range.start,
10312 428 : k1.key_range.end,
10313 428 : k1.lsn_range.start,
10314 428 : k1.lsn_range.end,
10315 428 : )
10316 428 : .cmp(&(
10317 428 : k2.is_delta,
10318 428 : k2.key_range.start,
10319 428 : k2.key_range.end,
10320 428 : k2.lsn_range.start,
10321 428 : k2.lsn_range.end,
10322 428 : ))
10323 428 : }
10324 :
10325 48 : async fn inspect_and_sort(
10326 48 : tline: &Arc<Timeline>,
10327 48 : filter: Option<std::ops::Range<Key>>,
10328 48 : ) -> Vec<PersistentLayerKey> {
10329 48 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
10330 48 : if let Some(filter) = filter {
10331 216 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
10332 44 : }
10333 48 : all_layers.sort_by(sort_layer_key);
10334 48 : all_layers
10335 48 : }
10336 :
10337 : #[cfg(feature = "testing")]
10338 44 : fn check_layer_map_key_eq(
10339 44 : mut left: Vec<PersistentLayerKey>,
10340 44 : mut right: Vec<PersistentLayerKey>,
10341 44 : ) {
10342 44 : left.sort_by(sort_layer_key);
10343 44 : right.sort_by(sort_layer_key);
10344 44 : if left != right {
10345 0 : eprintln!("---LEFT---");
10346 0 : for left in left.iter() {
10347 0 : eprintln!("{}", left);
10348 0 : }
10349 0 : eprintln!("---RIGHT---");
10350 0 : for right in right.iter() {
10351 0 : eprintln!("{}", right);
10352 0 : }
10353 0 : assert_eq!(left, right);
10354 44 : }
10355 44 : }
10356 :
10357 : #[cfg(feature = "testing")]
10358 : #[tokio::test]
10359 4 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
10360 4 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
10361 4 : let (tenant, ctx) = harness.load().await;
10362 4 :
10363 364 : fn get_key(id: u32) -> Key {
10364 364 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10365 364 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10366 364 : key.field6 = id;
10367 364 : key
10368 364 : }
10369 4 :
10370 4 : // img layer at 0x10
10371 4 : let img_layer = (0..10)
10372 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10373 4 : .collect_vec();
10374 4 :
10375 4 : let delta1 = vec![
10376 4 : (
10377 4 : get_key(1),
10378 4 : Lsn(0x20),
10379 4 : Value::Image(Bytes::from("value 1@0x20")),
10380 4 : ),
10381 4 : (
10382 4 : get_key(2),
10383 4 : Lsn(0x30),
10384 4 : Value::Image(Bytes::from("value 2@0x30")),
10385 4 : ),
10386 4 : (
10387 4 : get_key(3),
10388 4 : Lsn(0x40),
10389 4 : Value::Image(Bytes::from("value 3@0x40")),
10390 4 : ),
10391 4 : ];
10392 4 : let delta2 = vec![
10393 4 : (
10394 4 : get_key(5),
10395 4 : Lsn(0x20),
10396 4 : Value::Image(Bytes::from("value 5@0x20")),
10397 4 : ),
10398 4 : (
10399 4 : get_key(6),
10400 4 : Lsn(0x20),
10401 4 : Value::Image(Bytes::from("value 6@0x20")),
10402 4 : ),
10403 4 : ];
10404 4 : let delta3 = vec![
10405 4 : (
10406 4 : get_key(8),
10407 4 : Lsn(0x48),
10408 4 : Value::Image(Bytes::from("value 8@0x48")),
10409 4 : ),
10410 4 : (
10411 4 : get_key(9),
10412 4 : Lsn(0x48),
10413 4 : Value::Image(Bytes::from("value 9@0x48")),
10414 4 : ),
10415 4 : ];
10416 4 :
10417 4 : let tline = tenant
10418 4 : .create_test_timeline_with_layers(
10419 4 : TIMELINE_ID,
10420 4 : Lsn(0x10),
10421 4 : DEFAULT_PG_VERSION,
10422 4 : &ctx,
10423 4 : vec![
10424 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10425 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10426 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10427 4 : ], // delta layers
10428 4 : vec![(Lsn(0x10), img_layer)], // image layers
10429 4 : Lsn(0x50),
10430 4 : )
10431 4 : .await?;
10432 4 :
10433 4 : {
10434 4 : tline
10435 4 : .applied_gc_cutoff_lsn
10436 4 : .lock_for_write()
10437 4 : .store_and_unlock(Lsn(0x30))
10438 4 : .wait()
10439 4 : .await;
10440 4 : // Update GC info
10441 4 : let mut guard = tline.gc_info.write().unwrap();
10442 4 : *guard = GcInfo {
10443 4 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
10444 4 : cutoffs: GcCutoffs {
10445 4 : time: Lsn(0x30),
10446 4 : space: Lsn(0x30),
10447 4 : },
10448 4 : leases: Default::default(),
10449 4 : within_ancestor_pitr: false,
10450 4 : };
10451 4 : }
10452 4 :
10453 4 : let cancel = CancellationToken::new();
10454 4 :
10455 4 : // Do a partial compaction on key range 0..2
10456 4 : tline
10457 4 : .compact_with_gc(
10458 4 : &cancel,
10459 4 : CompactOptions {
10460 4 : flags: EnumSet::new(),
10461 4 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
10462 4 : ..Default::default()
10463 4 : },
10464 4 : &ctx,
10465 4 : )
10466 4 : .await
10467 4 : .unwrap();
10468 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10469 4 : check_layer_map_key_eq(
10470 4 : all_layers,
10471 4 : vec![
10472 4 : // newly-generated image layer for the partial compaction range 0-2
10473 4 : PersistentLayerKey {
10474 4 : key_range: get_key(0)..get_key(2),
10475 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10476 4 : is_delta: false,
10477 4 : },
10478 4 : PersistentLayerKey {
10479 4 : key_range: get_key(0)..get_key(10),
10480 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10481 4 : is_delta: false,
10482 4 : },
10483 4 : // delta1 is split and the second part is rewritten
10484 4 : PersistentLayerKey {
10485 4 : key_range: get_key(2)..get_key(4),
10486 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10487 4 : is_delta: true,
10488 4 : },
10489 4 : PersistentLayerKey {
10490 4 : key_range: get_key(5)..get_key(7),
10491 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10492 4 : is_delta: true,
10493 4 : },
10494 4 : PersistentLayerKey {
10495 4 : key_range: get_key(8)..get_key(10),
10496 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10497 4 : is_delta: true,
10498 4 : },
10499 4 : ],
10500 4 : );
10501 4 :
10502 4 : // Do a partial compaction on key range 2..4
10503 4 : tline
10504 4 : .compact_with_gc(
10505 4 : &cancel,
10506 4 : CompactOptions {
10507 4 : flags: EnumSet::new(),
10508 4 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
10509 4 : ..Default::default()
10510 4 : },
10511 4 : &ctx,
10512 4 : )
10513 4 : .await
10514 4 : .unwrap();
10515 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10516 4 : check_layer_map_key_eq(
10517 4 : all_layers,
10518 4 : vec![
10519 4 : PersistentLayerKey {
10520 4 : key_range: get_key(0)..get_key(2),
10521 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10522 4 : is_delta: false,
10523 4 : },
10524 4 : PersistentLayerKey {
10525 4 : key_range: get_key(0)..get_key(10),
10526 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10527 4 : is_delta: false,
10528 4 : },
10529 4 : // image layer generated for the compaction range 2-4
10530 4 : PersistentLayerKey {
10531 4 : key_range: get_key(2)..get_key(4),
10532 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10533 4 : is_delta: false,
10534 4 : },
10535 4 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
10536 4 : PersistentLayerKey {
10537 4 : key_range: get_key(2)..get_key(4),
10538 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10539 4 : is_delta: true,
10540 4 : },
10541 4 : PersistentLayerKey {
10542 4 : key_range: get_key(5)..get_key(7),
10543 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10544 4 : is_delta: true,
10545 4 : },
10546 4 : PersistentLayerKey {
10547 4 : key_range: get_key(8)..get_key(10),
10548 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10549 4 : is_delta: true,
10550 4 : },
10551 4 : ],
10552 4 : );
10553 4 :
10554 4 : // Do a partial compaction on key range 4..9
10555 4 : tline
10556 4 : .compact_with_gc(
10557 4 : &cancel,
10558 4 : CompactOptions {
10559 4 : flags: EnumSet::new(),
10560 4 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
10561 4 : ..Default::default()
10562 4 : },
10563 4 : &ctx,
10564 4 : )
10565 4 : .await
10566 4 : .unwrap();
10567 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10568 4 : check_layer_map_key_eq(
10569 4 : all_layers,
10570 4 : vec![
10571 4 : PersistentLayerKey {
10572 4 : key_range: get_key(0)..get_key(2),
10573 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10574 4 : is_delta: false,
10575 4 : },
10576 4 : PersistentLayerKey {
10577 4 : key_range: get_key(0)..get_key(10),
10578 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10579 4 : is_delta: false,
10580 4 : },
10581 4 : PersistentLayerKey {
10582 4 : key_range: get_key(2)..get_key(4),
10583 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10584 4 : is_delta: false,
10585 4 : },
10586 4 : PersistentLayerKey {
10587 4 : key_range: get_key(2)..get_key(4),
10588 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10589 4 : is_delta: true,
10590 4 : },
10591 4 : // image layer generated for this compaction range
10592 4 : PersistentLayerKey {
10593 4 : key_range: get_key(4)..get_key(9),
10594 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10595 4 : is_delta: false,
10596 4 : },
10597 4 : PersistentLayerKey {
10598 4 : key_range: get_key(8)..get_key(10),
10599 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10600 4 : is_delta: true,
10601 4 : },
10602 4 : ],
10603 4 : );
10604 4 :
10605 4 : // Do a partial compaction on key range 9..10
10606 4 : tline
10607 4 : .compact_with_gc(
10608 4 : &cancel,
10609 4 : CompactOptions {
10610 4 : flags: EnumSet::new(),
10611 4 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
10612 4 : ..Default::default()
10613 4 : },
10614 4 : &ctx,
10615 4 : )
10616 4 : .await
10617 4 : .unwrap();
10618 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10619 4 : check_layer_map_key_eq(
10620 4 : all_layers,
10621 4 : vec![
10622 4 : PersistentLayerKey {
10623 4 : key_range: get_key(0)..get_key(2),
10624 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10625 4 : is_delta: false,
10626 4 : },
10627 4 : PersistentLayerKey {
10628 4 : key_range: get_key(0)..get_key(10),
10629 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10630 4 : is_delta: false,
10631 4 : },
10632 4 : PersistentLayerKey {
10633 4 : key_range: get_key(2)..get_key(4),
10634 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10635 4 : is_delta: false,
10636 4 : },
10637 4 : PersistentLayerKey {
10638 4 : key_range: get_key(2)..get_key(4),
10639 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10640 4 : is_delta: true,
10641 4 : },
10642 4 : PersistentLayerKey {
10643 4 : key_range: get_key(4)..get_key(9),
10644 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10645 4 : is_delta: false,
10646 4 : },
10647 4 : // image layer generated for the compaction range
10648 4 : PersistentLayerKey {
10649 4 : key_range: get_key(9)..get_key(10),
10650 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10651 4 : is_delta: false,
10652 4 : },
10653 4 : PersistentLayerKey {
10654 4 : key_range: get_key(8)..get_key(10),
10655 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10656 4 : is_delta: true,
10657 4 : },
10658 4 : ],
10659 4 : );
10660 4 :
10661 4 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
10662 4 : tline
10663 4 : .compact_with_gc(
10664 4 : &cancel,
10665 4 : CompactOptions {
10666 4 : flags: EnumSet::new(),
10667 4 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
10668 4 : ..Default::default()
10669 4 : },
10670 4 : &ctx,
10671 4 : )
10672 4 : .await
10673 4 : .unwrap();
10674 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10675 4 : check_layer_map_key_eq(
10676 4 : all_layers,
10677 4 : vec![
10678 4 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
10679 4 : PersistentLayerKey {
10680 4 : key_range: get_key(0)..get_key(10),
10681 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10682 4 : is_delta: false,
10683 4 : },
10684 4 : PersistentLayerKey {
10685 4 : key_range: get_key(2)..get_key(4),
10686 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10687 4 : is_delta: true,
10688 4 : },
10689 4 : PersistentLayerKey {
10690 4 : key_range: get_key(8)..get_key(10),
10691 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10692 4 : is_delta: true,
10693 4 : },
10694 4 : ],
10695 4 : );
10696 4 : Ok(())
10697 4 : }
10698 :
10699 : #[cfg(feature = "testing")]
10700 : #[tokio::test]
10701 4 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
10702 4 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
10703 4 : .await
10704 4 : .unwrap();
10705 4 : let (tenant, ctx) = harness.load().await;
10706 4 : let tline_parent = tenant
10707 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
10708 4 : .await
10709 4 : .unwrap();
10710 4 : let tline_child = tenant
10711 4 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
10712 4 : .await
10713 4 : .unwrap();
10714 4 : {
10715 4 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10716 4 : assert_eq!(
10717 4 : gc_info_parent.retain_lsns,
10718 4 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
10719 4 : );
10720 4 : }
10721 4 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
10722 4 : tline_child
10723 4 : .remote_client
10724 4 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
10725 4 : .unwrap();
10726 4 : tline_child.remote_client.wait_completion().await.unwrap();
10727 4 : offload_timeline(&tenant, &tline_child)
10728 4 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
10729 4 : .await.unwrap();
10730 4 : let child_timeline_id = tline_child.timeline_id;
10731 4 : Arc::try_unwrap(tline_child).unwrap();
10732 4 :
10733 4 : {
10734 4 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10735 4 : assert_eq!(
10736 4 : gc_info_parent.retain_lsns,
10737 4 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
10738 4 : );
10739 4 : }
10740 4 :
10741 4 : tenant
10742 4 : .get_offloaded_timeline(child_timeline_id)
10743 4 : .unwrap()
10744 4 : .defuse_for_tenant_drop();
10745 4 :
10746 4 : Ok(())
10747 4 : }
10748 :
10749 : #[cfg(feature = "testing")]
10750 : #[tokio::test]
10751 4 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
10752 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
10753 4 : let (tenant, ctx) = harness.load().await;
10754 4 :
10755 592 : fn get_key(id: u32) -> Key {
10756 592 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10757 592 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10758 592 : key.field6 = id;
10759 592 : key
10760 592 : }
10761 4 :
10762 4 : let img_layer = (0..10)
10763 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10764 4 : .collect_vec();
10765 4 :
10766 4 : let delta1 = vec![(
10767 4 : get_key(1),
10768 4 : Lsn(0x20),
10769 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10770 4 : )];
10771 4 : let delta4 = vec![(
10772 4 : get_key(1),
10773 4 : Lsn(0x28),
10774 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10775 4 : )];
10776 4 : let delta2 = vec![
10777 4 : (
10778 4 : get_key(1),
10779 4 : Lsn(0x30),
10780 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10781 4 : ),
10782 4 : (
10783 4 : get_key(1),
10784 4 : Lsn(0x38),
10785 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10786 4 : ),
10787 4 : ];
10788 4 : let delta3 = vec![
10789 4 : (
10790 4 : get_key(8),
10791 4 : Lsn(0x48),
10792 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10793 4 : ),
10794 4 : (
10795 4 : get_key(9),
10796 4 : Lsn(0x48),
10797 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10798 4 : ),
10799 4 : ];
10800 4 :
10801 4 : let tline = tenant
10802 4 : .create_test_timeline_with_layers(
10803 4 : TIMELINE_ID,
10804 4 : Lsn(0x10),
10805 4 : DEFAULT_PG_VERSION,
10806 4 : &ctx,
10807 4 : vec![
10808 4 : // delta1/2/4 only contain a single key but multiple updates
10809 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
10810 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10811 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
10812 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
10813 4 : ], // delta layers
10814 4 : vec![(Lsn(0x10), img_layer)], // image layers
10815 4 : Lsn(0x50),
10816 4 : )
10817 4 : .await?;
10818 4 : {
10819 4 : tline
10820 4 : .applied_gc_cutoff_lsn
10821 4 : .lock_for_write()
10822 4 : .store_and_unlock(Lsn(0x30))
10823 4 : .wait()
10824 4 : .await;
10825 4 : // Update GC info
10826 4 : let mut guard = tline.gc_info.write().unwrap();
10827 4 : *guard = GcInfo {
10828 4 : retain_lsns: vec![
10829 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10830 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10831 4 : ],
10832 4 : cutoffs: GcCutoffs {
10833 4 : time: Lsn(0x30),
10834 4 : space: Lsn(0x30),
10835 4 : },
10836 4 : leases: Default::default(),
10837 4 : within_ancestor_pitr: false,
10838 4 : };
10839 4 : }
10840 4 :
10841 4 : let expected_result = [
10842 4 : Bytes::from_static(b"value 0@0x10"),
10843 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10844 4 : Bytes::from_static(b"value 2@0x10"),
10845 4 : Bytes::from_static(b"value 3@0x10"),
10846 4 : Bytes::from_static(b"value 4@0x10"),
10847 4 : Bytes::from_static(b"value 5@0x10"),
10848 4 : Bytes::from_static(b"value 6@0x10"),
10849 4 : Bytes::from_static(b"value 7@0x10"),
10850 4 : Bytes::from_static(b"value 8@0x10@0x48"),
10851 4 : Bytes::from_static(b"value 9@0x10@0x48"),
10852 4 : ];
10853 4 :
10854 4 : let expected_result_at_gc_horizon = [
10855 4 : Bytes::from_static(b"value 0@0x10"),
10856 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10857 4 : Bytes::from_static(b"value 2@0x10"),
10858 4 : Bytes::from_static(b"value 3@0x10"),
10859 4 : Bytes::from_static(b"value 4@0x10"),
10860 4 : Bytes::from_static(b"value 5@0x10"),
10861 4 : Bytes::from_static(b"value 6@0x10"),
10862 4 : Bytes::from_static(b"value 7@0x10"),
10863 4 : Bytes::from_static(b"value 8@0x10"),
10864 4 : Bytes::from_static(b"value 9@0x10"),
10865 4 : ];
10866 4 :
10867 4 : let expected_result_at_lsn_20 = [
10868 4 : Bytes::from_static(b"value 0@0x10"),
10869 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10870 4 : Bytes::from_static(b"value 2@0x10"),
10871 4 : Bytes::from_static(b"value 3@0x10"),
10872 4 : Bytes::from_static(b"value 4@0x10"),
10873 4 : Bytes::from_static(b"value 5@0x10"),
10874 4 : Bytes::from_static(b"value 6@0x10"),
10875 4 : Bytes::from_static(b"value 7@0x10"),
10876 4 : Bytes::from_static(b"value 8@0x10"),
10877 4 : Bytes::from_static(b"value 9@0x10"),
10878 4 : ];
10879 4 :
10880 4 : let expected_result_at_lsn_10 = [
10881 4 : Bytes::from_static(b"value 0@0x10"),
10882 4 : Bytes::from_static(b"value 1@0x10"),
10883 4 : Bytes::from_static(b"value 2@0x10"),
10884 4 : Bytes::from_static(b"value 3@0x10"),
10885 4 : Bytes::from_static(b"value 4@0x10"),
10886 4 : Bytes::from_static(b"value 5@0x10"),
10887 4 : Bytes::from_static(b"value 6@0x10"),
10888 4 : Bytes::from_static(b"value 7@0x10"),
10889 4 : Bytes::from_static(b"value 8@0x10"),
10890 4 : Bytes::from_static(b"value 9@0x10"),
10891 4 : ];
10892 4 :
10893 12 : let verify_result = || async {
10894 12 : let gc_horizon = {
10895 12 : let gc_info = tline.gc_info.read().unwrap();
10896 12 : gc_info.cutoffs.time
10897 4 : };
10898 132 : for idx in 0..10 {
10899 120 : assert_eq!(
10900 120 : tline
10901 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10902 120 : .await
10903 120 : .unwrap(),
10904 120 : &expected_result[idx]
10905 4 : );
10906 120 : assert_eq!(
10907 120 : tline
10908 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
10909 120 : .await
10910 120 : .unwrap(),
10911 120 : &expected_result_at_gc_horizon[idx]
10912 4 : );
10913 120 : assert_eq!(
10914 120 : tline
10915 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10916 120 : .await
10917 120 : .unwrap(),
10918 120 : &expected_result_at_lsn_20[idx]
10919 4 : );
10920 120 : assert_eq!(
10921 120 : tline
10922 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10923 120 : .await
10924 120 : .unwrap(),
10925 120 : &expected_result_at_lsn_10[idx]
10926 4 : );
10927 4 : }
10928 24 : };
10929 4 :
10930 4 : verify_result().await;
10931 4 :
10932 4 : let cancel = CancellationToken::new();
10933 4 : tline
10934 4 : .compact_with_gc(
10935 4 : &cancel,
10936 4 : CompactOptions {
10937 4 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
10938 4 : ..Default::default()
10939 4 : },
10940 4 : &ctx,
10941 4 : )
10942 4 : .await
10943 4 : .unwrap();
10944 4 : verify_result().await;
10945 4 :
10946 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10947 4 : check_layer_map_key_eq(
10948 4 : all_layers,
10949 4 : vec![
10950 4 : // The original image layer, not compacted
10951 4 : PersistentLayerKey {
10952 4 : key_range: get_key(0)..get_key(10),
10953 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10954 4 : is_delta: false,
10955 4 : },
10956 4 : // Delta layer below the specified above_lsn not compacted
10957 4 : PersistentLayerKey {
10958 4 : key_range: get_key(1)..get_key(2),
10959 4 : lsn_range: Lsn(0x20)..Lsn(0x28),
10960 4 : is_delta: true,
10961 4 : },
10962 4 : // Delta layer compacted above the LSN
10963 4 : PersistentLayerKey {
10964 4 : key_range: get_key(1)..get_key(10),
10965 4 : lsn_range: Lsn(0x28)..Lsn(0x50),
10966 4 : is_delta: true,
10967 4 : },
10968 4 : ],
10969 4 : );
10970 4 :
10971 4 : // compact again
10972 4 : tline
10973 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10974 4 : .await
10975 4 : .unwrap();
10976 4 : verify_result().await;
10977 4 :
10978 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10979 4 : check_layer_map_key_eq(
10980 4 : all_layers,
10981 4 : vec![
10982 4 : // The compacted image layer (full key range)
10983 4 : PersistentLayerKey {
10984 4 : key_range: Key::MIN..Key::MAX,
10985 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10986 4 : is_delta: false,
10987 4 : },
10988 4 : // All other data in the delta layer
10989 4 : PersistentLayerKey {
10990 4 : key_range: get_key(1)..get_key(10),
10991 4 : lsn_range: Lsn(0x10)..Lsn(0x50),
10992 4 : is_delta: true,
10993 4 : },
10994 4 : ],
10995 4 : );
10996 4 :
10997 4 : Ok(())
10998 4 : }
10999 :
11000 : #[cfg(feature = "testing")]
11001 : #[tokio::test]
11002 4 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
11003 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
11004 4 : let (tenant, ctx) = harness.load().await;
11005 4 :
11006 1016 : fn get_key(id: u32) -> Key {
11007 1016 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11008 1016 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11009 1016 : key.field6 = id;
11010 1016 : key
11011 1016 : }
11012 4 :
11013 4 : let img_layer = (0..10)
11014 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11015 4 : .collect_vec();
11016 4 :
11017 4 : let delta1 = vec![(
11018 4 : get_key(1),
11019 4 : Lsn(0x20),
11020 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11021 4 : )];
11022 4 : let delta4 = vec![(
11023 4 : get_key(1),
11024 4 : Lsn(0x28),
11025 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11026 4 : )];
11027 4 : let delta2 = vec![
11028 4 : (
11029 4 : get_key(1),
11030 4 : Lsn(0x30),
11031 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11032 4 : ),
11033 4 : (
11034 4 : get_key(1),
11035 4 : Lsn(0x38),
11036 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11037 4 : ),
11038 4 : ];
11039 4 : let delta3 = vec![
11040 4 : (
11041 4 : get_key(8),
11042 4 : Lsn(0x48),
11043 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11044 4 : ),
11045 4 : (
11046 4 : get_key(9),
11047 4 : Lsn(0x48),
11048 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11049 4 : ),
11050 4 : ];
11051 4 :
11052 4 : let tline = tenant
11053 4 : .create_test_timeline_with_layers(
11054 4 : TIMELINE_ID,
11055 4 : Lsn(0x10),
11056 4 : DEFAULT_PG_VERSION,
11057 4 : &ctx,
11058 4 : vec![
11059 4 : // delta1/2/4 only contain a single key but multiple updates
11060 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11061 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11062 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11063 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11064 4 : ], // delta layers
11065 4 : vec![(Lsn(0x10), img_layer)], // image layers
11066 4 : Lsn(0x50),
11067 4 : )
11068 4 : .await?;
11069 4 : {
11070 4 : tline
11071 4 : .applied_gc_cutoff_lsn
11072 4 : .lock_for_write()
11073 4 : .store_and_unlock(Lsn(0x30))
11074 4 : .wait()
11075 4 : .await;
11076 4 : // Update GC info
11077 4 : let mut guard = tline.gc_info.write().unwrap();
11078 4 : *guard = GcInfo {
11079 4 : retain_lsns: vec![
11080 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11081 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11082 4 : ],
11083 4 : cutoffs: GcCutoffs {
11084 4 : time: Lsn(0x30),
11085 4 : space: Lsn(0x30),
11086 4 : },
11087 4 : leases: Default::default(),
11088 4 : within_ancestor_pitr: false,
11089 4 : };
11090 4 : }
11091 4 :
11092 4 : let expected_result = [
11093 4 : Bytes::from_static(b"value 0@0x10"),
11094 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11095 4 : Bytes::from_static(b"value 2@0x10"),
11096 4 : Bytes::from_static(b"value 3@0x10"),
11097 4 : Bytes::from_static(b"value 4@0x10"),
11098 4 : Bytes::from_static(b"value 5@0x10"),
11099 4 : Bytes::from_static(b"value 6@0x10"),
11100 4 : Bytes::from_static(b"value 7@0x10"),
11101 4 : Bytes::from_static(b"value 8@0x10@0x48"),
11102 4 : Bytes::from_static(b"value 9@0x10@0x48"),
11103 4 : ];
11104 4 :
11105 4 : let expected_result_at_gc_horizon = [
11106 4 : Bytes::from_static(b"value 0@0x10"),
11107 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11108 4 : Bytes::from_static(b"value 2@0x10"),
11109 4 : Bytes::from_static(b"value 3@0x10"),
11110 4 : Bytes::from_static(b"value 4@0x10"),
11111 4 : Bytes::from_static(b"value 5@0x10"),
11112 4 : Bytes::from_static(b"value 6@0x10"),
11113 4 : Bytes::from_static(b"value 7@0x10"),
11114 4 : Bytes::from_static(b"value 8@0x10"),
11115 4 : Bytes::from_static(b"value 9@0x10"),
11116 4 : ];
11117 4 :
11118 4 : let expected_result_at_lsn_20 = [
11119 4 : Bytes::from_static(b"value 0@0x10"),
11120 4 : Bytes::from_static(b"value 1@0x10@0x20"),
11121 4 : Bytes::from_static(b"value 2@0x10"),
11122 4 : Bytes::from_static(b"value 3@0x10"),
11123 4 : Bytes::from_static(b"value 4@0x10"),
11124 4 : Bytes::from_static(b"value 5@0x10"),
11125 4 : Bytes::from_static(b"value 6@0x10"),
11126 4 : Bytes::from_static(b"value 7@0x10"),
11127 4 : Bytes::from_static(b"value 8@0x10"),
11128 4 : Bytes::from_static(b"value 9@0x10"),
11129 4 : ];
11130 4 :
11131 4 : let expected_result_at_lsn_10 = [
11132 4 : Bytes::from_static(b"value 0@0x10"),
11133 4 : Bytes::from_static(b"value 1@0x10"),
11134 4 : Bytes::from_static(b"value 2@0x10"),
11135 4 : Bytes::from_static(b"value 3@0x10"),
11136 4 : Bytes::from_static(b"value 4@0x10"),
11137 4 : Bytes::from_static(b"value 5@0x10"),
11138 4 : Bytes::from_static(b"value 6@0x10"),
11139 4 : Bytes::from_static(b"value 7@0x10"),
11140 4 : Bytes::from_static(b"value 8@0x10"),
11141 4 : Bytes::from_static(b"value 9@0x10"),
11142 4 : ];
11143 4 :
11144 20 : let verify_result = || async {
11145 20 : let gc_horizon = {
11146 20 : let gc_info = tline.gc_info.read().unwrap();
11147 20 : gc_info.cutoffs.time
11148 4 : };
11149 220 : for idx in 0..10 {
11150 200 : assert_eq!(
11151 200 : tline
11152 200 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
11153 200 : .await
11154 200 : .unwrap(),
11155 200 : &expected_result[idx]
11156 4 : );
11157 200 : assert_eq!(
11158 200 : tline
11159 200 : .get(get_key(idx as u32), gc_horizon, &ctx)
11160 200 : .await
11161 200 : .unwrap(),
11162 200 : &expected_result_at_gc_horizon[idx]
11163 4 : );
11164 200 : assert_eq!(
11165 200 : tline
11166 200 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
11167 200 : .await
11168 200 : .unwrap(),
11169 200 : &expected_result_at_lsn_20[idx]
11170 4 : );
11171 200 : assert_eq!(
11172 200 : tline
11173 200 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
11174 200 : .await
11175 200 : .unwrap(),
11176 200 : &expected_result_at_lsn_10[idx]
11177 4 : );
11178 4 : }
11179 40 : };
11180 4 :
11181 4 : verify_result().await;
11182 4 :
11183 4 : let cancel = CancellationToken::new();
11184 4 :
11185 4 : tline
11186 4 : .compact_with_gc(
11187 4 : &cancel,
11188 4 : CompactOptions {
11189 4 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11190 4 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
11191 4 : ..Default::default()
11192 4 : },
11193 4 : &ctx,
11194 4 : )
11195 4 : .await
11196 4 : .unwrap();
11197 4 : verify_result().await;
11198 4 :
11199 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11200 4 : check_layer_map_key_eq(
11201 4 : all_layers,
11202 4 : vec![
11203 4 : // The original image layer, not compacted
11204 4 : PersistentLayerKey {
11205 4 : key_range: get_key(0)..get_key(10),
11206 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11207 4 : is_delta: false,
11208 4 : },
11209 4 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
11210 4 : // the layer 0x28-0x30 into one.
11211 4 : PersistentLayerKey {
11212 4 : key_range: get_key(1)..get_key(2),
11213 4 : lsn_range: Lsn(0x20)..Lsn(0x30),
11214 4 : is_delta: true,
11215 4 : },
11216 4 : // Above the upper bound and untouched
11217 4 : PersistentLayerKey {
11218 4 : key_range: get_key(1)..get_key(2),
11219 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11220 4 : is_delta: true,
11221 4 : },
11222 4 : // This layer is untouched
11223 4 : PersistentLayerKey {
11224 4 : key_range: get_key(8)..get_key(10),
11225 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11226 4 : is_delta: true,
11227 4 : },
11228 4 : ],
11229 4 : );
11230 4 :
11231 4 : tline
11232 4 : .compact_with_gc(
11233 4 : &cancel,
11234 4 : CompactOptions {
11235 4 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
11236 4 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
11237 4 : ..Default::default()
11238 4 : },
11239 4 : &ctx,
11240 4 : )
11241 4 : .await
11242 4 : .unwrap();
11243 4 : verify_result().await;
11244 4 :
11245 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11246 4 : check_layer_map_key_eq(
11247 4 : all_layers,
11248 4 : vec![
11249 4 : // The original image layer, not compacted
11250 4 : PersistentLayerKey {
11251 4 : key_range: get_key(0)..get_key(10),
11252 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11253 4 : is_delta: false,
11254 4 : },
11255 4 : // Not in the compaction key range, uncompacted
11256 4 : PersistentLayerKey {
11257 4 : key_range: get_key(1)..get_key(2),
11258 4 : lsn_range: Lsn(0x20)..Lsn(0x30),
11259 4 : is_delta: true,
11260 4 : },
11261 4 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
11262 4 : PersistentLayerKey {
11263 4 : key_range: get_key(1)..get_key(2),
11264 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11265 4 : is_delta: true,
11266 4 : },
11267 4 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
11268 4 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
11269 4 : // becomes 0x50.
11270 4 : PersistentLayerKey {
11271 4 : key_range: get_key(8)..get_key(10),
11272 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11273 4 : is_delta: true,
11274 4 : },
11275 4 : ],
11276 4 : );
11277 4 :
11278 4 : // compact again
11279 4 : tline
11280 4 : .compact_with_gc(
11281 4 : &cancel,
11282 4 : CompactOptions {
11283 4 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
11284 4 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
11285 4 : ..Default::default()
11286 4 : },
11287 4 : &ctx,
11288 4 : )
11289 4 : .await
11290 4 : .unwrap();
11291 4 : verify_result().await;
11292 4 :
11293 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11294 4 : check_layer_map_key_eq(
11295 4 : all_layers,
11296 4 : vec![
11297 4 : // The original image layer, not compacted
11298 4 : PersistentLayerKey {
11299 4 : key_range: get_key(0)..get_key(10),
11300 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11301 4 : is_delta: false,
11302 4 : },
11303 4 : // The range gets compacted
11304 4 : PersistentLayerKey {
11305 4 : key_range: get_key(1)..get_key(2),
11306 4 : lsn_range: Lsn(0x20)..Lsn(0x50),
11307 4 : is_delta: true,
11308 4 : },
11309 4 : // Not touched during this iteration of compaction
11310 4 : PersistentLayerKey {
11311 4 : key_range: get_key(8)..get_key(10),
11312 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11313 4 : is_delta: true,
11314 4 : },
11315 4 : ],
11316 4 : );
11317 4 :
11318 4 : // final full compaction
11319 4 : tline
11320 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11321 4 : .await
11322 4 : .unwrap();
11323 4 : verify_result().await;
11324 4 :
11325 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11326 4 : check_layer_map_key_eq(
11327 4 : all_layers,
11328 4 : vec![
11329 4 : // The compacted image layer (full key range)
11330 4 : PersistentLayerKey {
11331 4 : key_range: Key::MIN..Key::MAX,
11332 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11333 4 : is_delta: false,
11334 4 : },
11335 4 : // All other data in the delta layer
11336 4 : PersistentLayerKey {
11337 4 : key_range: get_key(1)..get_key(10),
11338 4 : lsn_range: Lsn(0x10)..Lsn(0x50),
11339 4 : is_delta: true,
11340 4 : },
11341 4 : ],
11342 4 : );
11343 4 :
11344 4 : Ok(())
11345 4 : }
11346 : }
|
