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 : pub use pageserver_api::models::TenantState;
35 : use pageserver_api::models::{self, RelSizeMigration};
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;
81 : use crate::context::RequestContextBuilder;
82 : use crate::context::{DownloadBehavior, RequestContext};
83 : use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
84 : use crate::l0_flush::L0FlushGlobalState;
85 : use crate::metrics::{
86 : BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
87 : INITDB_RUN_TIME, INITDB_SEMAPHORE_ACQUISITION_TIME, TENANT, TENANT_STATE_METRIC,
88 : TENANT_SYNTHETIC_SIZE_METRIC, remove_tenant_metrics,
89 : };
90 : use crate::task_mgr::TaskKind;
91 : use crate::tenant::config::{LocationMode, TenantConfOpt};
92 : use crate::tenant::gc_result::GcResult;
93 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
94 : use crate::tenant::remote_timeline_client::{
95 : INITDB_PATH, MaybeDeletedIndexPart, remote_initdb_archive_path,
96 : };
97 : use crate::tenant::storage_layer::{DeltaLayer, ImageLayer};
98 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
99 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
100 : use crate::virtual_file::VirtualFile;
101 : use crate::walingest::WalLagCooldown;
102 : use crate::walredo::PostgresRedoManager;
103 : use crate::{InitializationOrder, TEMP_FILE_SUFFIX, import_datadir, span, task_mgr, walredo};
104 :
105 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
106 : use utils::crashsafe;
107 : use utils::generation::Generation;
108 : use utils::id::TimelineId;
109 : use utils::lsn::{Lsn, RecordLsn};
110 :
111 : pub mod blob_io;
112 : pub mod block_io;
113 : pub mod vectored_blob_io;
114 :
115 : pub mod disk_btree;
116 : pub(crate) mod ephemeral_file;
117 : pub mod layer_map;
118 :
119 : pub mod metadata;
120 : pub mod remote_timeline_client;
121 : pub mod storage_layer;
122 :
123 : pub mod checks;
124 : pub mod config;
125 : pub mod mgr;
126 : pub mod secondary;
127 : pub mod tasks;
128 : pub mod upload_queue;
129 :
130 : pub(crate) mod timeline;
131 :
132 : pub mod size;
133 :
134 : mod gc_block;
135 : mod gc_result;
136 : pub(crate) mod throttle;
137 :
138 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
139 :
140 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
141 : // re-export for use in walreceiver
142 : pub use crate::tenant::timeline::WalReceiverInfo;
143 :
144 : /// The "tenants" part of `tenants/<tenant>/timelines...`
145 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
146 :
147 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
148 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
149 :
150 : /// References to shared objects that are passed into each tenant, such
151 : /// as the shared remote storage client and process initialization state.
152 : #[derive(Clone)]
153 : pub struct TenantSharedResources {
154 : pub broker_client: storage_broker::BrokerClientChannel,
155 : pub remote_storage: GenericRemoteStorage,
156 : pub deletion_queue_client: DeletionQueueClient,
157 : pub l0_flush_global_state: L0FlushGlobalState,
158 : }
159 :
160 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
161 : /// for an attached tenant is a subset of the [`LocationConf`], represented
162 : /// in this struct.
163 : #[derive(Clone)]
164 : pub(super) struct AttachedTenantConf {
165 : tenant_conf: TenantConfOpt,
166 : location: AttachedLocationConfig,
167 : /// The deadline before which we are blocked from GC so that
168 : /// leases have a chance to be renewed.
169 : lsn_lease_deadline: Option<tokio::time::Instant>,
170 : }
171 :
172 : impl AttachedTenantConf {
173 452 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
174 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
175 : //
176 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
177 : // length, we guarantee that all the leases we granted before will have a chance to renew
178 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
179 452 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
180 452 : Some(
181 452 : tokio::time::Instant::now()
182 452 : + tenant_conf
183 452 : .lsn_lease_length
184 452 : .unwrap_or(LsnLease::DEFAULT_LENGTH),
185 452 : )
186 : } else {
187 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
188 : // because we don't do GC in these modes.
189 0 : None
190 : };
191 :
192 452 : Self {
193 452 : tenant_conf,
194 452 : location,
195 452 : lsn_lease_deadline,
196 452 : }
197 452 : }
198 :
199 452 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
200 452 : match &location_conf.mode {
201 452 : LocationMode::Attached(attach_conf) => {
202 452 : Ok(Self::new(location_conf.tenant_conf, *attach_conf))
203 : }
204 : LocationMode::Secondary(_) => {
205 0 : anyhow::bail!(
206 0 : "Attempted to construct AttachedTenantConf from a LocationConf in secondary mode"
207 0 : )
208 : }
209 : }
210 452 : }
211 :
212 1524 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
213 1524 : self.lsn_lease_deadline
214 1524 : .map(|d| tokio::time::Instant::now() < d)
215 1524 : .unwrap_or(false)
216 1524 : }
217 : }
218 : struct TimelinePreload {
219 : timeline_id: TimelineId,
220 : client: RemoteTimelineClient,
221 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
222 : previous_heatmap: Option<PreviousHeatmap>,
223 : }
224 :
225 : pub(crate) struct TenantPreload {
226 : tenant_manifest: TenantManifest,
227 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
228 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
229 : }
230 :
231 : /// When we spawn a tenant, there is a special mode for tenant creation that
232 : /// avoids trying to read anything from remote storage.
233 : pub(crate) enum SpawnMode {
234 : /// Activate as soon as possible
235 : Eager,
236 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
237 : Lazy,
238 : }
239 :
240 : ///
241 : /// Tenant consists of multiple timelines. Keep them in a hash table.
242 : ///
243 : pub struct Tenant {
244 : // Global pageserver config parameters
245 : pub conf: &'static PageServerConf,
246 :
247 : /// The value creation timestamp, used to measure activation delay, see:
248 : /// <https://github.com/neondatabase/neon/issues/4025>
249 : constructed_at: Instant,
250 :
251 : state: watch::Sender<TenantState>,
252 :
253 : // Overridden tenant-specific config parameters.
254 : // We keep TenantConfOpt sturct here to preserve the information
255 : // about parameters that are not set.
256 : // This is necessary to allow global config updates.
257 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
258 :
259 : tenant_shard_id: TenantShardId,
260 :
261 : // The detailed sharding information, beyond the number/count in tenant_shard_id
262 : shard_identity: ShardIdentity,
263 :
264 : /// The remote storage generation, used to protect S3 objects from split-brain.
265 : /// Does not change over the lifetime of the [`Tenant`] object.
266 : ///
267 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
268 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
269 : generation: Generation,
270 :
271 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
272 :
273 : /// During timeline creation, we first insert the TimelineId to the
274 : /// creating map, then `timelines`, then remove it from the creating map.
275 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
276 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
277 :
278 : /// Possibly offloaded and archived timelines
279 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
280 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
281 :
282 : /// Serialize writes of the tenant manifest to remote storage. If there are concurrent operations
283 : /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
284 : /// each other (this could be optimized to coalesce writes if necessary).
285 : ///
286 : /// The contents of the Mutex are the last manifest we successfully uploaded
287 : tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
288 :
289 : // This mutex prevents creation of new timelines during GC.
290 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
291 : // `timelines` mutex during all GC iteration
292 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
293 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
294 : // timeout...
295 : gc_cs: tokio::sync::Mutex<()>,
296 : walredo_mgr: Option<Arc<WalRedoManager>>,
297 :
298 : // provides access to timeline data sitting in the remote storage
299 : pub(crate) remote_storage: GenericRemoteStorage,
300 :
301 : // Access to global deletion queue for when this tenant wants to schedule a deletion
302 : deletion_queue_client: DeletionQueueClient,
303 :
304 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
305 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
306 : cached_synthetic_tenant_size: Arc<AtomicU64>,
307 :
308 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
309 :
310 : /// Track repeated failures to compact, so that we can back off.
311 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
312 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
313 :
314 : /// Signals the tenant compaction loop that there is L0 compaction work to be done.
315 : pub(crate) l0_compaction_trigger: Arc<Notify>,
316 :
317 : /// Scheduled gc-compaction tasks.
318 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
319 :
320 : /// If the tenant is in Activating state, notify this to encourage it
321 : /// to proceed to Active as soon as possible, rather than waiting for lazy
322 : /// background warmup.
323 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
324 :
325 : /// Time it took for the tenant to activate. Zero if not active yet.
326 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
327 :
328 : // Cancellation token fires when we have entered shutdown(). This is a parent of
329 : // Timelines' cancellation token.
330 : pub(crate) cancel: CancellationToken,
331 :
332 : // Users of the Tenant such as the page service must take this Gate to avoid
333 : // trying to use a Tenant which is shutting down.
334 : pub(crate) gate: Gate,
335 :
336 : /// Throttle applied at the top of [`Timeline::get`].
337 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
338 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
339 :
340 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
341 :
342 : /// An ongoing timeline detach concurrency limiter.
343 : ///
344 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
345 : /// to have two running at the same time. A different one can be started if an earlier one
346 : /// has failed for whatever reason.
347 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
348 :
349 : /// `index_part.json` based gc blocking reason tracking.
350 : ///
351 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
352 : /// proceeding.
353 : pub(crate) gc_block: gc_block::GcBlock,
354 :
355 : l0_flush_global_state: L0FlushGlobalState,
356 : }
357 : impl std::fmt::Debug for Tenant {
358 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
359 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
360 0 : }
361 : }
362 :
363 : pub(crate) enum WalRedoManager {
364 : Prod(WalredoManagerId, PostgresRedoManager),
365 : #[cfg(test)]
366 : Test(harness::TestRedoManager),
367 : }
368 :
369 : #[derive(thiserror::Error, Debug)]
370 : #[error("pageserver is shutting down")]
371 : pub(crate) struct GlobalShutDown;
372 :
373 : impl WalRedoManager {
374 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
375 0 : let id = WalredoManagerId::next();
376 0 : let arc = Arc::new(Self::Prod(id, mgr));
377 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
378 0 : match &mut *guard {
379 0 : Some(map) => {
380 0 : map.insert(id, Arc::downgrade(&arc));
381 0 : Ok(arc)
382 : }
383 0 : None => Err(GlobalShutDown),
384 : }
385 0 : }
386 : }
387 :
388 : impl Drop for WalRedoManager {
389 20 : fn drop(&mut self) {
390 20 : match self {
391 0 : Self::Prod(id, _) => {
392 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
393 0 : if let Some(map) = &mut *guard {
394 0 : map.remove(id).expect("new() registers, drop() unregisters");
395 0 : }
396 : }
397 : #[cfg(test)]
398 20 : Self::Test(_) => {
399 20 : // Not applicable to test redo manager
400 20 : }
401 : }
402 20 : }
403 : }
404 :
405 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
406 : /// the walredo processes outside of the regular order.
407 : ///
408 : /// This is necessary to work around a systemd bug where it freezes if there are
409 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
410 : #[allow(clippy::type_complexity)]
411 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
412 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
413 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
414 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
415 : pub(crate) struct WalredoManagerId(u64);
416 : impl WalredoManagerId {
417 0 : pub fn next() -> Self {
418 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
419 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
420 0 : if id == 0 {
421 0 : panic!(
422 0 : "WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique"
423 0 : );
424 0 : }
425 0 : Self(id)
426 0 : }
427 : }
428 :
429 : #[cfg(test)]
430 : impl From<harness::TestRedoManager> for WalRedoManager {
431 452 : fn from(mgr: harness::TestRedoManager) -> Self {
432 452 : Self::Test(mgr)
433 452 : }
434 : }
435 :
436 : impl WalRedoManager {
437 12 : pub(crate) async fn shutdown(&self) -> bool {
438 12 : match self {
439 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
440 : #[cfg(test)]
441 : Self::Test(_) => {
442 : // Not applicable to test redo manager
443 12 : true
444 : }
445 : }
446 12 : }
447 :
448 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
449 0 : match self {
450 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
451 0 : #[cfg(test)]
452 0 : Self::Test(_) => {
453 0 : // Not applicable to test redo manager
454 0 : }
455 0 : }
456 0 : }
457 :
458 : /// # Cancel-Safety
459 : ///
460 : /// This method is cancellation-safe.
461 1676 : pub async fn request_redo(
462 1676 : &self,
463 1676 : key: pageserver_api::key::Key,
464 1676 : lsn: Lsn,
465 1676 : base_img: Option<(Lsn, bytes::Bytes)>,
466 1676 : records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
467 1676 : pg_version: u32,
468 1676 : ) -> Result<bytes::Bytes, walredo::Error> {
469 1676 : match self {
470 0 : Self::Prod(_, mgr) => {
471 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
472 0 : .await
473 : }
474 : #[cfg(test)]
475 1676 : Self::Test(mgr) => {
476 1676 : mgr.request_redo(key, lsn, base_img, records, pg_version)
477 1676 : .await
478 : }
479 : }
480 1676 : }
481 :
482 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
483 0 : match self {
484 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
485 0 : #[cfg(test)]
486 0 : WalRedoManager::Test(_) => None,
487 0 : }
488 0 : }
489 : }
490 :
491 : /// A very lightweight memory representation of an offloaded timeline.
492 : ///
493 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
494 : /// like unoffloading them, or (at a later date), decide to perform flattening.
495 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
496 : /// more offloaded timelines than we can manage ones that aren't.
497 : pub struct OffloadedTimeline {
498 : pub tenant_shard_id: TenantShardId,
499 : pub timeline_id: TimelineId,
500 : pub ancestor_timeline_id: Option<TimelineId>,
501 : /// Whether to retain the branch lsn at the ancestor or not
502 : pub ancestor_retain_lsn: Option<Lsn>,
503 :
504 : /// When the timeline was archived.
505 : ///
506 : /// Present for future flattening deliberations.
507 : pub archived_at: NaiveDateTime,
508 :
509 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
510 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
511 : pub delete_progress: TimelineDeleteProgress,
512 :
513 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
514 : pub deleted_from_ancestor: AtomicBool,
515 : }
516 :
517 : impl OffloadedTimeline {
518 : /// Obtains an offloaded timeline from a given timeline object.
519 : ///
520 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
521 : /// the timeline is not in a stopped state.
522 : /// Panics if the timeline is not archived.
523 4 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
524 4 : let (ancestor_retain_lsn, ancestor_timeline_id) =
525 4 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
526 4 : let ancestor_lsn = timeline.get_ancestor_lsn();
527 4 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
528 4 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
529 4 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
530 4 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
531 : } else {
532 0 : (None, None)
533 : };
534 4 : let archived_at = timeline
535 4 : .remote_client
536 4 : .archived_at_stopped_queue()?
537 4 : .expect("must be called on an archived timeline");
538 4 : Ok(Self {
539 4 : tenant_shard_id: timeline.tenant_shard_id,
540 4 : timeline_id: timeline.timeline_id,
541 4 : ancestor_timeline_id,
542 4 : ancestor_retain_lsn,
543 4 : archived_at,
544 4 :
545 4 : delete_progress: timeline.delete_progress.clone(),
546 4 : deleted_from_ancestor: AtomicBool::new(false),
547 4 : })
548 4 : }
549 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
550 0 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
551 0 : // by the `initialize_gc_info` function.
552 0 : let OffloadedTimelineManifest {
553 0 : timeline_id,
554 0 : ancestor_timeline_id,
555 0 : ancestor_retain_lsn,
556 0 : archived_at,
557 0 : } = *manifest;
558 0 : Self {
559 0 : tenant_shard_id,
560 0 : timeline_id,
561 0 : ancestor_timeline_id,
562 0 : ancestor_retain_lsn,
563 0 : archived_at,
564 0 : delete_progress: TimelineDeleteProgress::default(),
565 0 : deleted_from_ancestor: AtomicBool::new(false),
566 0 : }
567 0 : }
568 4 : fn manifest(&self) -> OffloadedTimelineManifest {
569 4 : let Self {
570 4 : timeline_id,
571 4 : ancestor_timeline_id,
572 4 : ancestor_retain_lsn,
573 4 : archived_at,
574 4 : ..
575 4 : } = self;
576 4 : OffloadedTimelineManifest {
577 4 : timeline_id: *timeline_id,
578 4 : ancestor_timeline_id: *ancestor_timeline_id,
579 4 : ancestor_retain_lsn: *ancestor_retain_lsn,
580 4 : archived_at: *archived_at,
581 4 : }
582 4 : }
583 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
584 0 : fn delete_from_ancestor_with_timelines(
585 0 : &self,
586 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
587 0 : ) {
588 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
589 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
590 : {
591 0 : if let Some((_, ancestor_timeline)) = timelines
592 0 : .iter()
593 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
594 : {
595 0 : let removal_happened = ancestor_timeline
596 0 : .gc_info
597 0 : .write()
598 0 : .unwrap()
599 0 : .remove_child_offloaded(self.timeline_id);
600 0 : if !removal_happened {
601 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
602 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
603 0 : }
604 0 : }
605 0 : }
606 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
607 0 : }
608 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
609 : ///
610 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
611 4 : fn defuse_for_tenant_drop(&self) {
612 4 : self.deleted_from_ancestor.store(true, Ordering::Release);
613 4 : }
614 : }
615 :
616 : impl fmt::Debug for OffloadedTimeline {
617 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
618 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
619 0 : }
620 : }
621 :
622 : impl Drop for OffloadedTimeline {
623 4 : fn drop(&mut self) {
624 4 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
625 0 : tracing::warn!(
626 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
627 : self.timeline_id
628 : );
629 4 : }
630 4 : }
631 : }
632 :
633 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
634 : pub enum MaybeOffloaded {
635 : Yes,
636 : No,
637 : }
638 :
639 : #[derive(Clone, Debug)]
640 : pub enum TimelineOrOffloaded {
641 : Timeline(Arc<Timeline>),
642 : Offloaded(Arc<OffloadedTimeline>),
643 : }
644 :
645 : impl TimelineOrOffloaded {
646 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
647 0 : match self {
648 0 : TimelineOrOffloaded::Timeline(timeline) => {
649 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
650 : }
651 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
652 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
653 : }
654 : }
655 0 : }
656 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
657 0 : self.arc_ref().tenant_shard_id()
658 0 : }
659 0 : pub fn timeline_id(&self) -> TimelineId {
660 0 : self.arc_ref().timeline_id()
661 0 : }
662 4 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
663 4 : match self {
664 4 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
665 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
666 : }
667 4 : }
668 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
669 0 : match self {
670 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
671 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
672 : }
673 0 : }
674 : }
675 :
676 : pub enum TimelineOrOffloadedArcRef<'a> {
677 : Timeline(&'a Arc<Timeline>),
678 : Offloaded(&'a Arc<OffloadedTimeline>),
679 : }
680 :
681 : impl TimelineOrOffloadedArcRef<'_> {
682 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
683 0 : match self {
684 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
685 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
686 : }
687 0 : }
688 0 : pub fn timeline_id(&self) -> TimelineId {
689 0 : match self {
690 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
691 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
692 : }
693 0 : }
694 : }
695 :
696 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
697 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
698 0 : Self::Timeline(timeline)
699 0 : }
700 : }
701 :
702 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
703 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
704 0 : Self::Offloaded(timeline)
705 0 : }
706 : }
707 :
708 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
709 : pub enum GetTimelineError {
710 : #[error("Timeline is shutting down")]
711 : ShuttingDown,
712 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
713 : NotActive {
714 : tenant_id: TenantShardId,
715 : timeline_id: TimelineId,
716 : state: TimelineState,
717 : },
718 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
719 : NotFound {
720 : tenant_id: TenantShardId,
721 : timeline_id: TimelineId,
722 : },
723 : }
724 :
725 : #[derive(Debug, thiserror::Error)]
726 : pub enum LoadLocalTimelineError {
727 : #[error("FailedToLoad")]
728 : Load(#[source] anyhow::Error),
729 : #[error("FailedToResumeDeletion")]
730 : ResumeDeletion(#[source] anyhow::Error),
731 : }
732 :
733 : #[derive(thiserror::Error)]
734 : pub enum DeleteTimelineError {
735 : #[error("NotFound")]
736 : NotFound,
737 :
738 : #[error("HasChildren")]
739 : HasChildren(Vec<TimelineId>),
740 :
741 : #[error("Timeline deletion is already in progress")]
742 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
743 :
744 : #[error("Cancelled")]
745 : Cancelled,
746 :
747 : #[error(transparent)]
748 : Other(#[from] anyhow::Error),
749 : }
750 :
751 : impl Debug for DeleteTimelineError {
752 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
753 0 : match self {
754 0 : Self::NotFound => write!(f, "NotFound"),
755 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
756 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
757 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
758 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
759 : }
760 0 : }
761 : }
762 :
763 : #[derive(thiserror::Error)]
764 : pub enum TimelineArchivalError {
765 : #[error("NotFound")]
766 : NotFound,
767 :
768 : #[error("Timeout")]
769 : Timeout,
770 :
771 : #[error("Cancelled")]
772 : Cancelled,
773 :
774 : #[error("ancestor is archived: {}", .0)]
775 : HasArchivedParent(TimelineId),
776 :
777 : #[error("HasUnarchivedChildren")]
778 : HasUnarchivedChildren(Vec<TimelineId>),
779 :
780 : #[error("Timeline archival is already in progress")]
781 : AlreadyInProgress,
782 :
783 : #[error(transparent)]
784 : Other(anyhow::Error),
785 : }
786 :
787 : #[derive(thiserror::Error, Debug)]
788 : pub(crate) enum TenantManifestError {
789 : #[error("Remote storage error: {0}")]
790 : RemoteStorage(anyhow::Error),
791 :
792 : #[error("Cancelled")]
793 : Cancelled,
794 : }
795 :
796 : impl From<TenantManifestError> for TimelineArchivalError {
797 0 : fn from(e: TenantManifestError) -> Self {
798 0 : match e {
799 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
800 0 : TenantManifestError::Cancelled => Self::Cancelled,
801 : }
802 0 : }
803 : }
804 :
805 : impl Debug for TimelineArchivalError {
806 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
807 0 : match self {
808 0 : Self::NotFound => write!(f, "NotFound"),
809 0 : Self::Timeout => write!(f, "Timeout"),
810 0 : Self::Cancelled => write!(f, "Cancelled"),
811 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
812 0 : Self::HasUnarchivedChildren(c) => {
813 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
814 : }
815 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
816 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
817 : }
818 0 : }
819 : }
820 :
821 : pub enum SetStoppingError {
822 : AlreadyStopping(completion::Barrier),
823 : Broken,
824 : }
825 :
826 : impl Debug for SetStoppingError {
827 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
828 0 : match self {
829 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
830 0 : Self::Broken => write!(f, "Broken"),
831 : }
832 0 : }
833 : }
834 :
835 : /// Arguments to [`Tenant::create_timeline`].
836 : ///
837 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
838 : /// is `None`, the result of the timeline create call is not deterministic.
839 : ///
840 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
841 : #[derive(Debug)]
842 : pub(crate) enum CreateTimelineParams {
843 : Bootstrap(CreateTimelineParamsBootstrap),
844 : Branch(CreateTimelineParamsBranch),
845 : ImportPgdata(CreateTimelineParamsImportPgdata),
846 : }
847 :
848 : #[derive(Debug)]
849 : pub(crate) struct CreateTimelineParamsBootstrap {
850 : pub(crate) new_timeline_id: TimelineId,
851 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
852 : pub(crate) pg_version: u32,
853 : }
854 :
855 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
856 : #[derive(Debug)]
857 : pub(crate) struct CreateTimelineParamsBranch {
858 : pub(crate) new_timeline_id: TimelineId,
859 : pub(crate) ancestor_timeline_id: TimelineId,
860 : pub(crate) ancestor_start_lsn: Option<Lsn>,
861 : }
862 :
863 : #[derive(Debug)]
864 : pub(crate) struct CreateTimelineParamsImportPgdata {
865 : pub(crate) new_timeline_id: TimelineId,
866 : pub(crate) location: import_pgdata::index_part_format::Location,
867 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
868 : }
869 :
870 : /// What is used to determine idempotency of a [`Tenant::create_timeline`] call in [`Tenant::start_creating_timeline`] in [`Tenant::start_creating_timeline`].
871 : ///
872 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
873 : ///
874 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
875 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
876 : ///
877 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
878 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
879 : ///
880 : /// Notes:
881 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
882 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
883 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
884 : ///
885 : #[derive(Debug, Clone, PartialEq, Eq)]
886 : pub(crate) enum CreateTimelineIdempotency {
887 : /// NB: special treatment, see comment in [`Self`].
888 : FailWithConflict,
889 : Bootstrap {
890 : pg_version: u32,
891 : },
892 : /// NB: branches always have the same `pg_version` as their ancestor.
893 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
894 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
895 : /// determining the child branch pg_version.
896 : Branch {
897 : ancestor_timeline_id: TimelineId,
898 : ancestor_start_lsn: Lsn,
899 : },
900 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
901 : }
902 :
903 : #[derive(Debug, Clone, PartialEq, Eq)]
904 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
905 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
906 : }
907 :
908 : /// What is returned by [`Tenant::start_creating_timeline`].
909 : #[must_use]
910 : enum StartCreatingTimelineResult {
911 : CreateGuard(TimelineCreateGuard),
912 : Idempotent(Arc<Timeline>),
913 : }
914 :
915 : enum TimelineInitAndSyncResult {
916 : ReadyToActivate(Arc<Timeline>),
917 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
918 : }
919 :
920 : impl TimelineInitAndSyncResult {
921 0 : fn ready_to_activate(self) -> Option<Arc<Timeline>> {
922 0 : match self {
923 0 : Self::ReadyToActivate(timeline) => Some(timeline),
924 0 : _ => None,
925 : }
926 0 : }
927 : }
928 :
929 : #[must_use]
930 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
931 : timeline: Arc<Timeline>,
932 : import_pgdata: import_pgdata::index_part_format::Root,
933 : guard: TimelineCreateGuard,
934 : }
935 :
936 : /// What is returned by [`Tenant::create_timeline`].
937 : enum CreateTimelineResult {
938 : Created(Arc<Timeline>),
939 : Idempotent(Arc<Timeline>),
940 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`Tenant::timelines`] when
941 : /// we return this result, nor will this concrete object ever be added there.
942 : /// Cf method comment on [`Tenant::create_timeline_import_pgdata`].
943 : ImportSpawned(Arc<Timeline>),
944 : }
945 :
946 : impl CreateTimelineResult {
947 0 : fn discriminant(&self) -> &'static str {
948 0 : match self {
949 0 : Self::Created(_) => "Created",
950 0 : Self::Idempotent(_) => "Idempotent",
951 0 : Self::ImportSpawned(_) => "ImportSpawned",
952 : }
953 0 : }
954 0 : fn timeline(&self) -> &Arc<Timeline> {
955 0 : match self {
956 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
957 0 : }
958 0 : }
959 : /// Unit test timelines aren't activated, test has to do it if it needs to.
960 : #[cfg(test)]
961 460 : fn into_timeline_for_test(self) -> Arc<Timeline> {
962 460 : match self {
963 460 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
964 460 : }
965 460 : }
966 : }
967 :
968 : #[derive(thiserror::Error, Debug)]
969 : pub enum CreateTimelineError {
970 : #[error("creation of timeline with the given ID is in progress")]
971 : AlreadyCreating,
972 : #[error("timeline already exists with different parameters")]
973 : Conflict,
974 : #[error(transparent)]
975 : AncestorLsn(anyhow::Error),
976 : #[error("ancestor timeline is not active")]
977 : AncestorNotActive,
978 : #[error("ancestor timeline is archived")]
979 : AncestorArchived,
980 : #[error("tenant shutting down")]
981 : ShuttingDown,
982 : #[error(transparent)]
983 : Other(#[from] anyhow::Error),
984 : }
985 :
986 : #[derive(thiserror::Error, Debug)]
987 : pub enum InitdbError {
988 : #[error("Operation was cancelled")]
989 : Cancelled,
990 : #[error(transparent)]
991 : Other(anyhow::Error),
992 : #[error(transparent)]
993 : Inner(postgres_initdb::Error),
994 : }
995 :
996 : enum CreateTimelineCause {
997 : Load,
998 : Delete,
999 : }
1000 :
1001 : enum LoadTimelineCause {
1002 : Attach,
1003 : Unoffload,
1004 : ImportPgdata {
1005 : create_guard: TimelineCreateGuard,
1006 : activate: ActivateTimelineArgs,
1007 : },
1008 : }
1009 :
1010 : #[derive(thiserror::Error, Debug)]
1011 : pub(crate) enum GcError {
1012 : // The tenant is shutting down
1013 : #[error("tenant shutting down")]
1014 : TenantCancelled,
1015 :
1016 : // The tenant is shutting down
1017 : #[error("timeline shutting down")]
1018 : TimelineCancelled,
1019 :
1020 : // The tenant is in a state inelegible to run GC
1021 : #[error("not active")]
1022 : NotActive,
1023 :
1024 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1025 : #[error("not active")]
1026 : BadLsn { why: String },
1027 :
1028 : // A remote storage error while scheduling updates after compaction
1029 : #[error(transparent)]
1030 : Remote(anyhow::Error),
1031 :
1032 : // An error reading while calculating GC cutoffs
1033 : #[error(transparent)]
1034 : GcCutoffs(PageReconstructError),
1035 :
1036 : // If GC was invoked for a particular timeline, this error means it didn't exist
1037 : #[error("timeline not found")]
1038 : TimelineNotFound,
1039 : }
1040 :
1041 : impl From<PageReconstructError> for GcError {
1042 0 : fn from(value: PageReconstructError) -> Self {
1043 0 : match value {
1044 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1045 0 : other => Self::GcCutoffs(other),
1046 : }
1047 0 : }
1048 : }
1049 :
1050 : impl From<NotInitialized> for GcError {
1051 0 : fn from(value: NotInitialized) -> Self {
1052 0 : match value {
1053 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1054 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1055 : }
1056 0 : }
1057 : }
1058 :
1059 : impl From<timeline::layer_manager::Shutdown> for GcError {
1060 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1061 0 : GcError::TimelineCancelled
1062 0 : }
1063 : }
1064 :
1065 : #[derive(thiserror::Error, Debug)]
1066 : pub(crate) enum LoadConfigError {
1067 : #[error("TOML deserialization error: '{0}'")]
1068 : DeserializeToml(#[from] toml_edit::de::Error),
1069 :
1070 : #[error("Config not found at {0}")]
1071 : NotFound(Utf8PathBuf),
1072 : }
1073 :
1074 : impl Tenant {
1075 : /// Yet another helper for timeline initialization.
1076 : ///
1077 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1078 : /// - Scans the local timeline directory for layer files and builds the layer map
1079 : /// - Downloads remote index file and adds remote files to the layer map
1080 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1081 : ///
1082 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1083 : /// it is marked as Active.
1084 : #[allow(clippy::too_many_arguments)]
1085 12 : async fn timeline_init_and_sync(
1086 12 : self: &Arc<Self>,
1087 12 : timeline_id: TimelineId,
1088 12 : resources: TimelineResources,
1089 12 : mut index_part: IndexPart,
1090 12 : metadata: TimelineMetadata,
1091 12 : previous_heatmap: Option<PreviousHeatmap>,
1092 12 : ancestor: Option<Arc<Timeline>>,
1093 12 : cause: LoadTimelineCause,
1094 12 : ctx: &RequestContext,
1095 12 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1096 12 : let tenant_id = self.tenant_shard_id;
1097 12 :
1098 12 : let import_pgdata = index_part.import_pgdata.take();
1099 12 : let idempotency = match &import_pgdata {
1100 0 : Some(import_pgdata) => {
1101 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1102 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1103 0 : })
1104 : }
1105 : None => {
1106 12 : if metadata.ancestor_timeline().is_none() {
1107 8 : CreateTimelineIdempotency::Bootstrap {
1108 8 : pg_version: metadata.pg_version(),
1109 8 : }
1110 : } else {
1111 4 : CreateTimelineIdempotency::Branch {
1112 4 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1113 4 : ancestor_start_lsn: metadata.ancestor_lsn(),
1114 4 : }
1115 : }
1116 : }
1117 : };
1118 :
1119 12 : let (timeline, timeline_ctx) = self.create_timeline_struct(
1120 12 : timeline_id,
1121 12 : &metadata,
1122 12 : previous_heatmap,
1123 12 : ancestor.clone(),
1124 12 : resources,
1125 12 : CreateTimelineCause::Load,
1126 12 : idempotency.clone(),
1127 12 : index_part.gc_compaction.clone(),
1128 12 : index_part.rel_size_migration.clone(),
1129 12 : ctx,
1130 12 : )?;
1131 12 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1132 12 : anyhow::ensure!(
1133 12 : disk_consistent_lsn.is_valid(),
1134 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
1135 : );
1136 12 : assert_eq!(
1137 12 : disk_consistent_lsn,
1138 12 : metadata.disk_consistent_lsn(),
1139 0 : "these are used interchangeably"
1140 : );
1141 :
1142 12 : timeline.remote_client.init_upload_queue(&index_part)?;
1143 :
1144 12 : timeline
1145 12 : .load_layer_map(disk_consistent_lsn, index_part)
1146 12 : .await
1147 12 : .with_context(|| {
1148 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1149 12 : })?;
1150 :
1151 : // When unarchiving, we've mostly likely lost the heatmap generated prior
1152 : // to the archival operation. To allow warming this timeline up, generate
1153 : // a previous heatmap which contains all visible layers in the layer map.
1154 : // This previous heatmap will be used whenever a fresh heatmap is generated
1155 : // for the timeline.
1156 12 : if self.conf.generate_unarchival_heatmap && matches!(cause, LoadTimelineCause::Unoffload) {
1157 0 : let mut tline_ending_at = Some((&timeline, timeline.get_last_record_lsn()));
1158 0 : while let Some((tline, end_lsn)) = tline_ending_at {
1159 0 : let unarchival_heatmap = tline.generate_unarchival_heatmap(end_lsn).await;
1160 : // Another unearchived timeline might have generated a heatmap for this ancestor.
1161 : // If the current branch point greater than the previous one use the the heatmap
1162 : // we just generated - it should include more layers.
1163 0 : if !tline.should_keep_previous_heatmap(end_lsn) {
1164 0 : tline
1165 0 : .previous_heatmap
1166 0 : .store(Some(Arc::new(unarchival_heatmap)));
1167 0 : } else {
1168 0 : tracing::info!("Previous heatmap preferred. Dropping unarchival heatmap.")
1169 : }
1170 :
1171 0 : match tline.ancestor_timeline() {
1172 0 : Some(ancestor) => {
1173 0 : if ancestor.update_layer_visibility().await.is_err() {
1174 : // Ancestor timeline is shutting down.
1175 0 : break;
1176 0 : }
1177 0 :
1178 0 : tline_ending_at = Some((ancestor, tline.get_ancestor_lsn()));
1179 : }
1180 0 : None => {
1181 0 : tline_ending_at = None;
1182 0 : }
1183 : }
1184 : }
1185 12 : }
1186 :
1187 0 : match import_pgdata {
1188 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1189 0 : match cause {
1190 0 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1191 : LoadTimelineCause::ImportPgdata { .. } => {
1192 0 : unreachable!(
1193 0 : "ImportPgdata should not be reloading timeline import is done and persisted as such in s3"
1194 0 : )
1195 : }
1196 : }
1197 0 : let mut guard = self.timelines_creating.lock().unwrap();
1198 0 : if !guard.insert(timeline_id) {
1199 : // We should never try and load the same timeline twice during startup
1200 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1201 0 : }
1202 0 : let timeline_create_guard = TimelineCreateGuard {
1203 0 : _tenant_gate_guard: self.gate.enter()?,
1204 0 : owning_tenant: self.clone(),
1205 0 : timeline_id,
1206 0 : idempotency,
1207 0 : // The users of this specific return value don't need the timline_path in there.
1208 0 : timeline_path: timeline
1209 0 : .conf
1210 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1211 0 : };
1212 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1213 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1214 0 : timeline,
1215 0 : import_pgdata,
1216 0 : guard: timeline_create_guard,
1217 0 : },
1218 0 : ))
1219 : }
1220 : Some(_) | None => {
1221 : {
1222 12 : let mut timelines_accessor = self.timelines.lock().unwrap();
1223 12 : match timelines_accessor.entry(timeline_id) {
1224 : // We should never try and load the same timeline twice during startup
1225 : Entry::Occupied(_) => {
1226 0 : unreachable!(
1227 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1228 0 : );
1229 : }
1230 12 : Entry::Vacant(v) => {
1231 12 : v.insert(Arc::clone(&timeline));
1232 12 : timeline.maybe_spawn_flush_loop();
1233 12 : }
1234 : }
1235 : }
1236 :
1237 : // Sanity check: a timeline should have some content.
1238 12 : anyhow::ensure!(
1239 12 : ancestor.is_some()
1240 8 : || timeline
1241 8 : .layers
1242 8 : .read()
1243 8 : .await
1244 8 : .layer_map()
1245 8 : .expect("currently loading, layer manager cannot be shutdown already")
1246 8 : .iter_historic_layers()
1247 8 : .next()
1248 8 : .is_some(),
1249 0 : "Timeline has no ancestor and no layer files"
1250 : );
1251 :
1252 12 : match cause {
1253 12 : LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
1254 : LoadTimelineCause::ImportPgdata {
1255 0 : create_guard,
1256 0 : activate,
1257 0 : } => {
1258 0 : // TODO: see the comment in the task code above how I'm not so certain
1259 0 : // it is safe to activate here because of concurrent shutdowns.
1260 0 : match activate {
1261 0 : ActivateTimelineArgs::Yes { broker_client } => {
1262 0 : info!("activating timeline after reload from pgdata import task");
1263 0 : timeline.activate(self.clone(), broker_client, None, &timeline_ctx);
1264 : }
1265 0 : ActivateTimelineArgs::No => (),
1266 : }
1267 0 : drop(create_guard);
1268 : }
1269 : }
1270 :
1271 12 : Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
1272 : }
1273 : }
1274 12 : }
1275 :
1276 : /// Attach a tenant that's available in cloud storage.
1277 : ///
1278 : /// This returns quickly, after just creating the in-memory object
1279 : /// Tenant struct and launching a background task to download
1280 : /// the remote index files. On return, the tenant is most likely still in
1281 : /// Attaching state, and it will become Active once the background task
1282 : /// finishes. You can use wait_until_active() to wait for the task to
1283 : /// complete.
1284 : ///
1285 : #[allow(clippy::too_many_arguments)]
1286 0 : pub(crate) fn spawn(
1287 0 : conf: &'static PageServerConf,
1288 0 : tenant_shard_id: TenantShardId,
1289 0 : resources: TenantSharedResources,
1290 0 : attached_conf: AttachedTenantConf,
1291 0 : shard_identity: ShardIdentity,
1292 0 : init_order: Option<InitializationOrder>,
1293 0 : mode: SpawnMode,
1294 0 : ctx: &RequestContext,
1295 0 : ) -> Result<Arc<Tenant>, GlobalShutDown> {
1296 0 : let wal_redo_manager =
1297 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1298 :
1299 : let TenantSharedResources {
1300 0 : broker_client,
1301 0 : remote_storage,
1302 0 : deletion_queue_client,
1303 0 : l0_flush_global_state,
1304 0 : } = resources;
1305 0 :
1306 0 : let attach_mode = attached_conf.location.attach_mode;
1307 0 : let generation = attached_conf.location.generation;
1308 0 :
1309 0 : let tenant = Arc::new(Tenant::new(
1310 0 : TenantState::Attaching,
1311 0 : conf,
1312 0 : attached_conf,
1313 0 : shard_identity,
1314 0 : Some(wal_redo_manager),
1315 0 : tenant_shard_id,
1316 0 : remote_storage.clone(),
1317 0 : deletion_queue_client,
1318 0 : l0_flush_global_state,
1319 0 : ));
1320 0 :
1321 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1322 0 : // we shut down while attaching.
1323 0 : let attach_gate_guard = tenant
1324 0 : .gate
1325 0 : .enter()
1326 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
1327 0 :
1328 0 : // Do all the hard work in the background
1329 0 : let tenant_clone = Arc::clone(&tenant);
1330 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1331 0 : task_mgr::spawn(
1332 0 : &tokio::runtime::Handle::current(),
1333 0 : TaskKind::Attach,
1334 0 : tenant_shard_id,
1335 0 : None,
1336 0 : "attach tenant",
1337 0 : async move {
1338 0 :
1339 0 : info!(
1340 : ?attach_mode,
1341 0 : "Attaching tenant"
1342 : );
1343 :
1344 0 : let _gate_guard = attach_gate_guard;
1345 0 :
1346 0 : // Is this tenant being spawned as part of process startup?
1347 0 : let starting_up = init_order.is_some();
1348 0 : scopeguard::defer! {
1349 0 : if starting_up {
1350 0 : TENANT.startup_complete.inc();
1351 0 : }
1352 0 : }
1353 :
1354 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
1355 : enum BrokenVerbosity {
1356 : Error,
1357 : Info
1358 : }
1359 0 : let make_broken =
1360 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
1361 0 : match verbosity {
1362 : BrokenVerbosity::Info => {
1363 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
1364 : },
1365 : BrokenVerbosity::Error => {
1366 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
1367 : }
1368 : }
1369 0 : t.state.send_modify(|state| {
1370 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
1371 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
1372 0 : assert!(
1373 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
1374 0 : "the attach task owns the tenant state until activation is complete"
1375 : );
1376 :
1377 0 : *state = TenantState::broken_from_reason(err.to_string());
1378 0 : });
1379 0 : };
1380 :
1381 : // TODO: should also be rejecting tenant conf changes that violate this check.
1382 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1383 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1384 0 : return Ok(());
1385 0 : }
1386 0 :
1387 0 : let mut init_order = init_order;
1388 0 : // take the completion because initial tenant loading will complete when all of
1389 0 : // these tasks complete.
1390 0 : let _completion = init_order
1391 0 : .as_mut()
1392 0 : .and_then(|x| x.initial_tenant_load.take());
1393 0 : let remote_load_completion = init_order
1394 0 : .as_mut()
1395 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1396 :
1397 : enum AttachType<'a> {
1398 : /// We are attaching this tenant lazily in the background.
1399 : Warmup {
1400 : _permit: tokio::sync::SemaphorePermit<'a>,
1401 : during_startup: bool
1402 : },
1403 : /// We are attaching this tenant as soon as we can, because for example an
1404 : /// endpoint tried to access it.
1405 : OnDemand,
1406 : /// During normal operations after startup, we are attaching a tenant, and
1407 : /// eager attach was requested.
1408 : Normal,
1409 : }
1410 :
1411 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1412 : // Before doing any I/O, wait for at least one of:
1413 : // - A client attempting to access to this tenant (on-demand loading)
1414 : // - A permit becoming available in the warmup semaphore (background warmup)
1415 :
1416 0 : tokio::select!(
1417 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1418 0 : let _ = permit.expect("activate_now_sem is never closed");
1419 0 : tracing::info!("Activating tenant (on-demand)");
1420 0 : AttachType::OnDemand
1421 : },
1422 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1423 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1424 0 : tracing::info!("Activating tenant (warmup)");
1425 0 : AttachType::Warmup {
1426 0 : _permit,
1427 0 : during_startup: init_order.is_some()
1428 0 : }
1429 : }
1430 0 : _ = tenant_clone.cancel.cancelled() => {
1431 : // This is safe, but should be pretty rare: it is interesting if a tenant
1432 : // stayed in Activating for such a long time that shutdown found it in
1433 : // that state.
1434 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1435 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
1436 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
1437 : // just shutting down), but ensures progress.
1438 0 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
1439 0 : return Ok(());
1440 : },
1441 : )
1442 : } else {
1443 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1444 : // concurrent_tenant_warmup queue
1445 0 : AttachType::Normal
1446 : };
1447 :
1448 0 : let preload = match &mode {
1449 : SpawnMode::Eager | SpawnMode::Lazy => {
1450 0 : let _preload_timer = TENANT.preload.start_timer();
1451 0 : let res = tenant_clone
1452 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1453 0 : .await;
1454 0 : match res {
1455 0 : Ok(p) => Some(p),
1456 0 : Err(e) => {
1457 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1458 0 : return Ok(());
1459 : }
1460 : }
1461 : }
1462 :
1463 : };
1464 :
1465 : // Remote preload is complete.
1466 0 : drop(remote_load_completion);
1467 0 :
1468 0 :
1469 0 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1470 0 : let attach_start = std::time::Instant::now();
1471 0 : let attached = {
1472 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1473 0 : tenant_clone.attach(preload, &ctx).await
1474 : };
1475 0 : let attach_duration = attach_start.elapsed();
1476 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1477 0 :
1478 0 : match attached {
1479 : Ok(()) => {
1480 0 : info!("attach finished, activating");
1481 0 : tenant_clone.activate(broker_client, None, &ctx);
1482 : }
1483 0 : Err(e) => {
1484 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
1485 0 : }
1486 : }
1487 :
1488 : // If we are doing an opportunistic warmup attachment at startup, initialize
1489 : // logical size at the same time. This is better than starting a bunch of idle tenants
1490 : // with cold caches and then coming back later to initialize their logical sizes.
1491 : //
1492 : // It also prevents the warmup proccess competing with the concurrency limit on
1493 : // logical size calculations: if logical size calculation semaphore is saturated,
1494 : // then warmup will wait for that before proceeding to the next tenant.
1495 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1496 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1497 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1498 0 : while futs.next().await.is_some() {}
1499 0 : tracing::info!("Warm-up complete");
1500 0 : }
1501 :
1502 0 : Ok(())
1503 0 : }
1504 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1505 : );
1506 0 : Ok(tenant)
1507 0 : }
1508 :
1509 : #[instrument(skip_all)]
1510 : pub(crate) async fn preload(
1511 : self: &Arc<Self>,
1512 : remote_storage: &GenericRemoteStorage,
1513 : cancel: CancellationToken,
1514 : ) -> anyhow::Result<TenantPreload> {
1515 : span::debug_assert_current_span_has_tenant_id();
1516 : // Get list of remote timelines
1517 : // download index files for every tenant timeline
1518 : info!("listing remote timelines");
1519 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1520 : remote_storage,
1521 : self.tenant_shard_id,
1522 : cancel.clone(),
1523 : )
1524 : .await?;
1525 : let (offloaded_add, tenant_manifest) =
1526 : match remote_timeline_client::download_tenant_manifest(
1527 : remote_storage,
1528 : &self.tenant_shard_id,
1529 : self.generation,
1530 : &cancel,
1531 : )
1532 : .await
1533 : {
1534 : Ok((tenant_manifest, _generation, _manifest_mtime)) => (
1535 : format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
1536 : tenant_manifest,
1537 : ),
1538 : Err(DownloadError::NotFound) => {
1539 : ("no manifest".to_string(), TenantManifest::empty())
1540 : }
1541 : Err(e) => Err(e)?,
1542 : };
1543 :
1544 : info!(
1545 : "found {} timelines, and {offloaded_add}",
1546 : remote_timeline_ids.len()
1547 : );
1548 :
1549 : for k in other_keys {
1550 : warn!("Unexpected non timeline key {k}");
1551 : }
1552 :
1553 : // Avoid downloading IndexPart of offloaded timelines.
1554 : let mut offloaded_with_prefix = HashSet::new();
1555 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1556 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1557 : offloaded_with_prefix.insert(offloaded.timeline_id);
1558 : } else {
1559 : // We'll take care later of timelines in the manifest without a prefix
1560 : }
1561 : }
1562 :
1563 : // TODO(vlad): Could go to S3 if the secondary is freezing cold and hasn't even
1564 : // pulled the first heatmap. Not entirely necessary since the storage controller
1565 : // will kick the secondary in any case and cause a download.
1566 : let maybe_heatmap_at = self.read_on_disk_heatmap().await;
1567 :
1568 : let timelines = self
1569 : .load_timelines_metadata(
1570 : remote_timeline_ids,
1571 : remote_storage,
1572 : maybe_heatmap_at,
1573 : cancel,
1574 : )
1575 : .await?;
1576 :
1577 : Ok(TenantPreload {
1578 : tenant_manifest,
1579 : timelines: timelines
1580 : .into_iter()
1581 12 : .map(|(id, tl)| (id, Some(tl)))
1582 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1583 : .collect(),
1584 : })
1585 : }
1586 :
1587 452 : async fn read_on_disk_heatmap(&self) -> Option<(HeatMapTenant, std::time::Instant)> {
1588 452 : if !self.conf.load_previous_heatmap {
1589 0 : return None;
1590 452 : }
1591 452 :
1592 452 : let on_disk_heatmap_path = self.conf.tenant_heatmap_path(&self.tenant_shard_id);
1593 452 : match tokio::fs::read_to_string(on_disk_heatmap_path).await {
1594 0 : Ok(heatmap) => match serde_json::from_str::<HeatMapTenant>(&heatmap) {
1595 0 : Ok(heatmap) => Some((heatmap, std::time::Instant::now())),
1596 0 : Err(err) => {
1597 0 : error!("Failed to deserialize old heatmap: {err}");
1598 0 : None
1599 : }
1600 : },
1601 452 : Err(err) => match err.kind() {
1602 452 : std::io::ErrorKind::NotFound => None,
1603 : _ => {
1604 0 : error!("Unexpected IO error reading old heatmap: {err}");
1605 0 : None
1606 : }
1607 : },
1608 : }
1609 452 : }
1610 :
1611 : ///
1612 : /// Background task that downloads all data for a tenant and brings it to Active state.
1613 : ///
1614 : /// No background tasks are started as part of this routine.
1615 : ///
1616 452 : async fn attach(
1617 452 : self: &Arc<Tenant>,
1618 452 : preload: Option<TenantPreload>,
1619 452 : ctx: &RequestContext,
1620 452 : ) -> anyhow::Result<()> {
1621 452 : span::debug_assert_current_span_has_tenant_id();
1622 452 :
1623 452 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1624 :
1625 452 : let Some(preload) = preload else {
1626 0 : anyhow::bail!(
1627 0 : "local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624"
1628 0 : );
1629 : };
1630 :
1631 452 : let mut offloaded_timeline_ids = HashSet::new();
1632 452 : let mut offloaded_timelines_list = Vec::new();
1633 452 : for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
1634 0 : let timeline_id = timeline_manifest.timeline_id;
1635 0 : let offloaded_timeline =
1636 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1637 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1638 0 : offloaded_timeline_ids.insert(timeline_id);
1639 0 : }
1640 : // Complete deletions for offloaded timeline id's from manifest.
1641 : // The manifest will be uploaded later in this function.
1642 452 : offloaded_timelines_list
1643 452 : .retain(|(offloaded_id, offloaded)| {
1644 0 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1645 0 : // If there is dangling references in another location, they need to be cleaned up.
1646 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1647 0 : if delete {
1648 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1649 0 : offloaded.defuse_for_tenant_drop();
1650 0 : }
1651 0 : !delete
1652 452 : });
1653 452 :
1654 452 : let mut timelines_to_resume_deletions = vec![];
1655 452 :
1656 452 : let mut remote_index_and_client = HashMap::new();
1657 452 : let mut timeline_ancestors = HashMap::new();
1658 452 : let mut existent_timelines = HashSet::new();
1659 464 : for (timeline_id, preload) in preload.timelines {
1660 12 : let Some(preload) = preload else { continue };
1661 : // This is an invariant of the `preload` function's API
1662 12 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1663 12 : let index_part = match preload.index_part {
1664 12 : Ok(i) => {
1665 12 : debug!("remote index part exists for timeline {timeline_id}");
1666 : // We found index_part on the remote, this is the standard case.
1667 12 : existent_timelines.insert(timeline_id);
1668 12 : i
1669 : }
1670 : Err(DownloadError::NotFound) => {
1671 : // There is no index_part on the remote. We only get here
1672 : // if there is some prefix for the timeline in the remote storage.
1673 : // This can e.g. be the initdb.tar.zst archive, maybe a
1674 : // remnant from a prior incomplete creation or deletion attempt.
1675 : // Delete the local directory as the deciding criterion for a
1676 : // timeline's existence is presence of index_part.
1677 0 : info!(%timeline_id, "index_part not found on remote");
1678 0 : continue;
1679 : }
1680 0 : Err(DownloadError::Fatal(why)) => {
1681 0 : // If, while loading one remote timeline, we saw an indication that our generation
1682 0 : // number is likely invalid, then we should not load the whole tenant.
1683 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1684 0 : anyhow::bail!(why.to_string());
1685 : }
1686 0 : Err(e) => {
1687 0 : // Some (possibly ephemeral) error happened during index_part download.
1688 0 : // Pretend the timeline exists to not delete the timeline directory,
1689 0 : // as it might be a temporary issue and we don't want to re-download
1690 0 : // everything after it resolves.
1691 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1692 :
1693 0 : existent_timelines.insert(timeline_id);
1694 0 : continue;
1695 : }
1696 : };
1697 12 : match index_part {
1698 12 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1699 12 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1700 12 : remote_index_and_client.insert(
1701 12 : timeline_id,
1702 12 : (index_part, preload.client, preload.previous_heatmap),
1703 12 : );
1704 12 : }
1705 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1706 0 : info!(
1707 0 : "timeline {} is deleted, picking to resume deletion",
1708 : timeline_id
1709 : );
1710 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1711 : }
1712 : }
1713 : }
1714 :
1715 452 : let mut gc_blocks = HashMap::new();
1716 :
1717 : // For every timeline, download the metadata file, scan the local directory,
1718 : // and build a layer map that contains an entry for each remote and local
1719 : // layer file.
1720 452 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1721 464 : for (timeline_id, remote_metadata) in sorted_timelines {
1722 12 : let (index_part, remote_client, previous_heatmap) = remote_index_and_client
1723 12 : .remove(&timeline_id)
1724 12 : .expect("just put it in above");
1725 :
1726 12 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1727 : // could just filter these away, but it helps while testing
1728 0 : anyhow::ensure!(
1729 0 : !blocking.reasons.is_empty(),
1730 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1731 : );
1732 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1733 0 : assert!(prev.is_none());
1734 12 : }
1735 :
1736 : // TODO again handle early failure
1737 12 : let effect = self
1738 12 : .load_remote_timeline(
1739 12 : timeline_id,
1740 12 : index_part,
1741 12 : remote_metadata,
1742 12 : previous_heatmap,
1743 12 : self.get_timeline_resources_for(remote_client),
1744 12 : LoadTimelineCause::Attach,
1745 12 : ctx,
1746 12 : )
1747 12 : .await
1748 12 : .with_context(|| {
1749 0 : format!(
1750 0 : "failed to load remote timeline {} for tenant {}",
1751 0 : timeline_id, self.tenant_shard_id
1752 0 : )
1753 12 : })?;
1754 :
1755 12 : match effect {
1756 12 : TimelineInitAndSyncResult::ReadyToActivate(_) => {
1757 12 : // activation happens later, on Tenant::activate
1758 12 : }
1759 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1760 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1761 0 : timeline,
1762 0 : import_pgdata,
1763 0 : guard,
1764 0 : },
1765 0 : ) => {
1766 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1767 0 : timeline,
1768 0 : import_pgdata,
1769 0 : ActivateTimelineArgs::No,
1770 0 : guard,
1771 0 : ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
1772 0 : ));
1773 0 : }
1774 : }
1775 : }
1776 :
1777 : // Walk through deleted timelines, resume deletion
1778 452 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1779 0 : remote_timeline_client
1780 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1781 0 : .context("init queue stopped")
1782 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1783 :
1784 0 : DeleteTimelineFlow::resume_deletion(
1785 0 : Arc::clone(self),
1786 0 : timeline_id,
1787 0 : &index_part.metadata,
1788 0 : remote_timeline_client,
1789 0 : ctx,
1790 0 : )
1791 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1792 0 : .await
1793 0 : .context("resume_deletion")
1794 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1795 : }
1796 452 : let needs_manifest_upload =
1797 452 : offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
1798 452 : {
1799 452 : let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
1800 452 : offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
1801 452 : }
1802 452 : if needs_manifest_upload {
1803 0 : self.store_tenant_manifest().await?;
1804 452 : }
1805 :
1806 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1807 : // IndexPart is the source of truth.
1808 452 : self.clean_up_timelines(&existent_timelines)?;
1809 :
1810 452 : self.gc_block.set_scanned(gc_blocks);
1811 452 :
1812 452 : fail::fail_point!("attach-before-activate", |_| {
1813 0 : anyhow::bail!("attach-before-activate");
1814 452 : });
1815 452 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1816 :
1817 452 : info!("Done");
1818 :
1819 452 : Ok(())
1820 452 : }
1821 :
1822 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1823 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1824 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1825 452 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1826 452 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1827 :
1828 452 : let entries = match timelines_dir.read_dir_utf8() {
1829 452 : Ok(d) => d,
1830 0 : Err(e) => {
1831 0 : if e.kind() == std::io::ErrorKind::NotFound {
1832 0 : return Ok(());
1833 : } else {
1834 0 : return Err(e).context("list timelines directory for tenant");
1835 : }
1836 : }
1837 : };
1838 :
1839 468 : for entry in entries {
1840 16 : let entry = entry.context("read timeline dir entry")?;
1841 16 : let entry_path = entry.path();
1842 :
1843 16 : let purge = if crate::is_temporary(entry_path) {
1844 0 : true
1845 : } else {
1846 16 : match TimelineId::try_from(entry_path.file_name()) {
1847 16 : Ok(i) => {
1848 16 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1849 16 : !existent_timelines.contains(&i)
1850 : }
1851 0 : Err(e) => {
1852 0 : tracing::warn!(
1853 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1854 : );
1855 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1856 0 : false
1857 : }
1858 : }
1859 : };
1860 :
1861 16 : if purge {
1862 4 : tracing::info!("Purging stale timeline dentry {entry_path}");
1863 4 : if let Err(e) = match entry.file_type() {
1864 4 : Ok(t) => if t.is_dir() {
1865 4 : std::fs::remove_dir_all(entry_path)
1866 : } else {
1867 0 : std::fs::remove_file(entry_path)
1868 : }
1869 4 : .or_else(fs_ext::ignore_not_found),
1870 0 : Err(e) => Err(e),
1871 : } {
1872 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1873 4 : }
1874 12 : }
1875 : }
1876 :
1877 452 : Ok(())
1878 452 : }
1879 :
1880 : /// Get sum of all remote timelines sizes
1881 : ///
1882 : /// This function relies on the index_part instead of listing the remote storage
1883 0 : pub fn remote_size(&self) -> u64 {
1884 0 : let mut size = 0;
1885 :
1886 0 : for timeline in self.list_timelines() {
1887 0 : size += timeline.remote_client.get_remote_physical_size();
1888 0 : }
1889 :
1890 0 : size
1891 0 : }
1892 :
1893 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1894 : #[allow(clippy::too_many_arguments)]
1895 : async fn load_remote_timeline(
1896 : self: &Arc<Self>,
1897 : timeline_id: TimelineId,
1898 : index_part: IndexPart,
1899 : remote_metadata: TimelineMetadata,
1900 : previous_heatmap: Option<PreviousHeatmap>,
1901 : resources: TimelineResources,
1902 : cause: LoadTimelineCause,
1903 : ctx: &RequestContext,
1904 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1905 : span::debug_assert_current_span_has_tenant_id();
1906 :
1907 : info!("downloading index file for timeline {}", timeline_id);
1908 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1909 : .await
1910 : .context("Failed to create new timeline directory")?;
1911 :
1912 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1913 : let timelines = self.timelines.lock().unwrap();
1914 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1915 0 : || {
1916 0 : anyhow::anyhow!(
1917 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1918 0 : )
1919 0 : },
1920 : )?))
1921 : } else {
1922 : None
1923 : };
1924 :
1925 : self.timeline_init_and_sync(
1926 : timeline_id,
1927 : resources,
1928 : index_part,
1929 : remote_metadata,
1930 : previous_heatmap,
1931 : ancestor,
1932 : cause,
1933 : ctx,
1934 : )
1935 : .await
1936 : }
1937 :
1938 452 : async fn load_timelines_metadata(
1939 452 : self: &Arc<Tenant>,
1940 452 : timeline_ids: HashSet<TimelineId>,
1941 452 : remote_storage: &GenericRemoteStorage,
1942 452 : heatmap: Option<(HeatMapTenant, std::time::Instant)>,
1943 452 : cancel: CancellationToken,
1944 452 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1945 452 : let mut timeline_heatmaps = heatmap.map(|h| (h.0.into_timelines_index(), h.1));
1946 452 :
1947 452 : let mut part_downloads = JoinSet::new();
1948 464 : for timeline_id in timeline_ids {
1949 12 : let cancel_clone = cancel.clone();
1950 12 :
1951 12 : let previous_timeline_heatmap = timeline_heatmaps.as_mut().and_then(|hs| {
1952 0 : hs.0.remove(&timeline_id).map(|h| PreviousHeatmap::Active {
1953 0 : heatmap: h,
1954 0 : read_at: hs.1,
1955 0 : end_lsn: None,
1956 0 : })
1957 12 : });
1958 12 : part_downloads.spawn(
1959 12 : self.load_timeline_metadata(
1960 12 : timeline_id,
1961 12 : remote_storage.clone(),
1962 12 : previous_timeline_heatmap,
1963 12 : cancel_clone,
1964 12 : )
1965 12 : .instrument(info_span!("download_index_part", %timeline_id)),
1966 : );
1967 : }
1968 :
1969 452 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1970 :
1971 : loop {
1972 464 : tokio::select!(
1973 464 : next = part_downloads.join_next() => {
1974 464 : match next {
1975 12 : Some(result) => {
1976 12 : let preload = result.context("join preload task")?;
1977 12 : timeline_preloads.insert(preload.timeline_id, preload);
1978 : },
1979 : None => {
1980 452 : break;
1981 : }
1982 : }
1983 : },
1984 464 : _ = cancel.cancelled() => {
1985 0 : anyhow::bail!("Cancelled while waiting for remote index download")
1986 : }
1987 : )
1988 : }
1989 :
1990 452 : Ok(timeline_preloads)
1991 452 : }
1992 :
1993 12 : fn build_timeline_client(
1994 12 : &self,
1995 12 : timeline_id: TimelineId,
1996 12 : remote_storage: GenericRemoteStorage,
1997 12 : ) -> RemoteTimelineClient {
1998 12 : RemoteTimelineClient::new(
1999 12 : remote_storage.clone(),
2000 12 : self.deletion_queue_client.clone(),
2001 12 : self.conf,
2002 12 : self.tenant_shard_id,
2003 12 : timeline_id,
2004 12 : self.generation,
2005 12 : &self.tenant_conf.load().location,
2006 12 : )
2007 12 : }
2008 :
2009 12 : fn load_timeline_metadata(
2010 12 : self: &Arc<Tenant>,
2011 12 : timeline_id: TimelineId,
2012 12 : remote_storage: GenericRemoteStorage,
2013 12 : previous_heatmap: Option<PreviousHeatmap>,
2014 12 : cancel: CancellationToken,
2015 12 : ) -> impl Future<Output = TimelinePreload> + use<> {
2016 12 : let client = self.build_timeline_client(timeline_id, remote_storage);
2017 12 : async move {
2018 12 : debug_assert_current_span_has_tenant_and_timeline_id();
2019 12 : debug!("starting index part download");
2020 :
2021 12 : let index_part = client.download_index_file(&cancel).await;
2022 :
2023 12 : debug!("finished index part download");
2024 :
2025 12 : TimelinePreload {
2026 12 : client,
2027 12 : timeline_id,
2028 12 : index_part,
2029 12 : previous_heatmap,
2030 12 : }
2031 12 : }
2032 12 : }
2033 :
2034 0 : fn check_to_be_archived_has_no_unarchived_children(
2035 0 : timeline_id: TimelineId,
2036 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2037 0 : ) -> Result<(), TimelineArchivalError> {
2038 0 : let children: Vec<TimelineId> = timelines
2039 0 : .iter()
2040 0 : .filter_map(|(id, entry)| {
2041 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
2042 0 : return None;
2043 0 : }
2044 0 : if entry.is_archived() == Some(true) {
2045 0 : return None;
2046 0 : }
2047 0 : Some(*id)
2048 0 : })
2049 0 : .collect();
2050 0 :
2051 0 : if !children.is_empty() {
2052 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
2053 0 : }
2054 0 : Ok(())
2055 0 : }
2056 :
2057 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
2058 0 : ancestor_timeline_id: TimelineId,
2059 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2060 0 : offloaded_timelines: &std::sync::MutexGuard<
2061 0 : '_,
2062 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
2063 0 : >,
2064 0 : ) -> Result<(), TimelineArchivalError> {
2065 0 : let has_archived_parent =
2066 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2067 0 : ancestor_timeline.is_archived() == Some(true)
2068 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2069 0 : true
2070 : } else {
2071 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2072 0 : if cfg!(debug_assertions) {
2073 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2074 0 : }
2075 0 : return Err(TimelineArchivalError::NotFound);
2076 : };
2077 0 : if has_archived_parent {
2078 0 : return Err(TimelineArchivalError::HasArchivedParent(
2079 0 : ancestor_timeline_id,
2080 0 : ));
2081 0 : }
2082 0 : Ok(())
2083 0 : }
2084 :
2085 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2086 0 : timeline: &Arc<Timeline>,
2087 0 : ) -> Result<(), TimelineArchivalError> {
2088 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2089 0 : if ancestor_timeline.is_archived() == Some(true) {
2090 0 : return Err(TimelineArchivalError::HasArchivedParent(
2091 0 : ancestor_timeline.timeline_id,
2092 0 : ));
2093 0 : }
2094 0 : }
2095 0 : Ok(())
2096 0 : }
2097 :
2098 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2099 : ///
2100 : /// Counterpart to [`offload_timeline`].
2101 0 : async fn unoffload_timeline(
2102 0 : self: &Arc<Self>,
2103 0 : timeline_id: TimelineId,
2104 0 : broker_client: storage_broker::BrokerClientChannel,
2105 0 : ctx: RequestContext,
2106 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2107 0 : info!("unoffloading timeline");
2108 :
2109 : // We activate the timeline below manually, so this must be called on an active tenant.
2110 : // We expect callers of this function to ensure this.
2111 0 : match self.current_state() {
2112 : TenantState::Activating { .. }
2113 : | TenantState::Attaching
2114 : | TenantState::Broken { .. } => {
2115 0 : panic!("Timeline expected to be active")
2116 : }
2117 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2118 0 : TenantState::Active => {}
2119 0 : }
2120 0 : let cancel = self.cancel.clone();
2121 0 :
2122 0 : // Protect against concurrent attempts to use this TimelineId
2123 0 : // We don't care much about idempotency, as it's ensured a layer above.
2124 0 : let allow_offloaded = true;
2125 0 : let _create_guard = self
2126 0 : .create_timeline_create_guard(
2127 0 : timeline_id,
2128 0 : CreateTimelineIdempotency::FailWithConflict,
2129 0 : allow_offloaded,
2130 0 : )
2131 0 : .map_err(|err| match err {
2132 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2133 : TimelineExclusionError::AlreadyExists { .. } => {
2134 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2135 : }
2136 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2137 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2138 0 : })?;
2139 :
2140 0 : let timeline_preload = self
2141 0 : .load_timeline_metadata(
2142 0 : timeline_id,
2143 0 : self.remote_storage.clone(),
2144 0 : None,
2145 0 : cancel.clone(),
2146 0 : )
2147 0 : .await;
2148 :
2149 0 : let index_part = match timeline_preload.index_part {
2150 0 : Ok(index_part) => {
2151 0 : debug!("remote index part exists for timeline {timeline_id}");
2152 0 : index_part
2153 : }
2154 : Err(DownloadError::NotFound) => {
2155 0 : error!(%timeline_id, "index_part not found on remote");
2156 0 : return Err(TimelineArchivalError::NotFound);
2157 : }
2158 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2159 0 : Err(e) => {
2160 0 : // Some (possibly ephemeral) error happened during index_part download.
2161 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2162 0 : return Err(TimelineArchivalError::Other(
2163 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2164 0 : ));
2165 : }
2166 : };
2167 0 : let index_part = match index_part {
2168 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2169 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2170 0 : info!("timeline is deleted according to index_part.json");
2171 0 : return Err(TimelineArchivalError::NotFound);
2172 : }
2173 : };
2174 0 : let remote_metadata = index_part.metadata.clone();
2175 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2176 0 : self.load_remote_timeline(
2177 0 : timeline_id,
2178 0 : index_part,
2179 0 : remote_metadata,
2180 0 : None,
2181 0 : timeline_resources,
2182 0 : LoadTimelineCause::Unoffload,
2183 0 : &ctx,
2184 0 : )
2185 0 : .await
2186 0 : .with_context(|| {
2187 0 : format!(
2188 0 : "failed to load remote timeline {} for tenant {}",
2189 0 : timeline_id, self.tenant_shard_id
2190 0 : )
2191 0 : })
2192 0 : .map_err(TimelineArchivalError::Other)?;
2193 :
2194 0 : let timeline = {
2195 0 : let timelines = self.timelines.lock().unwrap();
2196 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2197 0 : warn!("timeline not available directly after attach");
2198 : // This is not a panic because no locks are held between `load_remote_timeline`
2199 : // which puts the timeline into timelines, and our look into the timeline map.
2200 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2201 0 : "timeline not available directly after attach"
2202 0 : )));
2203 : };
2204 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2205 0 : match offloaded_timelines.remove(&timeline_id) {
2206 0 : Some(offloaded) => {
2207 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2208 0 : }
2209 0 : None => warn!("timeline already removed from offloaded timelines"),
2210 : }
2211 :
2212 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2213 0 :
2214 0 : Arc::clone(timeline)
2215 0 : };
2216 0 :
2217 0 : // Upload new list of offloaded timelines to S3
2218 0 : self.store_tenant_manifest().await?;
2219 :
2220 : // Activate the timeline (if it makes sense)
2221 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2222 0 : let background_jobs_can_start = None;
2223 0 : timeline.activate(
2224 0 : self.clone(),
2225 0 : broker_client.clone(),
2226 0 : background_jobs_can_start,
2227 0 : &ctx.with_scope_timeline(&timeline),
2228 0 : );
2229 0 : }
2230 :
2231 0 : info!("timeline unoffloading complete");
2232 0 : Ok(timeline)
2233 0 : }
2234 :
2235 0 : pub(crate) async fn apply_timeline_archival_config(
2236 0 : self: &Arc<Self>,
2237 0 : timeline_id: TimelineId,
2238 0 : new_state: TimelineArchivalState,
2239 0 : broker_client: storage_broker::BrokerClientChannel,
2240 0 : ctx: RequestContext,
2241 0 : ) -> Result<(), TimelineArchivalError> {
2242 0 : info!("setting timeline archival config");
2243 : // First part: figure out what is needed to do, and do validation
2244 0 : let timeline_or_unarchive_offloaded = 'outer: {
2245 0 : let timelines = self.timelines.lock().unwrap();
2246 :
2247 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2248 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2249 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2250 0 : return Err(TimelineArchivalError::NotFound);
2251 : };
2252 0 : if new_state == TimelineArchivalState::Archived {
2253 : // It's offloaded already, so nothing to do
2254 0 : return Ok(());
2255 0 : }
2256 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2257 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2258 0 : ancestor_timeline_id,
2259 0 : &timelines,
2260 0 : &offloaded_timelines,
2261 0 : )?;
2262 0 : }
2263 0 : break 'outer None;
2264 : };
2265 :
2266 : // Do some validation. We release the timelines lock below, so there is potential
2267 : // for race conditions: these checks are more present to prevent misunderstandings of
2268 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2269 0 : match new_state {
2270 : TimelineArchivalState::Unarchived => {
2271 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2272 : }
2273 : TimelineArchivalState::Archived => {
2274 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2275 : }
2276 : }
2277 0 : Some(Arc::clone(timeline))
2278 : };
2279 :
2280 : // Second part: unoffload timeline (if needed)
2281 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2282 0 : timeline
2283 : } else {
2284 : // Turn offloaded timeline into a non-offloaded one
2285 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2286 0 : .await?
2287 : };
2288 :
2289 : // Third part: upload new timeline archival state and block until it is present in S3
2290 0 : let upload_needed = match timeline
2291 0 : .remote_client
2292 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2293 : {
2294 0 : Ok(upload_needed) => upload_needed,
2295 0 : Err(e) => {
2296 0 : if timeline.cancel.is_cancelled() {
2297 0 : return Err(TimelineArchivalError::Cancelled);
2298 : } else {
2299 0 : return Err(TimelineArchivalError::Other(e));
2300 : }
2301 : }
2302 : };
2303 :
2304 0 : if upload_needed {
2305 0 : info!("Uploading new state");
2306 : const MAX_WAIT: Duration = Duration::from_secs(10);
2307 0 : let Ok(v) =
2308 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2309 : else {
2310 0 : tracing::warn!("reached timeout for waiting on upload queue");
2311 0 : return Err(TimelineArchivalError::Timeout);
2312 : };
2313 0 : v.map_err(|e| match e {
2314 0 : WaitCompletionError::NotInitialized(e) => {
2315 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2316 : }
2317 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2318 0 : TimelineArchivalError::Cancelled
2319 : }
2320 0 : })?;
2321 0 : }
2322 0 : Ok(())
2323 0 : }
2324 :
2325 4 : pub fn get_offloaded_timeline(
2326 4 : &self,
2327 4 : timeline_id: TimelineId,
2328 4 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2329 4 : self.timelines_offloaded
2330 4 : .lock()
2331 4 : .unwrap()
2332 4 : .get(&timeline_id)
2333 4 : .map(Arc::clone)
2334 4 : .ok_or(GetTimelineError::NotFound {
2335 4 : tenant_id: self.tenant_shard_id,
2336 4 : timeline_id,
2337 4 : })
2338 4 : }
2339 :
2340 8 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2341 8 : self.tenant_shard_id
2342 8 : }
2343 :
2344 : /// Get Timeline handle for given Neon timeline ID.
2345 : /// This function is idempotent. It doesn't change internal state in any way.
2346 444 : pub fn get_timeline(
2347 444 : &self,
2348 444 : timeline_id: TimelineId,
2349 444 : active_only: bool,
2350 444 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2351 444 : let timelines_accessor = self.timelines.lock().unwrap();
2352 444 : let timeline = timelines_accessor
2353 444 : .get(&timeline_id)
2354 444 : .ok_or(GetTimelineError::NotFound {
2355 444 : tenant_id: self.tenant_shard_id,
2356 444 : timeline_id,
2357 444 : })?;
2358 :
2359 440 : if active_only && !timeline.is_active() {
2360 0 : Err(GetTimelineError::NotActive {
2361 0 : tenant_id: self.tenant_shard_id,
2362 0 : timeline_id,
2363 0 : state: timeline.current_state(),
2364 0 : })
2365 : } else {
2366 440 : Ok(Arc::clone(timeline))
2367 : }
2368 444 : }
2369 :
2370 : /// Lists timelines the tenant contains.
2371 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2372 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2373 0 : self.timelines
2374 0 : .lock()
2375 0 : .unwrap()
2376 0 : .values()
2377 0 : .map(Arc::clone)
2378 0 : .collect()
2379 0 : }
2380 :
2381 : /// Lists timelines the tenant manages, including offloaded ones.
2382 : ///
2383 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2384 0 : pub fn list_timelines_and_offloaded(
2385 0 : &self,
2386 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2387 0 : let timelines = self
2388 0 : .timelines
2389 0 : .lock()
2390 0 : .unwrap()
2391 0 : .values()
2392 0 : .map(Arc::clone)
2393 0 : .collect();
2394 0 : let offloaded = self
2395 0 : .timelines_offloaded
2396 0 : .lock()
2397 0 : .unwrap()
2398 0 : .values()
2399 0 : .map(Arc::clone)
2400 0 : .collect();
2401 0 : (timelines, offloaded)
2402 0 : }
2403 :
2404 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2405 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2406 0 : }
2407 :
2408 : /// This is used by tests & import-from-basebackup.
2409 : ///
2410 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2411 : /// a state that will fail [`Tenant::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2412 : ///
2413 : /// The caller is responsible for getting the timeline into a state that will be accepted
2414 : /// by [`Tenant::load_remote_timeline`] / [`Tenant::attach`].
2415 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2416 : /// to the [`Tenant::timelines`].
2417 : ///
2418 : /// Tests should use `Tenant::create_test_timeline` to set up the minimum required metadata keys.
2419 436 : pub(crate) async fn create_empty_timeline(
2420 436 : self: &Arc<Self>,
2421 436 : new_timeline_id: TimelineId,
2422 436 : initdb_lsn: Lsn,
2423 436 : pg_version: u32,
2424 436 : ctx: &RequestContext,
2425 436 : ) -> anyhow::Result<(UninitializedTimeline, RequestContext)> {
2426 436 : anyhow::ensure!(
2427 436 : self.is_active(),
2428 0 : "Cannot create empty timelines on inactive tenant"
2429 : );
2430 :
2431 : // Protect against concurrent attempts to use this TimelineId
2432 436 : let create_guard = match self
2433 436 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2434 436 : .await?
2435 : {
2436 432 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2437 : StartCreatingTimelineResult::Idempotent(_) => {
2438 0 : unreachable!("FailWithConflict implies we get an error instead")
2439 : }
2440 : };
2441 :
2442 432 : let new_metadata = TimelineMetadata::new(
2443 432 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2444 432 : // make it valid, before calling finish_creation()
2445 432 : Lsn(0),
2446 432 : None,
2447 432 : None,
2448 432 : Lsn(0),
2449 432 : initdb_lsn,
2450 432 : initdb_lsn,
2451 432 : pg_version,
2452 432 : );
2453 432 : self.prepare_new_timeline(
2454 432 : new_timeline_id,
2455 432 : &new_metadata,
2456 432 : create_guard,
2457 432 : initdb_lsn,
2458 432 : None,
2459 432 : None,
2460 432 : ctx,
2461 432 : )
2462 432 : .await
2463 436 : }
2464 :
2465 : /// Helper for unit tests to create an empty timeline.
2466 : ///
2467 : /// The timeline is has state value `Active` but its background loops are not running.
2468 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2469 : // Our current tests don't need the background loops.
2470 : #[cfg(test)]
2471 416 : pub async fn create_test_timeline(
2472 416 : self: &Arc<Self>,
2473 416 : new_timeline_id: TimelineId,
2474 416 : initdb_lsn: Lsn,
2475 416 : pg_version: u32,
2476 416 : ctx: &RequestContext,
2477 416 : ) -> anyhow::Result<Arc<Timeline>> {
2478 416 : let (uninit_tl, ctx) = self
2479 416 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2480 416 : .await?;
2481 416 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2482 416 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2483 :
2484 : // Setup minimum keys required for the timeline to be usable.
2485 416 : let mut modification = tline.begin_modification(initdb_lsn);
2486 416 : modification
2487 416 : .init_empty_test_timeline()
2488 416 : .context("init_empty_test_timeline")?;
2489 416 : modification
2490 416 : .commit(&ctx)
2491 416 : .await
2492 416 : .context("commit init_empty_test_timeline modification")?;
2493 :
2494 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2495 416 : tline.maybe_spawn_flush_loop();
2496 416 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2497 :
2498 : // Make sure the freeze_and_flush reaches remote storage.
2499 416 : tline.remote_client.wait_completion().await.unwrap();
2500 :
2501 416 : let tl = uninit_tl.finish_creation().await?;
2502 : // The non-test code would call tl.activate() here.
2503 416 : tl.set_state(TimelineState::Active);
2504 416 : Ok(tl)
2505 416 : }
2506 :
2507 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2508 : #[cfg(test)]
2509 : #[allow(clippy::too_many_arguments)]
2510 84 : pub async fn create_test_timeline_with_layers(
2511 84 : self: &Arc<Self>,
2512 84 : new_timeline_id: TimelineId,
2513 84 : initdb_lsn: Lsn,
2514 84 : pg_version: u32,
2515 84 : ctx: &RequestContext,
2516 84 : in_memory_layer_desc: Vec<timeline::InMemoryLayerTestDesc>,
2517 84 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2518 84 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2519 84 : end_lsn: Lsn,
2520 84 : ) -> anyhow::Result<Arc<Timeline>> {
2521 : use checks::check_valid_layermap;
2522 : use itertools::Itertools;
2523 :
2524 84 : let tline = self
2525 84 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2526 84 : .await?;
2527 84 : tline.force_advance_lsn(end_lsn);
2528 256 : for deltas in delta_layer_desc {
2529 172 : tline
2530 172 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2531 172 : .await?;
2532 : }
2533 204 : for (lsn, images) in image_layer_desc {
2534 120 : tline
2535 120 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2536 120 : .await?;
2537 : }
2538 92 : for in_memory in in_memory_layer_desc {
2539 8 : tline
2540 8 : .force_create_in_memory_layer(in_memory, Some(initdb_lsn), ctx)
2541 8 : .await?;
2542 : }
2543 84 : let layer_names = tline
2544 84 : .layers
2545 84 : .read()
2546 84 : .await
2547 84 : .layer_map()
2548 84 : .unwrap()
2549 84 : .iter_historic_layers()
2550 376 : .map(|layer| layer.layer_name())
2551 84 : .collect_vec();
2552 84 : if let Some(err) = check_valid_layermap(&layer_names) {
2553 0 : bail!("invalid layermap: {err}");
2554 84 : }
2555 84 : Ok(tline)
2556 84 : }
2557 :
2558 : /// Create a new timeline.
2559 : ///
2560 : /// Returns the new timeline ID and reference to its Timeline object.
2561 : ///
2562 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2563 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2564 : #[allow(clippy::too_many_arguments)]
2565 0 : pub(crate) async fn create_timeline(
2566 0 : self: &Arc<Tenant>,
2567 0 : params: CreateTimelineParams,
2568 0 : broker_client: storage_broker::BrokerClientChannel,
2569 0 : ctx: &RequestContext,
2570 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2571 0 : if !self.is_active() {
2572 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2573 0 : return Err(CreateTimelineError::ShuttingDown);
2574 : } else {
2575 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2576 0 : "Cannot create timelines on inactive tenant"
2577 0 : )));
2578 : }
2579 0 : }
2580 :
2581 0 : let _gate = self
2582 0 : .gate
2583 0 : .enter()
2584 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2585 :
2586 0 : let result: CreateTimelineResult = match params {
2587 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2588 0 : new_timeline_id,
2589 0 : existing_initdb_timeline_id,
2590 0 : pg_version,
2591 0 : }) => {
2592 0 : self.bootstrap_timeline(
2593 0 : new_timeline_id,
2594 0 : pg_version,
2595 0 : existing_initdb_timeline_id,
2596 0 : ctx,
2597 0 : )
2598 0 : .await?
2599 : }
2600 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2601 0 : new_timeline_id,
2602 0 : ancestor_timeline_id,
2603 0 : mut ancestor_start_lsn,
2604 : }) => {
2605 0 : let ancestor_timeline = self
2606 0 : .get_timeline(ancestor_timeline_id, false)
2607 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2608 :
2609 : // instead of waiting around, just deny the request because ancestor is not yet
2610 : // ready for other purposes either.
2611 0 : if !ancestor_timeline.is_active() {
2612 0 : return Err(CreateTimelineError::AncestorNotActive);
2613 0 : }
2614 0 :
2615 0 : if ancestor_timeline.is_archived() == Some(true) {
2616 0 : info!("tried to branch archived timeline");
2617 0 : return Err(CreateTimelineError::AncestorArchived);
2618 0 : }
2619 :
2620 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2621 0 : *lsn = lsn.align();
2622 0 :
2623 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2624 0 : if ancestor_ancestor_lsn > *lsn {
2625 : // can we safely just branch from the ancestor instead?
2626 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2627 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2628 0 : lsn,
2629 0 : ancestor_timeline_id,
2630 0 : ancestor_ancestor_lsn,
2631 0 : )));
2632 0 : }
2633 0 :
2634 0 : // Wait for the WAL to arrive and be processed on the parent branch up
2635 0 : // to the requested branch point. The repository code itself doesn't
2636 0 : // require it, but if we start to receive WAL on the new timeline,
2637 0 : // decoding the new WAL might need to look up previous pages, relation
2638 0 : // sizes etc. and that would get confused if the previous page versions
2639 0 : // are not in the repository yet.
2640 0 : ancestor_timeline
2641 0 : .wait_lsn(
2642 0 : *lsn,
2643 0 : timeline::WaitLsnWaiter::Tenant,
2644 0 : timeline::WaitLsnTimeout::Default,
2645 0 : ctx,
2646 0 : )
2647 0 : .await
2648 0 : .map_err(|e| match e {
2649 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2650 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2651 : }
2652 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2653 0 : })?;
2654 0 : }
2655 :
2656 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2657 0 : .await?
2658 : }
2659 0 : CreateTimelineParams::ImportPgdata(params) => {
2660 0 : self.create_timeline_import_pgdata(
2661 0 : params,
2662 0 : ActivateTimelineArgs::Yes {
2663 0 : broker_client: broker_client.clone(),
2664 0 : },
2665 0 : ctx,
2666 0 : )
2667 0 : .await?
2668 : }
2669 : };
2670 :
2671 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2672 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2673 : // not send a success to the caller until it is. The same applies to idempotent retries.
2674 : //
2675 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2676 : // assume that, because they can see the timeline via API, that the creation is done and
2677 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2678 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2679 : // interacts with UninitializedTimeline and is generally a bit tricky.
2680 : //
2681 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2682 : // creation API until it returns success. Only then is durability guaranteed.
2683 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2684 0 : result
2685 0 : .timeline()
2686 0 : .remote_client
2687 0 : .wait_completion()
2688 0 : .await
2689 0 : .map_err(|e| match e {
2690 : WaitCompletionError::NotInitialized(
2691 0 : e, // If the queue is already stopped, it's a shutdown error.
2692 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2693 : WaitCompletionError::NotInitialized(_) => {
2694 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2695 0 : debug_assert!(false);
2696 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2697 : }
2698 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2699 0 : CreateTimelineError::ShuttingDown
2700 : }
2701 0 : })?;
2702 :
2703 : // The creating task is responsible for activating the timeline.
2704 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2705 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2706 0 : let activated_timeline = match result {
2707 0 : CreateTimelineResult::Created(timeline) => {
2708 0 : timeline.activate(
2709 0 : self.clone(),
2710 0 : broker_client,
2711 0 : None,
2712 0 : &ctx.with_scope_timeline(&timeline),
2713 0 : );
2714 0 : timeline
2715 : }
2716 0 : CreateTimelineResult::Idempotent(timeline) => {
2717 0 : info!(
2718 0 : "request was deemed idempotent, activation will be done by the creating task"
2719 : );
2720 0 : timeline
2721 : }
2722 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2723 0 : info!(
2724 0 : "import task spawned, timeline will become visible and activated once the import is done"
2725 : );
2726 0 : timeline
2727 : }
2728 : };
2729 :
2730 0 : Ok(activated_timeline)
2731 0 : }
2732 :
2733 : /// The returned [`Arc<Timeline>`] is NOT in the [`Tenant::timelines`] map until the import
2734 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2735 : /// [`Tenant::timelines`] map when the import completes.
2736 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2737 : /// for the response.
2738 0 : async fn create_timeline_import_pgdata(
2739 0 : self: &Arc<Tenant>,
2740 0 : params: CreateTimelineParamsImportPgdata,
2741 0 : activate: ActivateTimelineArgs,
2742 0 : ctx: &RequestContext,
2743 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2744 0 : let CreateTimelineParamsImportPgdata {
2745 0 : new_timeline_id,
2746 0 : location,
2747 0 : idempotency_key,
2748 0 : } = params;
2749 0 :
2750 0 : let started_at = chrono::Utc::now().naive_utc();
2751 :
2752 : //
2753 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2754 : // is the canonical way we do it.
2755 : // - create an empty timeline in-memory
2756 : // - use its remote_timeline_client to do the upload
2757 : // - dispose of the uninit timeline
2758 : // - keep the creation guard alive
2759 :
2760 0 : let timeline_create_guard = match self
2761 0 : .start_creating_timeline(
2762 0 : new_timeline_id,
2763 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2764 0 : idempotency_key: idempotency_key.clone(),
2765 0 : }),
2766 0 : )
2767 0 : .await?
2768 : {
2769 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2770 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2771 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
2772 : }
2773 : };
2774 :
2775 0 : let (mut uninit_timeline, timeline_ctx) = {
2776 0 : let this = &self;
2777 0 : let initdb_lsn = Lsn(0);
2778 0 : async move {
2779 0 : let new_metadata = TimelineMetadata::new(
2780 0 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2781 0 : // make it valid, before calling finish_creation()
2782 0 : Lsn(0),
2783 0 : None,
2784 0 : None,
2785 0 : Lsn(0),
2786 0 : initdb_lsn,
2787 0 : initdb_lsn,
2788 0 : 15,
2789 0 : );
2790 0 : this.prepare_new_timeline(
2791 0 : new_timeline_id,
2792 0 : &new_metadata,
2793 0 : timeline_create_guard,
2794 0 : initdb_lsn,
2795 0 : None,
2796 0 : None,
2797 0 : ctx,
2798 0 : )
2799 0 : .await
2800 0 : }
2801 0 : }
2802 0 : .await?;
2803 :
2804 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2805 0 : idempotency_key,
2806 0 : location,
2807 0 : started_at,
2808 0 : };
2809 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2810 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2811 0 : );
2812 0 : uninit_timeline
2813 0 : .raw_timeline()
2814 0 : .unwrap()
2815 0 : .remote_client
2816 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2817 :
2818 : // wait_completion happens in caller
2819 :
2820 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2821 0 :
2822 0 : tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2823 0 : timeline.clone(),
2824 0 : index_part,
2825 0 : activate,
2826 0 : timeline_create_guard,
2827 0 : timeline_ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
2828 0 : ));
2829 0 :
2830 0 : // NB: the timeline doesn't exist in self.timelines at this point
2831 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2832 0 : }
2833 :
2834 : #[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))]
2835 : async fn create_timeline_import_pgdata_task(
2836 : self: Arc<Tenant>,
2837 : timeline: Arc<Timeline>,
2838 : index_part: import_pgdata::index_part_format::Root,
2839 : activate: ActivateTimelineArgs,
2840 : timeline_create_guard: TimelineCreateGuard,
2841 : ctx: RequestContext,
2842 : ) {
2843 : debug_assert_current_span_has_tenant_and_timeline_id();
2844 : info!("starting");
2845 : scopeguard::defer! {info!("exiting")};
2846 :
2847 : let res = self
2848 : .create_timeline_import_pgdata_task_impl(
2849 : timeline,
2850 : index_part,
2851 : activate,
2852 : timeline_create_guard,
2853 : ctx,
2854 : )
2855 : .await;
2856 : if let Err(err) = &res {
2857 : error!(?err, "task failed");
2858 : // TODO sleep & retry, sensitive to tenant shutdown
2859 : // TODO: allow timeline deletion requests => should cancel the task
2860 : }
2861 : }
2862 :
2863 0 : async fn create_timeline_import_pgdata_task_impl(
2864 0 : self: Arc<Tenant>,
2865 0 : timeline: Arc<Timeline>,
2866 0 : index_part: import_pgdata::index_part_format::Root,
2867 0 : activate: ActivateTimelineArgs,
2868 0 : timeline_create_guard: TimelineCreateGuard,
2869 0 : ctx: RequestContext,
2870 0 : ) -> Result<(), anyhow::Error> {
2871 0 : info!("importing pgdata");
2872 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
2873 0 : .await
2874 0 : .context("import")?;
2875 0 : info!("import done");
2876 :
2877 : //
2878 : // Reload timeline from remote.
2879 : // This proves that the remote state is attachable, and it reuses the code.
2880 : //
2881 : // TODO: think about whether this is safe to do with concurrent Tenant::shutdown.
2882 : // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
2883 : // But our activate() call might launch new background tasks after Tenant::shutdown
2884 : // already went past shutting down the Tenant::timelines, which this timeline here is no part of.
2885 : // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
2886 : // down while bootstrapping/branching + activating), but, the race condition is much more likely
2887 : // to manifest because of the long runtime of this import task.
2888 :
2889 : // in theory this shouldn't even .await anything except for coop yield
2890 0 : info!("shutting down timeline");
2891 0 : timeline.shutdown(ShutdownMode::Hard).await;
2892 0 : info!("timeline shut down, reloading from remote");
2893 : // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
2894 : // let Some(timeline) = Arc::into_inner(timeline) else {
2895 : // anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
2896 : // };
2897 0 : let timeline_id = timeline.timeline_id;
2898 0 :
2899 0 : // load from object storage like Tenant::attach does
2900 0 : let resources = self.build_timeline_resources(timeline_id);
2901 0 : let index_part = resources
2902 0 : .remote_client
2903 0 : .download_index_file(&self.cancel)
2904 0 : .await?;
2905 0 : let index_part = match index_part {
2906 : MaybeDeletedIndexPart::Deleted(_) => {
2907 : // likely concurrent delete call, cplane should prevent this
2908 0 : anyhow::bail!(
2909 0 : "index part says deleted but we are not done creating yet, this should not happen but"
2910 0 : )
2911 : }
2912 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2913 0 : };
2914 0 : let metadata = index_part.metadata.clone();
2915 0 : self
2916 0 : .load_remote_timeline(timeline_id, index_part, metadata, None, resources, LoadTimelineCause::ImportPgdata{
2917 0 : create_guard: timeline_create_guard, activate, }, &ctx)
2918 0 : .await?
2919 0 : .ready_to_activate()
2920 0 : .context("implementation error: reloaded timeline still needs import after import reported success")?;
2921 :
2922 0 : anyhow::Ok(())
2923 0 : }
2924 :
2925 0 : pub(crate) async fn delete_timeline(
2926 0 : self: Arc<Self>,
2927 0 : timeline_id: TimelineId,
2928 0 : ) -> Result<(), DeleteTimelineError> {
2929 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
2930 :
2931 0 : Ok(())
2932 0 : }
2933 :
2934 : /// perform one garbage collection iteration, removing old data files from disk.
2935 : /// this function is periodically called by gc task.
2936 : /// also it can be explicitly requested through page server api 'do_gc' command.
2937 : ///
2938 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
2939 : ///
2940 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
2941 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
2942 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
2943 : /// `pitr` specifies the same as a time difference from the current time. The effective
2944 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
2945 : /// requires more history to be retained.
2946 : //
2947 1508 : pub(crate) async fn gc_iteration(
2948 1508 : &self,
2949 1508 : target_timeline_id: Option<TimelineId>,
2950 1508 : horizon: u64,
2951 1508 : pitr: Duration,
2952 1508 : cancel: &CancellationToken,
2953 1508 : ctx: &RequestContext,
2954 1508 : ) -> Result<GcResult, GcError> {
2955 1508 : // Don't start doing work during shutdown
2956 1508 : if let TenantState::Stopping { .. } = self.current_state() {
2957 0 : return Ok(GcResult::default());
2958 1508 : }
2959 1508 :
2960 1508 : // there is a global allowed_error for this
2961 1508 : if !self.is_active() {
2962 0 : return Err(GcError::NotActive);
2963 1508 : }
2964 1508 :
2965 1508 : {
2966 1508 : let conf = self.tenant_conf.load();
2967 1508 :
2968 1508 : // If we may not delete layers, then simply skip GC. Even though a tenant
2969 1508 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
2970 1508 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
2971 1508 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
2972 1508 : if !conf.location.may_delete_layers_hint() {
2973 0 : info!("Skipping GC in location state {:?}", conf.location);
2974 0 : return Ok(GcResult::default());
2975 1508 : }
2976 1508 :
2977 1508 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
2978 1500 : info!("Skipping GC because lsn lease deadline is not reached");
2979 1500 : return Ok(GcResult::default());
2980 8 : }
2981 : }
2982 :
2983 8 : let _guard = match self.gc_block.start().await {
2984 8 : Ok(guard) => guard,
2985 0 : Err(reasons) => {
2986 0 : info!("Skipping GC: {reasons}");
2987 0 : return Ok(GcResult::default());
2988 : }
2989 : };
2990 :
2991 8 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2992 8 : .await
2993 1508 : }
2994 :
2995 : /// Performs one compaction iteration. Called periodically from the compaction loop. Returns
2996 : /// whether another compaction is needed, if we still have pending work or if we yield for
2997 : /// immediate L0 compaction.
2998 : ///
2999 : /// Compaction can also be explicitly requested for a timeline via the HTTP API.
3000 0 : async fn compaction_iteration(
3001 0 : self: &Arc<Self>,
3002 0 : cancel: &CancellationToken,
3003 0 : ctx: &RequestContext,
3004 0 : ) -> Result<CompactionOutcome, CompactionError> {
3005 0 : // Don't compact inactive tenants.
3006 0 : if !self.is_active() {
3007 0 : return Ok(CompactionOutcome::Skipped);
3008 0 : }
3009 0 :
3010 0 : // Don't compact tenants that can't upload layers. We don't check `may_delete_layers_hint`,
3011 0 : // since we need to compact L0 even in AttachedMulti to bound read amplification.
3012 0 : let location = self.tenant_conf.load().location;
3013 0 : if !location.may_upload_layers_hint() {
3014 0 : info!("skipping compaction in location state {location:?}");
3015 0 : return Ok(CompactionOutcome::Skipped);
3016 0 : }
3017 0 :
3018 0 : // Don't compact if the circuit breaker is tripped.
3019 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
3020 0 : info!("skipping compaction due to previous failures");
3021 0 : return Ok(CompactionOutcome::Skipped);
3022 0 : }
3023 0 :
3024 0 : // Collect all timelines to compact, along with offload instructions and L0 counts.
3025 0 : let mut compact: Vec<Arc<Timeline>> = Vec::new();
3026 0 : let mut offload: HashSet<TimelineId> = HashSet::new();
3027 0 : let mut l0_counts: HashMap<TimelineId, usize> = HashMap::new();
3028 0 :
3029 0 : {
3030 0 : let offload_enabled = self.get_timeline_offloading_enabled();
3031 0 : let timelines = self.timelines.lock().unwrap();
3032 0 : for (&timeline_id, timeline) in timelines.iter() {
3033 : // Skip inactive timelines.
3034 0 : if !timeline.is_active() {
3035 0 : continue;
3036 0 : }
3037 0 :
3038 0 : // Schedule the timeline for compaction.
3039 0 : compact.push(timeline.clone());
3040 :
3041 : // Schedule the timeline for offloading if eligible.
3042 0 : let can_offload = offload_enabled
3043 0 : && timeline.can_offload().0
3044 0 : && !timelines
3045 0 : .iter()
3046 0 : .any(|(_, tli)| tli.get_ancestor_timeline_id() == Some(timeline_id));
3047 0 : if can_offload {
3048 0 : offload.insert(timeline_id);
3049 0 : }
3050 : }
3051 : } // release timelines lock
3052 :
3053 0 : for timeline in &compact {
3054 : // Collect L0 counts. Can't await while holding lock above.
3055 0 : if let Ok(lm) = timeline.layers.read().await.layer_map() {
3056 0 : l0_counts.insert(timeline.timeline_id, lm.level0_deltas().len());
3057 0 : }
3058 : }
3059 :
3060 : // Pass 1: L0 compaction across all timelines, in order of L0 count. We prioritize this to
3061 : // bound read amplification.
3062 : //
3063 : // TODO: this may spin on one or more ingest-heavy timelines, starving out image/GC
3064 : // compaction and offloading. We leave that as a potential problem to solve later. Consider
3065 : // splitting L0 and image/GC compaction to separate background jobs.
3066 0 : if self.get_compaction_l0_first() {
3067 0 : let compaction_threshold = self.get_compaction_threshold();
3068 0 : let compact_l0 = compact
3069 0 : .iter()
3070 0 : .map(|tli| (tli, l0_counts.get(&tli.timeline_id).copied().unwrap_or(0)))
3071 0 : .filter(|&(_, l0)| l0 >= compaction_threshold)
3072 0 : .sorted_by_key(|&(_, l0)| l0)
3073 0 : .rev()
3074 0 : .map(|(tli, _)| tli.clone())
3075 0 : .collect_vec();
3076 0 :
3077 0 : let mut has_pending_l0 = false;
3078 0 : for timeline in compact_l0 {
3079 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3080 0 : let outcome = timeline
3081 0 : .compact(cancel, CompactFlags::OnlyL0Compaction.into(), ctx)
3082 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3083 0 : .await
3084 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3085 0 : match outcome {
3086 0 : CompactionOutcome::Done => {}
3087 0 : CompactionOutcome::Skipped => {}
3088 0 : CompactionOutcome::Pending => has_pending_l0 = true,
3089 0 : CompactionOutcome::YieldForL0 => has_pending_l0 = true,
3090 : }
3091 : }
3092 0 : if has_pending_l0 {
3093 0 : return Ok(CompactionOutcome::YieldForL0); // do another pass
3094 0 : }
3095 0 : }
3096 :
3097 : // Pass 2: image compaction and timeline offloading. If any timelines have accumulated
3098 : // more L0 layers, they may also be compacted here.
3099 : //
3100 : // NB: image compaction may yield if there is pending L0 compaction.
3101 : //
3102 : // TODO: it will only yield if there is pending L0 compaction on the same timeline. If a
3103 : // different timeline needs compaction, it won't. It should check `l0_compaction_trigger`.
3104 : // We leave this for a later PR.
3105 : //
3106 : // TODO: consider ordering timelines by some priority, e.g. time since last full compaction,
3107 : // amount of L1 delta debt or garbage, offload-eligible timelines first, etc.
3108 0 : let mut has_pending = false;
3109 0 : for timeline in compact {
3110 0 : if !timeline.is_active() {
3111 0 : continue;
3112 0 : }
3113 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3114 :
3115 0 : let mut outcome = timeline
3116 0 : .compact(cancel, EnumSet::default(), ctx)
3117 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3118 0 : .await
3119 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3120 :
3121 : // If we're done compacting, check the scheduled GC compaction queue for more work.
3122 0 : if outcome == CompactionOutcome::Done {
3123 0 : let queue = {
3124 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3125 0 : guard
3126 0 : .entry(timeline.timeline_id)
3127 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()))
3128 0 : .clone()
3129 0 : };
3130 0 : outcome = queue
3131 0 : .iteration(cancel, ctx, &self.gc_block, &timeline)
3132 0 : .instrument(
3133 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id),
3134 : )
3135 0 : .await?;
3136 0 : }
3137 :
3138 : // If we're done compacting, offload the timeline if requested.
3139 0 : if outcome == CompactionOutcome::Done && offload.contains(&timeline.timeline_id) {
3140 0 : pausable_failpoint!("before-timeline-auto-offload");
3141 0 : offload_timeline(self, &timeline)
3142 0 : .instrument(info_span!("offload_timeline", timeline_id = %timeline.timeline_id))
3143 0 : .await
3144 0 : .or_else(|err| match err {
3145 : // Ignore this, we likely raced with unarchival.
3146 0 : OffloadError::NotArchived => Ok(()),
3147 0 : err => Err(err),
3148 0 : })?;
3149 0 : }
3150 :
3151 0 : match outcome {
3152 0 : CompactionOutcome::Done => {}
3153 0 : CompactionOutcome::Skipped => {}
3154 0 : CompactionOutcome::Pending => has_pending = true,
3155 : // This mostly makes sense when the L0-only pass above is enabled, since there's
3156 : // otherwise no guarantee that we'll start with the timeline that has high L0.
3157 0 : CompactionOutcome::YieldForL0 => return Ok(CompactionOutcome::YieldForL0),
3158 : }
3159 : }
3160 :
3161 : // Success! Untrip the breaker if necessary.
3162 0 : self.compaction_circuit_breaker
3163 0 : .lock()
3164 0 : .unwrap()
3165 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3166 0 :
3167 0 : match has_pending {
3168 0 : true => Ok(CompactionOutcome::Pending),
3169 0 : false => Ok(CompactionOutcome::Done),
3170 : }
3171 0 : }
3172 :
3173 : /// Trips the compaction circuit breaker if appropriate.
3174 0 : pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
3175 0 : match err {
3176 0 : err if err.is_cancel() => {}
3177 0 : CompactionError::ShuttingDown => (),
3178 : // Offload failures don't trip the circuit breaker, since they're cheap to retry and
3179 : // shouldn't block compaction.
3180 0 : CompactionError::Offload(_) => {}
3181 0 : CompactionError::CollectKeySpaceError(err) => {
3182 0 : // CollectKeySpaceError::Cancelled and PageRead::Cancelled are handled in `err.is_cancel` branch.
3183 0 : self.compaction_circuit_breaker
3184 0 : .lock()
3185 0 : .unwrap()
3186 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3187 0 : }
3188 0 : CompactionError::Other(err) => {
3189 0 : self.compaction_circuit_breaker
3190 0 : .lock()
3191 0 : .unwrap()
3192 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3193 0 : }
3194 0 : CompactionError::AlreadyRunning(_) => {}
3195 : }
3196 0 : }
3197 :
3198 : /// Cancel scheduled compaction tasks
3199 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3200 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3201 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3202 0 : q.cancel_scheduled();
3203 0 : }
3204 0 : }
3205 :
3206 0 : pub(crate) fn get_scheduled_compaction_tasks(
3207 0 : &self,
3208 0 : timeline_id: TimelineId,
3209 0 : ) -> Vec<CompactInfoResponse> {
3210 0 : let res = {
3211 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3212 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3213 : };
3214 0 : let Some((running, remaining)) = res else {
3215 0 : return Vec::new();
3216 : };
3217 0 : let mut result = Vec::new();
3218 0 : if let Some((id, running)) = running {
3219 0 : result.extend(running.into_compact_info_resp(id, true));
3220 0 : }
3221 0 : for (id, job) in remaining {
3222 0 : result.extend(job.into_compact_info_resp(id, false));
3223 0 : }
3224 0 : result
3225 0 : }
3226 :
3227 : /// Schedule a compaction task for a timeline.
3228 0 : pub(crate) async fn schedule_compaction(
3229 0 : &self,
3230 0 : timeline_id: TimelineId,
3231 0 : options: CompactOptions,
3232 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3233 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3234 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3235 0 : let q = guard
3236 0 : .entry(timeline_id)
3237 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3238 0 : q.schedule_manual_compaction(options, Some(tx));
3239 0 : Ok(rx)
3240 0 : }
3241 :
3242 : /// Performs periodic housekeeping, via the tenant housekeeping background task.
3243 0 : async fn housekeeping(&self) {
3244 0 : // Call through to all timelines to freeze ephemeral layers as needed. This usually happens
3245 0 : // during ingest, but we don't want idle timelines to hold open layers for too long.
3246 0 : let timelines = self
3247 0 : .timelines
3248 0 : .lock()
3249 0 : .unwrap()
3250 0 : .values()
3251 0 : .filter(|tli| tli.is_active())
3252 0 : .cloned()
3253 0 : .collect_vec();
3254 :
3255 0 : for timeline in timelines {
3256 0 : timeline.maybe_freeze_ephemeral_layer().await;
3257 : }
3258 :
3259 : // Shut down walredo if idle.
3260 : const WALREDO_IDLE_TIMEOUT: Duration = Duration::from_secs(180);
3261 0 : if let Some(ref walredo_mgr) = self.walredo_mgr {
3262 0 : walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
3263 0 : }
3264 0 : }
3265 :
3266 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3267 0 : let timelines = self.timelines.lock().unwrap();
3268 0 : !timelines
3269 0 : .iter()
3270 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3271 0 : }
3272 :
3273 3476 : pub fn current_state(&self) -> TenantState {
3274 3476 : self.state.borrow().clone()
3275 3476 : }
3276 :
3277 1952 : pub fn is_active(&self) -> bool {
3278 1952 : self.current_state() == TenantState::Active
3279 1952 : }
3280 :
3281 0 : pub fn generation(&self) -> Generation {
3282 0 : self.generation
3283 0 : }
3284 :
3285 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3286 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3287 0 : }
3288 :
3289 : /// Changes tenant status to active, unless shutdown was already requested.
3290 : ///
3291 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3292 : /// to delay background jobs. Background jobs can be started right away when None is given.
3293 0 : fn activate(
3294 0 : self: &Arc<Self>,
3295 0 : broker_client: BrokerClientChannel,
3296 0 : background_jobs_can_start: Option<&completion::Barrier>,
3297 0 : ctx: &RequestContext,
3298 0 : ) {
3299 0 : span::debug_assert_current_span_has_tenant_id();
3300 0 :
3301 0 : let mut activating = false;
3302 0 : self.state.send_modify(|current_state| {
3303 : use pageserver_api::models::ActivatingFrom;
3304 0 : match &*current_state {
3305 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3306 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
3307 : }
3308 0 : TenantState::Attaching => {
3309 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3310 0 : }
3311 0 : }
3312 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3313 0 : activating = true;
3314 0 : // Continue outside the closure. We need to grab timelines.lock()
3315 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3316 0 : });
3317 0 :
3318 0 : if activating {
3319 0 : let timelines_accessor = self.timelines.lock().unwrap();
3320 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3321 0 : let timelines_to_activate = timelines_accessor
3322 0 : .values()
3323 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3324 0 :
3325 0 : // Before activation, populate each Timeline's GcInfo with information about its children
3326 0 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
3327 0 :
3328 0 : // Spawn gc and compaction loops. The loops will shut themselves
3329 0 : // down when they notice that the tenant is inactive.
3330 0 : tasks::start_background_loops(self, background_jobs_can_start);
3331 0 :
3332 0 : let mut activated_timelines = 0;
3333 :
3334 0 : for timeline in timelines_to_activate {
3335 0 : timeline.activate(
3336 0 : self.clone(),
3337 0 : broker_client.clone(),
3338 0 : background_jobs_can_start,
3339 0 : &ctx.with_scope_timeline(timeline),
3340 0 : );
3341 0 : activated_timelines += 1;
3342 0 : }
3343 :
3344 0 : self.state.send_modify(move |current_state| {
3345 0 : assert!(
3346 0 : matches!(current_state, TenantState::Activating(_)),
3347 0 : "set_stopping and set_broken wait for us to leave Activating state",
3348 : );
3349 0 : *current_state = TenantState::Active;
3350 0 :
3351 0 : let elapsed = self.constructed_at.elapsed();
3352 0 : let total_timelines = timelines_accessor.len();
3353 0 :
3354 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3355 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3356 0 : info!(
3357 0 : since_creation_millis = elapsed.as_millis(),
3358 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3359 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3360 0 : activated_timelines,
3361 0 : total_timelines,
3362 0 : post_state = <&'static str>::from(&*current_state),
3363 0 : "activation attempt finished"
3364 : );
3365 :
3366 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3367 0 : });
3368 0 : }
3369 0 : }
3370 :
3371 : /// Shutdown the tenant and join all of the spawned tasks.
3372 : ///
3373 : /// The method caters for all use-cases:
3374 : /// - pageserver shutdown (freeze_and_flush == true)
3375 : /// - detach + ignore (freeze_and_flush == false)
3376 : ///
3377 : /// This will attempt to shutdown even if tenant is broken.
3378 : ///
3379 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3380 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3381 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3382 : /// the ongoing shutdown.
3383 12 : async fn shutdown(
3384 12 : &self,
3385 12 : shutdown_progress: completion::Barrier,
3386 12 : shutdown_mode: timeline::ShutdownMode,
3387 12 : ) -> Result<(), completion::Barrier> {
3388 12 : span::debug_assert_current_span_has_tenant_id();
3389 :
3390 : // Set tenant (and its timlines) to Stoppping state.
3391 : //
3392 : // Since we can only transition into Stopping state after activation is complete,
3393 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3394 : //
3395 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3396 : // 1. Lock out any new requests to the tenants.
3397 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3398 : // 3. Signal cancellation for other tenant background loops.
3399 : // 4. ???
3400 : //
3401 : // The waiting for the cancellation is not done uniformly.
3402 : // We certainly wait for WAL receivers to shut down.
3403 : // That is necessary so that no new data comes in before the freeze_and_flush.
3404 : // But the tenant background loops are joined-on in our caller.
3405 : // It's mesed up.
3406 : // we just ignore the failure to stop
3407 :
3408 : // If we're still attaching, fire the cancellation token early to drop out: this
3409 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3410 : // is very slow.
3411 12 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3412 0 : self.cancel.cancel();
3413 0 :
3414 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3415 0 : // are children of ours, so their flush loops will have shut down already
3416 0 : timeline::ShutdownMode::Hard
3417 : } else {
3418 12 : shutdown_mode
3419 : };
3420 :
3421 12 : match self.set_stopping(shutdown_progress, false, false).await {
3422 12 : Ok(()) => {}
3423 0 : Err(SetStoppingError::Broken) => {
3424 0 : // assume that this is acceptable
3425 0 : }
3426 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3427 0 : // give caller the option to wait for this this shutdown
3428 0 : info!("Tenant::shutdown: AlreadyStopping");
3429 0 : return Err(other);
3430 : }
3431 : };
3432 :
3433 12 : let mut js = tokio::task::JoinSet::new();
3434 12 : {
3435 12 : let timelines = self.timelines.lock().unwrap();
3436 12 : timelines.values().for_each(|timeline| {
3437 12 : let timeline = Arc::clone(timeline);
3438 12 : let timeline_id = timeline.timeline_id;
3439 12 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3440 12 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3441 12 : });
3442 12 : }
3443 12 : {
3444 12 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3445 12 : timelines_offloaded.values().for_each(|timeline| {
3446 0 : timeline.defuse_for_tenant_drop();
3447 12 : });
3448 12 : }
3449 12 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3450 12 : tracing::info!("Waiting for timelines...");
3451 24 : while let Some(res) = js.join_next().await {
3452 0 : match res {
3453 12 : Ok(()) => {}
3454 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3455 0 : Err(je) if je.is_panic() => { /* logged already */ }
3456 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3457 : }
3458 : }
3459 :
3460 12 : if let ShutdownMode::Reload = shutdown_mode {
3461 0 : tracing::info!("Flushing deletion queue");
3462 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3463 0 : match e {
3464 0 : DeletionQueueError::ShuttingDown => {
3465 0 : // This is the only error we expect for now. In the future, if more error
3466 0 : // variants are added, we should handle them here.
3467 0 : }
3468 : }
3469 0 : }
3470 12 : }
3471 :
3472 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3473 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3474 12 : tracing::debug!("Cancelling CancellationToken");
3475 12 : self.cancel.cancel();
3476 12 :
3477 12 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3478 12 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3479 12 : //
3480 12 : // this will additionally shutdown and await all timeline tasks.
3481 12 : tracing::debug!("Waiting for tasks...");
3482 12 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3483 :
3484 12 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3485 12 : walredo_mgr.shutdown().await;
3486 0 : }
3487 :
3488 : // Wait for any in-flight operations to complete
3489 12 : self.gate.close().await;
3490 :
3491 12 : remove_tenant_metrics(&self.tenant_shard_id);
3492 12 :
3493 12 : Ok(())
3494 12 : }
3495 :
3496 : /// Change tenant status to Stopping, to mark that it is being shut down.
3497 : ///
3498 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3499 : ///
3500 : /// This function is not cancel-safe!
3501 : ///
3502 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
3503 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
3504 12 : async fn set_stopping(
3505 12 : &self,
3506 12 : progress: completion::Barrier,
3507 12 : _allow_transition_from_loading: bool,
3508 12 : allow_transition_from_attaching: bool,
3509 12 : ) -> Result<(), SetStoppingError> {
3510 12 : let mut rx = self.state.subscribe();
3511 12 :
3512 12 : // cannot stop before we're done activating, so wait out until we're done activating
3513 12 : rx.wait_for(|state| match state {
3514 0 : TenantState::Attaching if allow_transition_from_attaching => true,
3515 : TenantState::Activating(_) | TenantState::Attaching => {
3516 0 : info!(
3517 0 : "waiting for {} to turn Active|Broken|Stopping",
3518 0 : <&'static str>::from(state)
3519 : );
3520 0 : false
3521 : }
3522 12 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3523 12 : })
3524 12 : .await
3525 12 : .expect("cannot drop self.state while on a &self method");
3526 12 :
3527 12 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3528 12 : let mut err = None;
3529 12 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3530 : TenantState::Activating(_) => {
3531 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
3532 : }
3533 : TenantState::Attaching => {
3534 0 : if !allow_transition_from_attaching {
3535 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
3536 0 : };
3537 0 : *current_state = TenantState::Stopping { progress };
3538 0 : true
3539 : }
3540 : TenantState::Active => {
3541 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3542 : // are created after the transition to Stopping. That's harmless, as the Timelines
3543 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3544 12 : *current_state = TenantState::Stopping { progress };
3545 12 : // Continue stopping outside the closure. We need to grab timelines.lock()
3546 12 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3547 12 : true
3548 : }
3549 0 : TenantState::Broken { reason, .. } => {
3550 0 : info!(
3551 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3552 : );
3553 0 : err = Some(SetStoppingError::Broken);
3554 0 : false
3555 : }
3556 0 : TenantState::Stopping { progress } => {
3557 0 : info!("Tenant is already in Stopping state");
3558 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3559 0 : false
3560 : }
3561 12 : });
3562 12 : match (stopping, err) {
3563 12 : (true, None) => {} // continue
3564 0 : (false, Some(err)) => return Err(err),
3565 0 : (true, Some(_)) => unreachable!(
3566 0 : "send_if_modified closure must error out if not transitioning to Stopping"
3567 0 : ),
3568 0 : (false, None) => unreachable!(
3569 0 : "send_if_modified closure must return true if transitioning to Stopping"
3570 0 : ),
3571 : }
3572 :
3573 12 : let timelines_accessor = self.timelines.lock().unwrap();
3574 12 : let not_broken_timelines = timelines_accessor
3575 12 : .values()
3576 12 : .filter(|timeline| !timeline.is_broken());
3577 24 : for timeline in not_broken_timelines {
3578 12 : timeline.set_state(TimelineState::Stopping);
3579 12 : }
3580 12 : Ok(())
3581 12 : }
3582 :
3583 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3584 : /// `remove_tenant_from_memory`
3585 : ///
3586 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3587 : ///
3588 : /// In tests, we also use this to set tenants to Broken state on purpose.
3589 0 : pub(crate) async fn set_broken(&self, reason: String) {
3590 0 : let mut rx = self.state.subscribe();
3591 0 :
3592 0 : // The load & attach routines own the tenant state until it has reached `Active`.
3593 0 : // So, wait until it's done.
3594 0 : rx.wait_for(|state| match state {
3595 : TenantState::Activating(_) | TenantState::Attaching => {
3596 0 : info!(
3597 0 : "waiting for {} to turn Active|Broken|Stopping",
3598 0 : <&'static str>::from(state)
3599 : );
3600 0 : false
3601 : }
3602 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3603 0 : })
3604 0 : .await
3605 0 : .expect("cannot drop self.state while on a &self method");
3606 0 :
3607 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3608 0 : self.set_broken_no_wait(reason)
3609 0 : }
3610 :
3611 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3612 0 : let reason = reason.to_string();
3613 0 : self.state.send_modify(|current_state| {
3614 0 : match *current_state {
3615 : TenantState::Activating(_) | TenantState::Attaching => {
3616 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3617 : }
3618 : TenantState::Active => {
3619 0 : if cfg!(feature = "testing") {
3620 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3621 0 : *current_state = TenantState::broken_from_reason(reason);
3622 : } else {
3623 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3624 : }
3625 : }
3626 : TenantState::Broken { .. } => {
3627 0 : warn!("Tenant is already in Broken state");
3628 : }
3629 : // This is the only "expected" path, any other path is a bug.
3630 : TenantState::Stopping { .. } => {
3631 0 : warn!(
3632 0 : "Marking Stopping tenant as Broken state, reason: {}",
3633 : reason
3634 : );
3635 0 : *current_state = TenantState::broken_from_reason(reason);
3636 : }
3637 : }
3638 0 : });
3639 0 : }
3640 :
3641 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3642 0 : self.state.subscribe()
3643 0 : }
3644 :
3645 : /// The activate_now semaphore is initialized with zero units. As soon as
3646 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3647 0 : pub(crate) fn activate_now(&self) {
3648 0 : self.activate_now_sem.add_permits(1);
3649 0 : }
3650 :
3651 0 : pub(crate) async fn wait_to_become_active(
3652 0 : &self,
3653 0 : timeout: Duration,
3654 0 : ) -> Result<(), GetActiveTenantError> {
3655 0 : let mut receiver = self.state.subscribe();
3656 : loop {
3657 0 : let current_state = receiver.borrow_and_update().clone();
3658 0 : match current_state {
3659 : TenantState::Attaching | TenantState::Activating(_) => {
3660 : // in these states, there's a chance that we can reach ::Active
3661 0 : self.activate_now();
3662 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3663 0 : Ok(r) => {
3664 0 : r.map_err(
3665 0 : |_e: tokio::sync::watch::error::RecvError|
3666 : // Tenant existed but was dropped: report it as non-existent
3667 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3668 0 : )?
3669 : }
3670 : Err(TimeoutCancellableError::Cancelled) => {
3671 0 : return Err(GetActiveTenantError::Cancelled);
3672 : }
3673 : Err(TimeoutCancellableError::Timeout) => {
3674 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3675 0 : latest_state: Some(self.current_state()),
3676 0 : wait_time: timeout,
3677 0 : });
3678 : }
3679 : }
3680 : }
3681 : TenantState::Active { .. } => {
3682 0 : return Ok(());
3683 : }
3684 0 : TenantState::Broken { reason, .. } => {
3685 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3686 0 : // it's logically a 500 to external API users (broken is always a bug).
3687 0 : return Err(GetActiveTenantError::Broken(reason));
3688 : }
3689 : TenantState::Stopping { .. } => {
3690 : // There's no chance the tenant can transition back into ::Active
3691 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3692 : }
3693 : }
3694 : }
3695 0 : }
3696 :
3697 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3698 0 : self.tenant_conf.load().location.attach_mode
3699 0 : }
3700 :
3701 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3702 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3703 : /// rare external API calls, like a reconciliation at startup.
3704 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3705 0 : let conf = self.tenant_conf.load();
3706 :
3707 0 : let location_config_mode = match conf.location.attach_mode {
3708 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3709 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3710 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3711 : };
3712 :
3713 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
3714 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
3715 0 :
3716 0 : models::LocationConfig {
3717 0 : mode: location_config_mode,
3718 0 : generation: self.generation.into(),
3719 0 : secondary_conf: None,
3720 0 : shard_number: self.shard_identity.number.0,
3721 0 : shard_count: self.shard_identity.count.literal(),
3722 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3723 0 : tenant_conf: tenant_config,
3724 0 : }
3725 0 : }
3726 :
3727 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3728 0 : &self.tenant_shard_id
3729 0 : }
3730 :
3731 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3732 0 : self.shard_identity.stripe_size
3733 0 : }
3734 :
3735 0 : pub(crate) fn get_generation(&self) -> Generation {
3736 0 : self.generation
3737 0 : }
3738 :
3739 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3740 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3741 : /// resetting this tenant to a valid state if we fail.
3742 0 : pub(crate) async fn split_prepare(
3743 0 : &self,
3744 0 : child_shards: &Vec<TenantShardId>,
3745 0 : ) -> anyhow::Result<()> {
3746 0 : let (timelines, offloaded) = {
3747 0 : let timelines = self.timelines.lock().unwrap();
3748 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3749 0 : (timelines.clone(), offloaded.clone())
3750 0 : };
3751 0 : let timelines_iter = timelines
3752 0 : .values()
3753 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3754 0 : .chain(
3755 0 : offloaded
3756 0 : .values()
3757 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3758 0 : );
3759 0 : for timeline in timelines_iter {
3760 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3761 : // to ensure that they do not start a split if currently in the process of doing these.
3762 :
3763 0 : let timeline_id = timeline.timeline_id();
3764 :
3765 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3766 : // Upload an index from the parent: this is partly to provide freshness for the
3767 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3768 : // always be a parent shard index in the same generation as we wrote the child shard index.
3769 0 : tracing::info!(%timeline_id, "Uploading index");
3770 0 : timeline
3771 0 : .remote_client
3772 0 : .schedule_index_upload_for_file_changes()?;
3773 0 : timeline.remote_client.wait_completion().await?;
3774 0 : }
3775 :
3776 0 : let remote_client = match timeline {
3777 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3778 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3779 0 : let remote_client = self
3780 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3781 0 : Arc::new(remote_client)
3782 : }
3783 : };
3784 :
3785 : // Shut down the timeline's remote client: this means that the indices we write
3786 : // for child shards will not be invalidated by the parent shard deleting layers.
3787 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3788 0 : remote_client.shutdown().await;
3789 :
3790 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3791 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3792 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3793 : // we use here really is the remotely persistent one).
3794 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3795 0 : let result = remote_client
3796 0 : .download_index_file(&self.cancel)
3797 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))
3798 0 : .await?;
3799 0 : let index_part = match result {
3800 : MaybeDeletedIndexPart::Deleted(_) => {
3801 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3802 : }
3803 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3804 : };
3805 :
3806 0 : for child_shard in child_shards {
3807 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3808 0 : upload_index_part(
3809 0 : &self.remote_storage,
3810 0 : child_shard,
3811 0 : &timeline_id,
3812 0 : self.generation,
3813 0 : &index_part,
3814 0 : &self.cancel,
3815 0 : )
3816 0 : .await?;
3817 : }
3818 : }
3819 :
3820 0 : let tenant_manifest = self.build_tenant_manifest();
3821 0 : for child_shard in child_shards {
3822 0 : tracing::info!(
3823 0 : "Uploading tenant manifest for child {}",
3824 0 : child_shard.to_index()
3825 : );
3826 0 : upload_tenant_manifest(
3827 0 : &self.remote_storage,
3828 0 : child_shard,
3829 0 : self.generation,
3830 0 : &tenant_manifest,
3831 0 : &self.cancel,
3832 0 : )
3833 0 : .await?;
3834 : }
3835 :
3836 0 : Ok(())
3837 0 : }
3838 :
3839 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
3840 0 : let mut result = TopTenantShardItem {
3841 0 : id: self.tenant_shard_id,
3842 0 : resident_size: 0,
3843 0 : physical_size: 0,
3844 0 : max_logical_size: 0,
3845 0 : max_logical_size_per_shard: 0,
3846 0 : };
3847 :
3848 0 : for timeline in self.timelines.lock().unwrap().values() {
3849 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
3850 0 :
3851 0 : result.physical_size += timeline
3852 0 : .remote_client
3853 0 : .metrics
3854 0 : .remote_physical_size_gauge
3855 0 : .get();
3856 0 : result.max_logical_size = std::cmp::max(
3857 0 : result.max_logical_size,
3858 0 : timeline.metrics.current_logical_size_gauge.get(),
3859 0 : );
3860 0 : }
3861 :
3862 0 : result.max_logical_size_per_shard = result
3863 0 : .max_logical_size
3864 0 : .div_ceil(self.tenant_shard_id.shard_count.count() as u64);
3865 0 :
3866 0 : result
3867 0 : }
3868 : }
3869 :
3870 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
3871 : /// perform a topological sort, so that the parent of each timeline comes
3872 : /// before the children.
3873 : /// E extracts the ancestor from T
3874 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
3875 452 : fn tree_sort_timelines<T, E>(
3876 452 : timelines: HashMap<TimelineId, T>,
3877 452 : extractor: E,
3878 452 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
3879 452 : where
3880 452 : E: Fn(&T) -> Option<TimelineId>,
3881 452 : {
3882 452 : let mut result = Vec::with_capacity(timelines.len());
3883 452 :
3884 452 : let mut now = Vec::with_capacity(timelines.len());
3885 452 : // (ancestor, children)
3886 452 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
3887 452 : HashMap::with_capacity(timelines.len());
3888 :
3889 464 : for (timeline_id, value) in timelines {
3890 12 : if let Some(ancestor_id) = extractor(&value) {
3891 4 : let children = later.entry(ancestor_id).or_default();
3892 4 : children.push((timeline_id, value));
3893 8 : } else {
3894 8 : now.push((timeline_id, value));
3895 8 : }
3896 : }
3897 :
3898 464 : while let Some((timeline_id, metadata)) = now.pop() {
3899 12 : result.push((timeline_id, metadata));
3900 : // All children of this can be loaded now
3901 12 : if let Some(mut children) = later.remove(&timeline_id) {
3902 4 : now.append(&mut children);
3903 8 : }
3904 : }
3905 :
3906 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
3907 452 : if !later.is_empty() {
3908 0 : for (missing_id, orphan_ids) in later {
3909 0 : for (orphan_id, _) in orphan_ids {
3910 0 : error!(
3911 0 : "could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded"
3912 : );
3913 : }
3914 : }
3915 0 : bail!("could not load tenant because some timelines are missing ancestors");
3916 452 : }
3917 452 :
3918 452 : Ok(result)
3919 452 : }
3920 :
3921 : enum ActivateTimelineArgs {
3922 : Yes {
3923 : broker_client: storage_broker::BrokerClientChannel,
3924 : },
3925 : No,
3926 : }
3927 :
3928 : impl Tenant {
3929 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
3930 0 : self.tenant_conf.load().tenant_conf.clone()
3931 0 : }
3932 :
3933 0 : pub fn effective_config(&self) -> TenantConf {
3934 0 : self.tenant_specific_overrides()
3935 0 : .merge(self.conf.default_tenant_conf.clone())
3936 0 : }
3937 :
3938 0 : pub fn get_checkpoint_distance(&self) -> u64 {
3939 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3940 0 : tenant_conf
3941 0 : .checkpoint_distance
3942 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
3943 0 : }
3944 :
3945 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
3946 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3947 0 : tenant_conf
3948 0 : .checkpoint_timeout
3949 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
3950 0 : }
3951 :
3952 0 : pub fn get_compaction_target_size(&self) -> u64 {
3953 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3954 0 : tenant_conf
3955 0 : .compaction_target_size
3956 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
3957 0 : }
3958 :
3959 0 : pub fn get_compaction_period(&self) -> Duration {
3960 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3961 0 : tenant_conf
3962 0 : .compaction_period
3963 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
3964 0 : }
3965 :
3966 0 : pub fn get_compaction_threshold(&self) -> usize {
3967 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3968 0 : tenant_conf
3969 0 : .compaction_threshold
3970 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
3971 0 : }
3972 :
3973 0 : pub fn get_rel_size_v2_enabled(&self) -> bool {
3974 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3975 0 : tenant_conf
3976 0 : .rel_size_v2_enabled
3977 0 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
3978 0 : }
3979 :
3980 0 : pub fn get_compaction_upper_limit(&self) -> usize {
3981 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3982 0 : tenant_conf
3983 0 : .compaction_upper_limit
3984 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
3985 0 : }
3986 :
3987 0 : pub fn get_compaction_l0_first(&self) -> bool {
3988 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3989 0 : tenant_conf
3990 0 : .compaction_l0_first
3991 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
3992 0 : }
3993 :
3994 0 : pub fn get_gc_horizon(&self) -> u64 {
3995 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
3996 0 : tenant_conf
3997 0 : .gc_horizon
3998 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
3999 0 : }
4000 :
4001 0 : pub fn get_gc_period(&self) -> Duration {
4002 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4003 0 : tenant_conf
4004 0 : .gc_period
4005 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
4006 0 : }
4007 :
4008 0 : pub fn get_image_creation_threshold(&self) -> usize {
4009 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4010 0 : tenant_conf
4011 0 : .image_creation_threshold
4012 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
4013 0 : }
4014 :
4015 0 : pub fn get_pitr_interval(&self) -> Duration {
4016 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4017 0 : tenant_conf
4018 0 : .pitr_interval
4019 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
4020 0 : }
4021 :
4022 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
4023 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4024 0 : tenant_conf
4025 0 : .min_resident_size_override
4026 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
4027 0 : }
4028 :
4029 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
4030 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4031 0 : let heatmap_period = tenant_conf
4032 0 : .heatmap_period
4033 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
4034 0 : if heatmap_period.is_zero() {
4035 0 : None
4036 : } else {
4037 0 : Some(heatmap_period)
4038 : }
4039 0 : }
4040 :
4041 8 : pub fn get_lsn_lease_length(&self) -> Duration {
4042 8 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4043 8 : tenant_conf
4044 8 : .lsn_lease_length
4045 8 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
4046 8 : }
4047 :
4048 0 : pub fn get_timeline_offloading_enabled(&self) -> bool {
4049 0 : if self.conf.timeline_offloading {
4050 0 : return true;
4051 0 : }
4052 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4053 0 : tenant_conf
4054 0 : .timeline_offloading
4055 0 : .unwrap_or(self.conf.default_tenant_conf.timeline_offloading)
4056 0 : }
4057 :
4058 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
4059 4 : fn build_tenant_manifest(&self) -> TenantManifest {
4060 4 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
4061 4 :
4062 4 : let mut timeline_manifests = timelines_offloaded
4063 4 : .iter()
4064 4 : .map(|(_timeline_id, offloaded)| offloaded.manifest())
4065 4 : .collect::<Vec<_>>();
4066 4 : // Sort the manifests so that our output is deterministic
4067 4 : timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
4068 4 :
4069 4 : TenantManifest {
4070 4 : version: LATEST_TENANT_MANIFEST_VERSION,
4071 4 : offloaded_timelines: timeline_manifests,
4072 4 : }
4073 4 : }
4074 :
4075 0 : pub fn update_tenant_config<F: Fn(TenantConfOpt) -> anyhow::Result<TenantConfOpt>>(
4076 0 : &self,
4077 0 : update: F,
4078 0 : ) -> anyhow::Result<TenantConfOpt> {
4079 0 : // Use read-copy-update in order to avoid overwriting the location config
4080 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
4081 0 : // this race is not possible if both request types come from the storage
4082 0 : // controller (as they should!) because an exclusive op lock is required
4083 0 : // on the storage controller side.
4084 0 :
4085 0 : self.tenant_conf
4086 0 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
4087 0 : Ok(Arc::new(AttachedTenantConf {
4088 0 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
4089 0 : location: attached_conf.location,
4090 0 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
4091 : }))
4092 0 : })?;
4093 :
4094 0 : let updated = self.tenant_conf.load();
4095 0 :
4096 0 : self.tenant_conf_updated(&updated.tenant_conf);
4097 0 : // Don't hold self.timelines.lock() during the notifies.
4098 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4099 0 : // mutexes in struct Timeline in the future.
4100 0 : let timelines = self.list_timelines();
4101 0 : for timeline in timelines {
4102 0 : timeline.tenant_conf_updated(&updated);
4103 0 : }
4104 :
4105 0 : Ok(updated.tenant_conf.clone())
4106 0 : }
4107 :
4108 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4109 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4110 0 :
4111 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4112 0 :
4113 0 : self.tenant_conf_updated(&new_tenant_conf);
4114 0 : // Don't hold self.timelines.lock() during the notifies.
4115 0 : // There's no risk of deadlock right now, but there could be if we consolidate
4116 0 : // mutexes in struct Timeline in the future.
4117 0 : let timelines = self.list_timelines();
4118 0 : for timeline in timelines {
4119 0 : timeline.tenant_conf_updated(&new_conf);
4120 0 : }
4121 0 : }
4122 :
4123 452 : fn get_pagestream_throttle_config(
4124 452 : psconf: &'static PageServerConf,
4125 452 : overrides: &TenantConfOpt,
4126 452 : ) -> throttle::Config {
4127 452 : overrides
4128 452 : .timeline_get_throttle
4129 452 : .clone()
4130 452 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4131 452 : }
4132 :
4133 0 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
4134 0 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4135 0 : self.pagestream_throttle.reconfigure(conf)
4136 0 : }
4137 :
4138 : /// Helper function to create a new Timeline struct.
4139 : ///
4140 : /// The returned Timeline is in Loading state. The caller is responsible for
4141 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4142 : /// map.
4143 : ///
4144 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4145 : /// and we might not have the ancestor present anymore which is fine for to be
4146 : /// deleted timelines.
4147 : #[allow(clippy::too_many_arguments)]
4148 904 : fn create_timeline_struct(
4149 904 : &self,
4150 904 : new_timeline_id: TimelineId,
4151 904 : new_metadata: &TimelineMetadata,
4152 904 : previous_heatmap: Option<PreviousHeatmap>,
4153 904 : ancestor: Option<Arc<Timeline>>,
4154 904 : resources: TimelineResources,
4155 904 : cause: CreateTimelineCause,
4156 904 : create_idempotency: CreateTimelineIdempotency,
4157 904 : gc_compaction_state: Option<GcCompactionState>,
4158 904 : rel_size_v2_status: Option<RelSizeMigration>,
4159 904 : ctx: &RequestContext,
4160 904 : ) -> anyhow::Result<(Arc<Timeline>, RequestContext)> {
4161 904 : let state = match cause {
4162 : CreateTimelineCause::Load => {
4163 904 : let ancestor_id = new_metadata.ancestor_timeline();
4164 904 : anyhow::ensure!(
4165 904 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4166 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4167 : );
4168 904 : TimelineState::Loading
4169 : }
4170 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4171 : };
4172 :
4173 904 : let pg_version = new_metadata.pg_version();
4174 904 :
4175 904 : let timeline = Timeline::new(
4176 904 : self.conf,
4177 904 : Arc::clone(&self.tenant_conf),
4178 904 : new_metadata,
4179 904 : previous_heatmap,
4180 904 : ancestor,
4181 904 : new_timeline_id,
4182 904 : self.tenant_shard_id,
4183 904 : self.generation,
4184 904 : self.shard_identity,
4185 904 : self.walredo_mgr.clone(),
4186 904 : resources,
4187 904 : pg_version,
4188 904 : state,
4189 904 : self.attach_wal_lag_cooldown.clone(),
4190 904 : create_idempotency,
4191 904 : gc_compaction_state,
4192 904 : rel_size_v2_status,
4193 904 : self.cancel.child_token(),
4194 904 : );
4195 904 :
4196 904 : let timeline_ctx = RequestContextBuilder::extend(ctx)
4197 904 : .scope(context::Scope::new_timeline(&timeline))
4198 904 : .build();
4199 904 :
4200 904 : Ok((timeline, timeline_ctx))
4201 904 : }
4202 :
4203 : /// [`Tenant::shutdown`] must be called before dropping the returned [`Tenant`] object
4204 : /// to ensure proper cleanup of background tasks and metrics.
4205 : //
4206 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4207 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4208 : #[allow(clippy::too_many_arguments)]
4209 452 : fn new(
4210 452 : state: TenantState,
4211 452 : conf: &'static PageServerConf,
4212 452 : attached_conf: AttachedTenantConf,
4213 452 : shard_identity: ShardIdentity,
4214 452 : walredo_mgr: Option<Arc<WalRedoManager>>,
4215 452 : tenant_shard_id: TenantShardId,
4216 452 : remote_storage: GenericRemoteStorage,
4217 452 : deletion_queue_client: DeletionQueueClient,
4218 452 : l0_flush_global_state: L0FlushGlobalState,
4219 452 : ) -> Tenant {
4220 452 : debug_assert!(
4221 452 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
4222 : );
4223 :
4224 452 : let (state, mut rx) = watch::channel(state);
4225 452 :
4226 452 : tokio::spawn(async move {
4227 451 : // reflect tenant state in metrics:
4228 451 : // - global per tenant state: TENANT_STATE_METRIC
4229 451 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4230 451 : //
4231 451 : // set of broken tenants should not have zero counts so that it remains accessible for
4232 451 : // alerting.
4233 451 :
4234 451 : let tid = tenant_shard_id.to_string();
4235 451 : let shard_id = tenant_shard_id.shard_slug().to_string();
4236 451 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4237 :
4238 903 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4239 903 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4240 903 : }
4241 :
4242 451 : let mut tuple = inspect_state(&rx.borrow_and_update());
4243 451 :
4244 451 : let is_broken = tuple.1;
4245 451 : let mut counted_broken = if is_broken {
4246 : // add the id to the set right away, there should not be any updates on the channel
4247 : // after before tenant is removed, if ever
4248 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4249 0 : true
4250 : } else {
4251 451 : false
4252 : };
4253 :
4254 : loop {
4255 903 : let labels = &tuple.0;
4256 903 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4257 903 : current.inc();
4258 903 :
4259 903 : if rx.changed().await.is_err() {
4260 : // tenant has been dropped
4261 28 : current.dec();
4262 28 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4263 28 : break;
4264 452 : }
4265 452 :
4266 452 : current.dec();
4267 452 : tuple = inspect_state(&rx.borrow_and_update());
4268 452 :
4269 452 : let is_broken = tuple.1;
4270 452 : if is_broken && !counted_broken {
4271 0 : counted_broken = true;
4272 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4273 0 : // access
4274 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4275 452 : }
4276 : }
4277 452 : });
4278 452 :
4279 452 : Tenant {
4280 452 : tenant_shard_id,
4281 452 : shard_identity,
4282 452 : generation: attached_conf.location.generation,
4283 452 : conf,
4284 452 : // using now here is good enough approximation to catch tenants with really long
4285 452 : // activation times.
4286 452 : constructed_at: Instant::now(),
4287 452 : timelines: Mutex::new(HashMap::new()),
4288 452 : timelines_creating: Mutex::new(HashSet::new()),
4289 452 : timelines_offloaded: Mutex::new(HashMap::new()),
4290 452 : tenant_manifest_upload: Default::default(),
4291 452 : gc_cs: tokio::sync::Mutex::new(()),
4292 452 : walredo_mgr,
4293 452 : remote_storage,
4294 452 : deletion_queue_client,
4295 452 : state,
4296 452 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4297 452 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4298 452 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4299 452 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4300 452 : format!("compaction-{tenant_shard_id}"),
4301 452 : 5,
4302 452 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4303 452 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4304 452 : // use an extremely long backoff.
4305 452 : Some(Duration::from_secs(3600 * 24)),
4306 452 : )),
4307 452 : l0_compaction_trigger: Arc::new(Notify::new()),
4308 452 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4309 452 : activate_now_sem: tokio::sync::Semaphore::new(0),
4310 452 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4311 452 : cancel: CancellationToken::default(),
4312 452 : gate: Gate::default(),
4313 452 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4314 452 : Tenant::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4315 452 : )),
4316 452 : pagestream_throttle_metrics: Arc::new(
4317 452 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4318 452 : ),
4319 452 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4320 452 : ongoing_timeline_detach: std::sync::Mutex::default(),
4321 452 : gc_block: Default::default(),
4322 452 : l0_flush_global_state,
4323 452 : }
4324 452 : }
4325 :
4326 : /// Locate and load config
4327 0 : pub(super) fn load_tenant_config(
4328 0 : conf: &'static PageServerConf,
4329 0 : tenant_shard_id: &TenantShardId,
4330 0 : ) -> Result<LocationConf, LoadConfigError> {
4331 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4332 0 :
4333 0 : info!("loading tenant configuration from {config_path}");
4334 :
4335 : // load and parse file
4336 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4337 0 : match e.kind() {
4338 : std::io::ErrorKind::NotFound => {
4339 : // The config should almost always exist for a tenant directory:
4340 : // - When attaching a tenant, the config is the first thing we write
4341 : // - When detaching a tenant, we atomically move the directory to a tmp location
4342 : // before deleting contents.
4343 : //
4344 : // The very rare edge case that can result in a missing config is if we crash during attach
4345 : // between creating directory and writing config. Callers should handle that as if the
4346 : // directory didn't exist.
4347 :
4348 0 : LoadConfigError::NotFound(config_path)
4349 : }
4350 : _ => {
4351 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4352 : // that we cannot cleanly recover
4353 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4354 : }
4355 : }
4356 0 : })?;
4357 :
4358 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4359 0 : }
4360 :
4361 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4362 : pub(super) async fn persist_tenant_config(
4363 : conf: &'static PageServerConf,
4364 : tenant_shard_id: &TenantShardId,
4365 : location_conf: &LocationConf,
4366 : ) -> std::io::Result<()> {
4367 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4368 :
4369 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4370 : }
4371 :
4372 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4373 : pub(super) async fn persist_tenant_config_at(
4374 : tenant_shard_id: &TenantShardId,
4375 : config_path: &Utf8Path,
4376 : location_conf: &LocationConf,
4377 : ) -> std::io::Result<()> {
4378 : debug!("persisting tenantconf to {config_path}");
4379 :
4380 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4381 : # It is read in case of pageserver restart.
4382 : "#
4383 : .to_string();
4384 :
4385 0 : fail::fail_point!("tenant-config-before-write", |_| {
4386 0 : Err(std::io::Error::new(
4387 0 : std::io::ErrorKind::Other,
4388 0 : "tenant-config-before-write",
4389 0 : ))
4390 0 : });
4391 :
4392 : // Convert the config to a toml file.
4393 : conf_content +=
4394 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4395 :
4396 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4397 :
4398 : let conf_content = conf_content.into_bytes();
4399 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4400 : }
4401 :
4402 : //
4403 : // How garbage collection works:
4404 : //
4405 : // +--bar------------->
4406 : // /
4407 : // +----+-----foo---------------->
4408 : // /
4409 : // ----main--+-------------------------->
4410 : // \
4411 : // +-----baz-------->
4412 : //
4413 : //
4414 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4415 : // `gc_infos` are being refreshed
4416 : // 2. Scan collected timelines, and on each timeline, make note of the
4417 : // all the points where other timelines have been branched off.
4418 : // We will refrain from removing page versions at those LSNs.
4419 : // 3. For each timeline, scan all layer files on the timeline.
4420 : // Remove all files for which a newer file exists and which
4421 : // don't cover any branch point LSNs.
4422 : //
4423 : // TODO:
4424 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4425 : // don't need to keep that in the parent anymore. But currently
4426 : // we do.
4427 8 : async fn gc_iteration_internal(
4428 8 : &self,
4429 8 : target_timeline_id: Option<TimelineId>,
4430 8 : horizon: u64,
4431 8 : pitr: Duration,
4432 8 : cancel: &CancellationToken,
4433 8 : ctx: &RequestContext,
4434 8 : ) -> Result<GcResult, GcError> {
4435 8 : let mut totals: GcResult = Default::default();
4436 8 : let now = Instant::now();
4437 :
4438 8 : let gc_timelines = self
4439 8 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4440 8 : .await?;
4441 :
4442 8 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4443 :
4444 : // If there is nothing to GC, we don't want any messages in the INFO log.
4445 8 : if !gc_timelines.is_empty() {
4446 8 : info!("{} timelines need GC", gc_timelines.len());
4447 : } else {
4448 0 : debug!("{} timelines need GC", gc_timelines.len());
4449 : }
4450 :
4451 : // Perform GC for each timeline.
4452 : //
4453 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
4454 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4455 : // with branch creation.
4456 : //
4457 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
4458 : // creation task can run concurrently with timeline's GC iteration.
4459 16 : for timeline in gc_timelines {
4460 8 : if cancel.is_cancelled() {
4461 : // We were requested to shut down. Stop and return with the progress we
4462 : // made.
4463 0 : break;
4464 8 : }
4465 8 : let result = match timeline.gc().await {
4466 : Err(GcError::TimelineCancelled) => {
4467 0 : if target_timeline_id.is_some() {
4468 : // If we were targetting this specific timeline, surface cancellation to caller
4469 0 : return Err(GcError::TimelineCancelled);
4470 : } else {
4471 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4472 : // skip past this and proceed to try GC on other timelines.
4473 0 : continue;
4474 : }
4475 : }
4476 8 : r => r?,
4477 : };
4478 8 : totals += result;
4479 : }
4480 :
4481 8 : totals.elapsed = now.elapsed();
4482 8 : Ok(totals)
4483 8 : }
4484 :
4485 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4486 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4487 : /// [`Tenant::get_gc_horizon`].
4488 : ///
4489 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4490 0 : pub(crate) async fn refresh_gc_info(
4491 0 : &self,
4492 0 : cancel: &CancellationToken,
4493 0 : ctx: &RequestContext,
4494 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4495 0 : // since this method can now be called at different rates than the configured gc loop, it
4496 0 : // might be that these configuration values get applied faster than what it was previously,
4497 0 : // since these were only read from the gc task.
4498 0 : let horizon = self.get_gc_horizon();
4499 0 : let pitr = self.get_pitr_interval();
4500 0 :
4501 0 : // refresh all timelines
4502 0 : let target_timeline_id = None;
4503 0 :
4504 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4505 0 : .await
4506 0 : }
4507 :
4508 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4509 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4510 : ///
4511 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4512 0 : fn initialize_gc_info(
4513 0 : &self,
4514 0 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4515 0 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4516 0 : restrict_to_timeline: Option<TimelineId>,
4517 0 : ) {
4518 0 : if restrict_to_timeline.is_none() {
4519 : // This function must be called before activation: after activation timeline create/delete operations
4520 : // might happen, and this function is not safe to run concurrently with those.
4521 0 : assert!(!self.is_active());
4522 0 : }
4523 :
4524 : // Scan all timelines. For each timeline, remember the timeline ID and
4525 : // the branch point where it was created.
4526 0 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4527 0 : BTreeMap::new();
4528 0 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4529 0 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4530 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4531 0 : ancestor_children.push((
4532 0 : timeline_entry.get_ancestor_lsn(),
4533 0 : *timeline_id,
4534 0 : MaybeOffloaded::No,
4535 0 : ));
4536 0 : }
4537 0 : });
4538 0 : timelines_offloaded
4539 0 : .iter()
4540 0 : .for_each(|(timeline_id, timeline_entry)| {
4541 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4542 0 : return;
4543 : };
4544 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4545 0 : return;
4546 : };
4547 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4548 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4549 0 : });
4550 0 :
4551 0 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4552 0 : let horizon = self.get_gc_horizon();
4553 :
4554 : // Populate each timeline's GcInfo with information about its child branches
4555 0 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4556 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4557 : } else {
4558 0 : itertools::Either::Right(timelines.values())
4559 : };
4560 0 : for timeline in timelines_to_write {
4561 0 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4562 0 : .remove(&timeline.timeline_id)
4563 0 : .unwrap_or_default();
4564 0 :
4565 0 : branchpoints.sort_by_key(|b| b.0);
4566 0 :
4567 0 : let mut target = timeline.gc_info.write().unwrap();
4568 0 :
4569 0 : target.retain_lsns = branchpoints;
4570 0 :
4571 0 : let space_cutoff = timeline
4572 0 : .get_last_record_lsn()
4573 0 : .checked_sub(horizon)
4574 0 : .unwrap_or(Lsn(0));
4575 0 :
4576 0 : target.cutoffs = GcCutoffs {
4577 0 : space: space_cutoff,
4578 0 : time: Lsn::INVALID,
4579 0 : };
4580 0 : }
4581 0 : }
4582 :
4583 8 : async fn refresh_gc_info_internal(
4584 8 : &self,
4585 8 : target_timeline_id: Option<TimelineId>,
4586 8 : horizon: u64,
4587 8 : pitr: Duration,
4588 8 : cancel: &CancellationToken,
4589 8 : ctx: &RequestContext,
4590 8 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4591 8 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4592 8 : // currently visible timelines.
4593 8 : let timelines = self
4594 8 : .timelines
4595 8 : .lock()
4596 8 : .unwrap()
4597 8 : .values()
4598 8 : .filter(|tl| match target_timeline_id.as_ref() {
4599 8 : Some(target) => &tl.timeline_id == target,
4600 0 : None => true,
4601 8 : })
4602 8 : .cloned()
4603 8 : .collect::<Vec<_>>();
4604 8 :
4605 8 : if target_timeline_id.is_some() && timelines.is_empty() {
4606 : // We were to act on a particular timeline and it wasn't found
4607 0 : return Err(GcError::TimelineNotFound);
4608 8 : }
4609 8 :
4610 8 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4611 8 : HashMap::with_capacity(timelines.len());
4612 8 :
4613 8 : // Ensures all timelines use the same start time when computing the time cutoff.
4614 8 : let now_ts_for_pitr_calc = SystemTime::now();
4615 8 : for timeline in timelines.iter() {
4616 8 : let ctx = &ctx.with_scope_timeline(timeline);
4617 8 : let cutoff = timeline
4618 8 : .get_last_record_lsn()
4619 8 : .checked_sub(horizon)
4620 8 : .unwrap_or(Lsn(0));
4621 :
4622 8 : let cutoffs = timeline
4623 8 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4624 8 : .await?;
4625 8 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4626 8 : assert!(old.is_none());
4627 : }
4628 :
4629 8 : if !self.is_active() || self.cancel.is_cancelled() {
4630 0 : return Err(GcError::TenantCancelled);
4631 8 : }
4632 :
4633 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4634 : // because that will stall branch creation.
4635 8 : let gc_cs = self.gc_cs.lock().await;
4636 :
4637 : // Ok, we now know all the branch points.
4638 : // Update the GC information for each timeline.
4639 8 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4640 16 : for timeline in timelines {
4641 : // We filtered the timeline list above
4642 8 : if let Some(target_timeline_id) = target_timeline_id {
4643 8 : assert_eq!(target_timeline_id, timeline.timeline_id);
4644 0 : }
4645 :
4646 : {
4647 8 : let mut target = timeline.gc_info.write().unwrap();
4648 8 :
4649 8 : // Cull any expired leases
4650 8 : let now = SystemTime::now();
4651 12 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4652 8 :
4653 8 : timeline
4654 8 : .metrics
4655 8 : .valid_lsn_lease_count_gauge
4656 8 : .set(target.leases.len() as u64);
4657 :
4658 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4659 8 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4660 0 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4661 0 : target.within_ancestor_pitr =
4662 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
4663 0 : }
4664 8 : }
4665 :
4666 : // Update metrics that depend on GC state
4667 8 : timeline
4668 8 : .metrics
4669 8 : .archival_size
4670 8 : .set(if target.within_ancestor_pitr {
4671 0 : timeline.metrics.current_logical_size_gauge.get()
4672 : } else {
4673 8 : 0
4674 : });
4675 8 : timeline.metrics.pitr_history_size.set(
4676 8 : timeline
4677 8 : .get_last_record_lsn()
4678 8 : .checked_sub(target.cutoffs.time)
4679 8 : .unwrap_or(Lsn(0))
4680 8 : .0,
4681 8 : );
4682 :
4683 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4684 : // - this timeline was created while we were finding cutoffs
4685 : // - lsn for timestamp search fails for this timeline repeatedly
4686 8 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4687 8 : let original_cutoffs = target.cutoffs.clone();
4688 8 : // GC cutoffs should never go back
4689 8 : target.cutoffs = GcCutoffs {
4690 8 : space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
4691 8 : time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
4692 8 : }
4693 0 : }
4694 : }
4695 :
4696 8 : gc_timelines.push(timeline);
4697 : }
4698 8 : drop(gc_cs);
4699 8 : Ok(gc_timelines)
4700 8 : }
4701 :
4702 : /// A substitute for `branch_timeline` for use in unit tests.
4703 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4704 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4705 : /// timeline background tasks are launched, except the flush loop.
4706 : #[cfg(test)]
4707 464 : async fn branch_timeline_test(
4708 464 : self: &Arc<Self>,
4709 464 : src_timeline: &Arc<Timeline>,
4710 464 : dst_id: TimelineId,
4711 464 : ancestor_lsn: Option<Lsn>,
4712 464 : ctx: &RequestContext,
4713 464 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4714 464 : let tl = self
4715 464 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4716 464 : .await?
4717 456 : .into_timeline_for_test();
4718 456 : tl.set_state(TimelineState::Active);
4719 456 : Ok(tl)
4720 464 : }
4721 :
4722 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4723 : #[cfg(test)]
4724 : #[allow(clippy::too_many_arguments)]
4725 12 : pub async fn branch_timeline_test_with_layers(
4726 12 : self: &Arc<Self>,
4727 12 : src_timeline: &Arc<Timeline>,
4728 12 : dst_id: TimelineId,
4729 12 : ancestor_lsn: Option<Lsn>,
4730 12 : ctx: &RequestContext,
4731 12 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4732 12 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4733 12 : end_lsn: Lsn,
4734 12 : ) -> anyhow::Result<Arc<Timeline>> {
4735 : use checks::check_valid_layermap;
4736 : use itertools::Itertools;
4737 :
4738 12 : let tline = self
4739 12 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4740 12 : .await?;
4741 12 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4742 12 : ancestor_lsn
4743 : } else {
4744 0 : tline.get_last_record_lsn()
4745 : };
4746 12 : assert!(end_lsn >= ancestor_lsn);
4747 12 : tline.force_advance_lsn(end_lsn);
4748 24 : for deltas in delta_layer_desc {
4749 12 : tline
4750 12 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4751 12 : .await?;
4752 : }
4753 20 : for (lsn, images) in image_layer_desc {
4754 8 : tline
4755 8 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4756 8 : .await?;
4757 : }
4758 12 : let layer_names = tline
4759 12 : .layers
4760 12 : .read()
4761 12 : .await
4762 12 : .layer_map()
4763 12 : .unwrap()
4764 12 : .iter_historic_layers()
4765 20 : .map(|layer| layer.layer_name())
4766 12 : .collect_vec();
4767 12 : if let Some(err) = check_valid_layermap(&layer_names) {
4768 0 : bail!("invalid layermap: {err}");
4769 12 : }
4770 12 : Ok(tline)
4771 12 : }
4772 :
4773 : /// Branch an existing timeline.
4774 0 : async fn branch_timeline(
4775 0 : self: &Arc<Self>,
4776 0 : src_timeline: &Arc<Timeline>,
4777 0 : dst_id: TimelineId,
4778 0 : start_lsn: Option<Lsn>,
4779 0 : ctx: &RequestContext,
4780 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4781 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4782 0 : .await
4783 0 : }
4784 :
4785 464 : async fn branch_timeline_impl(
4786 464 : self: &Arc<Self>,
4787 464 : src_timeline: &Arc<Timeline>,
4788 464 : dst_id: TimelineId,
4789 464 : start_lsn: Option<Lsn>,
4790 464 : ctx: &RequestContext,
4791 464 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4792 464 : let src_id = src_timeline.timeline_id;
4793 :
4794 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4795 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4796 : // valid while we are creating the branch.
4797 464 : let _gc_cs = self.gc_cs.lock().await;
4798 :
4799 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4800 464 : let start_lsn = start_lsn.unwrap_or_else(|| {
4801 4 : let lsn = src_timeline.get_last_record_lsn();
4802 4 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4803 4 : lsn
4804 464 : });
4805 :
4806 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4807 464 : let timeline_create_guard = match self
4808 464 : .start_creating_timeline(
4809 464 : dst_id,
4810 464 : CreateTimelineIdempotency::Branch {
4811 464 : ancestor_timeline_id: src_timeline.timeline_id,
4812 464 : ancestor_start_lsn: start_lsn,
4813 464 : },
4814 464 : )
4815 464 : .await?
4816 : {
4817 464 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
4818 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
4819 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
4820 : }
4821 : };
4822 :
4823 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
4824 : // horizon on the source timeline
4825 : //
4826 : // We check it against both the planned GC cutoff stored in 'gc_info',
4827 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
4828 : // planned GC cutoff in 'gc_info' is normally larger than
4829 : // 'applied_gc_cutoff_lsn', but beware of corner cases like if you just
4830 : // changed the GC settings for the tenant to make the PITR window
4831 : // larger, but some of the data was already removed by an earlier GC
4832 : // iteration.
4833 :
4834 : // check against last actual 'latest_gc_cutoff' first
4835 464 : let applied_gc_cutoff_lsn = src_timeline.get_applied_gc_cutoff_lsn();
4836 464 : {
4837 464 : let gc_info = src_timeline.gc_info.read().unwrap();
4838 464 : let planned_cutoff = gc_info.min_cutoff();
4839 464 : if gc_info.lsn_covered_by_lease(start_lsn) {
4840 0 : tracing::info!(
4841 0 : "skipping comparison of {start_lsn} with gc cutoff {} and planned gc cutoff {planned_cutoff} due to lsn lease",
4842 0 : *applied_gc_cutoff_lsn
4843 : );
4844 : } else {
4845 464 : src_timeline
4846 464 : .check_lsn_is_in_scope(start_lsn, &applied_gc_cutoff_lsn)
4847 464 : .context(format!(
4848 464 : "invalid branch start lsn: less than latest GC cutoff {}",
4849 464 : *applied_gc_cutoff_lsn,
4850 464 : ))
4851 464 : .map_err(CreateTimelineError::AncestorLsn)?;
4852 :
4853 : // and then the planned GC cutoff
4854 456 : if start_lsn < planned_cutoff {
4855 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
4856 0 : "invalid branch start lsn: less than planned GC cutoff {planned_cutoff}"
4857 0 : )));
4858 456 : }
4859 : }
4860 : }
4861 :
4862 : //
4863 : // The branch point is valid, and we are still holding the 'gc_cs' lock
4864 : // so that GC cannot advance the GC cutoff until we are finished.
4865 : // Proceed with the branch creation.
4866 : //
4867 :
4868 : // Determine prev-LSN for the new timeline. We can only determine it if
4869 : // the timeline was branched at the current end of the source timeline.
4870 : let RecordLsn {
4871 456 : last: src_last,
4872 456 : prev: src_prev,
4873 456 : } = src_timeline.get_last_record_rlsn();
4874 456 : let dst_prev = if src_last == start_lsn {
4875 432 : Some(src_prev)
4876 : } else {
4877 24 : None
4878 : };
4879 :
4880 : // Create the metadata file, noting the ancestor of the new timeline.
4881 : // There is initially no data in it, but all the read-calls know to look
4882 : // into the ancestor.
4883 456 : let metadata = TimelineMetadata::new(
4884 456 : start_lsn,
4885 456 : dst_prev,
4886 456 : Some(src_id),
4887 456 : start_lsn,
4888 456 : *src_timeline.applied_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
4889 456 : src_timeline.initdb_lsn,
4890 456 : src_timeline.pg_version,
4891 456 : );
4892 :
4893 456 : let (uninitialized_timeline, _timeline_ctx) = self
4894 456 : .prepare_new_timeline(
4895 456 : dst_id,
4896 456 : &metadata,
4897 456 : timeline_create_guard,
4898 456 : start_lsn + 1,
4899 456 : Some(Arc::clone(src_timeline)),
4900 456 : Some(src_timeline.get_rel_size_v2_status()),
4901 456 : ctx,
4902 456 : )
4903 456 : .await?;
4904 :
4905 456 : let new_timeline = uninitialized_timeline.finish_creation().await?;
4906 :
4907 : // Root timeline gets its layers during creation and uploads them along with the metadata.
4908 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
4909 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
4910 : // could get incorrect information and remove more layers, than needed.
4911 : // See also https://github.com/neondatabase/neon/issues/3865
4912 456 : new_timeline
4913 456 : .remote_client
4914 456 : .schedule_index_upload_for_full_metadata_update(&metadata)
4915 456 : .context("branch initial metadata upload")?;
4916 :
4917 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
4918 :
4919 456 : Ok(CreateTimelineResult::Created(new_timeline))
4920 464 : }
4921 :
4922 : /// For unit tests, make this visible so that other modules can directly create timelines
4923 : #[cfg(test)]
4924 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
4925 : pub(crate) async fn bootstrap_timeline_test(
4926 : self: &Arc<Self>,
4927 : timeline_id: TimelineId,
4928 : pg_version: u32,
4929 : load_existing_initdb: Option<TimelineId>,
4930 : ctx: &RequestContext,
4931 : ) -> anyhow::Result<Arc<Timeline>> {
4932 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
4933 : .await
4934 : .map_err(anyhow::Error::new)
4935 4 : .map(|r| r.into_timeline_for_test())
4936 : }
4937 :
4938 : /// Get exclusive access to the timeline ID for creation.
4939 : ///
4940 : /// Timeline-creating code paths must use this function before making changes
4941 : /// to in-memory or persistent state.
4942 : ///
4943 : /// The `state` parameter is a description of the timeline creation operation
4944 : /// we intend to perform.
4945 : /// If the timeline was already created in the meantime, we check whether this
4946 : /// request conflicts or is idempotent , based on `state`.
4947 904 : async fn start_creating_timeline(
4948 904 : self: &Arc<Self>,
4949 904 : new_timeline_id: TimelineId,
4950 904 : idempotency: CreateTimelineIdempotency,
4951 904 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
4952 904 : let allow_offloaded = false;
4953 904 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
4954 900 : Ok(create_guard) => {
4955 900 : pausable_failpoint!("timeline-creation-after-uninit");
4956 900 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
4957 : }
4958 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
4959 : Err(TimelineExclusionError::AlreadyCreating) => {
4960 : // Creation is in progress, we cannot create it again, and we cannot
4961 : // check if this request matches the existing one, so caller must try
4962 : // again later.
4963 0 : Err(CreateTimelineError::AlreadyCreating)
4964 : }
4965 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
4966 : Err(TimelineExclusionError::AlreadyExists {
4967 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
4968 0 : ..
4969 0 : }) => {
4970 0 : info!("timeline already exists but is offloaded");
4971 0 : Err(CreateTimelineError::Conflict)
4972 : }
4973 : Err(TimelineExclusionError::AlreadyExists {
4974 4 : existing: TimelineOrOffloaded::Timeline(existing),
4975 4 : arg,
4976 4 : }) => {
4977 4 : {
4978 4 : let existing = &existing.create_idempotency;
4979 4 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
4980 4 : debug!("timeline already exists");
4981 :
4982 4 : match (existing, &arg) {
4983 : // FailWithConflict => no idempotency check
4984 : (CreateTimelineIdempotency::FailWithConflict, _)
4985 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
4986 4 : warn!("timeline already exists, failing request");
4987 4 : return Err(CreateTimelineError::Conflict);
4988 : }
4989 : // Idempotent <=> CreateTimelineIdempotency is identical
4990 0 : (x, y) if x == y => {
4991 0 : info!(
4992 0 : "timeline already exists and idempotency matches, succeeding request"
4993 : );
4994 : // fallthrough
4995 : }
4996 : (_, _) => {
4997 0 : warn!("idempotency conflict, failing request");
4998 0 : return Err(CreateTimelineError::Conflict);
4999 : }
5000 : }
5001 : }
5002 :
5003 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
5004 : }
5005 : }
5006 904 : }
5007 :
5008 0 : async fn upload_initdb(
5009 0 : &self,
5010 0 : timelines_path: &Utf8PathBuf,
5011 0 : pgdata_path: &Utf8PathBuf,
5012 0 : timeline_id: &TimelineId,
5013 0 : ) -> anyhow::Result<()> {
5014 0 : let temp_path = timelines_path.join(format!(
5015 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
5016 0 : ));
5017 0 :
5018 0 : scopeguard::defer! {
5019 0 : if let Err(e) = fs::remove_file(&temp_path) {
5020 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
5021 0 : }
5022 0 : }
5023 :
5024 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
5025 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
5026 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
5027 0 : warn!(
5028 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
5029 : );
5030 0 : }
5031 :
5032 0 : pausable_failpoint!("before-initdb-upload");
5033 :
5034 0 : backoff::retry(
5035 0 : || async {
5036 0 : self::remote_timeline_client::upload_initdb_dir(
5037 0 : &self.remote_storage,
5038 0 : &self.tenant_shard_id.tenant_id,
5039 0 : timeline_id,
5040 0 : pgdata_zstd.try_clone().await?,
5041 0 : tar_zst_size,
5042 0 : &self.cancel,
5043 0 : )
5044 0 : .await
5045 0 : },
5046 0 : |_| false,
5047 0 : 3,
5048 0 : u32::MAX,
5049 0 : "persist_initdb_tar_zst",
5050 0 : &self.cancel,
5051 0 : )
5052 0 : .await
5053 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
5054 0 : .and_then(|x| x)
5055 0 : }
5056 :
5057 : /// - run initdb to init temporary instance and get bootstrap data
5058 : /// - after initialization completes, tar up the temp dir and upload it to S3.
5059 4 : async fn bootstrap_timeline(
5060 4 : self: &Arc<Self>,
5061 4 : timeline_id: TimelineId,
5062 4 : pg_version: u32,
5063 4 : load_existing_initdb: Option<TimelineId>,
5064 4 : ctx: &RequestContext,
5065 4 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
5066 4 : let timeline_create_guard = match self
5067 4 : .start_creating_timeline(
5068 4 : timeline_id,
5069 4 : CreateTimelineIdempotency::Bootstrap { pg_version },
5070 4 : )
5071 4 : .await?
5072 : {
5073 4 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5074 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5075 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5076 : }
5077 : };
5078 :
5079 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
5080 : // temporary directory for basebackup files for the given timeline.
5081 :
5082 4 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
5083 4 : let pgdata_path = path_with_suffix_extension(
5084 4 : timelines_path.join(format!("basebackup-{timeline_id}")),
5085 4 : TEMP_FILE_SUFFIX,
5086 4 : );
5087 4 :
5088 4 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
5089 4 : // we won't race with other creations or existent timelines with the same path.
5090 4 : if pgdata_path.exists() {
5091 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
5092 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
5093 0 : })?;
5094 4 : }
5095 :
5096 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
5097 4 : let pgdata_path_deferred = pgdata_path.clone();
5098 4 : scopeguard::defer! {
5099 4 : if let Err(e) = fs::remove_dir_all(&pgdata_path_deferred) {
5100 4 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
5101 4 : error!("Failed to remove temporary initdb directory '{pgdata_path_deferred}': {e}");
5102 4 : }
5103 4 : }
5104 4 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
5105 4 : if existing_initdb_timeline_id != timeline_id {
5106 0 : let source_path = &remote_initdb_archive_path(
5107 0 : &self.tenant_shard_id.tenant_id,
5108 0 : &existing_initdb_timeline_id,
5109 0 : );
5110 0 : let dest_path =
5111 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
5112 0 :
5113 0 : // if this fails, it will get retried by retried control plane requests
5114 0 : self.remote_storage
5115 0 : .copy_object(source_path, dest_path, &self.cancel)
5116 0 : .await
5117 0 : .context("copy initdb tar")?;
5118 4 : }
5119 4 : let (initdb_tar_zst_path, initdb_tar_zst) =
5120 4 : self::remote_timeline_client::download_initdb_tar_zst(
5121 4 : self.conf,
5122 4 : &self.remote_storage,
5123 4 : &self.tenant_shard_id,
5124 4 : &existing_initdb_timeline_id,
5125 4 : &self.cancel,
5126 4 : )
5127 4 : .await
5128 4 : .context("download initdb tar")?;
5129 :
5130 4 : scopeguard::defer! {
5131 4 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
5132 4 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
5133 4 : }
5134 4 : }
5135 4 :
5136 4 : let buf_read =
5137 4 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5138 4 : extract_zst_tarball(&pgdata_path, buf_read)
5139 4 : .await
5140 4 : .context("extract initdb tar")?;
5141 : } else {
5142 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5143 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5144 0 : .await
5145 0 : .context("run initdb")?;
5146 :
5147 : // Upload the created data dir to S3
5148 0 : if self.tenant_shard_id().is_shard_zero() {
5149 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5150 0 : .await?;
5151 0 : }
5152 : }
5153 4 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5154 4 :
5155 4 : // Import the contents of the data directory at the initial checkpoint
5156 4 : // LSN, and any WAL after that.
5157 4 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5158 4 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5159 4 : let new_metadata = TimelineMetadata::new(
5160 4 : Lsn(0),
5161 4 : None,
5162 4 : None,
5163 4 : Lsn(0),
5164 4 : pgdata_lsn,
5165 4 : pgdata_lsn,
5166 4 : pg_version,
5167 4 : );
5168 4 : let (mut raw_timeline, timeline_ctx) = self
5169 4 : .prepare_new_timeline(
5170 4 : timeline_id,
5171 4 : &new_metadata,
5172 4 : timeline_create_guard,
5173 4 : pgdata_lsn,
5174 4 : None,
5175 4 : None,
5176 4 : ctx,
5177 4 : )
5178 4 : .await?;
5179 :
5180 4 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5181 4 : raw_timeline
5182 4 : .write(|unfinished_timeline| async move {
5183 4 : import_datadir::import_timeline_from_postgres_datadir(
5184 4 : &unfinished_timeline,
5185 4 : &pgdata_path,
5186 4 : pgdata_lsn,
5187 4 : &timeline_ctx,
5188 4 : )
5189 4 : .await
5190 4 : .with_context(|| {
5191 0 : format!(
5192 0 : "Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}"
5193 0 : )
5194 4 : })?;
5195 :
5196 4 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5197 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5198 0 : "failpoint before-checkpoint-new-timeline"
5199 0 : )))
5200 4 : });
5201 :
5202 4 : Ok(())
5203 8 : })
5204 4 : .await?;
5205 :
5206 : // All done!
5207 4 : let timeline = raw_timeline.finish_creation().await?;
5208 :
5209 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5210 :
5211 4 : Ok(CreateTimelineResult::Created(timeline))
5212 4 : }
5213 :
5214 892 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5215 892 : RemoteTimelineClient::new(
5216 892 : self.remote_storage.clone(),
5217 892 : self.deletion_queue_client.clone(),
5218 892 : self.conf,
5219 892 : self.tenant_shard_id,
5220 892 : timeline_id,
5221 892 : self.generation,
5222 892 : &self.tenant_conf.load().location,
5223 892 : )
5224 892 : }
5225 :
5226 : /// Builds required resources for a new timeline.
5227 892 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5228 892 : let remote_client = self.build_timeline_remote_client(timeline_id);
5229 892 : self.get_timeline_resources_for(remote_client)
5230 892 : }
5231 :
5232 : /// Builds timeline resources for the given remote client.
5233 904 : fn get_timeline_resources_for(&self, remote_client: RemoteTimelineClient) -> TimelineResources {
5234 904 : TimelineResources {
5235 904 : remote_client,
5236 904 : pagestream_throttle: self.pagestream_throttle.clone(),
5237 904 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5238 904 : l0_compaction_trigger: self.l0_compaction_trigger.clone(),
5239 904 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5240 904 : }
5241 904 : }
5242 :
5243 : /// Creates intermediate timeline structure and its files.
5244 : ///
5245 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5246 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5247 : /// `finish_creation` to insert the Timeline into the timelines map.
5248 : #[allow(clippy::too_many_arguments)]
5249 892 : async fn prepare_new_timeline<'a>(
5250 892 : &'a self,
5251 892 : new_timeline_id: TimelineId,
5252 892 : new_metadata: &TimelineMetadata,
5253 892 : create_guard: TimelineCreateGuard,
5254 892 : start_lsn: Lsn,
5255 892 : ancestor: Option<Arc<Timeline>>,
5256 892 : rel_size_v2_status: Option<RelSizeMigration>,
5257 892 : ctx: &RequestContext,
5258 892 : ) -> anyhow::Result<(UninitializedTimeline<'a>, RequestContext)> {
5259 892 : let tenant_shard_id = self.tenant_shard_id;
5260 892 :
5261 892 : let resources = self.build_timeline_resources(new_timeline_id);
5262 892 : resources
5263 892 : .remote_client
5264 892 : .init_upload_queue_for_empty_remote(new_metadata, rel_size_v2_status.clone())?;
5265 :
5266 892 : let (timeline_struct, timeline_ctx) = self
5267 892 : .create_timeline_struct(
5268 892 : new_timeline_id,
5269 892 : new_metadata,
5270 892 : None,
5271 892 : ancestor,
5272 892 : resources,
5273 892 : CreateTimelineCause::Load,
5274 892 : create_guard.idempotency.clone(),
5275 892 : None,
5276 892 : rel_size_v2_status,
5277 892 : ctx,
5278 892 : )
5279 892 : .context("Failed to create timeline data structure")?;
5280 :
5281 892 : timeline_struct.init_empty_layer_map(start_lsn);
5282 :
5283 892 : if let Err(e) = self
5284 892 : .create_timeline_files(&create_guard.timeline_path)
5285 892 : .await
5286 : {
5287 0 : error!(
5288 0 : "Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}"
5289 : );
5290 0 : cleanup_timeline_directory(create_guard);
5291 0 : return Err(e);
5292 892 : }
5293 892 :
5294 892 : debug!(
5295 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5296 : );
5297 :
5298 892 : Ok((
5299 892 : UninitializedTimeline::new(
5300 892 : self,
5301 892 : new_timeline_id,
5302 892 : Some((timeline_struct, create_guard)),
5303 892 : ),
5304 892 : timeline_ctx,
5305 892 : ))
5306 892 : }
5307 :
5308 892 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5309 892 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5310 :
5311 892 : fail::fail_point!("after-timeline-dir-creation", |_| {
5312 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5313 892 : });
5314 :
5315 892 : Ok(())
5316 892 : }
5317 :
5318 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5319 : /// concurrent attempts to create the same timeline.
5320 : ///
5321 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5322 : /// offloaded timelines or not.
5323 904 : fn create_timeline_create_guard(
5324 904 : self: &Arc<Self>,
5325 904 : timeline_id: TimelineId,
5326 904 : idempotency: CreateTimelineIdempotency,
5327 904 : allow_offloaded: bool,
5328 904 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5329 904 : let tenant_shard_id = self.tenant_shard_id;
5330 904 :
5331 904 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5332 :
5333 904 : let create_guard = TimelineCreateGuard::new(
5334 904 : self,
5335 904 : timeline_id,
5336 904 : timeline_path.clone(),
5337 904 : idempotency,
5338 904 : allow_offloaded,
5339 904 : )?;
5340 :
5341 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5342 : // for creation.
5343 : // A timeline directory should never exist on disk already:
5344 : // - a previous failed creation would have cleaned up after itself
5345 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5346 : //
5347 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5348 : // this error may indicate a bug in cleanup on failed creations.
5349 900 : if timeline_path.exists() {
5350 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5351 0 : "Timeline directory already exists! This is a bug."
5352 0 : )));
5353 900 : }
5354 900 :
5355 900 : Ok(create_guard)
5356 904 : }
5357 :
5358 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5359 : ///
5360 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5361 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5362 : pub async fn gather_size_inputs(
5363 : &self,
5364 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5365 : // (only if it is shorter than the real cutoff).
5366 : max_retention_period: Option<u64>,
5367 : cause: LogicalSizeCalculationCause,
5368 : cancel: &CancellationToken,
5369 : ctx: &RequestContext,
5370 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5371 : let logical_sizes_at_once = self
5372 : .conf
5373 : .concurrent_tenant_size_logical_size_queries
5374 : .inner();
5375 :
5376 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5377 : //
5378 : // But the only case where we need to run multiple of these at once is when we
5379 : // request a size for a tenant manually via API, while another background calculation
5380 : // is in progress (which is not a common case).
5381 : //
5382 : // See more for on the issue #2748 condenced out of the initial PR review.
5383 : let mut shared_cache = tokio::select! {
5384 : locked = self.cached_logical_sizes.lock() => locked,
5385 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5386 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5387 : };
5388 :
5389 : size::gather_inputs(
5390 : self,
5391 : logical_sizes_at_once,
5392 : max_retention_period,
5393 : &mut shared_cache,
5394 : cause,
5395 : cancel,
5396 : ctx,
5397 : )
5398 : .await
5399 : }
5400 :
5401 : /// Calculate synthetic tenant size and cache the result.
5402 : /// This is periodically called by background worker.
5403 : /// result is cached in tenant struct
5404 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5405 : pub async fn calculate_synthetic_size(
5406 : &self,
5407 : cause: LogicalSizeCalculationCause,
5408 : cancel: &CancellationToken,
5409 : ctx: &RequestContext,
5410 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5411 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5412 :
5413 : let size = inputs.calculate();
5414 :
5415 : self.set_cached_synthetic_size(size);
5416 :
5417 : Ok(size)
5418 : }
5419 :
5420 : /// Cache given synthetic size and update the metric value
5421 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5422 0 : self.cached_synthetic_tenant_size
5423 0 : .store(size, Ordering::Relaxed);
5424 0 :
5425 0 : // Only shard zero should be calculating synthetic sizes
5426 0 : debug_assert!(self.shard_identity.is_shard_zero());
5427 :
5428 0 : TENANT_SYNTHETIC_SIZE_METRIC
5429 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5430 0 : .unwrap()
5431 0 : .set(size);
5432 0 : }
5433 :
5434 0 : pub fn cached_synthetic_size(&self) -> u64 {
5435 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5436 0 : }
5437 :
5438 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5439 : ///
5440 : /// This function can take a long time: callers should wrap it in a timeout if calling
5441 : /// from an external API handler.
5442 : ///
5443 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5444 : /// still bounded by tenant/timeline shutdown.
5445 : #[tracing::instrument(skip_all)]
5446 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5447 : let timelines = self.timelines.lock().unwrap().clone();
5448 :
5449 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5450 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5451 0 : timeline.freeze_and_flush().await?;
5452 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5453 0 : timeline.remote_client.wait_completion().await?;
5454 :
5455 0 : Ok(())
5456 0 : }
5457 :
5458 : // We do not use a JoinSet for these tasks, because we don't want them to be
5459 : // aborted when this function's future is cancelled: they should stay alive
5460 : // holding their GateGuard until they complete, to ensure their I/Os complete
5461 : // before Timeline shutdown completes.
5462 : let mut results = FuturesUnordered::new();
5463 :
5464 : for (_timeline_id, timeline) in timelines {
5465 : // Run each timeline's flush in a task holding the timeline's gate: this
5466 : // means that if this function's future is cancelled, the Timeline shutdown
5467 : // will still wait for any I/O in here to complete.
5468 : let Ok(gate) = timeline.gate.enter() else {
5469 : continue;
5470 : };
5471 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5472 : results.push(jh);
5473 : }
5474 :
5475 : while let Some(r) = results.next().await {
5476 : if let Err(e) = r {
5477 : if !e.is_cancelled() && !e.is_panic() {
5478 : tracing::error!("unexpected join error: {e:?}");
5479 : }
5480 : }
5481 : }
5482 :
5483 : // The flushes we did above were just writes, but the Tenant might have had
5484 : // pending deletions as well from recent compaction/gc: we want to flush those
5485 : // as well. This requires flushing the global delete queue. This is cheap
5486 : // because it's typically a no-op.
5487 : match self.deletion_queue_client.flush_execute().await {
5488 : Ok(_) => {}
5489 : Err(DeletionQueueError::ShuttingDown) => {}
5490 : }
5491 :
5492 : Ok(())
5493 : }
5494 :
5495 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
5496 0 : self.tenant_conf.load().tenant_conf.clone()
5497 0 : }
5498 :
5499 : /// How much local storage would this tenant like to have? It can cope with
5500 : /// less than this (via eviction and on-demand downloads), but this function enables
5501 : /// the Tenant to advertise how much storage it would prefer to have to provide fast I/O
5502 : /// by keeping important things on local disk.
5503 : ///
5504 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5505 : /// than they report here, due to layer eviction. Tenants with many active branches may
5506 : /// actually use more than they report here.
5507 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5508 0 : let timelines = self.timelines.lock().unwrap();
5509 0 :
5510 0 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5511 0 : // reflects the observation that on tenants with multiple large branches, typically only one
5512 0 : // of them is used actively enough to occupy space on disk.
5513 0 : timelines
5514 0 : .values()
5515 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5516 0 : .max()
5517 0 : .unwrap_or(0)
5518 0 : }
5519 :
5520 : /// Serialize and write the latest TenantManifest to remote storage.
5521 4 : pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5522 : // Only one manifest write may be done at at time, and the contents of the manifest
5523 : // must be loaded while holding this lock. This makes it safe to call this function
5524 : // from anywhere without worrying about colliding updates.
5525 4 : let mut guard = tokio::select! {
5526 4 : g = self.tenant_manifest_upload.lock() => {
5527 4 : g
5528 : },
5529 4 : _ = self.cancel.cancelled() => {
5530 0 : return Err(TenantManifestError::Cancelled);
5531 : }
5532 : };
5533 :
5534 4 : let manifest = self.build_tenant_manifest();
5535 4 : if Some(&manifest) == (*guard).as_ref() {
5536 : // Optimisation: skip uploads that don't change anything.
5537 0 : return Ok(());
5538 4 : }
5539 4 :
5540 4 : // Remote storage does no retries internally, so wrap it
5541 4 : match backoff::retry(
5542 4 : || async {
5543 4 : upload_tenant_manifest(
5544 4 : &self.remote_storage,
5545 4 : &self.tenant_shard_id,
5546 4 : self.generation,
5547 4 : &manifest,
5548 4 : &self.cancel,
5549 4 : )
5550 4 : .await
5551 8 : },
5552 4 : |_e| self.cancel.is_cancelled(),
5553 4 : FAILED_UPLOAD_WARN_THRESHOLD,
5554 4 : FAILED_REMOTE_OP_RETRIES,
5555 4 : "uploading tenant manifest",
5556 4 : &self.cancel,
5557 4 : )
5558 4 : .await
5559 : {
5560 0 : None => Err(TenantManifestError::Cancelled),
5561 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5562 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5563 : Some(Ok(_)) => {
5564 : // Store the successfully uploaded manifest, so that future callers can avoid
5565 : // re-uploading the same thing.
5566 4 : *guard = Some(manifest);
5567 4 :
5568 4 : Ok(())
5569 : }
5570 : }
5571 4 : }
5572 : }
5573 :
5574 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5575 : /// to get bootstrap data for timeline initialization.
5576 0 : async fn run_initdb(
5577 0 : conf: &'static PageServerConf,
5578 0 : initdb_target_dir: &Utf8Path,
5579 0 : pg_version: u32,
5580 0 : cancel: &CancellationToken,
5581 0 : ) -> Result<(), InitdbError> {
5582 0 : let initdb_bin_path = conf
5583 0 : .pg_bin_dir(pg_version)
5584 0 : .map_err(InitdbError::Other)?
5585 0 : .join("initdb");
5586 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5587 0 : info!(
5588 0 : "running {} in {}, libdir: {}",
5589 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5590 : );
5591 :
5592 0 : let _permit = {
5593 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5594 0 : INIT_DB_SEMAPHORE.acquire().await
5595 : };
5596 :
5597 0 : CONCURRENT_INITDBS.inc();
5598 0 : scopeguard::defer! {
5599 0 : CONCURRENT_INITDBS.dec();
5600 0 : }
5601 0 :
5602 0 : let _timer = INITDB_RUN_TIME.start_timer();
5603 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5604 0 : superuser: &conf.superuser,
5605 0 : locale: &conf.locale,
5606 0 : initdb_bin: &initdb_bin_path,
5607 0 : pg_version,
5608 0 : library_search_path: &initdb_lib_dir,
5609 0 : pgdata: initdb_target_dir,
5610 0 : })
5611 0 : .await
5612 0 : .map_err(InitdbError::Inner);
5613 0 :
5614 0 : // This isn't true cancellation support, see above. Still return an error to
5615 0 : // excercise the cancellation code path.
5616 0 : if cancel.is_cancelled() {
5617 0 : return Err(InitdbError::Cancelled);
5618 0 : }
5619 0 :
5620 0 : res
5621 0 : }
5622 :
5623 : /// Dump contents of a layer file to stdout.
5624 0 : pub async fn dump_layerfile_from_path(
5625 0 : path: &Utf8Path,
5626 0 : verbose: bool,
5627 0 : ctx: &RequestContext,
5628 0 : ) -> anyhow::Result<()> {
5629 : use std::os::unix::fs::FileExt;
5630 :
5631 : // All layer files start with a two-byte "magic" value, to identify the kind of
5632 : // file.
5633 0 : let file = File::open(path)?;
5634 0 : let mut header_buf = [0u8; 2];
5635 0 : file.read_exact_at(&mut header_buf, 0)?;
5636 :
5637 0 : match u16::from_be_bytes(header_buf) {
5638 : crate::IMAGE_FILE_MAGIC => {
5639 0 : ImageLayer::new_for_path(path, file)?
5640 0 : .dump(verbose, ctx)
5641 0 : .await?
5642 : }
5643 : crate::DELTA_FILE_MAGIC => {
5644 0 : DeltaLayer::new_for_path(path, file)?
5645 0 : .dump(verbose, ctx)
5646 0 : .await?
5647 : }
5648 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5649 : }
5650 :
5651 0 : Ok(())
5652 0 : }
5653 :
5654 : #[cfg(test)]
5655 : pub(crate) mod harness {
5656 : use bytes::{Bytes, BytesMut};
5657 : use hex_literal::hex;
5658 : use once_cell::sync::OnceCell;
5659 : use pageserver_api::key::Key;
5660 : use pageserver_api::models::ShardParameters;
5661 : use pageserver_api::record::NeonWalRecord;
5662 : use pageserver_api::shard::ShardIndex;
5663 : use utils::id::TenantId;
5664 : use utils::logging;
5665 :
5666 : use super::*;
5667 : use crate::deletion_queue::mock::MockDeletionQueue;
5668 : use crate::l0_flush::L0FlushConfig;
5669 : use crate::walredo::apply_neon;
5670 :
5671 : pub const TIMELINE_ID: TimelineId =
5672 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5673 : pub const NEW_TIMELINE_ID: TimelineId =
5674 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5675 :
5676 : /// Convenience function to create a page image with given string as the only content
5677 10057462 : pub fn test_img(s: &str) -> Bytes {
5678 10057462 : let mut buf = BytesMut::new();
5679 10057462 : buf.extend_from_slice(s.as_bytes());
5680 10057462 : buf.resize(64, 0);
5681 10057462 :
5682 10057462 : buf.freeze()
5683 10057462 : }
5684 :
5685 : impl From<TenantConf> for TenantConfOpt {
5686 452 : fn from(tenant_conf: TenantConf) -> Self {
5687 452 : Self {
5688 452 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
5689 452 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
5690 452 : compaction_target_size: Some(tenant_conf.compaction_target_size),
5691 452 : compaction_period: Some(tenant_conf.compaction_period),
5692 452 : compaction_threshold: Some(tenant_conf.compaction_threshold),
5693 452 : compaction_upper_limit: Some(tenant_conf.compaction_upper_limit),
5694 452 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
5695 452 : compaction_l0_first: Some(tenant_conf.compaction_l0_first),
5696 452 : compaction_l0_semaphore: Some(tenant_conf.compaction_l0_semaphore),
5697 452 : l0_flush_delay_threshold: tenant_conf.l0_flush_delay_threshold,
5698 452 : l0_flush_stall_threshold: tenant_conf.l0_flush_stall_threshold,
5699 452 : l0_flush_wait_upload: Some(tenant_conf.l0_flush_wait_upload),
5700 452 : gc_horizon: Some(tenant_conf.gc_horizon),
5701 452 : gc_period: Some(tenant_conf.gc_period),
5702 452 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
5703 452 : pitr_interval: Some(tenant_conf.pitr_interval),
5704 452 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
5705 452 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
5706 452 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
5707 452 : eviction_policy: Some(tenant_conf.eviction_policy),
5708 452 : min_resident_size_override: tenant_conf.min_resident_size_override,
5709 452 : evictions_low_residence_duration_metric_threshold: Some(
5710 452 : tenant_conf.evictions_low_residence_duration_metric_threshold,
5711 452 : ),
5712 452 : heatmap_period: Some(tenant_conf.heatmap_period),
5713 452 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
5714 452 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
5715 452 : image_layer_creation_check_threshold: Some(
5716 452 : tenant_conf.image_layer_creation_check_threshold,
5717 452 : ),
5718 452 : image_creation_preempt_threshold: Some(
5719 452 : tenant_conf.image_creation_preempt_threshold,
5720 452 : ),
5721 452 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
5722 452 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
5723 452 : timeline_offloading: Some(tenant_conf.timeline_offloading),
5724 452 : wal_receiver_protocol_override: tenant_conf.wal_receiver_protocol_override,
5725 452 : rel_size_v2_enabled: Some(tenant_conf.rel_size_v2_enabled),
5726 452 : gc_compaction_enabled: Some(tenant_conf.gc_compaction_enabled),
5727 452 : gc_compaction_initial_threshold_kb: Some(
5728 452 : tenant_conf.gc_compaction_initial_threshold_kb,
5729 452 : ),
5730 452 : gc_compaction_ratio_percent: Some(tenant_conf.gc_compaction_ratio_percent),
5731 452 : }
5732 452 : }
5733 : }
5734 :
5735 : pub struct TenantHarness {
5736 : pub conf: &'static PageServerConf,
5737 : pub tenant_conf: TenantConf,
5738 : pub tenant_shard_id: TenantShardId,
5739 : pub generation: Generation,
5740 : pub shard: ShardIndex,
5741 : pub remote_storage: GenericRemoteStorage,
5742 : pub remote_fs_dir: Utf8PathBuf,
5743 : pub deletion_queue: MockDeletionQueue,
5744 : }
5745 :
5746 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5747 :
5748 500 : pub(crate) fn setup_logging() {
5749 500 : LOG_HANDLE.get_or_init(|| {
5750 476 : logging::init(
5751 476 : logging::LogFormat::Test,
5752 476 : // enable it in case the tests exercise code paths that use
5753 476 : // debug_assert_current_span_has_tenant_and_timeline_id
5754 476 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5755 476 : logging::Output::Stdout,
5756 476 : )
5757 476 : .expect("Failed to init test logging")
5758 500 : });
5759 500 : }
5760 :
5761 : impl TenantHarness {
5762 452 : pub async fn create_custom(
5763 452 : test_name: &'static str,
5764 452 : tenant_conf: TenantConf,
5765 452 : tenant_id: TenantId,
5766 452 : shard_identity: ShardIdentity,
5767 452 : generation: Generation,
5768 452 : ) -> anyhow::Result<Self> {
5769 452 : setup_logging();
5770 452 :
5771 452 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5772 452 : let _ = fs::remove_dir_all(&repo_dir);
5773 452 : fs::create_dir_all(&repo_dir)?;
5774 :
5775 452 : let conf = PageServerConf::dummy_conf(repo_dir);
5776 452 : // Make a static copy of the config. This can never be free'd, but that's
5777 452 : // OK in a test.
5778 452 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5779 452 :
5780 452 : let shard = shard_identity.shard_index();
5781 452 : let tenant_shard_id = TenantShardId {
5782 452 : tenant_id,
5783 452 : shard_number: shard.shard_number,
5784 452 : shard_count: shard.shard_count,
5785 452 : };
5786 452 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5787 452 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5788 :
5789 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5790 452 : let remote_fs_dir = conf.workdir.join("localfs");
5791 452 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5792 452 : let config = RemoteStorageConfig {
5793 452 : storage: RemoteStorageKind::LocalFs {
5794 452 : local_path: remote_fs_dir.clone(),
5795 452 : },
5796 452 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5797 452 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5798 452 : };
5799 452 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5800 452 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5801 452 :
5802 452 : Ok(Self {
5803 452 : conf,
5804 452 : tenant_conf,
5805 452 : tenant_shard_id,
5806 452 : generation,
5807 452 : shard,
5808 452 : remote_storage,
5809 452 : remote_fs_dir,
5810 452 : deletion_queue,
5811 452 : })
5812 452 : }
5813 :
5814 428 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5815 428 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5816 428 : // The tests perform them manually if needed.
5817 428 : let tenant_conf = TenantConf {
5818 428 : gc_period: Duration::ZERO,
5819 428 : compaction_period: Duration::ZERO,
5820 428 : ..TenantConf::default()
5821 428 : };
5822 428 : let tenant_id = TenantId::generate();
5823 428 : let shard = ShardIdentity::unsharded();
5824 428 : Self::create_custom(
5825 428 : test_name,
5826 428 : tenant_conf,
5827 428 : tenant_id,
5828 428 : shard,
5829 428 : Generation::new(0xdeadbeef),
5830 428 : )
5831 428 : .await
5832 428 : }
5833 :
5834 40 : pub fn span(&self) -> tracing::Span {
5835 40 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
5836 40 : }
5837 :
5838 452 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
5839 452 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error)
5840 452 : .with_scope_unit_test();
5841 452 : (
5842 452 : self.do_try_load(&ctx)
5843 452 : .await
5844 452 : .expect("failed to load test tenant"),
5845 452 : ctx,
5846 452 : )
5847 452 : }
5848 :
5849 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5850 : pub(crate) async fn do_try_load(
5851 : &self,
5852 : ctx: &RequestContext,
5853 : ) -> anyhow::Result<Arc<Tenant>> {
5854 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
5855 :
5856 : let tenant = Arc::new(Tenant::new(
5857 : TenantState::Attaching,
5858 : self.conf,
5859 : AttachedTenantConf::try_from(LocationConf::attached_single(
5860 : TenantConfOpt::from(self.tenant_conf.clone()),
5861 : self.generation,
5862 : &ShardParameters::default(),
5863 : ))
5864 : .unwrap(),
5865 : // This is a legacy/test code path: sharding isn't supported here.
5866 : ShardIdentity::unsharded(),
5867 : Some(walredo_mgr),
5868 : self.tenant_shard_id,
5869 : self.remote_storage.clone(),
5870 : self.deletion_queue.new_client(),
5871 : // TODO: ideally we should run all unit tests with both configs
5872 : L0FlushGlobalState::new(L0FlushConfig::default()),
5873 : ));
5874 :
5875 : let preload = tenant
5876 : .preload(&self.remote_storage, CancellationToken::new())
5877 : .await?;
5878 : tenant.attach(Some(preload), ctx).await?;
5879 :
5880 : tenant.state.send_replace(TenantState::Active);
5881 : for timeline in tenant.timelines.lock().unwrap().values() {
5882 : timeline.set_state(TimelineState::Active);
5883 : }
5884 : Ok(tenant)
5885 : }
5886 :
5887 4 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
5888 4 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
5889 4 : }
5890 : }
5891 :
5892 : // Mock WAL redo manager that doesn't do much
5893 : pub(crate) struct TestRedoManager;
5894 :
5895 : impl TestRedoManager {
5896 : /// # Cancel-Safety
5897 : ///
5898 : /// This method is cancellation-safe.
5899 1676 : pub async fn request_redo(
5900 1676 : &self,
5901 1676 : key: Key,
5902 1676 : lsn: Lsn,
5903 1676 : base_img: Option<(Lsn, Bytes)>,
5904 1676 : records: Vec<(Lsn, NeonWalRecord)>,
5905 1676 : _pg_version: u32,
5906 1676 : ) -> Result<Bytes, walredo::Error> {
5907 2472 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
5908 1676 : if records_neon {
5909 : // For Neon wal records, we can decode without spawning postgres, so do so.
5910 1676 : let mut page = match (base_img, records.first()) {
5911 1536 : (Some((_lsn, img)), _) => {
5912 1536 : let mut page = BytesMut::new();
5913 1536 : page.extend_from_slice(&img);
5914 1536 : page
5915 : }
5916 140 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
5917 : _ => {
5918 0 : panic!("Neon WAL redo requires base image or will init record");
5919 : }
5920 : };
5921 :
5922 4148 : for (record_lsn, record) in records {
5923 2472 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
5924 : }
5925 1676 : Ok(page.freeze())
5926 : } else {
5927 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
5928 0 : let s = format!(
5929 0 : "redo for {} to get to {}, with {} and {} records",
5930 0 : key,
5931 0 : lsn,
5932 0 : if base_img.is_some() {
5933 0 : "base image"
5934 : } else {
5935 0 : "no base image"
5936 : },
5937 0 : records.len()
5938 0 : );
5939 0 : println!("{s}");
5940 0 :
5941 0 : Ok(test_img(&s))
5942 : }
5943 1676 : }
5944 : }
5945 : }
5946 :
5947 : #[cfg(test)]
5948 : mod tests {
5949 : use std::collections::{BTreeMap, BTreeSet};
5950 :
5951 : use bytes::{Bytes, BytesMut};
5952 : use hex_literal::hex;
5953 : use itertools::Itertools;
5954 : #[cfg(feature = "testing")]
5955 : use models::CompactLsnRange;
5956 : use pageserver_api::key::{AUX_KEY_PREFIX, Key, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX};
5957 : use pageserver_api::keyspace::KeySpace;
5958 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
5959 : #[cfg(feature = "testing")]
5960 : use pageserver_api::record::NeonWalRecord;
5961 : use pageserver_api::value::Value;
5962 : use pageserver_compaction::helpers::overlaps_with;
5963 : use rand::{Rng, thread_rng};
5964 : use storage_layer::{IoConcurrency, PersistentLayerKey};
5965 : use tests::storage_layer::ValuesReconstructState;
5966 : use tests::timeline::{GetVectoredError, ShutdownMode};
5967 : #[cfg(feature = "testing")]
5968 : use timeline::GcInfo;
5969 : #[cfg(feature = "testing")]
5970 : use timeline::InMemoryLayerTestDesc;
5971 : #[cfg(feature = "testing")]
5972 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
5973 : use timeline::{CompactOptions, DeltaLayerTestDesc};
5974 : use utils::id::TenantId;
5975 :
5976 : use super::*;
5977 : use crate::DEFAULT_PG_VERSION;
5978 : use crate::keyspace::KeySpaceAccum;
5979 : use crate::tenant::harness::*;
5980 : use crate::tenant::timeline::CompactFlags;
5981 :
5982 : static TEST_KEY: Lazy<Key> =
5983 36 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
5984 :
5985 : #[tokio::test]
5986 4 : async fn test_basic() -> anyhow::Result<()> {
5987 4 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
5988 4 : let tline = tenant
5989 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5990 4 : .await?;
5991 4 :
5992 4 : let mut writer = tline.writer().await;
5993 4 : writer
5994 4 : .put(
5995 4 : *TEST_KEY,
5996 4 : Lsn(0x10),
5997 4 : &Value::Image(test_img("foo at 0x10")),
5998 4 : &ctx,
5999 4 : )
6000 4 : .await?;
6001 4 : writer.finish_write(Lsn(0x10));
6002 4 : drop(writer);
6003 4 :
6004 4 : let mut writer = tline.writer().await;
6005 4 : writer
6006 4 : .put(
6007 4 : *TEST_KEY,
6008 4 : Lsn(0x20),
6009 4 : &Value::Image(test_img("foo at 0x20")),
6010 4 : &ctx,
6011 4 : )
6012 4 : .await?;
6013 4 : writer.finish_write(Lsn(0x20));
6014 4 : drop(writer);
6015 4 :
6016 4 : assert_eq!(
6017 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6018 4 : test_img("foo at 0x10")
6019 4 : );
6020 4 : assert_eq!(
6021 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6022 4 : test_img("foo at 0x10")
6023 4 : );
6024 4 : assert_eq!(
6025 4 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6026 4 : test_img("foo at 0x20")
6027 4 : );
6028 4 :
6029 4 : Ok(())
6030 4 : }
6031 :
6032 : #[tokio::test]
6033 4 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
6034 4 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
6035 4 : .await?
6036 4 : .load()
6037 4 : .await;
6038 4 : let _ = tenant
6039 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6040 4 : .await?;
6041 4 :
6042 4 : match tenant
6043 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6044 4 : .await
6045 4 : {
6046 4 : Ok(_) => panic!("duplicate timeline creation should fail"),
6047 4 : Err(e) => assert_eq!(
6048 4 : e.to_string(),
6049 4 : "timeline already exists with different parameters".to_string()
6050 4 : ),
6051 4 : }
6052 4 :
6053 4 : Ok(())
6054 4 : }
6055 :
6056 : /// Convenience function to create a page image with given string as the only content
6057 20 : pub fn test_value(s: &str) -> Value {
6058 20 : let mut buf = BytesMut::new();
6059 20 : buf.extend_from_slice(s.as_bytes());
6060 20 : Value::Image(buf.freeze())
6061 20 : }
6062 :
6063 : ///
6064 : /// Test branch creation
6065 : ///
6066 : #[tokio::test]
6067 4 : async fn test_branch() -> anyhow::Result<()> {
6068 4 : use std::str::from_utf8;
6069 4 :
6070 4 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
6071 4 : let tline = tenant
6072 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6073 4 : .await?;
6074 4 : let mut writer = tline.writer().await;
6075 4 :
6076 4 : #[allow(non_snake_case)]
6077 4 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
6078 4 : #[allow(non_snake_case)]
6079 4 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
6080 4 :
6081 4 : // Insert a value on the timeline
6082 4 : writer
6083 4 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
6084 4 : .await?;
6085 4 : writer
6086 4 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
6087 4 : .await?;
6088 4 : writer.finish_write(Lsn(0x20));
6089 4 :
6090 4 : writer
6091 4 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
6092 4 : .await?;
6093 4 : writer.finish_write(Lsn(0x30));
6094 4 : writer
6095 4 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
6096 4 : .await?;
6097 4 : writer.finish_write(Lsn(0x40));
6098 4 :
6099 4 : //assert_current_logical_size(&tline, Lsn(0x40));
6100 4 :
6101 4 : // Branch the history, modify relation differently on the new timeline
6102 4 : tenant
6103 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
6104 4 : .await?;
6105 4 : let newtline = tenant
6106 4 : .get_timeline(NEW_TIMELINE_ID, true)
6107 4 : .expect("Should have a local timeline");
6108 4 : let mut new_writer = newtline.writer().await;
6109 4 : new_writer
6110 4 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
6111 4 : .await?;
6112 4 : new_writer.finish_write(Lsn(0x40));
6113 4 :
6114 4 : // Check page contents on both branches
6115 4 : assert_eq!(
6116 4 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6117 4 : "foo at 0x40"
6118 4 : );
6119 4 : assert_eq!(
6120 4 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6121 4 : "bar at 0x40"
6122 4 : );
6123 4 : assert_eq!(
6124 4 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
6125 4 : "foobar at 0x20"
6126 4 : );
6127 4 :
6128 4 : //assert_current_logical_size(&tline, Lsn(0x40));
6129 4 :
6130 4 : Ok(())
6131 4 : }
6132 :
6133 40 : async fn make_some_layers(
6134 40 : tline: &Timeline,
6135 40 : start_lsn: Lsn,
6136 40 : ctx: &RequestContext,
6137 40 : ) -> anyhow::Result<()> {
6138 40 : let mut lsn = start_lsn;
6139 : {
6140 40 : let mut writer = tline.writer().await;
6141 : // Create a relation on the timeline
6142 40 : writer
6143 40 : .put(
6144 40 : *TEST_KEY,
6145 40 : lsn,
6146 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6147 40 : ctx,
6148 40 : )
6149 40 : .await?;
6150 40 : writer.finish_write(lsn);
6151 40 : lsn += 0x10;
6152 40 : writer
6153 40 : .put(
6154 40 : *TEST_KEY,
6155 40 : lsn,
6156 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6157 40 : ctx,
6158 40 : )
6159 40 : .await?;
6160 40 : writer.finish_write(lsn);
6161 40 : lsn += 0x10;
6162 40 : }
6163 40 : tline.freeze_and_flush().await?;
6164 : {
6165 40 : let mut writer = tline.writer().await;
6166 40 : writer
6167 40 : .put(
6168 40 : *TEST_KEY,
6169 40 : lsn,
6170 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6171 40 : ctx,
6172 40 : )
6173 40 : .await?;
6174 40 : writer.finish_write(lsn);
6175 40 : lsn += 0x10;
6176 40 : writer
6177 40 : .put(
6178 40 : *TEST_KEY,
6179 40 : lsn,
6180 40 : &Value::Image(test_img(&format!("foo at {}", lsn))),
6181 40 : ctx,
6182 40 : )
6183 40 : .await?;
6184 40 : writer.finish_write(lsn);
6185 40 : }
6186 40 : tline.freeze_and_flush().await.map_err(|e| e.into())
6187 40 : }
6188 :
6189 : #[tokio::test(start_paused = true)]
6190 4 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6191 4 : let (tenant, ctx) =
6192 4 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6193 4 : .await?
6194 4 : .load()
6195 4 : .await;
6196 4 : // Advance to the lsn lease deadline so that GC is not blocked by
6197 4 : // initial transition into AttachedSingle.
6198 4 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
6199 4 : tokio::time::resume();
6200 4 : let tline = tenant
6201 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6202 4 : .await?;
6203 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6204 4 :
6205 4 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6206 4 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6207 4 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6208 4 : // below should fail.
6209 4 : tenant
6210 4 : .gc_iteration(
6211 4 : Some(TIMELINE_ID),
6212 4 : 0x10,
6213 4 : Duration::ZERO,
6214 4 : &CancellationToken::new(),
6215 4 : &ctx,
6216 4 : )
6217 4 : .await?;
6218 4 :
6219 4 : // try to branch at lsn 25, should fail because we already garbage collected the data
6220 4 : match tenant
6221 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6222 4 : .await
6223 4 : {
6224 4 : Ok(_) => panic!("branching should have failed"),
6225 4 : Err(err) => {
6226 4 : let CreateTimelineError::AncestorLsn(err) = err else {
6227 4 : panic!("wrong error type")
6228 4 : };
6229 4 : assert!(err.to_string().contains("invalid branch start lsn"));
6230 4 : assert!(
6231 4 : err.source()
6232 4 : .unwrap()
6233 4 : .to_string()
6234 4 : .contains("we might've already garbage collected needed data")
6235 4 : )
6236 4 : }
6237 4 : }
6238 4 :
6239 4 : Ok(())
6240 4 : }
6241 :
6242 : #[tokio::test]
6243 4 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6244 4 : let (tenant, ctx) =
6245 4 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6246 4 : .await?
6247 4 : .load()
6248 4 : .await;
6249 4 :
6250 4 : let tline = tenant
6251 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6252 4 : .await?;
6253 4 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6254 4 : match tenant
6255 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6256 4 : .await
6257 4 : {
6258 4 : Ok(_) => panic!("branching should have failed"),
6259 4 : Err(err) => {
6260 4 : let CreateTimelineError::AncestorLsn(err) = err else {
6261 4 : panic!("wrong error type");
6262 4 : };
6263 4 : assert!(&err.to_string().contains("invalid branch start lsn"));
6264 4 : assert!(
6265 4 : &err.source()
6266 4 : .unwrap()
6267 4 : .to_string()
6268 4 : .contains("is earlier than latest GC cutoff")
6269 4 : );
6270 4 : }
6271 4 : }
6272 4 :
6273 4 : Ok(())
6274 4 : }
6275 :
6276 : /*
6277 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6278 : // remove the old value, we'd need to work a little harder
6279 : #[tokio::test]
6280 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6281 : let repo =
6282 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6283 : .load();
6284 :
6285 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6286 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6287 :
6288 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6289 : let applied_gc_cutoff_lsn = tline.get_applied_gc_cutoff_lsn();
6290 : assert!(*applied_gc_cutoff_lsn > Lsn(0x25));
6291 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6292 : Ok(_) => panic!("request for page should have failed"),
6293 : Err(err) => assert!(err.to_string().contains("not found at")),
6294 : }
6295 : Ok(())
6296 : }
6297 : */
6298 :
6299 : #[tokio::test]
6300 4 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6301 4 : let (tenant, ctx) =
6302 4 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6303 4 : .await?
6304 4 : .load()
6305 4 : .await;
6306 4 : let tline = tenant
6307 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6308 4 : .await?;
6309 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6310 4 :
6311 4 : tenant
6312 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6313 4 : .await?;
6314 4 : let newtline = tenant
6315 4 : .get_timeline(NEW_TIMELINE_ID, true)
6316 4 : .expect("Should have a local timeline");
6317 4 :
6318 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6319 4 :
6320 4 : tline.set_broken("test".to_owned());
6321 4 :
6322 4 : tenant
6323 4 : .gc_iteration(
6324 4 : Some(TIMELINE_ID),
6325 4 : 0x10,
6326 4 : Duration::ZERO,
6327 4 : &CancellationToken::new(),
6328 4 : &ctx,
6329 4 : )
6330 4 : .await?;
6331 4 :
6332 4 : // The branchpoints should contain all timelines, even ones marked
6333 4 : // as Broken.
6334 4 : {
6335 4 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6336 4 : assert_eq!(branchpoints.len(), 1);
6337 4 : assert_eq!(
6338 4 : branchpoints[0],
6339 4 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6340 4 : );
6341 4 : }
6342 4 :
6343 4 : // You can read the key from the child branch even though the parent is
6344 4 : // Broken, as long as you don't need to access data from the parent.
6345 4 : assert_eq!(
6346 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6347 4 : test_img(&format!("foo at {}", Lsn(0x70)))
6348 4 : );
6349 4 :
6350 4 : // This needs to traverse to the parent, and fails.
6351 4 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6352 4 : assert!(
6353 4 : err.to_string().starts_with(&format!(
6354 4 : "bad state on timeline {}: Broken",
6355 4 : tline.timeline_id
6356 4 : )),
6357 4 : "{err}"
6358 4 : );
6359 4 :
6360 4 : Ok(())
6361 4 : }
6362 :
6363 : #[tokio::test]
6364 4 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6365 4 : let (tenant, ctx) =
6366 4 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6367 4 : .await?
6368 4 : .load()
6369 4 : .await;
6370 4 : let tline = tenant
6371 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6372 4 : .await?;
6373 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6374 4 :
6375 4 : tenant
6376 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6377 4 : .await?;
6378 4 : let newtline = tenant
6379 4 : .get_timeline(NEW_TIMELINE_ID, true)
6380 4 : .expect("Should have a local timeline");
6381 4 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6382 4 : tenant
6383 4 : .gc_iteration(
6384 4 : Some(TIMELINE_ID),
6385 4 : 0x10,
6386 4 : Duration::ZERO,
6387 4 : &CancellationToken::new(),
6388 4 : &ctx,
6389 4 : )
6390 4 : .await?;
6391 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6392 4 :
6393 4 : Ok(())
6394 4 : }
6395 : #[tokio::test]
6396 4 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6397 4 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6398 4 : .await?
6399 4 : .load()
6400 4 : .await;
6401 4 : let tline = tenant
6402 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6403 4 : .await?;
6404 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6405 4 :
6406 4 : tenant
6407 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6408 4 : .await?;
6409 4 : let newtline = tenant
6410 4 : .get_timeline(NEW_TIMELINE_ID, true)
6411 4 : .expect("Should have a local timeline");
6412 4 :
6413 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6414 4 :
6415 4 : // run gc on parent
6416 4 : tenant
6417 4 : .gc_iteration(
6418 4 : Some(TIMELINE_ID),
6419 4 : 0x10,
6420 4 : Duration::ZERO,
6421 4 : &CancellationToken::new(),
6422 4 : &ctx,
6423 4 : )
6424 4 : .await?;
6425 4 :
6426 4 : // Check that the data is still accessible on the branch.
6427 4 : assert_eq!(
6428 4 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6429 4 : test_img(&format!("foo at {}", Lsn(0x40)))
6430 4 : );
6431 4 :
6432 4 : Ok(())
6433 4 : }
6434 :
6435 : #[tokio::test]
6436 4 : async fn timeline_load() -> anyhow::Result<()> {
6437 4 : const TEST_NAME: &str = "timeline_load";
6438 4 : let harness = TenantHarness::create(TEST_NAME).await?;
6439 4 : {
6440 4 : let (tenant, ctx) = harness.load().await;
6441 4 : let tline = tenant
6442 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6443 4 : .await?;
6444 4 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6445 4 : // so that all uploads finish & we can call harness.load() below again
6446 4 : tenant
6447 4 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6448 4 : .instrument(harness.span())
6449 4 : .await
6450 4 : .ok()
6451 4 : .unwrap();
6452 4 : }
6453 4 :
6454 4 : let (tenant, _ctx) = harness.load().await;
6455 4 : tenant
6456 4 : .get_timeline(TIMELINE_ID, true)
6457 4 : .expect("cannot load timeline");
6458 4 :
6459 4 : Ok(())
6460 4 : }
6461 :
6462 : #[tokio::test]
6463 4 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6464 4 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6465 4 : let harness = TenantHarness::create(TEST_NAME).await?;
6466 4 : // create two timelines
6467 4 : {
6468 4 : let (tenant, ctx) = harness.load().await;
6469 4 : let tline = tenant
6470 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6471 4 : .await?;
6472 4 :
6473 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6474 4 :
6475 4 : let child_tline = tenant
6476 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6477 4 : .await?;
6478 4 : child_tline.set_state(TimelineState::Active);
6479 4 :
6480 4 : let newtline = tenant
6481 4 : .get_timeline(NEW_TIMELINE_ID, true)
6482 4 : .expect("Should have a local timeline");
6483 4 :
6484 4 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6485 4 :
6486 4 : // so that all uploads finish & we can call harness.load() below again
6487 4 : tenant
6488 4 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6489 4 : .instrument(harness.span())
6490 4 : .await
6491 4 : .ok()
6492 4 : .unwrap();
6493 4 : }
6494 4 :
6495 4 : // check that both of them are initially unloaded
6496 4 : let (tenant, _ctx) = harness.load().await;
6497 4 :
6498 4 : // check that both, child and ancestor are loaded
6499 4 : let _child_tline = tenant
6500 4 : .get_timeline(NEW_TIMELINE_ID, true)
6501 4 : .expect("cannot get child timeline loaded");
6502 4 :
6503 4 : let _ancestor_tline = tenant
6504 4 : .get_timeline(TIMELINE_ID, true)
6505 4 : .expect("cannot get ancestor timeline loaded");
6506 4 :
6507 4 : Ok(())
6508 4 : }
6509 :
6510 : #[tokio::test]
6511 4 : async fn delta_layer_dumping() -> anyhow::Result<()> {
6512 4 : use storage_layer::AsLayerDesc;
6513 4 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
6514 4 : .await?
6515 4 : .load()
6516 4 : .await;
6517 4 : let tline = tenant
6518 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6519 4 : .await?;
6520 4 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6521 4 :
6522 4 : let layer_map = tline.layers.read().await;
6523 4 : let level0_deltas = layer_map
6524 4 : .layer_map()?
6525 4 : .level0_deltas()
6526 4 : .iter()
6527 8 : .map(|desc| layer_map.get_from_desc(desc))
6528 4 : .collect::<Vec<_>>();
6529 4 :
6530 4 : assert!(!level0_deltas.is_empty());
6531 4 :
6532 12 : for delta in level0_deltas {
6533 4 : // Ensure we are dumping a delta layer here
6534 8 : assert!(delta.layer_desc().is_delta);
6535 8 : delta.dump(true, &ctx).await.unwrap();
6536 4 : }
6537 4 :
6538 4 : Ok(())
6539 4 : }
6540 :
6541 : #[tokio::test]
6542 4 : async fn test_images() -> anyhow::Result<()> {
6543 4 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
6544 4 : let tline = tenant
6545 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6546 4 : .await?;
6547 4 :
6548 4 : let mut writer = tline.writer().await;
6549 4 : writer
6550 4 : .put(
6551 4 : *TEST_KEY,
6552 4 : Lsn(0x10),
6553 4 : &Value::Image(test_img("foo at 0x10")),
6554 4 : &ctx,
6555 4 : )
6556 4 : .await?;
6557 4 : writer.finish_write(Lsn(0x10));
6558 4 : drop(writer);
6559 4 :
6560 4 : tline.freeze_and_flush().await?;
6561 4 : tline
6562 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6563 4 : .await?;
6564 4 :
6565 4 : let mut writer = tline.writer().await;
6566 4 : writer
6567 4 : .put(
6568 4 : *TEST_KEY,
6569 4 : Lsn(0x20),
6570 4 : &Value::Image(test_img("foo at 0x20")),
6571 4 : &ctx,
6572 4 : )
6573 4 : .await?;
6574 4 : writer.finish_write(Lsn(0x20));
6575 4 : drop(writer);
6576 4 :
6577 4 : tline.freeze_and_flush().await?;
6578 4 : tline
6579 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6580 4 : .await?;
6581 4 :
6582 4 : let mut writer = tline.writer().await;
6583 4 : writer
6584 4 : .put(
6585 4 : *TEST_KEY,
6586 4 : Lsn(0x30),
6587 4 : &Value::Image(test_img("foo at 0x30")),
6588 4 : &ctx,
6589 4 : )
6590 4 : .await?;
6591 4 : writer.finish_write(Lsn(0x30));
6592 4 : drop(writer);
6593 4 :
6594 4 : tline.freeze_and_flush().await?;
6595 4 : tline
6596 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6597 4 : .await?;
6598 4 :
6599 4 : let mut writer = tline.writer().await;
6600 4 : writer
6601 4 : .put(
6602 4 : *TEST_KEY,
6603 4 : Lsn(0x40),
6604 4 : &Value::Image(test_img("foo at 0x40")),
6605 4 : &ctx,
6606 4 : )
6607 4 : .await?;
6608 4 : writer.finish_write(Lsn(0x40));
6609 4 : drop(writer);
6610 4 :
6611 4 : tline.freeze_and_flush().await?;
6612 4 : tline
6613 4 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
6614 4 : .await?;
6615 4 :
6616 4 : assert_eq!(
6617 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6618 4 : test_img("foo at 0x10")
6619 4 : );
6620 4 : assert_eq!(
6621 4 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6622 4 : test_img("foo at 0x10")
6623 4 : );
6624 4 : assert_eq!(
6625 4 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6626 4 : test_img("foo at 0x20")
6627 4 : );
6628 4 : assert_eq!(
6629 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
6630 4 : test_img("foo at 0x30")
6631 4 : );
6632 4 : assert_eq!(
6633 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
6634 4 : test_img("foo at 0x40")
6635 4 : );
6636 4 :
6637 4 : Ok(())
6638 4 : }
6639 :
6640 8 : async fn bulk_insert_compact_gc(
6641 8 : tenant: &Tenant,
6642 8 : timeline: &Arc<Timeline>,
6643 8 : ctx: &RequestContext,
6644 8 : lsn: Lsn,
6645 8 : repeat: usize,
6646 8 : key_count: usize,
6647 8 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6648 8 : let compact = true;
6649 8 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
6650 8 : }
6651 :
6652 16 : async fn bulk_insert_maybe_compact_gc(
6653 16 : tenant: &Tenant,
6654 16 : timeline: &Arc<Timeline>,
6655 16 : ctx: &RequestContext,
6656 16 : mut lsn: Lsn,
6657 16 : repeat: usize,
6658 16 : key_count: usize,
6659 16 : compact: bool,
6660 16 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
6661 16 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
6662 16 :
6663 16 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6664 16 : let mut blknum = 0;
6665 16 :
6666 16 : // Enforce that key range is monotonously increasing
6667 16 : let mut keyspace = KeySpaceAccum::new();
6668 16 :
6669 16 : let cancel = CancellationToken::new();
6670 16 :
6671 16 : for _ in 0..repeat {
6672 800 : for _ in 0..key_count {
6673 8000000 : test_key.field6 = blknum;
6674 8000000 : let mut writer = timeline.writer().await;
6675 8000000 : writer
6676 8000000 : .put(
6677 8000000 : test_key,
6678 8000000 : lsn,
6679 8000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6680 8000000 : ctx,
6681 8000000 : )
6682 8000000 : .await?;
6683 8000000 : inserted.entry(test_key).or_default().insert(lsn);
6684 8000000 : writer.finish_write(lsn);
6685 8000000 : drop(writer);
6686 8000000 :
6687 8000000 : keyspace.add_key(test_key);
6688 8000000 :
6689 8000000 : lsn = Lsn(lsn.0 + 0x10);
6690 8000000 : blknum += 1;
6691 : }
6692 :
6693 800 : timeline.freeze_and_flush().await?;
6694 800 : if compact {
6695 : // this requires timeline to be &Arc<Timeline>
6696 400 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
6697 400 : }
6698 :
6699 : // this doesn't really need to use the timeline_id target, but it is closer to what it
6700 : // originally was.
6701 800 : let res = tenant
6702 800 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
6703 800 : .await?;
6704 :
6705 800 : assert_eq!(res.layers_removed, 0, "this never removes anything");
6706 : }
6707 :
6708 16 : Ok(inserted)
6709 16 : }
6710 :
6711 : //
6712 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
6713 : // Repeat 50 times.
6714 : //
6715 : #[tokio::test]
6716 4 : async fn test_bulk_insert() -> anyhow::Result<()> {
6717 4 : let harness = TenantHarness::create("test_bulk_insert").await?;
6718 4 : let (tenant, ctx) = harness.load().await;
6719 4 : let tline = tenant
6720 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6721 4 : .await?;
6722 4 :
6723 4 : let lsn = Lsn(0x10);
6724 4 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6725 4 :
6726 4 : Ok(())
6727 4 : }
6728 :
6729 : // Test the vectored get real implementation against a simple sequential implementation.
6730 : //
6731 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
6732 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
6733 : // grow to the right on the X axis.
6734 : // [Delta]
6735 : // [Delta]
6736 : // [Delta]
6737 : // [Delta]
6738 : // ------------ Image ---------------
6739 : //
6740 : // After layer generation we pick the ranges to query as follows:
6741 : // 1. The beginning of each delta layer
6742 : // 2. At the seam between two adjacent delta layers
6743 : //
6744 : // There's one major downside to this test: delta layers only contains images,
6745 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
6746 : #[tokio::test]
6747 4 : async fn test_get_vectored() -> anyhow::Result<()> {
6748 4 : let harness = TenantHarness::create("test_get_vectored").await?;
6749 4 : let (tenant, ctx) = harness.load().await;
6750 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6751 4 : let tline = tenant
6752 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6753 4 : .await?;
6754 4 :
6755 4 : let lsn = Lsn(0x10);
6756 4 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
6757 4 :
6758 4 : let guard = tline.layers.read().await;
6759 4 : let lm = guard.layer_map()?;
6760 4 :
6761 4 : lm.dump(true, &ctx).await?;
6762 4 :
6763 4 : let mut reads = Vec::new();
6764 4 : let mut prev = None;
6765 24 : lm.iter_historic_layers().for_each(|desc| {
6766 24 : if !desc.is_delta() {
6767 4 : prev = Some(desc.clone());
6768 4 : return;
6769 20 : }
6770 20 :
6771 20 : let start = desc.key_range.start;
6772 20 : let end = desc
6773 20 : .key_range
6774 20 : .start
6775 20 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
6776 20 : reads.push(KeySpace {
6777 20 : ranges: vec![start..end],
6778 20 : });
6779 4 :
6780 20 : if let Some(prev) = &prev {
6781 20 : if !prev.is_delta() {
6782 20 : return;
6783 4 : }
6784 0 :
6785 0 : let first_range = Key {
6786 0 : field6: prev.key_range.end.field6 - 4,
6787 0 : ..prev.key_range.end
6788 0 : }..prev.key_range.end;
6789 0 :
6790 0 : let second_range = desc.key_range.start..Key {
6791 0 : field6: desc.key_range.start.field6 + 4,
6792 0 : ..desc.key_range.start
6793 0 : };
6794 0 :
6795 0 : reads.push(KeySpace {
6796 0 : ranges: vec![first_range, second_range],
6797 0 : });
6798 4 : };
6799 4 :
6800 4 : prev = Some(desc.clone());
6801 24 : });
6802 4 :
6803 4 : drop(guard);
6804 4 :
6805 4 : // Pick a big LSN such that we query over all the changes.
6806 4 : let reads_lsn = Lsn(u64::MAX - 1);
6807 4 :
6808 24 : for read in reads {
6809 20 : info!("Doing vectored read on {:?}", read);
6810 4 :
6811 20 : let vectored_res = tline
6812 20 : .get_vectored_impl(
6813 20 : read.clone(),
6814 20 : reads_lsn,
6815 20 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6816 20 : &ctx,
6817 20 : )
6818 20 : .await;
6819 4 :
6820 20 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
6821 20 : let mut expect_missing = false;
6822 20 : let mut key = read.start().unwrap();
6823 660 : while key != read.end().unwrap() {
6824 640 : if let Some(lsns) = inserted.get(&key) {
6825 640 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
6826 640 : match expected_lsn {
6827 640 : Some(lsn) => {
6828 640 : expected_lsns.insert(key, *lsn);
6829 640 : }
6830 4 : None => {
6831 4 : expect_missing = true;
6832 0 : break;
6833 4 : }
6834 4 : }
6835 4 : } else {
6836 4 : expect_missing = true;
6837 0 : break;
6838 4 : }
6839 4 :
6840 640 : key = key.next();
6841 4 : }
6842 4 :
6843 20 : if expect_missing {
6844 4 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
6845 4 : } else {
6846 640 : for (key, image) in vectored_res? {
6847 640 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
6848 640 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
6849 640 : assert_eq!(image?, expected_image);
6850 4 : }
6851 4 : }
6852 4 : }
6853 4 :
6854 4 : Ok(())
6855 4 : }
6856 :
6857 : #[tokio::test]
6858 4 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
6859 4 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
6860 4 :
6861 4 : let (tenant, ctx) = harness.load().await;
6862 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6863 4 : let (tline, ctx) = tenant
6864 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
6865 4 : .await?;
6866 4 : let tline = tline.raw_timeline().unwrap();
6867 4 :
6868 4 : let mut modification = tline.begin_modification(Lsn(0x1000));
6869 4 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
6870 4 : modification.set_lsn(Lsn(0x1008))?;
6871 4 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
6872 4 : modification.commit(&ctx).await?;
6873 4 :
6874 4 : let child_timeline_id = TimelineId::generate();
6875 4 : tenant
6876 4 : .branch_timeline_test(
6877 4 : tline,
6878 4 : child_timeline_id,
6879 4 : Some(tline.get_last_record_lsn()),
6880 4 : &ctx,
6881 4 : )
6882 4 : .await?;
6883 4 :
6884 4 : let child_timeline = tenant
6885 4 : .get_timeline(child_timeline_id, true)
6886 4 : .expect("Should have the branched timeline");
6887 4 :
6888 4 : let aux_keyspace = KeySpace {
6889 4 : ranges: vec![NON_INHERITED_RANGE],
6890 4 : };
6891 4 : let read_lsn = child_timeline.get_last_record_lsn();
6892 4 :
6893 4 : let vectored_res = child_timeline
6894 4 : .get_vectored_impl(
6895 4 : aux_keyspace.clone(),
6896 4 : read_lsn,
6897 4 : &mut ValuesReconstructState::new(io_concurrency.clone()),
6898 4 : &ctx,
6899 4 : )
6900 4 : .await;
6901 4 :
6902 4 : let images = vectored_res?;
6903 4 : assert!(images.is_empty());
6904 4 : Ok(())
6905 4 : }
6906 :
6907 : // Test that vectored get handles layer gaps correctly
6908 : // by advancing into the next ancestor timeline if required.
6909 : //
6910 : // The test generates timelines that look like the diagram below.
6911 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
6912 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
6913 : //
6914 : // ```
6915 : //-------------------------------+
6916 : // ... |
6917 : // [ L1 ] |
6918 : // [ / L1 ] | Child Timeline
6919 : // ... |
6920 : // ------------------------------+
6921 : // [ X L1 ] | Parent Timeline
6922 : // ------------------------------+
6923 : // ```
6924 : #[tokio::test]
6925 4 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
6926 4 : let tenant_conf = TenantConf {
6927 4 : // Make compaction deterministic
6928 4 : gc_period: Duration::ZERO,
6929 4 : compaction_period: Duration::ZERO,
6930 4 : // Encourage creation of L1 layers
6931 4 : checkpoint_distance: 16 * 1024,
6932 4 : compaction_target_size: 8 * 1024,
6933 4 : ..TenantConf::default()
6934 4 : };
6935 4 :
6936 4 : let harness = TenantHarness::create_custom(
6937 4 : "test_get_vectored_key_gap",
6938 4 : tenant_conf,
6939 4 : TenantId::generate(),
6940 4 : ShardIdentity::unsharded(),
6941 4 : Generation::new(0xdeadbeef),
6942 4 : )
6943 4 : .await?;
6944 4 : let (tenant, ctx) = harness.load().await;
6945 4 : let io_concurrency = IoConcurrency::spawn_for_test();
6946 4 :
6947 4 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6948 4 : let gap_at_key = current_key.add(100);
6949 4 : let mut current_lsn = Lsn(0x10);
6950 4 :
6951 4 : const KEY_COUNT: usize = 10_000;
6952 4 :
6953 4 : let timeline_id = TimelineId::generate();
6954 4 : let current_timeline = tenant
6955 4 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
6956 4 : .await?;
6957 4 :
6958 4 : current_lsn += 0x100;
6959 4 :
6960 4 : let mut writer = current_timeline.writer().await;
6961 4 : writer
6962 4 : .put(
6963 4 : gap_at_key,
6964 4 : current_lsn,
6965 4 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
6966 4 : &ctx,
6967 4 : )
6968 4 : .await?;
6969 4 : writer.finish_write(current_lsn);
6970 4 : drop(writer);
6971 4 :
6972 4 : let mut latest_lsns = HashMap::new();
6973 4 : latest_lsns.insert(gap_at_key, current_lsn);
6974 4 :
6975 4 : current_timeline.freeze_and_flush().await?;
6976 4 :
6977 4 : let child_timeline_id = TimelineId::generate();
6978 4 :
6979 4 : tenant
6980 4 : .branch_timeline_test(
6981 4 : ¤t_timeline,
6982 4 : child_timeline_id,
6983 4 : Some(current_lsn),
6984 4 : &ctx,
6985 4 : )
6986 4 : .await?;
6987 4 : let child_timeline = tenant
6988 4 : .get_timeline(child_timeline_id, true)
6989 4 : .expect("Should have the branched timeline");
6990 4 :
6991 40004 : for i in 0..KEY_COUNT {
6992 40000 : if current_key == gap_at_key {
6993 4 : current_key = current_key.next();
6994 4 : continue;
6995 39996 : }
6996 39996 :
6997 39996 : current_lsn += 0x10;
6998 4 :
6999 39996 : let mut writer = child_timeline.writer().await;
7000 39996 : writer
7001 39996 : .put(
7002 39996 : current_key,
7003 39996 : current_lsn,
7004 39996 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
7005 39996 : &ctx,
7006 39996 : )
7007 39996 : .await?;
7008 39996 : writer.finish_write(current_lsn);
7009 39996 : drop(writer);
7010 39996 :
7011 39996 : latest_lsns.insert(current_key, current_lsn);
7012 39996 : current_key = current_key.next();
7013 39996 :
7014 39996 : // Flush every now and then to encourage layer file creation.
7015 39996 : if i % 500 == 0 {
7016 80 : child_timeline.freeze_and_flush().await?;
7017 39916 : }
7018 4 : }
7019 4 :
7020 4 : child_timeline.freeze_and_flush().await?;
7021 4 : let mut flags = EnumSet::new();
7022 4 : flags.insert(CompactFlags::ForceRepartition);
7023 4 : child_timeline
7024 4 : .compact(&CancellationToken::new(), flags, &ctx)
7025 4 : .await?;
7026 4 :
7027 4 : let key_near_end = {
7028 4 : let mut tmp = current_key;
7029 4 : tmp.field6 -= 10;
7030 4 : tmp
7031 4 : };
7032 4 :
7033 4 : let key_near_gap = {
7034 4 : let mut tmp = gap_at_key;
7035 4 : tmp.field6 -= 10;
7036 4 : tmp
7037 4 : };
7038 4 :
7039 4 : let read = KeySpace {
7040 4 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
7041 4 : };
7042 4 : let results = child_timeline
7043 4 : .get_vectored_impl(
7044 4 : read.clone(),
7045 4 : current_lsn,
7046 4 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7047 4 : &ctx,
7048 4 : )
7049 4 : .await?;
7050 4 :
7051 88 : for (key, img_res) in results {
7052 84 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
7053 84 : assert_eq!(img_res?, expected);
7054 4 : }
7055 4 :
7056 4 : Ok(())
7057 4 : }
7058 :
7059 : // Test that vectored get descends into ancestor timelines correctly and
7060 : // does not return an image that's newer than requested.
7061 : //
7062 : // The diagram below ilustrates an interesting case. We have a parent timeline
7063 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
7064 : // from the child timeline, so the parent timeline must be visited. When advacing into
7065 : // the child timeline, the read path needs to remember what the requested Lsn was in
7066 : // order to avoid returning an image that's too new. The test below constructs such
7067 : // a timeline setup and does a few queries around the Lsn of each page image.
7068 : // ```
7069 : // LSN
7070 : // ^
7071 : // |
7072 : // |
7073 : // 500 | --------------------------------------> branch point
7074 : // 400 | X
7075 : // 300 | X
7076 : // 200 | --------------------------------------> requested lsn
7077 : // 100 | X
7078 : // |---------------------------------------> Key
7079 : // |
7080 : // ------> requested key
7081 : //
7082 : // Legend:
7083 : // * X - page images
7084 : // ```
7085 : #[tokio::test]
7086 4 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
7087 4 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
7088 4 : let (tenant, ctx) = harness.load().await;
7089 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7090 4 :
7091 4 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7092 4 : let end_key = start_key.add(1000);
7093 4 : let child_gap_at_key = start_key.add(500);
7094 4 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
7095 4 :
7096 4 : let mut current_lsn = Lsn(0x10);
7097 4 :
7098 4 : let timeline_id = TimelineId::generate();
7099 4 : let parent_timeline = tenant
7100 4 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7101 4 : .await?;
7102 4 :
7103 4 : current_lsn += 0x100;
7104 4 :
7105 16 : for _ in 0..3 {
7106 12 : let mut key = start_key;
7107 12012 : while key < end_key {
7108 12000 : current_lsn += 0x10;
7109 12000 :
7110 12000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
7111 4 :
7112 12000 : let mut writer = parent_timeline.writer().await;
7113 12000 : writer
7114 12000 : .put(
7115 12000 : key,
7116 12000 : current_lsn,
7117 12000 : &Value::Image(test_img(&image_value)),
7118 12000 : &ctx,
7119 12000 : )
7120 12000 : .await?;
7121 12000 : writer.finish_write(current_lsn);
7122 12000 :
7123 12000 : if key == child_gap_at_key {
7124 12 : parent_gap_lsns.insert(current_lsn, image_value);
7125 11988 : }
7126 4 :
7127 12000 : key = key.next();
7128 4 : }
7129 4 :
7130 12 : parent_timeline.freeze_and_flush().await?;
7131 4 : }
7132 4 :
7133 4 : let child_timeline_id = TimelineId::generate();
7134 4 :
7135 4 : let child_timeline = tenant
7136 4 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
7137 4 : .await?;
7138 4 :
7139 4 : let mut key = start_key;
7140 4004 : while key < end_key {
7141 4000 : if key == child_gap_at_key {
7142 4 : key = key.next();
7143 4 : continue;
7144 3996 : }
7145 3996 :
7146 3996 : current_lsn += 0x10;
7147 4 :
7148 3996 : let mut writer = child_timeline.writer().await;
7149 3996 : writer
7150 3996 : .put(
7151 3996 : key,
7152 3996 : current_lsn,
7153 3996 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
7154 3996 : &ctx,
7155 3996 : )
7156 3996 : .await?;
7157 3996 : writer.finish_write(current_lsn);
7158 3996 :
7159 3996 : key = key.next();
7160 4 : }
7161 4 :
7162 4 : child_timeline.freeze_and_flush().await?;
7163 4 :
7164 4 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
7165 4 : let mut query_lsns = Vec::new();
7166 12 : for image_lsn in parent_gap_lsns.keys().rev() {
7167 72 : for offset in lsn_offsets {
7168 60 : query_lsns.push(Lsn(image_lsn
7169 60 : .0
7170 60 : .checked_add_signed(offset)
7171 60 : .expect("Shouldn't overflow")));
7172 60 : }
7173 4 : }
7174 4 :
7175 64 : for query_lsn in query_lsns {
7176 60 : let results = child_timeline
7177 60 : .get_vectored_impl(
7178 60 : KeySpace {
7179 60 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
7180 60 : },
7181 60 : query_lsn,
7182 60 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7183 60 : &ctx,
7184 60 : )
7185 60 : .await;
7186 4 :
7187 60 : let expected_item = parent_gap_lsns
7188 60 : .iter()
7189 60 : .rev()
7190 136 : .find(|(lsn, _)| **lsn <= query_lsn);
7191 60 :
7192 60 : info!(
7193 4 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7194 4 : query_lsn, expected_item
7195 4 : );
7196 4 :
7197 60 : match expected_item {
7198 52 : Some((_, img_value)) => {
7199 52 : let key_results = results.expect("No vectored get error expected");
7200 52 : let key_result = &key_results[&child_gap_at_key];
7201 52 : let returned_img = key_result
7202 52 : .as_ref()
7203 52 : .expect("No page reconstruct error expected");
7204 52 :
7205 52 : info!(
7206 4 : "Vectored read at LSN {} returned image {}",
7207 0 : query_lsn,
7208 0 : std::str::from_utf8(returned_img)?
7209 4 : );
7210 52 : assert_eq!(*returned_img, test_img(img_value));
7211 4 : }
7212 4 : None => {
7213 8 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7214 4 : }
7215 4 : }
7216 4 : }
7217 4 :
7218 4 : Ok(())
7219 4 : }
7220 :
7221 : #[tokio::test]
7222 4 : async fn test_random_updates() -> anyhow::Result<()> {
7223 4 : let names_algorithms = [
7224 4 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7225 4 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7226 4 : ];
7227 12 : for (name, algorithm) in names_algorithms {
7228 8 : test_random_updates_algorithm(name, algorithm).await?;
7229 4 : }
7230 4 : Ok(())
7231 4 : }
7232 :
7233 8 : async fn test_random_updates_algorithm(
7234 8 : name: &'static str,
7235 8 : compaction_algorithm: CompactionAlgorithm,
7236 8 : ) -> anyhow::Result<()> {
7237 8 : let mut harness = TenantHarness::create(name).await?;
7238 8 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7239 8 : kind: compaction_algorithm,
7240 8 : };
7241 8 : let (tenant, ctx) = harness.load().await;
7242 8 : let tline = tenant
7243 8 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7244 8 : .await?;
7245 :
7246 : const NUM_KEYS: usize = 1000;
7247 8 : let cancel = CancellationToken::new();
7248 8 :
7249 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7250 8 : let mut test_key_end = test_key;
7251 8 : test_key_end.field6 = NUM_KEYS as u32;
7252 8 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7253 8 :
7254 8 : let mut keyspace = KeySpaceAccum::new();
7255 8 :
7256 8 : // Track when each page was last modified. Used to assert that
7257 8 : // a read sees the latest page version.
7258 8 : let mut updated = [Lsn(0); NUM_KEYS];
7259 8 :
7260 8 : let mut lsn = Lsn(0x10);
7261 : #[allow(clippy::needless_range_loop)]
7262 8008 : for blknum in 0..NUM_KEYS {
7263 8000 : lsn = Lsn(lsn.0 + 0x10);
7264 8000 : test_key.field6 = blknum as u32;
7265 8000 : let mut writer = tline.writer().await;
7266 8000 : writer
7267 8000 : .put(
7268 8000 : test_key,
7269 8000 : lsn,
7270 8000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7271 8000 : &ctx,
7272 8000 : )
7273 8000 : .await?;
7274 8000 : writer.finish_write(lsn);
7275 8000 : updated[blknum] = lsn;
7276 8000 : drop(writer);
7277 8000 :
7278 8000 : keyspace.add_key(test_key);
7279 : }
7280 :
7281 408 : for _ in 0..50 {
7282 400400 : for _ in 0..NUM_KEYS {
7283 400000 : lsn = Lsn(lsn.0 + 0x10);
7284 400000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7285 400000 : test_key.field6 = blknum as u32;
7286 400000 : let mut writer = tline.writer().await;
7287 400000 : writer
7288 400000 : .put(
7289 400000 : test_key,
7290 400000 : lsn,
7291 400000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7292 400000 : &ctx,
7293 400000 : )
7294 400000 : .await?;
7295 400000 : writer.finish_write(lsn);
7296 400000 : drop(writer);
7297 400000 : updated[blknum] = lsn;
7298 : }
7299 :
7300 : // Read all the blocks
7301 400000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7302 400000 : test_key.field6 = blknum as u32;
7303 400000 : assert_eq!(
7304 400000 : tline.get(test_key, lsn, &ctx).await?,
7305 400000 : test_img(&format!("{} at {}", blknum, last_lsn))
7306 : );
7307 : }
7308 :
7309 : // Perform a cycle of flush, and GC
7310 400 : tline.freeze_and_flush().await?;
7311 400 : tenant
7312 400 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7313 400 : .await?;
7314 : }
7315 :
7316 8 : Ok(())
7317 8 : }
7318 :
7319 : #[tokio::test]
7320 4 : async fn test_traverse_branches() -> anyhow::Result<()> {
7321 4 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7322 4 : .await?
7323 4 : .load()
7324 4 : .await;
7325 4 : let mut tline = tenant
7326 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7327 4 : .await?;
7328 4 :
7329 4 : const NUM_KEYS: usize = 1000;
7330 4 :
7331 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7332 4 :
7333 4 : let mut keyspace = KeySpaceAccum::new();
7334 4 :
7335 4 : let cancel = CancellationToken::new();
7336 4 :
7337 4 : // Track when each page was last modified. Used to assert that
7338 4 : // a read sees the latest page version.
7339 4 : let mut updated = [Lsn(0); NUM_KEYS];
7340 4 :
7341 4 : let mut lsn = Lsn(0x10);
7342 4 : #[allow(clippy::needless_range_loop)]
7343 4004 : for blknum in 0..NUM_KEYS {
7344 4000 : lsn = Lsn(lsn.0 + 0x10);
7345 4000 : test_key.field6 = blknum as u32;
7346 4000 : let mut writer = tline.writer().await;
7347 4000 : writer
7348 4000 : .put(
7349 4000 : test_key,
7350 4000 : lsn,
7351 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7352 4000 : &ctx,
7353 4000 : )
7354 4000 : .await?;
7355 4000 : writer.finish_write(lsn);
7356 4000 : updated[blknum] = lsn;
7357 4000 : drop(writer);
7358 4000 :
7359 4000 : keyspace.add_key(test_key);
7360 4 : }
7361 4 :
7362 204 : for _ in 0..50 {
7363 200 : let new_tline_id = TimelineId::generate();
7364 200 : tenant
7365 200 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7366 200 : .await?;
7367 200 : tline = tenant
7368 200 : .get_timeline(new_tline_id, true)
7369 200 : .expect("Should have the branched timeline");
7370 4 :
7371 200200 : for _ in 0..NUM_KEYS {
7372 200000 : lsn = Lsn(lsn.0 + 0x10);
7373 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7374 200000 : test_key.field6 = blknum as u32;
7375 200000 : let mut writer = tline.writer().await;
7376 200000 : writer
7377 200000 : .put(
7378 200000 : test_key,
7379 200000 : lsn,
7380 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7381 200000 : &ctx,
7382 200000 : )
7383 200000 : .await?;
7384 200000 : println!("updating {} at {}", blknum, lsn);
7385 200000 : writer.finish_write(lsn);
7386 200000 : drop(writer);
7387 200000 : updated[blknum] = lsn;
7388 4 : }
7389 4 :
7390 4 : // Read all the blocks
7391 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7392 200000 : test_key.field6 = blknum as u32;
7393 200000 : assert_eq!(
7394 200000 : tline.get(test_key, lsn, &ctx).await?,
7395 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
7396 4 : );
7397 4 : }
7398 4 :
7399 4 : // Perform a cycle of flush, compact, and GC
7400 200 : tline.freeze_and_flush().await?;
7401 200 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7402 200 : tenant
7403 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7404 200 : .await?;
7405 4 : }
7406 4 :
7407 4 : Ok(())
7408 4 : }
7409 :
7410 : #[tokio::test]
7411 4 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7412 4 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7413 4 : .await?
7414 4 : .load()
7415 4 : .await;
7416 4 : let mut tline = tenant
7417 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7418 4 : .await?;
7419 4 :
7420 4 : const NUM_KEYS: usize = 100;
7421 4 : const NUM_TLINES: usize = 50;
7422 4 :
7423 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7424 4 : // Track page mutation lsns across different timelines.
7425 4 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7426 4 :
7427 4 : let mut lsn = Lsn(0x10);
7428 4 :
7429 4 : #[allow(clippy::needless_range_loop)]
7430 204 : for idx in 0..NUM_TLINES {
7431 200 : let new_tline_id = TimelineId::generate();
7432 200 : tenant
7433 200 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7434 200 : .await?;
7435 200 : tline = tenant
7436 200 : .get_timeline(new_tline_id, true)
7437 200 : .expect("Should have the branched timeline");
7438 4 :
7439 20200 : for _ in 0..NUM_KEYS {
7440 20000 : lsn = Lsn(lsn.0 + 0x10);
7441 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7442 20000 : test_key.field6 = blknum as u32;
7443 20000 : let mut writer = tline.writer().await;
7444 20000 : writer
7445 20000 : .put(
7446 20000 : test_key,
7447 20000 : lsn,
7448 20000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
7449 20000 : &ctx,
7450 20000 : )
7451 20000 : .await?;
7452 20000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
7453 20000 : writer.finish_write(lsn);
7454 20000 : drop(writer);
7455 20000 : updated[idx][blknum] = lsn;
7456 4 : }
7457 4 : }
7458 4 :
7459 4 : // Read pages from leaf timeline across all ancestors.
7460 200 : for (idx, lsns) in updated.iter().enumerate() {
7461 20000 : for (blknum, lsn) in lsns.iter().enumerate() {
7462 4 : // Skip empty mutations.
7463 20000 : if lsn.0 == 0 {
7464 7426 : continue;
7465 12574 : }
7466 12574 : println!("checking [{idx}][{blknum}] at {lsn}");
7467 12574 : test_key.field6 = blknum as u32;
7468 12574 : assert_eq!(
7469 12574 : tline.get(test_key, *lsn, &ctx).await?,
7470 12574 : test_img(&format!("{idx} {blknum} at {lsn}"))
7471 4 : );
7472 4 : }
7473 4 : }
7474 4 : Ok(())
7475 4 : }
7476 :
7477 : #[tokio::test]
7478 4 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7479 4 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7480 4 : .await?
7481 4 : .load()
7482 4 : .await;
7483 4 :
7484 4 : let initdb_lsn = Lsn(0x20);
7485 4 : let (utline, ctx) = tenant
7486 4 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7487 4 : .await?;
7488 4 : let tline = utline.raw_timeline().unwrap();
7489 4 :
7490 4 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7491 4 : tline.maybe_spawn_flush_loop();
7492 4 :
7493 4 : // Make sure the timeline has the minimum set of required keys for operation.
7494 4 : // The only operation you can always do on an empty timeline is to `put` new data.
7495 4 : // Except if you `put` at `initdb_lsn`.
7496 4 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7497 4 : // It uses `repartition()`, which assumes some keys to be present.
7498 4 : // Let's make sure the test timeline can handle that case.
7499 4 : {
7500 4 : let mut state = tline.flush_loop_state.lock().unwrap();
7501 4 : assert_eq!(
7502 4 : timeline::FlushLoopState::Running {
7503 4 : expect_initdb_optimization: false,
7504 4 : initdb_optimization_count: 0,
7505 4 : },
7506 4 : *state
7507 4 : );
7508 4 : *state = timeline::FlushLoopState::Running {
7509 4 : expect_initdb_optimization: true,
7510 4 : initdb_optimization_count: 0,
7511 4 : };
7512 4 : }
7513 4 :
7514 4 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
7515 4 : // As explained above, the optimization requires some keys to be present.
7516 4 : // As per `create_empty_timeline` documentation, use init_empty to set them.
7517 4 : // This is what `create_test_timeline` does, by the way.
7518 4 : let mut modification = tline.begin_modification(initdb_lsn);
7519 4 : modification
7520 4 : .init_empty_test_timeline()
7521 4 : .context("init_empty_test_timeline")?;
7522 4 : modification
7523 4 : .commit(&ctx)
7524 4 : .await
7525 4 : .context("commit init_empty_test_timeline modification")?;
7526 4 :
7527 4 : // Do the flush. The flush code will check the expectations that we set above.
7528 4 : tline.freeze_and_flush().await?;
7529 4 :
7530 4 : // assert freeze_and_flush exercised the initdb optimization
7531 4 : {
7532 4 : let state = tline.flush_loop_state.lock().unwrap();
7533 4 : let timeline::FlushLoopState::Running {
7534 4 : expect_initdb_optimization,
7535 4 : initdb_optimization_count,
7536 4 : } = *state
7537 4 : else {
7538 4 : panic!("unexpected state: {:?}", *state);
7539 4 : };
7540 4 : assert!(expect_initdb_optimization);
7541 4 : assert!(initdb_optimization_count > 0);
7542 4 : }
7543 4 : Ok(())
7544 4 : }
7545 :
7546 : #[tokio::test]
7547 4 : async fn test_create_guard_crash() -> anyhow::Result<()> {
7548 4 : let name = "test_create_guard_crash";
7549 4 : let harness = TenantHarness::create(name).await?;
7550 4 : {
7551 4 : let (tenant, ctx) = harness.load().await;
7552 4 : let (tline, _ctx) = tenant
7553 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7554 4 : .await?;
7555 4 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
7556 4 : let raw_tline = tline.raw_timeline().unwrap();
7557 4 : raw_tline
7558 4 : .shutdown(super::timeline::ShutdownMode::Hard)
7559 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))
7560 4 : .await;
7561 4 : std::mem::forget(tline);
7562 4 : }
7563 4 :
7564 4 : let (tenant, _) = harness.load().await;
7565 4 : match tenant.get_timeline(TIMELINE_ID, false) {
7566 4 : Ok(_) => panic!("timeline should've been removed during load"),
7567 4 : Err(e) => {
7568 4 : assert_eq!(
7569 4 : e,
7570 4 : GetTimelineError::NotFound {
7571 4 : tenant_id: tenant.tenant_shard_id,
7572 4 : timeline_id: TIMELINE_ID,
7573 4 : }
7574 4 : )
7575 4 : }
7576 4 : }
7577 4 :
7578 4 : assert!(
7579 4 : !harness
7580 4 : .conf
7581 4 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
7582 4 : .exists()
7583 4 : );
7584 4 :
7585 4 : Ok(())
7586 4 : }
7587 :
7588 : #[tokio::test]
7589 4 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
7590 4 : let names_algorithms = [
7591 4 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
7592 4 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
7593 4 : ];
7594 12 : for (name, algorithm) in names_algorithms {
7595 8 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
7596 4 : }
7597 4 : Ok(())
7598 4 : }
7599 :
7600 8 : async fn test_read_at_max_lsn_algorithm(
7601 8 : name: &'static str,
7602 8 : compaction_algorithm: CompactionAlgorithm,
7603 8 : ) -> anyhow::Result<()> {
7604 8 : let mut harness = TenantHarness::create(name).await?;
7605 8 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
7606 8 : kind: compaction_algorithm,
7607 8 : };
7608 8 : let (tenant, ctx) = harness.load().await;
7609 8 : let tline = tenant
7610 8 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7611 8 : .await?;
7612 :
7613 8 : let lsn = Lsn(0x10);
7614 8 : let compact = false;
7615 8 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
7616 :
7617 8 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7618 8 : let read_lsn = Lsn(u64::MAX - 1);
7619 :
7620 8 : let result = tline.get(test_key, read_lsn, &ctx).await;
7621 8 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
7622 :
7623 8 : Ok(())
7624 8 : }
7625 :
7626 : #[tokio::test]
7627 4 : async fn test_metadata_scan() -> anyhow::Result<()> {
7628 4 : let harness = TenantHarness::create("test_metadata_scan").await?;
7629 4 : let (tenant, ctx) = harness.load().await;
7630 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7631 4 : let tline = tenant
7632 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7633 4 : .await?;
7634 4 :
7635 4 : const NUM_KEYS: usize = 1000;
7636 4 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7637 4 :
7638 4 : let cancel = CancellationToken::new();
7639 4 :
7640 4 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7641 4 : base_key.field1 = AUX_KEY_PREFIX;
7642 4 : let mut test_key = base_key;
7643 4 :
7644 4 : // Track when each page was last modified. Used to assert that
7645 4 : // a read sees the latest page version.
7646 4 : let mut updated = [Lsn(0); NUM_KEYS];
7647 4 :
7648 4 : let mut lsn = Lsn(0x10);
7649 4 : #[allow(clippy::needless_range_loop)]
7650 4004 : for blknum in 0..NUM_KEYS {
7651 4000 : lsn = Lsn(lsn.0 + 0x10);
7652 4000 : test_key.field6 = (blknum * STEP) as u32;
7653 4000 : let mut writer = tline.writer().await;
7654 4000 : writer
7655 4000 : .put(
7656 4000 : test_key,
7657 4000 : lsn,
7658 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7659 4000 : &ctx,
7660 4000 : )
7661 4000 : .await?;
7662 4000 : writer.finish_write(lsn);
7663 4000 : updated[blknum] = lsn;
7664 4000 : drop(writer);
7665 4 : }
7666 4 :
7667 4 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7668 4 :
7669 48 : for iter in 0..=10 {
7670 4 : // Read all the blocks
7671 44000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7672 44000 : test_key.field6 = (blknum * STEP) as u32;
7673 44000 : assert_eq!(
7674 44000 : tline.get(test_key, lsn, &ctx).await?,
7675 44000 : test_img(&format!("{} at {}", blknum, last_lsn))
7676 4 : );
7677 4 : }
7678 4 :
7679 44 : let mut cnt = 0;
7680 44000 : for (key, value) in tline
7681 44 : .get_vectored_impl(
7682 44 : keyspace.clone(),
7683 44 : lsn,
7684 44 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7685 44 : &ctx,
7686 44 : )
7687 44 : .await?
7688 4 : {
7689 44000 : let blknum = key.field6 as usize;
7690 44000 : let value = value?;
7691 44000 : assert!(blknum % STEP == 0);
7692 44000 : let blknum = blknum / STEP;
7693 44000 : assert_eq!(
7694 44000 : value,
7695 44000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
7696 44000 : );
7697 44000 : cnt += 1;
7698 4 : }
7699 4 :
7700 44 : assert_eq!(cnt, NUM_KEYS);
7701 4 :
7702 44044 : for _ in 0..NUM_KEYS {
7703 44000 : lsn = Lsn(lsn.0 + 0x10);
7704 44000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7705 44000 : test_key.field6 = (blknum * STEP) as u32;
7706 44000 : let mut writer = tline.writer().await;
7707 44000 : writer
7708 44000 : .put(
7709 44000 : test_key,
7710 44000 : lsn,
7711 44000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7712 44000 : &ctx,
7713 44000 : )
7714 44000 : .await?;
7715 44000 : writer.finish_write(lsn);
7716 44000 : drop(writer);
7717 44000 : updated[blknum] = lsn;
7718 4 : }
7719 4 :
7720 4 : // Perform two cycles of flush, compact, and GC
7721 132 : for round in 0..2 {
7722 88 : tline.freeze_and_flush().await?;
7723 88 : tline
7724 88 : .compact(
7725 88 : &cancel,
7726 88 : if iter % 5 == 0 && round == 0 {
7727 12 : let mut flags = EnumSet::new();
7728 12 : flags.insert(CompactFlags::ForceImageLayerCreation);
7729 12 : flags.insert(CompactFlags::ForceRepartition);
7730 12 : flags
7731 4 : } else {
7732 76 : EnumSet::empty()
7733 4 : },
7734 88 : &ctx,
7735 88 : )
7736 88 : .await?;
7737 88 : tenant
7738 88 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7739 88 : .await?;
7740 4 : }
7741 4 : }
7742 4 :
7743 4 : Ok(())
7744 4 : }
7745 :
7746 : #[tokio::test]
7747 4 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
7748 4 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
7749 4 : let (tenant, ctx) = harness.load().await;
7750 4 : let tline = tenant
7751 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7752 4 : .await?;
7753 4 :
7754 4 : let cancel = CancellationToken::new();
7755 4 :
7756 4 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7757 4 : base_key.field1 = AUX_KEY_PREFIX;
7758 4 : let test_key = base_key;
7759 4 : let mut lsn = Lsn(0x10);
7760 4 :
7761 84 : for _ in 0..20 {
7762 80 : lsn = Lsn(lsn.0 + 0x10);
7763 80 : let mut writer = tline.writer().await;
7764 80 : writer
7765 80 : .put(
7766 80 : test_key,
7767 80 : lsn,
7768 80 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
7769 80 : &ctx,
7770 80 : )
7771 80 : .await?;
7772 80 : writer.finish_write(lsn);
7773 80 : drop(writer);
7774 80 : tline.freeze_and_flush().await?; // force create a delta layer
7775 4 : }
7776 4 :
7777 4 : let before_num_l0_delta_files =
7778 4 : tline.layers.read().await.layer_map()?.level0_deltas().len();
7779 4 :
7780 4 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
7781 4 :
7782 4 : let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
7783 4 :
7784 4 : assert!(
7785 4 : after_num_l0_delta_files < before_num_l0_delta_files,
7786 4 : "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}"
7787 4 : );
7788 4 :
7789 4 : assert_eq!(
7790 4 : tline.get(test_key, lsn, &ctx).await?,
7791 4 : test_img(&format!("{} at {}", 0, lsn))
7792 4 : );
7793 4 :
7794 4 : Ok(())
7795 4 : }
7796 :
7797 : #[tokio::test]
7798 4 : async fn test_aux_file_e2e() {
7799 4 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
7800 4 :
7801 4 : let (tenant, ctx) = harness.load().await;
7802 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7803 4 :
7804 4 : let mut lsn = Lsn(0x08);
7805 4 :
7806 4 : let tline: Arc<Timeline> = tenant
7807 4 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
7808 4 : .await
7809 4 : .unwrap();
7810 4 :
7811 4 : {
7812 4 : lsn += 8;
7813 4 : let mut modification = tline.begin_modification(lsn);
7814 4 : modification
7815 4 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
7816 4 : .await
7817 4 : .unwrap();
7818 4 : modification.commit(&ctx).await.unwrap();
7819 4 : }
7820 4 :
7821 4 : // we can read everything from the storage
7822 4 : let files = tline
7823 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7824 4 : .await
7825 4 : .unwrap();
7826 4 : assert_eq!(
7827 4 : files.get("pg_logical/mappings/test1"),
7828 4 : Some(&bytes::Bytes::from_static(b"first"))
7829 4 : );
7830 4 :
7831 4 : {
7832 4 : lsn += 8;
7833 4 : let mut modification = tline.begin_modification(lsn);
7834 4 : modification
7835 4 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
7836 4 : .await
7837 4 : .unwrap();
7838 4 : modification.commit(&ctx).await.unwrap();
7839 4 : }
7840 4 :
7841 4 : let files = tline
7842 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7843 4 : .await
7844 4 : .unwrap();
7845 4 : assert_eq!(
7846 4 : files.get("pg_logical/mappings/test2"),
7847 4 : Some(&bytes::Bytes::from_static(b"second"))
7848 4 : );
7849 4 :
7850 4 : let child = tenant
7851 4 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
7852 4 : .await
7853 4 : .unwrap();
7854 4 :
7855 4 : let files = child
7856 4 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
7857 4 : .await
7858 4 : .unwrap();
7859 4 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
7860 4 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
7861 4 : }
7862 :
7863 : #[tokio::test]
7864 4 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
7865 4 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
7866 4 : let (tenant, ctx) = harness.load().await;
7867 4 : let io_concurrency = IoConcurrency::spawn_for_test();
7868 4 : let tline = tenant
7869 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7870 4 : .await?;
7871 4 :
7872 4 : const NUM_KEYS: usize = 1000;
7873 4 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
7874 4 :
7875 4 : let cancel = CancellationToken::new();
7876 4 :
7877 4 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7878 4 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
7879 4 : let mut test_key = base_key;
7880 4 : let mut lsn = Lsn(0x10);
7881 4 :
7882 16 : async fn scan_with_statistics(
7883 16 : tline: &Timeline,
7884 16 : keyspace: &KeySpace,
7885 16 : lsn: Lsn,
7886 16 : ctx: &RequestContext,
7887 16 : io_concurrency: IoConcurrency,
7888 16 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
7889 16 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
7890 16 : let res = tline
7891 16 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
7892 16 : .await?;
7893 16 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
7894 16 : }
7895 4 :
7896 4 : #[allow(clippy::needless_range_loop)]
7897 4004 : for blknum in 0..NUM_KEYS {
7898 4000 : lsn = Lsn(lsn.0 + 0x10);
7899 4000 : test_key.field6 = (blknum * STEP) as u32;
7900 4000 : let mut writer = tline.writer().await;
7901 4000 : writer
7902 4000 : .put(
7903 4000 : test_key,
7904 4000 : lsn,
7905 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7906 4000 : &ctx,
7907 4000 : )
7908 4000 : .await?;
7909 4000 : writer.finish_write(lsn);
7910 4000 : drop(writer);
7911 4 : }
7912 4 :
7913 4 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
7914 4 :
7915 44 : for iter in 1..=10 {
7916 40040 : for _ in 0..NUM_KEYS {
7917 40000 : lsn = Lsn(lsn.0 + 0x10);
7918 40000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7919 40000 : test_key.field6 = (blknum * STEP) as u32;
7920 40000 : let mut writer = tline.writer().await;
7921 40000 : writer
7922 40000 : .put(
7923 40000 : test_key,
7924 40000 : lsn,
7925 40000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
7926 40000 : &ctx,
7927 40000 : )
7928 40000 : .await?;
7929 40000 : writer.finish_write(lsn);
7930 40000 : drop(writer);
7931 4 : }
7932 4 :
7933 40 : tline.freeze_and_flush().await?;
7934 4 :
7935 40 : if iter % 5 == 0 {
7936 8 : let (_, before_delta_file_accessed) =
7937 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
7938 8 : .await?;
7939 8 : tline
7940 8 : .compact(
7941 8 : &cancel,
7942 8 : {
7943 8 : let mut flags = EnumSet::new();
7944 8 : flags.insert(CompactFlags::ForceImageLayerCreation);
7945 8 : flags.insert(CompactFlags::ForceRepartition);
7946 8 : flags
7947 8 : },
7948 8 : &ctx,
7949 8 : )
7950 8 : .await?;
7951 8 : let (_, after_delta_file_accessed) =
7952 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
7953 8 : .await?;
7954 8 : assert!(
7955 8 : after_delta_file_accessed < before_delta_file_accessed,
7956 4 : "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}"
7957 4 : );
7958 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.
7959 8 : assert!(
7960 8 : after_delta_file_accessed <= 2,
7961 4 : "after_delta_file_accessed={after_delta_file_accessed}"
7962 4 : );
7963 32 : }
7964 4 : }
7965 4 :
7966 4 : Ok(())
7967 4 : }
7968 :
7969 : #[tokio::test]
7970 4 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
7971 4 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
7972 4 : let (tenant, ctx) = harness.load().await;
7973 4 :
7974 4 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
7975 4 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
7976 4 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
7977 4 :
7978 4 : let tline = tenant
7979 4 : .create_test_timeline_with_layers(
7980 4 : TIMELINE_ID,
7981 4 : Lsn(0x10),
7982 4 : DEFAULT_PG_VERSION,
7983 4 : &ctx,
7984 4 : Vec::new(), // in-memory layers
7985 4 : Vec::new(), // delta layers
7986 4 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
7987 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
7988 4 : )
7989 4 : .await?;
7990 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
7991 4 :
7992 4 : let child = tenant
7993 4 : .branch_timeline_test_with_layers(
7994 4 : &tline,
7995 4 : NEW_TIMELINE_ID,
7996 4 : Some(Lsn(0x20)),
7997 4 : &ctx,
7998 4 : Vec::new(), // delta layers
7999 4 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
8000 4 : Lsn(0x30),
8001 4 : )
8002 4 : .await
8003 4 : .unwrap();
8004 4 :
8005 4 : let lsn = Lsn(0x30);
8006 4 :
8007 4 : // test vectored get on parent timeline
8008 4 : assert_eq!(
8009 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8010 4 : Some(test_img("data key 1"))
8011 4 : );
8012 4 : assert!(
8013 4 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
8014 4 : .await
8015 4 : .unwrap_err()
8016 4 : .is_missing_key_error()
8017 4 : );
8018 4 : assert!(
8019 4 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
8020 4 : .await
8021 4 : .unwrap_err()
8022 4 : .is_missing_key_error()
8023 4 : );
8024 4 :
8025 4 : // test vectored get on child timeline
8026 4 : assert_eq!(
8027 4 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8028 4 : Some(test_img("data key 1"))
8029 4 : );
8030 4 : assert_eq!(
8031 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8032 4 : Some(test_img("data key 2"))
8033 4 : );
8034 4 : assert!(
8035 4 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
8036 4 : .await
8037 4 : .unwrap_err()
8038 4 : .is_missing_key_error()
8039 4 : );
8040 4 :
8041 4 : Ok(())
8042 4 : }
8043 :
8044 : #[tokio::test]
8045 4 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
8046 4 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
8047 4 : let (tenant, ctx) = harness.load().await;
8048 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8049 4 :
8050 4 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8051 4 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
8052 4 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
8053 4 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
8054 4 :
8055 4 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
8056 4 : let base_inherited_key_child =
8057 4 : Key::from_hex("610000000033333333444444445500000001").unwrap();
8058 4 : let base_inherited_key_nonexist =
8059 4 : Key::from_hex("610000000033333333444444445500000002").unwrap();
8060 4 : let base_inherited_key_overwrite =
8061 4 : Key::from_hex("610000000033333333444444445500000003").unwrap();
8062 4 :
8063 4 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8064 4 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
8065 4 :
8066 4 : let tline = tenant
8067 4 : .create_test_timeline_with_layers(
8068 4 : TIMELINE_ID,
8069 4 : Lsn(0x10),
8070 4 : DEFAULT_PG_VERSION,
8071 4 : &ctx,
8072 4 : Vec::new(), // in-memory layers
8073 4 : Vec::new(), // delta layers
8074 4 : vec![(
8075 4 : Lsn(0x20),
8076 4 : vec![
8077 4 : (base_inherited_key, test_img("metadata inherited key 1")),
8078 4 : (
8079 4 : base_inherited_key_overwrite,
8080 4 : test_img("metadata key overwrite 1a"),
8081 4 : ),
8082 4 : (base_key, test_img("metadata key 1")),
8083 4 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8084 4 : ],
8085 4 : )], // image layers
8086 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
8087 4 : )
8088 4 : .await?;
8089 4 :
8090 4 : let child = tenant
8091 4 : .branch_timeline_test_with_layers(
8092 4 : &tline,
8093 4 : NEW_TIMELINE_ID,
8094 4 : Some(Lsn(0x20)),
8095 4 : &ctx,
8096 4 : Vec::new(), // delta layers
8097 4 : vec![(
8098 4 : Lsn(0x30),
8099 4 : vec![
8100 4 : (
8101 4 : base_inherited_key_child,
8102 4 : test_img("metadata inherited key 2"),
8103 4 : ),
8104 4 : (
8105 4 : base_inherited_key_overwrite,
8106 4 : test_img("metadata key overwrite 2a"),
8107 4 : ),
8108 4 : (base_key_child, test_img("metadata key 2")),
8109 4 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8110 4 : ],
8111 4 : )], // image layers
8112 4 : Lsn(0x30),
8113 4 : )
8114 4 : .await
8115 4 : .unwrap();
8116 4 :
8117 4 : let lsn = Lsn(0x30);
8118 4 :
8119 4 : // test vectored get on parent timeline
8120 4 : assert_eq!(
8121 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8122 4 : Some(test_img("metadata key 1"))
8123 4 : );
8124 4 : assert_eq!(
8125 4 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
8126 4 : None
8127 4 : );
8128 4 : assert_eq!(
8129 4 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
8130 4 : None
8131 4 : );
8132 4 : assert_eq!(
8133 4 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
8134 4 : Some(test_img("metadata key overwrite 1b"))
8135 4 : );
8136 4 : assert_eq!(
8137 4 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
8138 4 : Some(test_img("metadata inherited key 1"))
8139 4 : );
8140 4 : assert_eq!(
8141 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
8142 4 : None
8143 4 : );
8144 4 : assert_eq!(
8145 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
8146 4 : None
8147 4 : );
8148 4 : assert_eq!(
8149 4 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
8150 4 : Some(test_img("metadata key overwrite 1a"))
8151 4 : );
8152 4 :
8153 4 : // test vectored get on child timeline
8154 4 : assert_eq!(
8155 4 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8156 4 : None
8157 4 : );
8158 4 : assert_eq!(
8159 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8160 4 : Some(test_img("metadata key 2"))
8161 4 : );
8162 4 : assert_eq!(
8163 4 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
8164 4 : None
8165 4 : );
8166 4 : assert_eq!(
8167 4 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
8168 4 : Some(test_img("metadata inherited key 1"))
8169 4 : );
8170 4 : assert_eq!(
8171 4 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
8172 4 : Some(test_img("metadata inherited key 2"))
8173 4 : );
8174 4 : assert_eq!(
8175 4 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
8176 4 : None
8177 4 : );
8178 4 : assert_eq!(
8179 4 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
8180 4 : Some(test_img("metadata key overwrite 2b"))
8181 4 : );
8182 4 : assert_eq!(
8183 4 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
8184 4 : Some(test_img("metadata key overwrite 2a"))
8185 4 : );
8186 4 :
8187 4 : // test vectored scan on parent timeline
8188 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8189 4 : let res = tline
8190 4 : .get_vectored_impl(
8191 4 : KeySpace::single(Key::metadata_key_range()),
8192 4 : lsn,
8193 4 : &mut reconstruct_state,
8194 4 : &ctx,
8195 4 : )
8196 4 : .await?;
8197 4 :
8198 4 : assert_eq!(
8199 4 : res.into_iter()
8200 16 : .map(|(k, v)| (k, v.unwrap()))
8201 4 : .collect::<Vec<_>>(),
8202 4 : vec![
8203 4 : (base_inherited_key, test_img("metadata inherited key 1")),
8204 4 : (
8205 4 : base_inherited_key_overwrite,
8206 4 : test_img("metadata key overwrite 1a")
8207 4 : ),
8208 4 : (base_key, test_img("metadata key 1")),
8209 4 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8210 4 : ]
8211 4 : );
8212 4 :
8213 4 : // test vectored scan on child timeline
8214 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8215 4 : let res = child
8216 4 : .get_vectored_impl(
8217 4 : KeySpace::single(Key::metadata_key_range()),
8218 4 : lsn,
8219 4 : &mut reconstruct_state,
8220 4 : &ctx,
8221 4 : )
8222 4 : .await?;
8223 4 :
8224 4 : assert_eq!(
8225 4 : res.into_iter()
8226 20 : .map(|(k, v)| (k, v.unwrap()))
8227 4 : .collect::<Vec<_>>(),
8228 4 : vec![
8229 4 : (base_inherited_key, test_img("metadata inherited key 1")),
8230 4 : (
8231 4 : base_inherited_key_child,
8232 4 : test_img("metadata inherited key 2")
8233 4 : ),
8234 4 : (
8235 4 : base_inherited_key_overwrite,
8236 4 : test_img("metadata key overwrite 2a")
8237 4 : ),
8238 4 : (base_key_child, test_img("metadata key 2")),
8239 4 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8240 4 : ]
8241 4 : );
8242 4 :
8243 4 : Ok(())
8244 4 : }
8245 :
8246 112 : async fn get_vectored_impl_wrapper(
8247 112 : tline: &Arc<Timeline>,
8248 112 : key: Key,
8249 112 : lsn: Lsn,
8250 112 : ctx: &RequestContext,
8251 112 : ) -> Result<Option<Bytes>, GetVectoredError> {
8252 112 : let io_concurrency =
8253 112 : IoConcurrency::spawn_from_conf(tline.conf, tline.gate.enter().unwrap());
8254 112 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8255 112 : let mut res = tline
8256 112 : .get_vectored_impl(
8257 112 : KeySpace::single(key..key.next()),
8258 112 : lsn,
8259 112 : &mut reconstruct_state,
8260 112 : ctx,
8261 112 : )
8262 112 : .await?;
8263 100 : Ok(res.pop_last().map(|(k, v)| {
8264 64 : assert_eq!(k, key);
8265 64 : v.unwrap()
8266 100 : }))
8267 112 : }
8268 :
8269 : #[tokio::test]
8270 4 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8271 4 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8272 4 : let (tenant, ctx) = harness.load().await;
8273 4 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8274 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8275 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8276 4 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8277 4 :
8278 4 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8279 4 : // Lsn 0x30 key0, key3, no key1+key2
8280 4 : // Lsn 0x20 key1+key2 tomestones
8281 4 : // Lsn 0x10 key1 in image, key2 in delta
8282 4 : let tline = tenant
8283 4 : .create_test_timeline_with_layers(
8284 4 : TIMELINE_ID,
8285 4 : Lsn(0x10),
8286 4 : DEFAULT_PG_VERSION,
8287 4 : &ctx,
8288 4 : Vec::new(), // in-memory layers
8289 4 : // delta layers
8290 4 : vec![
8291 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8292 4 : Lsn(0x10)..Lsn(0x20),
8293 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8294 4 : ),
8295 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8296 4 : Lsn(0x20)..Lsn(0x30),
8297 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8298 4 : ),
8299 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8300 4 : Lsn(0x20)..Lsn(0x30),
8301 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8302 4 : ),
8303 4 : ],
8304 4 : // image layers
8305 4 : vec![
8306 4 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8307 4 : (
8308 4 : Lsn(0x30),
8309 4 : vec![
8310 4 : (key0, test_img("metadata key 0")),
8311 4 : (key3, test_img("metadata key 3")),
8312 4 : ],
8313 4 : ),
8314 4 : ],
8315 4 : Lsn(0x30),
8316 4 : )
8317 4 : .await?;
8318 4 :
8319 4 : let lsn = Lsn(0x30);
8320 4 : let old_lsn = Lsn(0x20);
8321 4 :
8322 4 : assert_eq!(
8323 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8324 4 : Some(test_img("metadata key 0"))
8325 4 : );
8326 4 : assert_eq!(
8327 4 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8328 4 : None,
8329 4 : );
8330 4 : assert_eq!(
8331 4 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8332 4 : None,
8333 4 : );
8334 4 : assert_eq!(
8335 4 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8336 4 : Some(Bytes::new()),
8337 4 : );
8338 4 : assert_eq!(
8339 4 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8340 4 : Some(Bytes::new()),
8341 4 : );
8342 4 : assert_eq!(
8343 4 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8344 4 : Some(test_img("metadata key 3"))
8345 4 : );
8346 4 :
8347 4 : Ok(())
8348 4 : }
8349 :
8350 : #[tokio::test]
8351 4 : async fn test_metadata_tombstone_image_creation() {
8352 4 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8353 4 : .await
8354 4 : .unwrap();
8355 4 : let (tenant, ctx) = harness.load().await;
8356 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8357 4 :
8358 4 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8359 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8360 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8361 4 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8362 4 :
8363 4 : let tline = tenant
8364 4 : .create_test_timeline_with_layers(
8365 4 : TIMELINE_ID,
8366 4 : Lsn(0x10),
8367 4 : DEFAULT_PG_VERSION,
8368 4 : &ctx,
8369 4 : Vec::new(), // in-memory layers
8370 4 : // delta layers
8371 4 : vec![
8372 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8373 4 : Lsn(0x10)..Lsn(0x20),
8374 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8375 4 : ),
8376 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8377 4 : Lsn(0x20)..Lsn(0x30),
8378 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8379 4 : ),
8380 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8381 4 : Lsn(0x20)..Lsn(0x30),
8382 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8383 4 : ),
8384 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8385 4 : Lsn(0x30)..Lsn(0x40),
8386 4 : vec![
8387 4 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8388 4 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8389 4 : ],
8390 4 : ),
8391 4 : ],
8392 4 : // image layers
8393 4 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8394 4 : Lsn(0x40),
8395 4 : )
8396 4 : .await
8397 4 : .unwrap();
8398 4 :
8399 4 : let cancel = CancellationToken::new();
8400 4 :
8401 4 : tline
8402 4 : .compact(
8403 4 : &cancel,
8404 4 : {
8405 4 : let mut flags = EnumSet::new();
8406 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
8407 4 : flags.insert(CompactFlags::ForceRepartition);
8408 4 : flags
8409 4 : },
8410 4 : &ctx,
8411 4 : )
8412 4 : .await
8413 4 : .unwrap();
8414 4 :
8415 4 : // Image layers are created at last_record_lsn
8416 4 : let images = tline
8417 4 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
8418 4 : .await
8419 4 : .unwrap()
8420 4 : .into_iter()
8421 36 : .filter(|(k, _)| k.is_metadata_key())
8422 4 : .collect::<Vec<_>>();
8423 4 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8424 4 : }
8425 :
8426 : #[tokio::test]
8427 4 : async fn test_metadata_tombstone_empty_image_creation() {
8428 4 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8429 4 : .await
8430 4 : .unwrap();
8431 4 : let (tenant, ctx) = harness.load().await;
8432 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8433 4 :
8434 4 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8435 4 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8436 4 :
8437 4 : let tline = tenant
8438 4 : .create_test_timeline_with_layers(
8439 4 : TIMELINE_ID,
8440 4 : Lsn(0x10),
8441 4 : DEFAULT_PG_VERSION,
8442 4 : &ctx,
8443 4 : Vec::new(), // in-memory layers
8444 4 : // delta layers
8445 4 : vec![
8446 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8447 4 : Lsn(0x10)..Lsn(0x20),
8448 4 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8449 4 : ),
8450 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8451 4 : Lsn(0x20)..Lsn(0x30),
8452 4 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8453 4 : ),
8454 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
8455 4 : Lsn(0x20)..Lsn(0x30),
8456 4 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8457 4 : ),
8458 4 : ],
8459 4 : // image layers
8460 4 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8461 4 : Lsn(0x30),
8462 4 : )
8463 4 : .await
8464 4 : .unwrap();
8465 4 :
8466 4 : let cancel = CancellationToken::new();
8467 4 :
8468 4 : tline
8469 4 : .compact(
8470 4 : &cancel,
8471 4 : {
8472 4 : let mut flags = EnumSet::new();
8473 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
8474 4 : flags.insert(CompactFlags::ForceRepartition);
8475 4 : flags
8476 4 : },
8477 4 : &ctx,
8478 4 : )
8479 4 : .await
8480 4 : .unwrap();
8481 4 :
8482 4 : // Image layers are created at last_record_lsn
8483 4 : let images = tline
8484 4 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8485 4 : .await
8486 4 : .unwrap()
8487 4 : .into_iter()
8488 28 : .filter(|(k, _)| k.is_metadata_key())
8489 4 : .collect::<Vec<_>>();
8490 4 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
8491 4 : }
8492 :
8493 : #[tokio::test]
8494 4 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
8495 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
8496 4 : let (tenant, ctx) = harness.load().await;
8497 4 : let io_concurrency = IoConcurrency::spawn_for_test();
8498 4 :
8499 204 : fn get_key(id: u32) -> Key {
8500 204 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8501 204 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8502 204 : key.field6 = id;
8503 204 : key
8504 204 : }
8505 4 :
8506 4 : // We create
8507 4 : // - one bottom-most image layer,
8508 4 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8509 4 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8510 4 : // - a delta layer D3 above the horizon.
8511 4 : //
8512 4 : // | D3 |
8513 4 : // | D1 |
8514 4 : // -| |-- gc horizon -----------------
8515 4 : // | | | D2 |
8516 4 : // --------- img layer ------------------
8517 4 : //
8518 4 : // What we should expact from this compaction is:
8519 4 : // | D3 |
8520 4 : // | Part of D1 |
8521 4 : // --------- img layer with D1+D2 at GC horizon------------------
8522 4 :
8523 4 : // img layer at 0x10
8524 4 : let img_layer = (0..10)
8525 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8526 4 : .collect_vec();
8527 4 :
8528 4 : let delta1 = vec![
8529 4 : (
8530 4 : get_key(1),
8531 4 : Lsn(0x20),
8532 4 : Value::Image(Bytes::from("value 1@0x20")),
8533 4 : ),
8534 4 : (
8535 4 : get_key(2),
8536 4 : Lsn(0x30),
8537 4 : Value::Image(Bytes::from("value 2@0x30")),
8538 4 : ),
8539 4 : (
8540 4 : get_key(3),
8541 4 : Lsn(0x40),
8542 4 : Value::Image(Bytes::from("value 3@0x40")),
8543 4 : ),
8544 4 : ];
8545 4 : let delta2 = vec![
8546 4 : (
8547 4 : get_key(5),
8548 4 : Lsn(0x20),
8549 4 : Value::Image(Bytes::from("value 5@0x20")),
8550 4 : ),
8551 4 : (
8552 4 : get_key(6),
8553 4 : Lsn(0x20),
8554 4 : Value::Image(Bytes::from("value 6@0x20")),
8555 4 : ),
8556 4 : ];
8557 4 : let delta3 = vec![
8558 4 : (
8559 4 : get_key(8),
8560 4 : Lsn(0x48),
8561 4 : Value::Image(Bytes::from("value 8@0x48")),
8562 4 : ),
8563 4 : (
8564 4 : get_key(9),
8565 4 : Lsn(0x48),
8566 4 : Value::Image(Bytes::from("value 9@0x48")),
8567 4 : ),
8568 4 : ];
8569 4 :
8570 4 : let tline = tenant
8571 4 : .create_test_timeline_with_layers(
8572 4 : TIMELINE_ID,
8573 4 : Lsn(0x10),
8574 4 : DEFAULT_PG_VERSION,
8575 4 : &ctx,
8576 4 : Vec::new(), // in-memory layers
8577 4 : vec![
8578 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
8579 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
8580 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
8581 4 : ], // delta layers
8582 4 : vec![(Lsn(0x10), img_layer)], // image layers
8583 4 : Lsn(0x50),
8584 4 : )
8585 4 : .await?;
8586 4 : {
8587 4 : tline
8588 4 : .applied_gc_cutoff_lsn
8589 4 : .lock_for_write()
8590 4 : .store_and_unlock(Lsn(0x30))
8591 4 : .wait()
8592 4 : .await;
8593 4 : // Update GC info
8594 4 : let mut guard = tline.gc_info.write().unwrap();
8595 4 : guard.cutoffs.time = Lsn(0x30);
8596 4 : guard.cutoffs.space = Lsn(0x30);
8597 4 : }
8598 4 :
8599 4 : let expected_result = [
8600 4 : Bytes::from_static(b"value 0@0x10"),
8601 4 : Bytes::from_static(b"value 1@0x20"),
8602 4 : Bytes::from_static(b"value 2@0x30"),
8603 4 : Bytes::from_static(b"value 3@0x40"),
8604 4 : Bytes::from_static(b"value 4@0x10"),
8605 4 : Bytes::from_static(b"value 5@0x20"),
8606 4 : Bytes::from_static(b"value 6@0x20"),
8607 4 : Bytes::from_static(b"value 7@0x10"),
8608 4 : Bytes::from_static(b"value 8@0x48"),
8609 4 : Bytes::from_static(b"value 9@0x48"),
8610 4 : ];
8611 4 :
8612 40 : for (idx, expected) in expected_result.iter().enumerate() {
8613 40 : assert_eq!(
8614 40 : tline
8615 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8616 40 : .await
8617 40 : .unwrap(),
8618 4 : expected
8619 4 : );
8620 4 : }
8621 4 :
8622 4 : let cancel = CancellationToken::new();
8623 4 : tline
8624 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8625 4 : .await
8626 4 : .unwrap();
8627 4 :
8628 40 : for (idx, expected) in expected_result.iter().enumerate() {
8629 40 : assert_eq!(
8630 40 : tline
8631 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
8632 40 : .await
8633 40 : .unwrap(),
8634 4 : expected
8635 4 : );
8636 4 : }
8637 4 :
8638 4 : // Check if the image layer at the GC horizon contains exactly what we want
8639 4 : let image_at_gc_horizon = tline
8640 4 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
8641 4 : .await
8642 4 : .unwrap()
8643 4 : .into_iter()
8644 68 : .filter(|(k, _)| k.is_metadata_key())
8645 4 : .collect::<Vec<_>>();
8646 4 :
8647 4 : assert_eq!(image_at_gc_horizon.len(), 10);
8648 4 : let expected_result = [
8649 4 : Bytes::from_static(b"value 0@0x10"),
8650 4 : Bytes::from_static(b"value 1@0x20"),
8651 4 : Bytes::from_static(b"value 2@0x30"),
8652 4 : Bytes::from_static(b"value 3@0x10"),
8653 4 : Bytes::from_static(b"value 4@0x10"),
8654 4 : Bytes::from_static(b"value 5@0x20"),
8655 4 : Bytes::from_static(b"value 6@0x20"),
8656 4 : Bytes::from_static(b"value 7@0x10"),
8657 4 : Bytes::from_static(b"value 8@0x10"),
8658 4 : Bytes::from_static(b"value 9@0x10"),
8659 4 : ];
8660 44 : for idx in 0..10 {
8661 40 : assert_eq!(
8662 40 : image_at_gc_horizon[idx],
8663 40 : (get_key(idx as u32), expected_result[idx].clone())
8664 40 : );
8665 4 : }
8666 4 :
8667 4 : // Check if old layers are removed / new layers have the expected LSN
8668 4 : let all_layers = inspect_and_sort(&tline, None).await;
8669 4 : assert_eq!(
8670 4 : all_layers,
8671 4 : vec![
8672 4 : // Image layer at GC horizon
8673 4 : PersistentLayerKey {
8674 4 : key_range: Key::MIN..Key::MAX,
8675 4 : lsn_range: Lsn(0x30)..Lsn(0x31),
8676 4 : is_delta: false
8677 4 : },
8678 4 : // The delta layer below the horizon
8679 4 : PersistentLayerKey {
8680 4 : key_range: get_key(3)..get_key(4),
8681 4 : lsn_range: Lsn(0x30)..Lsn(0x48),
8682 4 : is_delta: true
8683 4 : },
8684 4 : // The delta3 layer that should not be picked for the compaction
8685 4 : PersistentLayerKey {
8686 4 : key_range: get_key(8)..get_key(10),
8687 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
8688 4 : is_delta: true
8689 4 : }
8690 4 : ]
8691 4 : );
8692 4 :
8693 4 : // increase GC horizon and compact again
8694 4 : {
8695 4 : tline
8696 4 : .applied_gc_cutoff_lsn
8697 4 : .lock_for_write()
8698 4 : .store_and_unlock(Lsn(0x40))
8699 4 : .wait()
8700 4 : .await;
8701 4 : // Update GC info
8702 4 : let mut guard = tline.gc_info.write().unwrap();
8703 4 : guard.cutoffs.time = Lsn(0x40);
8704 4 : guard.cutoffs.space = Lsn(0x40);
8705 4 : }
8706 4 : tline
8707 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
8708 4 : .await
8709 4 : .unwrap();
8710 4 :
8711 4 : Ok(())
8712 4 : }
8713 :
8714 : #[cfg(feature = "testing")]
8715 : #[tokio::test]
8716 4 : async fn test_neon_test_record() -> anyhow::Result<()> {
8717 4 : let harness = TenantHarness::create("test_neon_test_record").await?;
8718 4 : let (tenant, ctx) = harness.load().await;
8719 4 :
8720 48 : fn get_key(id: u32) -> Key {
8721 48 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8722 48 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8723 48 : key.field6 = id;
8724 48 : key
8725 48 : }
8726 4 :
8727 4 : let delta1 = vec![
8728 4 : (
8729 4 : get_key(1),
8730 4 : Lsn(0x20),
8731 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8732 4 : ),
8733 4 : (
8734 4 : get_key(1),
8735 4 : Lsn(0x30),
8736 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8737 4 : ),
8738 4 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
8739 4 : (
8740 4 : get_key(2),
8741 4 : Lsn(0x20),
8742 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
8743 4 : ),
8744 4 : (
8745 4 : get_key(2),
8746 4 : Lsn(0x30),
8747 4 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
8748 4 : ),
8749 4 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
8750 4 : (
8751 4 : get_key(3),
8752 4 : Lsn(0x20),
8753 4 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
8754 4 : ),
8755 4 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
8756 4 : (
8757 4 : get_key(4),
8758 4 : Lsn(0x20),
8759 4 : Value::WalRecord(NeonWalRecord::wal_init("i")),
8760 4 : ),
8761 4 : ];
8762 4 : let image1 = vec![(get_key(1), "0x10".into())];
8763 4 :
8764 4 : let tline = tenant
8765 4 : .create_test_timeline_with_layers(
8766 4 : TIMELINE_ID,
8767 4 : Lsn(0x10),
8768 4 : DEFAULT_PG_VERSION,
8769 4 : &ctx,
8770 4 : Vec::new(), // in-memory layers
8771 4 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
8772 4 : Lsn(0x10)..Lsn(0x40),
8773 4 : delta1,
8774 4 : )], // delta layers
8775 4 : vec![(Lsn(0x10), image1)], // image layers
8776 4 : Lsn(0x50),
8777 4 : )
8778 4 : .await?;
8779 4 :
8780 4 : assert_eq!(
8781 4 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
8782 4 : Bytes::from_static(b"0x10,0x20,0x30")
8783 4 : );
8784 4 : assert_eq!(
8785 4 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
8786 4 : Bytes::from_static(b"0x10,0x20,0x30")
8787 4 : );
8788 4 :
8789 4 : // Need to remove the limit of "Neon WAL redo requires base image".
8790 4 :
8791 4 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
8792 4 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
8793 4 :
8794 4 : Ok(())
8795 4 : }
8796 :
8797 : #[tokio::test(start_paused = true)]
8798 4 : async fn test_lsn_lease() -> anyhow::Result<()> {
8799 4 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
8800 4 : .await
8801 4 : .unwrap()
8802 4 : .load()
8803 4 : .await;
8804 4 : // Advance to the lsn lease deadline so that GC is not blocked by
8805 4 : // initial transition into AttachedSingle.
8806 4 : tokio::time::advance(tenant.get_lsn_lease_length()).await;
8807 4 : tokio::time::resume();
8808 4 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8809 4 :
8810 4 : let end_lsn = Lsn(0x100);
8811 4 : let image_layers = (0x20..=0x90)
8812 4 : .step_by(0x10)
8813 32 : .map(|n| {
8814 32 : (
8815 32 : Lsn(n),
8816 32 : vec![(key, test_img(&format!("data key at {:x}", n)))],
8817 32 : )
8818 32 : })
8819 4 : .collect();
8820 4 :
8821 4 : let timeline = tenant
8822 4 : .create_test_timeline_with_layers(
8823 4 : TIMELINE_ID,
8824 4 : Lsn(0x10),
8825 4 : DEFAULT_PG_VERSION,
8826 4 : &ctx,
8827 4 : Vec::new(), // in-memory layers
8828 4 : Vec::new(),
8829 4 : image_layers,
8830 4 : end_lsn,
8831 4 : )
8832 4 : .await?;
8833 4 :
8834 4 : let leased_lsns = [0x30, 0x50, 0x70];
8835 4 : let mut leases = Vec::new();
8836 12 : leased_lsns.iter().for_each(|n| {
8837 12 : leases.push(
8838 12 : timeline
8839 12 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
8840 12 : .expect("lease request should succeed"),
8841 12 : );
8842 12 : });
8843 4 :
8844 4 : let updated_lease_0 = timeline
8845 4 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
8846 4 : .expect("lease renewal should succeed");
8847 4 : assert_eq!(
8848 4 : updated_lease_0.valid_until, leases[0].valid_until,
8849 4 : " Renewing with shorter lease should not change the lease."
8850 4 : );
8851 4 :
8852 4 : let updated_lease_1 = timeline
8853 4 : .renew_lsn_lease(
8854 4 : Lsn(leased_lsns[1]),
8855 4 : timeline.get_lsn_lease_length() * 2,
8856 4 : &ctx,
8857 4 : )
8858 4 : .expect("lease renewal should succeed");
8859 4 : assert!(
8860 4 : updated_lease_1.valid_until > leases[1].valid_until,
8861 4 : "Renewing with a long lease should renew lease with later expiration time."
8862 4 : );
8863 4 :
8864 4 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
8865 4 : info!(
8866 4 : "applied_gc_cutoff_lsn: {}",
8867 0 : *timeline.get_applied_gc_cutoff_lsn()
8868 4 : );
8869 4 : timeline.force_set_disk_consistent_lsn(end_lsn);
8870 4 :
8871 4 : let res = tenant
8872 4 : .gc_iteration(
8873 4 : Some(TIMELINE_ID),
8874 4 : 0,
8875 4 : Duration::ZERO,
8876 4 : &CancellationToken::new(),
8877 4 : &ctx,
8878 4 : )
8879 4 : .await
8880 4 : .unwrap();
8881 4 :
8882 4 : // Keeping everything <= Lsn(0x80) b/c leases:
8883 4 : // 0/10: initdb layer
8884 4 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
8885 4 : assert_eq!(res.layers_needed_by_leases, 7);
8886 4 : // Keeping 0/90 b/c it is the latest layer.
8887 4 : assert_eq!(res.layers_not_updated, 1);
8888 4 : // Removed 0/80.
8889 4 : assert_eq!(res.layers_removed, 1);
8890 4 :
8891 4 : // Make lease on a already GC-ed LSN.
8892 4 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
8893 4 : assert!(Lsn(0x80) < *timeline.get_applied_gc_cutoff_lsn());
8894 4 : timeline
8895 4 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
8896 4 : .expect_err("lease request on GC-ed LSN should fail");
8897 4 :
8898 4 : // Should still be able to renew a currently valid lease
8899 4 : // Assumption: original lease to is still valid for 0/50.
8900 4 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
8901 4 : timeline
8902 4 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
8903 4 : .expect("lease renewal with validation should succeed");
8904 4 :
8905 4 : Ok(())
8906 4 : }
8907 :
8908 : #[cfg(feature = "testing")]
8909 : #[tokio::test]
8910 4 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
8911 4 : test_simple_bottom_most_compaction_deltas_helper(
8912 4 : "test_simple_bottom_most_compaction_deltas_1",
8913 4 : false,
8914 4 : )
8915 4 : .await
8916 4 : }
8917 :
8918 : #[cfg(feature = "testing")]
8919 : #[tokio::test]
8920 4 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
8921 4 : test_simple_bottom_most_compaction_deltas_helper(
8922 4 : "test_simple_bottom_most_compaction_deltas_2",
8923 4 : true,
8924 4 : )
8925 4 : .await
8926 4 : }
8927 :
8928 : #[cfg(feature = "testing")]
8929 8 : async fn test_simple_bottom_most_compaction_deltas_helper(
8930 8 : test_name: &'static str,
8931 8 : use_delta_bottom_layer: bool,
8932 8 : ) -> anyhow::Result<()> {
8933 8 : let harness = TenantHarness::create(test_name).await?;
8934 8 : let (tenant, ctx) = harness.load().await;
8935 :
8936 552 : fn get_key(id: u32) -> Key {
8937 552 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
8938 552 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8939 552 : key.field6 = id;
8940 552 : key
8941 552 : }
8942 :
8943 : // We create
8944 : // - one bottom-most image layer,
8945 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
8946 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
8947 : // - a delta layer D3 above the horizon.
8948 : //
8949 : // | D3 |
8950 : // | D1 |
8951 : // -| |-- gc horizon -----------------
8952 : // | | | D2 |
8953 : // --------- img layer ------------------
8954 : //
8955 : // What we should expact from this compaction is:
8956 : // | D3 |
8957 : // | Part of D1 |
8958 : // --------- img layer with D1+D2 at GC horizon------------------
8959 :
8960 : // img layer at 0x10
8961 8 : let img_layer = (0..10)
8962 80 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
8963 8 : .collect_vec();
8964 8 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
8965 8 : let delta4 = (0..10)
8966 80 : .map(|id| {
8967 80 : (
8968 80 : get_key(id),
8969 80 : Lsn(0x08),
8970 80 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
8971 80 : )
8972 80 : })
8973 8 : .collect_vec();
8974 8 :
8975 8 : let delta1 = vec![
8976 8 : (
8977 8 : get_key(1),
8978 8 : Lsn(0x20),
8979 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
8980 8 : ),
8981 8 : (
8982 8 : get_key(2),
8983 8 : Lsn(0x30),
8984 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8985 8 : ),
8986 8 : (
8987 8 : get_key(3),
8988 8 : Lsn(0x28),
8989 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
8990 8 : ),
8991 8 : (
8992 8 : get_key(3),
8993 8 : Lsn(0x30),
8994 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
8995 8 : ),
8996 8 : (
8997 8 : get_key(3),
8998 8 : Lsn(0x40),
8999 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9000 8 : ),
9001 8 : ];
9002 8 : let delta2 = vec![
9003 8 : (
9004 8 : get_key(5),
9005 8 : Lsn(0x20),
9006 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9007 8 : ),
9008 8 : (
9009 8 : get_key(6),
9010 8 : Lsn(0x20),
9011 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9012 8 : ),
9013 8 : ];
9014 8 : let delta3 = vec![
9015 8 : (
9016 8 : get_key(8),
9017 8 : Lsn(0x48),
9018 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9019 8 : ),
9020 8 : (
9021 8 : get_key(9),
9022 8 : Lsn(0x48),
9023 8 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9024 8 : ),
9025 8 : ];
9026 :
9027 8 : let tline = if use_delta_bottom_layer {
9028 4 : tenant
9029 4 : .create_test_timeline_with_layers(
9030 4 : TIMELINE_ID,
9031 4 : Lsn(0x08),
9032 4 : DEFAULT_PG_VERSION,
9033 4 : &ctx,
9034 4 : Vec::new(), // in-memory layers
9035 4 : vec![
9036 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
9037 4 : Lsn(0x08)..Lsn(0x10),
9038 4 : delta4,
9039 4 : ),
9040 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
9041 4 : Lsn(0x20)..Lsn(0x48),
9042 4 : delta1,
9043 4 : ),
9044 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
9045 4 : Lsn(0x20)..Lsn(0x48),
9046 4 : delta2,
9047 4 : ),
9048 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
9049 4 : Lsn(0x48)..Lsn(0x50),
9050 4 : delta3,
9051 4 : ),
9052 4 : ], // delta layers
9053 4 : vec![], // image layers
9054 4 : Lsn(0x50),
9055 4 : )
9056 4 : .await?
9057 : } else {
9058 4 : tenant
9059 4 : .create_test_timeline_with_layers(
9060 4 : TIMELINE_ID,
9061 4 : Lsn(0x10),
9062 4 : DEFAULT_PG_VERSION,
9063 4 : &ctx,
9064 4 : Vec::new(), // in-memory layers
9065 4 : vec![
9066 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
9067 4 : Lsn(0x10)..Lsn(0x48),
9068 4 : delta1,
9069 4 : ),
9070 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
9071 4 : Lsn(0x10)..Lsn(0x48),
9072 4 : delta2,
9073 4 : ),
9074 4 : DeltaLayerTestDesc::new_with_inferred_key_range(
9075 4 : Lsn(0x48)..Lsn(0x50),
9076 4 : delta3,
9077 4 : ),
9078 4 : ], // delta layers
9079 4 : vec![(Lsn(0x10), img_layer)], // image layers
9080 4 : Lsn(0x50),
9081 4 : )
9082 4 : .await?
9083 : };
9084 : {
9085 8 : tline
9086 8 : .applied_gc_cutoff_lsn
9087 8 : .lock_for_write()
9088 8 : .store_and_unlock(Lsn(0x30))
9089 8 : .wait()
9090 8 : .await;
9091 : // Update GC info
9092 8 : let mut guard = tline.gc_info.write().unwrap();
9093 8 : *guard = GcInfo {
9094 8 : retain_lsns: vec![],
9095 8 : cutoffs: GcCutoffs {
9096 8 : time: Lsn(0x30),
9097 8 : space: Lsn(0x30),
9098 8 : },
9099 8 : leases: Default::default(),
9100 8 : within_ancestor_pitr: false,
9101 8 : };
9102 8 : }
9103 8 :
9104 8 : let expected_result = [
9105 8 : Bytes::from_static(b"value 0@0x10"),
9106 8 : Bytes::from_static(b"value 1@0x10@0x20"),
9107 8 : Bytes::from_static(b"value 2@0x10@0x30"),
9108 8 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9109 8 : Bytes::from_static(b"value 4@0x10"),
9110 8 : Bytes::from_static(b"value 5@0x10@0x20"),
9111 8 : Bytes::from_static(b"value 6@0x10@0x20"),
9112 8 : Bytes::from_static(b"value 7@0x10"),
9113 8 : Bytes::from_static(b"value 8@0x10@0x48"),
9114 8 : Bytes::from_static(b"value 9@0x10@0x48"),
9115 8 : ];
9116 8 :
9117 8 : let expected_result_at_gc_horizon = [
9118 8 : Bytes::from_static(b"value 0@0x10"),
9119 8 : Bytes::from_static(b"value 1@0x10@0x20"),
9120 8 : Bytes::from_static(b"value 2@0x10@0x30"),
9121 8 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9122 8 : Bytes::from_static(b"value 4@0x10"),
9123 8 : Bytes::from_static(b"value 5@0x10@0x20"),
9124 8 : Bytes::from_static(b"value 6@0x10@0x20"),
9125 8 : Bytes::from_static(b"value 7@0x10"),
9126 8 : Bytes::from_static(b"value 8@0x10"),
9127 8 : Bytes::from_static(b"value 9@0x10"),
9128 8 : ];
9129 :
9130 88 : for idx in 0..10 {
9131 80 : assert_eq!(
9132 80 : tline
9133 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9134 80 : .await
9135 80 : .unwrap(),
9136 80 : &expected_result[idx]
9137 : );
9138 80 : assert_eq!(
9139 80 : tline
9140 80 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9141 80 : .await
9142 80 : .unwrap(),
9143 80 : &expected_result_at_gc_horizon[idx]
9144 : );
9145 : }
9146 :
9147 8 : let cancel = CancellationToken::new();
9148 8 : tline
9149 8 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9150 8 : .await
9151 8 : .unwrap();
9152 :
9153 88 : for idx in 0..10 {
9154 80 : assert_eq!(
9155 80 : tline
9156 80 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9157 80 : .await
9158 80 : .unwrap(),
9159 80 : &expected_result[idx]
9160 : );
9161 80 : assert_eq!(
9162 80 : tline
9163 80 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9164 80 : .await
9165 80 : .unwrap(),
9166 80 : &expected_result_at_gc_horizon[idx]
9167 : );
9168 : }
9169 :
9170 : // increase GC horizon and compact again
9171 : {
9172 8 : tline
9173 8 : .applied_gc_cutoff_lsn
9174 8 : .lock_for_write()
9175 8 : .store_and_unlock(Lsn(0x40))
9176 8 : .wait()
9177 8 : .await;
9178 : // Update GC info
9179 8 : let mut guard = tline.gc_info.write().unwrap();
9180 8 : guard.cutoffs.time = Lsn(0x40);
9181 8 : guard.cutoffs.space = Lsn(0x40);
9182 8 : }
9183 8 : tline
9184 8 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9185 8 : .await
9186 8 : .unwrap();
9187 8 :
9188 8 : Ok(())
9189 8 : }
9190 :
9191 : #[cfg(feature = "testing")]
9192 : #[tokio::test]
9193 4 : async fn test_generate_key_retention() -> anyhow::Result<()> {
9194 4 : let harness = TenantHarness::create("test_generate_key_retention").await?;
9195 4 : let (tenant, ctx) = harness.load().await;
9196 4 : let tline = tenant
9197 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9198 4 : .await?;
9199 4 : tline.force_advance_lsn(Lsn(0x70));
9200 4 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9201 4 : let history = vec![
9202 4 : (
9203 4 : key,
9204 4 : Lsn(0x10),
9205 4 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
9206 4 : ),
9207 4 : (
9208 4 : key,
9209 4 : Lsn(0x20),
9210 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9211 4 : ),
9212 4 : (
9213 4 : key,
9214 4 : Lsn(0x30),
9215 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9216 4 : ),
9217 4 : (
9218 4 : key,
9219 4 : Lsn(0x40),
9220 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9221 4 : ),
9222 4 : (
9223 4 : key,
9224 4 : Lsn(0x50),
9225 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9226 4 : ),
9227 4 : (
9228 4 : key,
9229 4 : Lsn(0x60),
9230 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9231 4 : ),
9232 4 : (
9233 4 : key,
9234 4 : Lsn(0x70),
9235 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9236 4 : ),
9237 4 : (
9238 4 : key,
9239 4 : Lsn(0x80),
9240 4 : Value::Image(Bytes::copy_from_slice(
9241 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9242 4 : )),
9243 4 : ),
9244 4 : (
9245 4 : key,
9246 4 : Lsn(0x90),
9247 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9248 4 : ),
9249 4 : ];
9250 4 : let res = tline
9251 4 : .generate_key_retention(
9252 4 : key,
9253 4 : &history,
9254 4 : Lsn(0x60),
9255 4 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9256 4 : 3,
9257 4 : None,
9258 4 : )
9259 4 : .await
9260 4 : .unwrap();
9261 4 : let expected_res = KeyHistoryRetention {
9262 4 : below_horizon: vec![
9263 4 : (
9264 4 : Lsn(0x20),
9265 4 : KeyLogAtLsn(vec![(
9266 4 : Lsn(0x20),
9267 4 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9268 4 : )]),
9269 4 : ),
9270 4 : (
9271 4 : Lsn(0x40),
9272 4 : KeyLogAtLsn(vec![
9273 4 : (
9274 4 : Lsn(0x30),
9275 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9276 4 : ),
9277 4 : (
9278 4 : Lsn(0x40),
9279 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9280 4 : ),
9281 4 : ]),
9282 4 : ),
9283 4 : (
9284 4 : Lsn(0x50),
9285 4 : KeyLogAtLsn(vec![(
9286 4 : Lsn(0x50),
9287 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9288 4 : )]),
9289 4 : ),
9290 4 : (
9291 4 : Lsn(0x60),
9292 4 : KeyLogAtLsn(vec![(
9293 4 : Lsn(0x60),
9294 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9295 4 : )]),
9296 4 : ),
9297 4 : ],
9298 4 : above_horizon: KeyLogAtLsn(vec![
9299 4 : (
9300 4 : Lsn(0x70),
9301 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9302 4 : ),
9303 4 : (
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 : Lsn(0x90),
9311 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9312 4 : ),
9313 4 : ]),
9314 4 : };
9315 4 : assert_eq!(res, expected_res);
9316 4 :
9317 4 : // We expect GC-compaction to run with the original GC. This would create a situation that
9318 4 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9319 4 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9320 4 : // For example, we have
9321 4 : // ```plain
9322 4 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9323 4 : // ```
9324 4 : // Now the GC horizon moves up, and we have
9325 4 : // ```plain
9326 4 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9327 4 : // ```
9328 4 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9329 4 : // We will end up with
9330 4 : // ```plain
9331 4 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9332 4 : // ```
9333 4 : // Now we run the GC-compaction, and this key does not have a full history.
9334 4 : // We should be able to handle this partial history and drop everything before the
9335 4 : // gc_horizon image.
9336 4 :
9337 4 : let history = vec![
9338 4 : (
9339 4 : key,
9340 4 : Lsn(0x20),
9341 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9342 4 : ),
9343 4 : (
9344 4 : key,
9345 4 : Lsn(0x30),
9346 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9347 4 : ),
9348 4 : (
9349 4 : key,
9350 4 : Lsn(0x40),
9351 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9352 4 : ),
9353 4 : (
9354 4 : key,
9355 4 : Lsn(0x50),
9356 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9357 4 : ),
9358 4 : (
9359 4 : key,
9360 4 : Lsn(0x60),
9361 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9362 4 : ),
9363 4 : (
9364 4 : key,
9365 4 : Lsn(0x70),
9366 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9367 4 : ),
9368 4 : (
9369 4 : key,
9370 4 : Lsn(0x80),
9371 4 : Value::Image(Bytes::copy_from_slice(
9372 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9373 4 : )),
9374 4 : ),
9375 4 : (
9376 4 : key,
9377 4 : Lsn(0x90),
9378 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9379 4 : ),
9380 4 : ];
9381 4 : let res = tline
9382 4 : .generate_key_retention(key, &history, Lsn(0x60), &[Lsn(0x40), Lsn(0x50)], 3, None)
9383 4 : .await
9384 4 : .unwrap();
9385 4 : let expected_res = KeyHistoryRetention {
9386 4 : below_horizon: vec![
9387 4 : (
9388 4 : Lsn(0x40),
9389 4 : KeyLogAtLsn(vec![(
9390 4 : Lsn(0x40),
9391 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
9392 4 : )]),
9393 4 : ),
9394 4 : (
9395 4 : Lsn(0x50),
9396 4 : KeyLogAtLsn(vec![(
9397 4 : Lsn(0x50),
9398 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9399 4 : )]),
9400 4 : ),
9401 4 : (
9402 4 : Lsn(0x60),
9403 4 : KeyLogAtLsn(vec![(
9404 4 : Lsn(0x60),
9405 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9406 4 : )]),
9407 4 : ),
9408 4 : ],
9409 4 : above_horizon: KeyLogAtLsn(vec![
9410 4 : (
9411 4 : Lsn(0x70),
9412 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9413 4 : ),
9414 4 : (
9415 4 : Lsn(0x80),
9416 4 : Value::Image(Bytes::copy_from_slice(
9417 4 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9418 4 : )),
9419 4 : ),
9420 4 : (
9421 4 : Lsn(0x90),
9422 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9423 4 : ),
9424 4 : ]),
9425 4 : };
9426 4 : assert_eq!(res, expected_res);
9427 4 :
9428 4 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
9429 4 : // the ancestor image in the test case.
9430 4 :
9431 4 : let history = vec![
9432 4 : (
9433 4 : key,
9434 4 : Lsn(0x20),
9435 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9436 4 : ),
9437 4 : (
9438 4 : key,
9439 4 : Lsn(0x30),
9440 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9441 4 : ),
9442 4 : (
9443 4 : key,
9444 4 : Lsn(0x40),
9445 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9446 4 : ),
9447 4 : (
9448 4 : key,
9449 4 : Lsn(0x70),
9450 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9451 4 : ),
9452 4 : ];
9453 4 : let res = tline
9454 4 : .generate_key_retention(
9455 4 : key,
9456 4 : &history,
9457 4 : Lsn(0x60),
9458 4 : &[],
9459 4 : 3,
9460 4 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9461 4 : )
9462 4 : .await
9463 4 : .unwrap();
9464 4 : let expected_res = KeyHistoryRetention {
9465 4 : below_horizon: vec![(
9466 4 : Lsn(0x60),
9467 4 : KeyLogAtLsn(vec![(
9468 4 : Lsn(0x60),
9469 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
9470 4 : )]),
9471 4 : )],
9472 4 : above_horizon: KeyLogAtLsn(vec![(
9473 4 : Lsn(0x70),
9474 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9475 4 : )]),
9476 4 : };
9477 4 : assert_eq!(res, expected_res);
9478 4 :
9479 4 : let history = vec![
9480 4 : (
9481 4 : key,
9482 4 : Lsn(0x20),
9483 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9484 4 : ),
9485 4 : (
9486 4 : key,
9487 4 : Lsn(0x40),
9488 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9489 4 : ),
9490 4 : (
9491 4 : key,
9492 4 : Lsn(0x60),
9493 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9494 4 : ),
9495 4 : (
9496 4 : key,
9497 4 : Lsn(0x70),
9498 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9499 4 : ),
9500 4 : ];
9501 4 : let res = tline
9502 4 : .generate_key_retention(
9503 4 : key,
9504 4 : &history,
9505 4 : Lsn(0x60),
9506 4 : &[Lsn(0x30)],
9507 4 : 3,
9508 4 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
9509 4 : )
9510 4 : .await
9511 4 : .unwrap();
9512 4 : let expected_res = KeyHistoryRetention {
9513 4 : below_horizon: vec![
9514 4 : (
9515 4 : Lsn(0x30),
9516 4 : KeyLogAtLsn(vec![(
9517 4 : Lsn(0x20),
9518 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9519 4 : )]),
9520 4 : ),
9521 4 : (
9522 4 : Lsn(0x60),
9523 4 : KeyLogAtLsn(vec![(
9524 4 : Lsn(0x60),
9525 4 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
9526 4 : )]),
9527 4 : ),
9528 4 : ],
9529 4 : above_horizon: KeyLogAtLsn(vec![(
9530 4 : Lsn(0x70),
9531 4 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9532 4 : )]),
9533 4 : };
9534 4 : assert_eq!(res, expected_res);
9535 4 :
9536 4 : Ok(())
9537 4 : }
9538 :
9539 : #[cfg(feature = "testing")]
9540 : #[tokio::test]
9541 4 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
9542 4 : let harness =
9543 4 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
9544 4 : let (tenant, ctx) = harness.load().await;
9545 4 :
9546 1036 : fn get_key(id: u32) -> Key {
9547 1036 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9548 1036 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9549 1036 : key.field6 = id;
9550 1036 : key
9551 1036 : }
9552 4 :
9553 4 : let img_layer = (0..10)
9554 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9555 4 : .collect_vec();
9556 4 :
9557 4 : let delta1 = vec![
9558 4 : (
9559 4 : get_key(1),
9560 4 : Lsn(0x20),
9561 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9562 4 : ),
9563 4 : (
9564 4 : get_key(2),
9565 4 : Lsn(0x30),
9566 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9567 4 : ),
9568 4 : (
9569 4 : get_key(3),
9570 4 : Lsn(0x28),
9571 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9572 4 : ),
9573 4 : (
9574 4 : get_key(3),
9575 4 : Lsn(0x30),
9576 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9577 4 : ),
9578 4 : (
9579 4 : get_key(3),
9580 4 : Lsn(0x40),
9581 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9582 4 : ),
9583 4 : ];
9584 4 : let delta2 = vec![
9585 4 : (
9586 4 : get_key(5),
9587 4 : Lsn(0x20),
9588 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9589 4 : ),
9590 4 : (
9591 4 : get_key(6),
9592 4 : Lsn(0x20),
9593 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9594 4 : ),
9595 4 : ];
9596 4 : let delta3 = vec![
9597 4 : (
9598 4 : get_key(8),
9599 4 : Lsn(0x48),
9600 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9601 4 : ),
9602 4 : (
9603 4 : get_key(9),
9604 4 : Lsn(0x48),
9605 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9606 4 : ),
9607 4 : ];
9608 4 :
9609 4 : let tline = tenant
9610 4 : .create_test_timeline_with_layers(
9611 4 : TIMELINE_ID,
9612 4 : Lsn(0x10),
9613 4 : DEFAULT_PG_VERSION,
9614 4 : &ctx,
9615 4 : Vec::new(), // in-memory layers
9616 4 : vec![
9617 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
9618 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
9619 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9620 4 : ], // delta layers
9621 4 : vec![(Lsn(0x10), img_layer)], // image layers
9622 4 : Lsn(0x50),
9623 4 : )
9624 4 : .await?;
9625 4 : {
9626 4 : tline
9627 4 : .applied_gc_cutoff_lsn
9628 4 : .lock_for_write()
9629 4 : .store_and_unlock(Lsn(0x30))
9630 4 : .wait()
9631 4 : .await;
9632 4 : // Update GC info
9633 4 : let mut guard = tline.gc_info.write().unwrap();
9634 4 : *guard = GcInfo {
9635 4 : retain_lsns: vec![
9636 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9637 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9638 4 : ],
9639 4 : cutoffs: GcCutoffs {
9640 4 : time: Lsn(0x30),
9641 4 : space: Lsn(0x30),
9642 4 : },
9643 4 : leases: Default::default(),
9644 4 : within_ancestor_pitr: false,
9645 4 : };
9646 4 : }
9647 4 :
9648 4 : let expected_result = [
9649 4 : Bytes::from_static(b"value 0@0x10"),
9650 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9651 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9652 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9653 4 : Bytes::from_static(b"value 4@0x10"),
9654 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9655 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9656 4 : Bytes::from_static(b"value 7@0x10"),
9657 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9658 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9659 4 : ];
9660 4 :
9661 4 : let expected_result_at_gc_horizon = [
9662 4 : Bytes::from_static(b"value 0@0x10"),
9663 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9664 4 : Bytes::from_static(b"value 2@0x10@0x30"),
9665 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9666 4 : Bytes::from_static(b"value 4@0x10"),
9667 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9668 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9669 4 : Bytes::from_static(b"value 7@0x10"),
9670 4 : Bytes::from_static(b"value 8@0x10"),
9671 4 : Bytes::from_static(b"value 9@0x10"),
9672 4 : ];
9673 4 :
9674 4 : let expected_result_at_lsn_20 = [
9675 4 : Bytes::from_static(b"value 0@0x10"),
9676 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9677 4 : Bytes::from_static(b"value 2@0x10"),
9678 4 : Bytes::from_static(b"value 3@0x10"),
9679 4 : Bytes::from_static(b"value 4@0x10"),
9680 4 : Bytes::from_static(b"value 5@0x10@0x20"),
9681 4 : Bytes::from_static(b"value 6@0x10@0x20"),
9682 4 : Bytes::from_static(b"value 7@0x10"),
9683 4 : Bytes::from_static(b"value 8@0x10"),
9684 4 : Bytes::from_static(b"value 9@0x10"),
9685 4 : ];
9686 4 :
9687 4 : let expected_result_at_lsn_10 = [
9688 4 : Bytes::from_static(b"value 0@0x10"),
9689 4 : Bytes::from_static(b"value 1@0x10"),
9690 4 : Bytes::from_static(b"value 2@0x10"),
9691 4 : Bytes::from_static(b"value 3@0x10"),
9692 4 : Bytes::from_static(b"value 4@0x10"),
9693 4 : Bytes::from_static(b"value 5@0x10"),
9694 4 : Bytes::from_static(b"value 6@0x10"),
9695 4 : Bytes::from_static(b"value 7@0x10"),
9696 4 : Bytes::from_static(b"value 8@0x10"),
9697 4 : Bytes::from_static(b"value 9@0x10"),
9698 4 : ];
9699 4 :
9700 24 : let verify_result = || async {
9701 24 : let gc_horizon = {
9702 24 : let gc_info = tline.gc_info.read().unwrap();
9703 24 : gc_info.cutoffs.time
9704 4 : };
9705 264 : for idx in 0..10 {
9706 240 : assert_eq!(
9707 240 : tline
9708 240 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9709 240 : .await
9710 240 : .unwrap(),
9711 240 : &expected_result[idx]
9712 4 : );
9713 240 : assert_eq!(
9714 240 : tline
9715 240 : .get(get_key(idx as u32), gc_horizon, &ctx)
9716 240 : .await
9717 240 : .unwrap(),
9718 240 : &expected_result_at_gc_horizon[idx]
9719 4 : );
9720 240 : assert_eq!(
9721 240 : tline
9722 240 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9723 240 : .await
9724 240 : .unwrap(),
9725 240 : &expected_result_at_lsn_20[idx]
9726 4 : );
9727 240 : assert_eq!(
9728 240 : tline
9729 240 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9730 240 : .await
9731 240 : .unwrap(),
9732 240 : &expected_result_at_lsn_10[idx]
9733 4 : );
9734 4 : }
9735 48 : };
9736 4 :
9737 4 : verify_result().await;
9738 4 :
9739 4 : let cancel = CancellationToken::new();
9740 4 : let mut dryrun_flags = EnumSet::new();
9741 4 : dryrun_flags.insert(CompactFlags::DryRun);
9742 4 :
9743 4 : tline
9744 4 : .compact_with_gc(
9745 4 : &cancel,
9746 4 : CompactOptions {
9747 4 : flags: dryrun_flags,
9748 4 : ..Default::default()
9749 4 : },
9750 4 : &ctx,
9751 4 : )
9752 4 : .await
9753 4 : .unwrap();
9754 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
9755 4 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
9756 4 : verify_result().await;
9757 4 :
9758 4 : tline
9759 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9760 4 : .await
9761 4 : .unwrap();
9762 4 : verify_result().await;
9763 4 :
9764 4 : // compact again
9765 4 : tline
9766 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9767 4 : .await
9768 4 : .unwrap();
9769 4 : verify_result().await;
9770 4 :
9771 4 : // increase GC horizon and compact again
9772 4 : {
9773 4 : tline
9774 4 : .applied_gc_cutoff_lsn
9775 4 : .lock_for_write()
9776 4 : .store_and_unlock(Lsn(0x38))
9777 4 : .wait()
9778 4 : .await;
9779 4 : // Update GC info
9780 4 : let mut guard = tline.gc_info.write().unwrap();
9781 4 : guard.cutoffs.time = Lsn(0x38);
9782 4 : guard.cutoffs.space = Lsn(0x38);
9783 4 : }
9784 4 : tline
9785 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9786 4 : .await
9787 4 : .unwrap();
9788 4 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
9789 4 :
9790 4 : // not increasing the GC horizon and compact again
9791 4 : tline
9792 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9793 4 : .await
9794 4 : .unwrap();
9795 4 : verify_result().await;
9796 4 :
9797 4 : Ok(())
9798 4 : }
9799 :
9800 : #[cfg(feature = "testing")]
9801 : #[tokio::test]
9802 4 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
9803 4 : {
9804 4 : let harness =
9805 4 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
9806 4 : .await?;
9807 4 : let (tenant, ctx) = harness.load().await;
9808 4 :
9809 704 : fn get_key(id: u32) -> Key {
9810 704 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9811 704 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9812 704 : key.field6 = id;
9813 704 : key
9814 704 : }
9815 4 :
9816 4 : let img_layer = (0..10)
9817 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9818 4 : .collect_vec();
9819 4 :
9820 4 : let delta1 = vec![
9821 4 : (
9822 4 : get_key(1),
9823 4 : Lsn(0x20),
9824 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9825 4 : ),
9826 4 : (
9827 4 : get_key(1),
9828 4 : Lsn(0x28),
9829 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9830 4 : ),
9831 4 : ];
9832 4 : let delta2 = vec![
9833 4 : (
9834 4 : get_key(1),
9835 4 : Lsn(0x30),
9836 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9837 4 : ),
9838 4 : (
9839 4 : get_key(1),
9840 4 : Lsn(0x38),
9841 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
9842 4 : ),
9843 4 : ];
9844 4 : let delta3 = vec![
9845 4 : (
9846 4 : get_key(8),
9847 4 : Lsn(0x48),
9848 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9849 4 : ),
9850 4 : (
9851 4 : get_key(9),
9852 4 : Lsn(0x48),
9853 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9854 4 : ),
9855 4 : ];
9856 4 :
9857 4 : let tline = tenant
9858 4 : .create_test_timeline_with_layers(
9859 4 : TIMELINE_ID,
9860 4 : Lsn(0x10),
9861 4 : DEFAULT_PG_VERSION,
9862 4 : &ctx,
9863 4 : Vec::new(), // in-memory layers
9864 4 : vec![
9865 4 : // delta1 and delta 2 only contain a single key but multiple updates
9866 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
9867 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
9868 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
9869 4 : ], // delta layers
9870 4 : vec![(Lsn(0x10), img_layer)], // image layers
9871 4 : Lsn(0x50),
9872 4 : )
9873 4 : .await?;
9874 4 : {
9875 4 : tline
9876 4 : .applied_gc_cutoff_lsn
9877 4 : .lock_for_write()
9878 4 : .store_and_unlock(Lsn(0x30))
9879 4 : .wait()
9880 4 : .await;
9881 4 : // Update GC info
9882 4 : let mut guard = tline.gc_info.write().unwrap();
9883 4 : *guard = GcInfo {
9884 4 : retain_lsns: vec![
9885 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
9886 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
9887 4 : ],
9888 4 : cutoffs: GcCutoffs {
9889 4 : time: Lsn(0x30),
9890 4 : space: Lsn(0x30),
9891 4 : },
9892 4 : leases: Default::default(),
9893 4 : within_ancestor_pitr: false,
9894 4 : };
9895 4 : }
9896 4 :
9897 4 : let expected_result = [
9898 4 : Bytes::from_static(b"value 0@0x10"),
9899 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
9900 4 : Bytes::from_static(b"value 2@0x10"),
9901 4 : Bytes::from_static(b"value 3@0x10"),
9902 4 : Bytes::from_static(b"value 4@0x10"),
9903 4 : Bytes::from_static(b"value 5@0x10"),
9904 4 : Bytes::from_static(b"value 6@0x10"),
9905 4 : Bytes::from_static(b"value 7@0x10"),
9906 4 : Bytes::from_static(b"value 8@0x10@0x48"),
9907 4 : Bytes::from_static(b"value 9@0x10@0x48"),
9908 4 : ];
9909 4 :
9910 4 : let expected_result_at_gc_horizon = [
9911 4 : Bytes::from_static(b"value 0@0x10"),
9912 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
9913 4 : Bytes::from_static(b"value 2@0x10"),
9914 4 : Bytes::from_static(b"value 3@0x10"),
9915 4 : Bytes::from_static(b"value 4@0x10"),
9916 4 : Bytes::from_static(b"value 5@0x10"),
9917 4 : Bytes::from_static(b"value 6@0x10"),
9918 4 : Bytes::from_static(b"value 7@0x10"),
9919 4 : Bytes::from_static(b"value 8@0x10"),
9920 4 : Bytes::from_static(b"value 9@0x10"),
9921 4 : ];
9922 4 :
9923 4 : let expected_result_at_lsn_20 = [
9924 4 : Bytes::from_static(b"value 0@0x10"),
9925 4 : Bytes::from_static(b"value 1@0x10@0x20"),
9926 4 : Bytes::from_static(b"value 2@0x10"),
9927 4 : Bytes::from_static(b"value 3@0x10"),
9928 4 : Bytes::from_static(b"value 4@0x10"),
9929 4 : Bytes::from_static(b"value 5@0x10"),
9930 4 : Bytes::from_static(b"value 6@0x10"),
9931 4 : Bytes::from_static(b"value 7@0x10"),
9932 4 : Bytes::from_static(b"value 8@0x10"),
9933 4 : Bytes::from_static(b"value 9@0x10"),
9934 4 : ];
9935 4 :
9936 4 : let expected_result_at_lsn_10 = [
9937 4 : Bytes::from_static(b"value 0@0x10"),
9938 4 : Bytes::from_static(b"value 1@0x10"),
9939 4 : Bytes::from_static(b"value 2@0x10"),
9940 4 : Bytes::from_static(b"value 3@0x10"),
9941 4 : Bytes::from_static(b"value 4@0x10"),
9942 4 : Bytes::from_static(b"value 5@0x10"),
9943 4 : Bytes::from_static(b"value 6@0x10"),
9944 4 : Bytes::from_static(b"value 7@0x10"),
9945 4 : Bytes::from_static(b"value 8@0x10"),
9946 4 : Bytes::from_static(b"value 9@0x10"),
9947 4 : ];
9948 4 :
9949 16 : let verify_result = || async {
9950 16 : let gc_horizon = {
9951 16 : let gc_info = tline.gc_info.read().unwrap();
9952 16 : gc_info.cutoffs.time
9953 4 : };
9954 176 : for idx in 0..10 {
9955 160 : assert_eq!(
9956 160 : tline
9957 160 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9958 160 : .await
9959 160 : .unwrap(),
9960 160 : &expected_result[idx]
9961 4 : );
9962 160 : assert_eq!(
9963 160 : tline
9964 160 : .get(get_key(idx as u32), gc_horizon, &ctx)
9965 160 : .await
9966 160 : .unwrap(),
9967 160 : &expected_result_at_gc_horizon[idx]
9968 4 : );
9969 160 : assert_eq!(
9970 160 : tline
9971 160 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
9972 160 : .await
9973 160 : .unwrap(),
9974 160 : &expected_result_at_lsn_20[idx]
9975 4 : );
9976 160 : assert_eq!(
9977 160 : tline
9978 160 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
9979 160 : .await
9980 160 : .unwrap(),
9981 160 : &expected_result_at_lsn_10[idx]
9982 4 : );
9983 4 : }
9984 32 : };
9985 4 :
9986 4 : verify_result().await;
9987 4 :
9988 4 : let cancel = CancellationToken::new();
9989 4 : let mut dryrun_flags = EnumSet::new();
9990 4 : dryrun_flags.insert(CompactFlags::DryRun);
9991 4 :
9992 4 : tline
9993 4 : .compact_with_gc(
9994 4 : &cancel,
9995 4 : CompactOptions {
9996 4 : flags: dryrun_flags,
9997 4 : ..Default::default()
9998 4 : },
9999 4 : &ctx,
10000 4 : )
10001 4 : .await
10002 4 : .unwrap();
10003 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
10004 4 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10005 4 : verify_result().await;
10006 4 :
10007 4 : tline
10008 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10009 4 : .await
10010 4 : .unwrap();
10011 4 : verify_result().await;
10012 4 :
10013 4 : // compact again
10014 4 : tline
10015 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10016 4 : .await
10017 4 : .unwrap();
10018 4 : verify_result().await;
10019 4 :
10020 4 : Ok(())
10021 4 : }
10022 :
10023 : #[cfg(feature = "testing")]
10024 : #[tokio::test]
10025 4 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
10026 4 : use models::CompactLsnRange;
10027 4 :
10028 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
10029 4 : let (tenant, ctx) = harness.load().await;
10030 4 :
10031 332 : fn get_key(id: u32) -> Key {
10032 332 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
10033 332 : key.field6 = id;
10034 332 : key
10035 332 : }
10036 4 :
10037 4 : let img_layer = (0..10)
10038 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10039 4 : .collect_vec();
10040 4 :
10041 4 : let delta1 = vec![
10042 4 : (
10043 4 : get_key(1),
10044 4 : Lsn(0x20),
10045 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10046 4 : ),
10047 4 : (
10048 4 : get_key(2),
10049 4 : Lsn(0x30),
10050 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10051 4 : ),
10052 4 : (
10053 4 : get_key(3),
10054 4 : Lsn(0x28),
10055 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10056 4 : ),
10057 4 : (
10058 4 : get_key(3),
10059 4 : Lsn(0x30),
10060 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10061 4 : ),
10062 4 : (
10063 4 : get_key(3),
10064 4 : Lsn(0x40),
10065 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10066 4 : ),
10067 4 : ];
10068 4 : let delta2 = vec![
10069 4 : (
10070 4 : get_key(5),
10071 4 : Lsn(0x20),
10072 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10073 4 : ),
10074 4 : (
10075 4 : get_key(6),
10076 4 : Lsn(0x20),
10077 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10078 4 : ),
10079 4 : ];
10080 4 : let delta3 = vec![
10081 4 : (
10082 4 : get_key(8),
10083 4 : Lsn(0x48),
10084 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10085 4 : ),
10086 4 : (
10087 4 : get_key(9),
10088 4 : Lsn(0x48),
10089 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10090 4 : ),
10091 4 : ];
10092 4 :
10093 4 : let parent_tline = tenant
10094 4 : .create_test_timeline_with_layers(
10095 4 : TIMELINE_ID,
10096 4 : Lsn(0x10),
10097 4 : DEFAULT_PG_VERSION,
10098 4 : &ctx,
10099 4 : vec![], // in-memory layers
10100 4 : vec![], // delta layers
10101 4 : vec![(Lsn(0x18), img_layer)], // image layers
10102 4 : Lsn(0x18),
10103 4 : )
10104 4 : .await?;
10105 4 :
10106 4 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10107 4 :
10108 4 : let branch_tline = tenant
10109 4 : .branch_timeline_test_with_layers(
10110 4 : &parent_tline,
10111 4 : NEW_TIMELINE_ID,
10112 4 : Some(Lsn(0x18)),
10113 4 : &ctx,
10114 4 : vec![
10115 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10116 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10117 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10118 4 : ], // delta layers
10119 4 : vec![], // image layers
10120 4 : Lsn(0x50),
10121 4 : )
10122 4 : .await?;
10123 4 :
10124 4 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10125 4 :
10126 4 : {
10127 4 : parent_tline
10128 4 : .applied_gc_cutoff_lsn
10129 4 : .lock_for_write()
10130 4 : .store_and_unlock(Lsn(0x10))
10131 4 : .wait()
10132 4 : .await;
10133 4 : // Update GC info
10134 4 : let mut guard = parent_tline.gc_info.write().unwrap();
10135 4 : *guard = GcInfo {
10136 4 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
10137 4 : cutoffs: GcCutoffs {
10138 4 : time: Lsn(0x10),
10139 4 : space: Lsn(0x10),
10140 4 : },
10141 4 : leases: Default::default(),
10142 4 : within_ancestor_pitr: false,
10143 4 : };
10144 4 : }
10145 4 :
10146 4 : {
10147 4 : branch_tline
10148 4 : .applied_gc_cutoff_lsn
10149 4 : .lock_for_write()
10150 4 : .store_and_unlock(Lsn(0x50))
10151 4 : .wait()
10152 4 : .await;
10153 4 : // Update GC info
10154 4 : let mut guard = branch_tline.gc_info.write().unwrap();
10155 4 : *guard = GcInfo {
10156 4 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
10157 4 : cutoffs: GcCutoffs {
10158 4 : time: Lsn(0x50),
10159 4 : space: Lsn(0x50),
10160 4 : },
10161 4 : leases: Default::default(),
10162 4 : within_ancestor_pitr: false,
10163 4 : };
10164 4 : }
10165 4 :
10166 4 : let expected_result_at_gc_horizon = [
10167 4 : Bytes::from_static(b"value 0@0x10"),
10168 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10169 4 : Bytes::from_static(b"value 2@0x10@0x30"),
10170 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10171 4 : Bytes::from_static(b"value 4@0x10"),
10172 4 : Bytes::from_static(b"value 5@0x10@0x20"),
10173 4 : Bytes::from_static(b"value 6@0x10@0x20"),
10174 4 : Bytes::from_static(b"value 7@0x10"),
10175 4 : Bytes::from_static(b"value 8@0x10@0x48"),
10176 4 : Bytes::from_static(b"value 9@0x10@0x48"),
10177 4 : ];
10178 4 :
10179 4 : let expected_result_at_lsn_40 = [
10180 4 : Bytes::from_static(b"value 0@0x10"),
10181 4 : Bytes::from_static(b"value 1@0x10@0x20"),
10182 4 : Bytes::from_static(b"value 2@0x10@0x30"),
10183 4 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10184 4 : Bytes::from_static(b"value 4@0x10"),
10185 4 : Bytes::from_static(b"value 5@0x10@0x20"),
10186 4 : Bytes::from_static(b"value 6@0x10@0x20"),
10187 4 : Bytes::from_static(b"value 7@0x10"),
10188 4 : Bytes::from_static(b"value 8@0x10"),
10189 4 : Bytes::from_static(b"value 9@0x10"),
10190 4 : ];
10191 4 :
10192 12 : let verify_result = || async {
10193 132 : for idx in 0..10 {
10194 120 : assert_eq!(
10195 120 : branch_tline
10196 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10197 120 : .await
10198 120 : .unwrap(),
10199 120 : &expected_result_at_gc_horizon[idx]
10200 4 : );
10201 120 : assert_eq!(
10202 120 : branch_tline
10203 120 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
10204 120 : .await
10205 120 : .unwrap(),
10206 120 : &expected_result_at_lsn_40[idx]
10207 4 : );
10208 4 : }
10209 24 : };
10210 4 :
10211 4 : verify_result().await;
10212 4 :
10213 4 : let cancel = CancellationToken::new();
10214 4 : branch_tline
10215 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10216 4 : .await
10217 4 : .unwrap();
10218 4 :
10219 4 : verify_result().await;
10220 4 :
10221 4 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
10222 4 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
10223 4 : branch_tline
10224 4 : .compact_with_gc(
10225 4 : &cancel,
10226 4 : CompactOptions {
10227 4 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
10228 4 : ..Default::default()
10229 4 : },
10230 4 : &ctx,
10231 4 : )
10232 4 : .await
10233 4 : .unwrap();
10234 4 :
10235 4 : verify_result().await;
10236 4 :
10237 4 : Ok(())
10238 4 : }
10239 :
10240 : // Regression test for https://github.com/neondatabase/neon/issues/9012
10241 : // Create an image arrangement where we have to read at different LSN ranges
10242 : // from a delta layer. This is achieved by overlapping an image layer on top of
10243 : // a delta layer. Like so:
10244 : //
10245 : // A B
10246 : // +----------------+ -> delta_layer
10247 : // | | ^ lsn
10248 : // | =========|-> nested_image_layer |
10249 : // | C | |
10250 : // +----------------+ |
10251 : // ======== -> baseline_image_layer +-------> key
10252 : //
10253 : //
10254 : // When querying the key range [A, B) we need to read at different LSN ranges
10255 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
10256 : #[cfg(feature = "testing")]
10257 : #[tokio::test]
10258 4 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
10259 4 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
10260 4 : let (tenant, ctx) = harness.load().await;
10261 4 :
10262 4 : let will_init_keys = [2, 6];
10263 88 : fn get_key(id: u32) -> Key {
10264 88 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10265 88 : key.field6 = id;
10266 88 : key
10267 88 : }
10268 4 :
10269 4 : let mut expected_key_values = HashMap::new();
10270 4 :
10271 4 : let baseline_image_layer_lsn = Lsn(0x10);
10272 4 : let mut baseline_img_layer = Vec::new();
10273 24 : for i in 0..5 {
10274 20 : let key = get_key(i);
10275 20 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10276 20 :
10277 20 : let removed = expected_key_values.insert(key, value.clone());
10278 20 : assert!(removed.is_none());
10279 4 :
10280 20 : baseline_img_layer.push((key, Bytes::from(value)));
10281 4 : }
10282 4 :
10283 4 : let nested_image_layer_lsn = Lsn(0x50);
10284 4 : let mut nested_img_layer = Vec::new();
10285 24 : for i in 5..10 {
10286 20 : let key = get_key(i);
10287 20 : let value = format!("value {i}@{nested_image_layer_lsn}");
10288 20 :
10289 20 : let removed = expected_key_values.insert(key, value.clone());
10290 20 : assert!(removed.is_none());
10291 4 :
10292 20 : nested_img_layer.push((key, Bytes::from(value)));
10293 4 : }
10294 4 :
10295 4 : let mut delta_layer_spec = Vec::default();
10296 4 : let delta_layer_start_lsn = Lsn(0x20);
10297 4 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10298 4 :
10299 44 : for i in 0..10 {
10300 40 : let key = get_key(i);
10301 40 : let key_in_nested = nested_img_layer
10302 40 : .iter()
10303 160 : .any(|(key_with_img, _)| *key_with_img == key);
10304 40 : let lsn = {
10305 40 : if key_in_nested {
10306 20 : Lsn(nested_image_layer_lsn.0 + 0x10)
10307 4 : } else {
10308 20 : delta_layer_start_lsn
10309 4 : }
10310 4 : };
10311 4 :
10312 40 : let will_init = will_init_keys.contains(&i);
10313 40 : if will_init {
10314 8 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10315 8 :
10316 8 : expected_key_values.insert(key, "".to_string());
10317 32 : } else {
10318 32 : let delta = format!("@{lsn}");
10319 32 : delta_layer_spec.push((
10320 32 : key,
10321 32 : lsn,
10322 32 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10323 32 : ));
10324 32 :
10325 32 : expected_key_values
10326 32 : .get_mut(&key)
10327 32 : .expect("An image exists for each key")
10328 32 : .push_str(delta.as_str());
10329 32 : }
10330 40 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
10331 4 : }
10332 4 :
10333 4 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
10334 4 :
10335 4 : assert!(
10336 4 : nested_image_layer_lsn > delta_layer_start_lsn
10337 4 : && nested_image_layer_lsn < delta_layer_end_lsn
10338 4 : );
10339 4 :
10340 4 : let tline = tenant
10341 4 : .create_test_timeline_with_layers(
10342 4 : TIMELINE_ID,
10343 4 : baseline_image_layer_lsn,
10344 4 : DEFAULT_PG_VERSION,
10345 4 : &ctx,
10346 4 : vec![], // in-memory layers
10347 4 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
10348 4 : delta_layer_start_lsn..delta_layer_end_lsn,
10349 4 : delta_layer_spec,
10350 4 : )], // delta layers
10351 4 : vec![
10352 4 : (baseline_image_layer_lsn, baseline_img_layer),
10353 4 : (nested_image_layer_lsn, nested_img_layer),
10354 4 : ], // image layers
10355 4 : delta_layer_end_lsn,
10356 4 : )
10357 4 : .await?;
10358 4 :
10359 4 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
10360 4 : let results = tline
10361 4 : .get_vectored(
10362 4 : keyspace,
10363 4 : delta_layer_end_lsn,
10364 4 : IoConcurrency::sequential(),
10365 4 : &ctx,
10366 4 : )
10367 4 : .await
10368 4 : .expect("No vectored errors");
10369 44 : for (key, res) in results {
10370 40 : let value = res.expect("No key errors");
10371 40 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10372 40 : assert_eq!(value, Bytes::from(expected_value));
10373 4 : }
10374 4 :
10375 4 : Ok(())
10376 4 : }
10377 :
10378 : #[cfg(feature = "testing")]
10379 : #[tokio::test]
10380 4 : async fn test_vectored_read_with_image_layer_inside_inmem() -> anyhow::Result<()> {
10381 4 : let harness =
10382 4 : TenantHarness::create("test_vectored_read_with_image_layer_inside_inmem").await?;
10383 4 : let (tenant, ctx) = harness.load().await;
10384 4 :
10385 4 : let will_init_keys = [2, 6];
10386 128 : fn get_key(id: u32) -> Key {
10387 128 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10388 128 : key.field6 = id;
10389 128 : key
10390 128 : }
10391 4 :
10392 4 : let mut expected_key_values = HashMap::new();
10393 4 :
10394 4 : let baseline_image_layer_lsn = Lsn(0x10);
10395 4 : let mut baseline_img_layer = Vec::new();
10396 24 : for i in 0..5 {
10397 20 : let key = get_key(i);
10398 20 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10399 20 :
10400 20 : let removed = expected_key_values.insert(key, value.clone());
10401 20 : assert!(removed.is_none());
10402 4 :
10403 20 : baseline_img_layer.push((key, Bytes::from(value)));
10404 4 : }
10405 4 :
10406 4 : let nested_image_layer_lsn = Lsn(0x50);
10407 4 : let mut nested_img_layer = Vec::new();
10408 24 : for i in 5..10 {
10409 20 : let key = get_key(i);
10410 20 : let value = format!("value {i}@{nested_image_layer_lsn}");
10411 20 :
10412 20 : let removed = expected_key_values.insert(key, value.clone());
10413 20 : assert!(removed.is_none());
10414 4 :
10415 20 : nested_img_layer.push((key, Bytes::from(value)));
10416 4 : }
10417 4 :
10418 4 : let frozen_layer = {
10419 4 : let lsn_range = Lsn(0x40)..Lsn(0x60);
10420 4 : let mut data = Vec::new();
10421 44 : for i in 0..10 {
10422 40 : let key = get_key(i);
10423 40 : let key_in_nested = nested_img_layer
10424 40 : .iter()
10425 160 : .any(|(key_with_img, _)| *key_with_img == key);
10426 40 : let lsn = {
10427 40 : if key_in_nested {
10428 20 : Lsn(nested_image_layer_lsn.0 + 5)
10429 4 : } else {
10430 20 : lsn_range.start
10431 4 : }
10432 4 : };
10433 4 :
10434 40 : let will_init = will_init_keys.contains(&i);
10435 40 : if will_init {
10436 8 : data.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10437 8 :
10438 8 : expected_key_values.insert(key, "".to_string());
10439 32 : } else {
10440 32 : let delta = format!("@{lsn}");
10441 32 : data.push((
10442 32 : key,
10443 32 : lsn,
10444 32 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10445 32 : ));
10446 32 :
10447 32 : expected_key_values
10448 32 : .get_mut(&key)
10449 32 : .expect("An image exists for each key")
10450 32 : .push_str(delta.as_str());
10451 32 : }
10452 4 : }
10453 4 :
10454 4 : InMemoryLayerTestDesc {
10455 4 : lsn_range,
10456 4 : is_open: false,
10457 4 : data,
10458 4 : }
10459 4 : };
10460 4 :
10461 4 : let (open_layer, last_record_lsn) = {
10462 4 : let start_lsn = Lsn(0x70);
10463 4 : let mut data = Vec::new();
10464 4 : let mut end_lsn = Lsn(0);
10465 44 : for i in 0..10 {
10466 40 : let key = get_key(i);
10467 40 : let lsn = Lsn(start_lsn.0 + i as u64);
10468 40 : let delta = format!("@{lsn}");
10469 40 : data.push((
10470 40 : key,
10471 40 : lsn,
10472 40 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10473 40 : ));
10474 40 :
10475 40 : expected_key_values
10476 40 : .get_mut(&key)
10477 40 : .expect("An image exists for each key")
10478 40 : .push_str(delta.as_str());
10479 40 :
10480 40 : end_lsn = std::cmp::max(end_lsn, lsn);
10481 40 : }
10482 4 :
10483 4 : (
10484 4 : InMemoryLayerTestDesc {
10485 4 : lsn_range: start_lsn..Lsn::MAX,
10486 4 : is_open: true,
10487 4 : data,
10488 4 : },
10489 4 : end_lsn,
10490 4 : )
10491 4 : };
10492 4 :
10493 4 : assert!(
10494 4 : nested_image_layer_lsn > frozen_layer.lsn_range.start
10495 4 : && nested_image_layer_lsn < frozen_layer.lsn_range.end
10496 4 : );
10497 4 :
10498 4 : let tline = tenant
10499 4 : .create_test_timeline_with_layers(
10500 4 : TIMELINE_ID,
10501 4 : baseline_image_layer_lsn,
10502 4 : DEFAULT_PG_VERSION,
10503 4 : &ctx,
10504 4 : vec![open_layer, frozen_layer], // in-memory layers
10505 4 : Vec::new(), // delta layers
10506 4 : vec![
10507 4 : (baseline_image_layer_lsn, baseline_img_layer),
10508 4 : (nested_image_layer_lsn, nested_img_layer),
10509 4 : ], // image layers
10510 4 : last_record_lsn,
10511 4 : )
10512 4 : .await?;
10513 4 :
10514 4 : let keyspace = KeySpace::single(get_key(0)..get_key(10));
10515 4 : let results = tline
10516 4 : .get_vectored(keyspace, last_record_lsn, IoConcurrency::sequential(), &ctx)
10517 4 : .await
10518 4 : .expect("No vectored errors");
10519 44 : for (key, res) in results {
10520 40 : let value = res.expect("No key errors");
10521 40 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
10522 40 : assert_eq!(value, Bytes::from(expected_value.clone()));
10523 4 :
10524 40 : tracing::info!("key={key} value={expected_value}");
10525 4 : }
10526 4 :
10527 4 : Ok(())
10528 4 : }
10529 :
10530 428 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
10531 428 : (
10532 428 : k1.is_delta,
10533 428 : k1.key_range.start,
10534 428 : k1.key_range.end,
10535 428 : k1.lsn_range.start,
10536 428 : k1.lsn_range.end,
10537 428 : )
10538 428 : .cmp(&(
10539 428 : k2.is_delta,
10540 428 : k2.key_range.start,
10541 428 : k2.key_range.end,
10542 428 : k2.lsn_range.start,
10543 428 : k2.lsn_range.end,
10544 428 : ))
10545 428 : }
10546 :
10547 48 : async fn inspect_and_sort(
10548 48 : tline: &Arc<Timeline>,
10549 48 : filter: Option<std::ops::Range<Key>>,
10550 48 : ) -> Vec<PersistentLayerKey> {
10551 48 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
10552 48 : if let Some(filter) = filter {
10553 216 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
10554 44 : }
10555 48 : all_layers.sort_by(sort_layer_key);
10556 48 : all_layers
10557 48 : }
10558 :
10559 : #[cfg(feature = "testing")]
10560 44 : fn check_layer_map_key_eq(
10561 44 : mut left: Vec<PersistentLayerKey>,
10562 44 : mut right: Vec<PersistentLayerKey>,
10563 44 : ) {
10564 44 : left.sort_by(sort_layer_key);
10565 44 : right.sort_by(sort_layer_key);
10566 44 : if left != right {
10567 0 : eprintln!("---LEFT---");
10568 0 : for left in left.iter() {
10569 0 : eprintln!("{}", left);
10570 0 : }
10571 0 : eprintln!("---RIGHT---");
10572 0 : for right in right.iter() {
10573 0 : eprintln!("{}", right);
10574 0 : }
10575 0 : assert_eq!(left, right);
10576 44 : }
10577 44 : }
10578 :
10579 : #[cfg(feature = "testing")]
10580 : #[tokio::test]
10581 4 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
10582 4 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
10583 4 : let (tenant, ctx) = harness.load().await;
10584 4 :
10585 364 : fn get_key(id: u32) -> Key {
10586 364 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10587 364 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10588 364 : key.field6 = id;
10589 364 : key
10590 364 : }
10591 4 :
10592 4 : // img layer at 0x10
10593 4 : let img_layer = (0..10)
10594 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10595 4 : .collect_vec();
10596 4 :
10597 4 : let delta1 = vec![
10598 4 : (
10599 4 : get_key(1),
10600 4 : Lsn(0x20),
10601 4 : Value::Image(Bytes::from("value 1@0x20")),
10602 4 : ),
10603 4 : (
10604 4 : get_key(2),
10605 4 : Lsn(0x30),
10606 4 : Value::Image(Bytes::from("value 2@0x30")),
10607 4 : ),
10608 4 : (
10609 4 : get_key(3),
10610 4 : Lsn(0x40),
10611 4 : Value::Image(Bytes::from("value 3@0x40")),
10612 4 : ),
10613 4 : ];
10614 4 : let delta2 = vec![
10615 4 : (
10616 4 : get_key(5),
10617 4 : Lsn(0x20),
10618 4 : Value::Image(Bytes::from("value 5@0x20")),
10619 4 : ),
10620 4 : (
10621 4 : get_key(6),
10622 4 : Lsn(0x20),
10623 4 : Value::Image(Bytes::from("value 6@0x20")),
10624 4 : ),
10625 4 : ];
10626 4 : let delta3 = vec![
10627 4 : (
10628 4 : get_key(8),
10629 4 : Lsn(0x48),
10630 4 : Value::Image(Bytes::from("value 8@0x48")),
10631 4 : ),
10632 4 : (
10633 4 : get_key(9),
10634 4 : Lsn(0x48),
10635 4 : Value::Image(Bytes::from("value 9@0x48")),
10636 4 : ),
10637 4 : ];
10638 4 :
10639 4 : let tline = tenant
10640 4 : .create_test_timeline_with_layers(
10641 4 : TIMELINE_ID,
10642 4 : Lsn(0x10),
10643 4 : DEFAULT_PG_VERSION,
10644 4 : &ctx,
10645 4 : vec![], // in-memory layers
10646 4 : vec![
10647 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10648 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10649 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10650 4 : ], // delta layers
10651 4 : vec![(Lsn(0x10), img_layer)], // image layers
10652 4 : Lsn(0x50),
10653 4 : )
10654 4 : .await?;
10655 4 :
10656 4 : {
10657 4 : tline
10658 4 : .applied_gc_cutoff_lsn
10659 4 : .lock_for_write()
10660 4 : .store_and_unlock(Lsn(0x30))
10661 4 : .wait()
10662 4 : .await;
10663 4 : // Update GC info
10664 4 : let mut guard = tline.gc_info.write().unwrap();
10665 4 : *guard = GcInfo {
10666 4 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
10667 4 : cutoffs: GcCutoffs {
10668 4 : time: Lsn(0x30),
10669 4 : space: Lsn(0x30),
10670 4 : },
10671 4 : leases: Default::default(),
10672 4 : within_ancestor_pitr: false,
10673 4 : };
10674 4 : }
10675 4 :
10676 4 : let cancel = CancellationToken::new();
10677 4 :
10678 4 : // Do a partial compaction on key range 0..2
10679 4 : tline
10680 4 : .compact_with_gc(
10681 4 : &cancel,
10682 4 : CompactOptions {
10683 4 : flags: EnumSet::new(),
10684 4 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
10685 4 : ..Default::default()
10686 4 : },
10687 4 : &ctx,
10688 4 : )
10689 4 : .await
10690 4 : .unwrap();
10691 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10692 4 : check_layer_map_key_eq(
10693 4 : all_layers,
10694 4 : vec![
10695 4 : // newly-generated image layer for the partial compaction range 0-2
10696 4 : PersistentLayerKey {
10697 4 : key_range: get_key(0)..get_key(2),
10698 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10699 4 : is_delta: false,
10700 4 : },
10701 4 : PersistentLayerKey {
10702 4 : key_range: get_key(0)..get_key(10),
10703 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10704 4 : is_delta: false,
10705 4 : },
10706 4 : // delta1 is split and the second part is rewritten
10707 4 : PersistentLayerKey {
10708 4 : key_range: get_key(2)..get_key(4),
10709 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10710 4 : is_delta: true,
10711 4 : },
10712 4 : PersistentLayerKey {
10713 4 : key_range: get_key(5)..get_key(7),
10714 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10715 4 : is_delta: true,
10716 4 : },
10717 4 : PersistentLayerKey {
10718 4 : key_range: get_key(8)..get_key(10),
10719 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10720 4 : is_delta: true,
10721 4 : },
10722 4 : ],
10723 4 : );
10724 4 :
10725 4 : // Do a partial compaction on key range 2..4
10726 4 : tline
10727 4 : .compact_with_gc(
10728 4 : &cancel,
10729 4 : CompactOptions {
10730 4 : flags: EnumSet::new(),
10731 4 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
10732 4 : ..Default::default()
10733 4 : },
10734 4 : &ctx,
10735 4 : )
10736 4 : .await
10737 4 : .unwrap();
10738 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10739 4 : check_layer_map_key_eq(
10740 4 : all_layers,
10741 4 : vec![
10742 4 : PersistentLayerKey {
10743 4 : key_range: get_key(0)..get_key(2),
10744 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10745 4 : is_delta: false,
10746 4 : },
10747 4 : PersistentLayerKey {
10748 4 : key_range: get_key(0)..get_key(10),
10749 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10750 4 : is_delta: false,
10751 4 : },
10752 4 : // image layer generated for the compaction range 2-4
10753 4 : PersistentLayerKey {
10754 4 : key_range: get_key(2)..get_key(4),
10755 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10756 4 : is_delta: false,
10757 4 : },
10758 4 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
10759 4 : PersistentLayerKey {
10760 4 : key_range: get_key(2)..get_key(4),
10761 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10762 4 : is_delta: true,
10763 4 : },
10764 4 : PersistentLayerKey {
10765 4 : key_range: get_key(5)..get_key(7),
10766 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10767 4 : is_delta: true,
10768 4 : },
10769 4 : PersistentLayerKey {
10770 4 : key_range: get_key(8)..get_key(10),
10771 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10772 4 : is_delta: true,
10773 4 : },
10774 4 : ],
10775 4 : );
10776 4 :
10777 4 : // Do a partial compaction on key range 4..9
10778 4 : tline
10779 4 : .compact_with_gc(
10780 4 : &cancel,
10781 4 : CompactOptions {
10782 4 : flags: EnumSet::new(),
10783 4 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
10784 4 : ..Default::default()
10785 4 : },
10786 4 : &ctx,
10787 4 : )
10788 4 : .await
10789 4 : .unwrap();
10790 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10791 4 : check_layer_map_key_eq(
10792 4 : all_layers,
10793 4 : vec![
10794 4 : PersistentLayerKey {
10795 4 : key_range: get_key(0)..get_key(2),
10796 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10797 4 : is_delta: false,
10798 4 : },
10799 4 : PersistentLayerKey {
10800 4 : key_range: get_key(0)..get_key(10),
10801 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10802 4 : is_delta: false,
10803 4 : },
10804 4 : PersistentLayerKey {
10805 4 : key_range: get_key(2)..get_key(4),
10806 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10807 4 : is_delta: false,
10808 4 : },
10809 4 : PersistentLayerKey {
10810 4 : key_range: get_key(2)..get_key(4),
10811 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10812 4 : is_delta: true,
10813 4 : },
10814 4 : // image layer generated for this compaction range
10815 4 : PersistentLayerKey {
10816 4 : key_range: get_key(4)..get_key(9),
10817 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10818 4 : is_delta: false,
10819 4 : },
10820 4 : PersistentLayerKey {
10821 4 : key_range: get_key(8)..get_key(10),
10822 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10823 4 : is_delta: true,
10824 4 : },
10825 4 : ],
10826 4 : );
10827 4 :
10828 4 : // Do a partial compaction on key range 9..10
10829 4 : tline
10830 4 : .compact_with_gc(
10831 4 : &cancel,
10832 4 : CompactOptions {
10833 4 : flags: EnumSet::new(),
10834 4 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
10835 4 : ..Default::default()
10836 4 : },
10837 4 : &ctx,
10838 4 : )
10839 4 : .await
10840 4 : .unwrap();
10841 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10842 4 : check_layer_map_key_eq(
10843 4 : all_layers,
10844 4 : vec![
10845 4 : PersistentLayerKey {
10846 4 : key_range: get_key(0)..get_key(2),
10847 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10848 4 : is_delta: false,
10849 4 : },
10850 4 : PersistentLayerKey {
10851 4 : key_range: get_key(0)..get_key(10),
10852 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
10853 4 : is_delta: false,
10854 4 : },
10855 4 : PersistentLayerKey {
10856 4 : key_range: get_key(2)..get_key(4),
10857 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10858 4 : is_delta: false,
10859 4 : },
10860 4 : PersistentLayerKey {
10861 4 : key_range: get_key(2)..get_key(4),
10862 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10863 4 : is_delta: true,
10864 4 : },
10865 4 : PersistentLayerKey {
10866 4 : key_range: get_key(4)..get_key(9),
10867 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10868 4 : is_delta: false,
10869 4 : },
10870 4 : // image layer generated for the compaction range
10871 4 : PersistentLayerKey {
10872 4 : key_range: get_key(9)..get_key(10),
10873 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10874 4 : is_delta: false,
10875 4 : },
10876 4 : PersistentLayerKey {
10877 4 : key_range: get_key(8)..get_key(10),
10878 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10879 4 : is_delta: true,
10880 4 : },
10881 4 : ],
10882 4 : );
10883 4 :
10884 4 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
10885 4 : tline
10886 4 : .compact_with_gc(
10887 4 : &cancel,
10888 4 : CompactOptions {
10889 4 : flags: EnumSet::new(),
10890 4 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
10891 4 : ..Default::default()
10892 4 : },
10893 4 : &ctx,
10894 4 : )
10895 4 : .await
10896 4 : .unwrap();
10897 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
10898 4 : check_layer_map_key_eq(
10899 4 : all_layers,
10900 4 : vec![
10901 4 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
10902 4 : PersistentLayerKey {
10903 4 : key_range: get_key(0)..get_key(10),
10904 4 : lsn_range: Lsn(0x20)..Lsn(0x21),
10905 4 : is_delta: false,
10906 4 : },
10907 4 : PersistentLayerKey {
10908 4 : key_range: get_key(2)..get_key(4),
10909 4 : lsn_range: Lsn(0x20)..Lsn(0x48),
10910 4 : is_delta: true,
10911 4 : },
10912 4 : PersistentLayerKey {
10913 4 : key_range: get_key(8)..get_key(10),
10914 4 : lsn_range: Lsn(0x48)..Lsn(0x50),
10915 4 : is_delta: true,
10916 4 : },
10917 4 : ],
10918 4 : );
10919 4 : Ok(())
10920 4 : }
10921 :
10922 : #[cfg(feature = "testing")]
10923 : #[tokio::test]
10924 4 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
10925 4 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
10926 4 : .await
10927 4 : .unwrap();
10928 4 : let (tenant, ctx) = harness.load().await;
10929 4 : let tline_parent = tenant
10930 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
10931 4 : .await
10932 4 : .unwrap();
10933 4 : let tline_child = tenant
10934 4 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
10935 4 : .await
10936 4 : .unwrap();
10937 4 : {
10938 4 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10939 4 : assert_eq!(
10940 4 : gc_info_parent.retain_lsns,
10941 4 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
10942 4 : );
10943 4 : }
10944 4 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
10945 4 : tline_child
10946 4 : .remote_client
10947 4 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
10948 4 : .unwrap();
10949 4 : tline_child.remote_client.wait_completion().await.unwrap();
10950 4 : offload_timeline(&tenant, &tline_child)
10951 4 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
10952 4 : .await.unwrap();
10953 4 : let child_timeline_id = tline_child.timeline_id;
10954 4 : Arc::try_unwrap(tline_child).unwrap();
10955 4 :
10956 4 : {
10957 4 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
10958 4 : assert_eq!(
10959 4 : gc_info_parent.retain_lsns,
10960 4 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
10961 4 : );
10962 4 : }
10963 4 :
10964 4 : tenant
10965 4 : .get_offloaded_timeline(child_timeline_id)
10966 4 : .unwrap()
10967 4 : .defuse_for_tenant_drop();
10968 4 :
10969 4 : Ok(())
10970 4 : }
10971 :
10972 : #[cfg(feature = "testing")]
10973 : #[tokio::test]
10974 4 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
10975 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
10976 4 : let (tenant, ctx) = harness.load().await;
10977 4 :
10978 592 : fn get_key(id: u32) -> Key {
10979 592 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10980 592 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10981 592 : key.field6 = id;
10982 592 : key
10983 592 : }
10984 4 :
10985 4 : let img_layer = (0..10)
10986 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10987 4 : .collect_vec();
10988 4 :
10989 4 : let delta1 = vec![(
10990 4 : get_key(1),
10991 4 : Lsn(0x20),
10992 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10993 4 : )];
10994 4 : let delta4 = vec![(
10995 4 : get_key(1),
10996 4 : Lsn(0x28),
10997 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10998 4 : )];
10999 4 : let delta2 = vec![
11000 4 : (
11001 4 : get_key(1),
11002 4 : Lsn(0x30),
11003 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11004 4 : ),
11005 4 : (
11006 4 : get_key(1),
11007 4 : Lsn(0x38),
11008 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11009 4 : ),
11010 4 : ];
11011 4 : let delta3 = vec![
11012 4 : (
11013 4 : get_key(8),
11014 4 : Lsn(0x48),
11015 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11016 4 : ),
11017 4 : (
11018 4 : get_key(9),
11019 4 : Lsn(0x48),
11020 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11021 4 : ),
11022 4 : ];
11023 4 :
11024 4 : let tline = tenant
11025 4 : .create_test_timeline_with_layers(
11026 4 : TIMELINE_ID,
11027 4 : Lsn(0x10),
11028 4 : DEFAULT_PG_VERSION,
11029 4 : &ctx,
11030 4 : vec![], // in-memory layers
11031 4 : vec![
11032 4 : // delta1/2/4 only contain a single key but multiple updates
11033 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11034 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11035 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11036 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11037 4 : ], // delta layers
11038 4 : vec![(Lsn(0x10), img_layer)], // image layers
11039 4 : Lsn(0x50),
11040 4 : )
11041 4 : .await?;
11042 4 : {
11043 4 : tline
11044 4 : .applied_gc_cutoff_lsn
11045 4 : .lock_for_write()
11046 4 : .store_and_unlock(Lsn(0x30))
11047 4 : .wait()
11048 4 : .await;
11049 4 : // Update GC info
11050 4 : let mut guard = tline.gc_info.write().unwrap();
11051 4 : *guard = GcInfo {
11052 4 : retain_lsns: vec![
11053 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11054 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11055 4 : ],
11056 4 : cutoffs: GcCutoffs {
11057 4 : time: Lsn(0x30),
11058 4 : space: Lsn(0x30),
11059 4 : },
11060 4 : leases: Default::default(),
11061 4 : within_ancestor_pitr: false,
11062 4 : };
11063 4 : }
11064 4 :
11065 4 : let expected_result = [
11066 4 : Bytes::from_static(b"value 0@0x10"),
11067 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11068 4 : Bytes::from_static(b"value 2@0x10"),
11069 4 : Bytes::from_static(b"value 3@0x10"),
11070 4 : Bytes::from_static(b"value 4@0x10"),
11071 4 : Bytes::from_static(b"value 5@0x10"),
11072 4 : Bytes::from_static(b"value 6@0x10"),
11073 4 : Bytes::from_static(b"value 7@0x10"),
11074 4 : Bytes::from_static(b"value 8@0x10@0x48"),
11075 4 : Bytes::from_static(b"value 9@0x10@0x48"),
11076 4 : ];
11077 4 :
11078 4 : let expected_result_at_gc_horizon = [
11079 4 : Bytes::from_static(b"value 0@0x10"),
11080 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11081 4 : Bytes::from_static(b"value 2@0x10"),
11082 4 : Bytes::from_static(b"value 3@0x10"),
11083 4 : Bytes::from_static(b"value 4@0x10"),
11084 4 : Bytes::from_static(b"value 5@0x10"),
11085 4 : Bytes::from_static(b"value 6@0x10"),
11086 4 : Bytes::from_static(b"value 7@0x10"),
11087 4 : Bytes::from_static(b"value 8@0x10"),
11088 4 : Bytes::from_static(b"value 9@0x10"),
11089 4 : ];
11090 4 :
11091 4 : let expected_result_at_lsn_20 = [
11092 4 : Bytes::from_static(b"value 0@0x10"),
11093 4 : Bytes::from_static(b"value 1@0x10@0x20"),
11094 4 : Bytes::from_static(b"value 2@0x10"),
11095 4 : Bytes::from_static(b"value 3@0x10"),
11096 4 : Bytes::from_static(b"value 4@0x10"),
11097 4 : Bytes::from_static(b"value 5@0x10"),
11098 4 : Bytes::from_static(b"value 6@0x10"),
11099 4 : Bytes::from_static(b"value 7@0x10"),
11100 4 : Bytes::from_static(b"value 8@0x10"),
11101 4 : Bytes::from_static(b"value 9@0x10"),
11102 4 : ];
11103 4 :
11104 4 : let expected_result_at_lsn_10 = [
11105 4 : Bytes::from_static(b"value 0@0x10"),
11106 4 : Bytes::from_static(b"value 1@0x10"),
11107 4 : Bytes::from_static(b"value 2@0x10"),
11108 4 : Bytes::from_static(b"value 3@0x10"),
11109 4 : Bytes::from_static(b"value 4@0x10"),
11110 4 : Bytes::from_static(b"value 5@0x10"),
11111 4 : Bytes::from_static(b"value 6@0x10"),
11112 4 : Bytes::from_static(b"value 7@0x10"),
11113 4 : Bytes::from_static(b"value 8@0x10"),
11114 4 : Bytes::from_static(b"value 9@0x10"),
11115 4 : ];
11116 4 :
11117 12 : let verify_result = || async {
11118 12 : let gc_horizon = {
11119 12 : let gc_info = tline.gc_info.read().unwrap();
11120 12 : gc_info.cutoffs.time
11121 4 : };
11122 132 : for idx in 0..10 {
11123 120 : assert_eq!(
11124 120 : tline
11125 120 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
11126 120 : .await
11127 120 : .unwrap(),
11128 120 : &expected_result[idx]
11129 4 : );
11130 120 : assert_eq!(
11131 120 : tline
11132 120 : .get(get_key(idx as u32), gc_horizon, &ctx)
11133 120 : .await
11134 120 : .unwrap(),
11135 120 : &expected_result_at_gc_horizon[idx]
11136 4 : );
11137 120 : assert_eq!(
11138 120 : tline
11139 120 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
11140 120 : .await
11141 120 : .unwrap(),
11142 120 : &expected_result_at_lsn_20[idx]
11143 4 : );
11144 120 : assert_eq!(
11145 120 : tline
11146 120 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
11147 120 : .await
11148 120 : .unwrap(),
11149 120 : &expected_result_at_lsn_10[idx]
11150 4 : );
11151 4 : }
11152 24 : };
11153 4 :
11154 4 : verify_result().await;
11155 4 :
11156 4 : let cancel = CancellationToken::new();
11157 4 : tline
11158 4 : .compact_with_gc(
11159 4 : &cancel,
11160 4 : CompactOptions {
11161 4 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
11162 4 : ..Default::default()
11163 4 : },
11164 4 : &ctx,
11165 4 : )
11166 4 : .await
11167 4 : .unwrap();
11168 4 : verify_result().await;
11169 4 :
11170 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11171 4 : check_layer_map_key_eq(
11172 4 : all_layers,
11173 4 : vec![
11174 4 : // The original image layer, not compacted
11175 4 : PersistentLayerKey {
11176 4 : key_range: get_key(0)..get_key(10),
11177 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11178 4 : is_delta: false,
11179 4 : },
11180 4 : // Delta layer below the specified above_lsn not compacted
11181 4 : PersistentLayerKey {
11182 4 : key_range: get_key(1)..get_key(2),
11183 4 : lsn_range: Lsn(0x20)..Lsn(0x28),
11184 4 : is_delta: true,
11185 4 : },
11186 4 : // Delta layer compacted above the LSN
11187 4 : PersistentLayerKey {
11188 4 : key_range: get_key(1)..get_key(10),
11189 4 : lsn_range: Lsn(0x28)..Lsn(0x50),
11190 4 : is_delta: true,
11191 4 : },
11192 4 : ],
11193 4 : );
11194 4 :
11195 4 : // compact again
11196 4 : tline
11197 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11198 4 : .await
11199 4 : .unwrap();
11200 4 : verify_result().await;
11201 4 :
11202 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11203 4 : check_layer_map_key_eq(
11204 4 : all_layers,
11205 4 : vec![
11206 4 : // The compacted image layer (full key range)
11207 4 : PersistentLayerKey {
11208 4 : key_range: Key::MIN..Key::MAX,
11209 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11210 4 : is_delta: false,
11211 4 : },
11212 4 : // All other data in the delta layer
11213 4 : PersistentLayerKey {
11214 4 : key_range: get_key(1)..get_key(10),
11215 4 : lsn_range: Lsn(0x10)..Lsn(0x50),
11216 4 : is_delta: true,
11217 4 : },
11218 4 : ],
11219 4 : );
11220 4 :
11221 4 : Ok(())
11222 4 : }
11223 :
11224 : #[cfg(feature = "testing")]
11225 : #[tokio::test]
11226 4 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
11227 4 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
11228 4 : let (tenant, ctx) = harness.load().await;
11229 4 :
11230 1016 : fn get_key(id: u32) -> Key {
11231 1016 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11232 1016 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11233 1016 : key.field6 = id;
11234 1016 : key
11235 1016 : }
11236 4 :
11237 4 : let img_layer = (0..10)
11238 40 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11239 4 : .collect_vec();
11240 4 :
11241 4 : let delta1 = vec![(
11242 4 : get_key(1),
11243 4 : Lsn(0x20),
11244 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11245 4 : )];
11246 4 : let delta4 = vec![(
11247 4 : get_key(1),
11248 4 : Lsn(0x28),
11249 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11250 4 : )];
11251 4 : let delta2 = vec![
11252 4 : (
11253 4 : get_key(1),
11254 4 : Lsn(0x30),
11255 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11256 4 : ),
11257 4 : (
11258 4 : get_key(1),
11259 4 : Lsn(0x38),
11260 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11261 4 : ),
11262 4 : ];
11263 4 : let delta3 = vec![
11264 4 : (
11265 4 : get_key(8),
11266 4 : Lsn(0x48),
11267 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11268 4 : ),
11269 4 : (
11270 4 : get_key(9),
11271 4 : Lsn(0x48),
11272 4 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11273 4 : ),
11274 4 : ];
11275 4 :
11276 4 : let tline = tenant
11277 4 : .create_test_timeline_with_layers(
11278 4 : TIMELINE_ID,
11279 4 : Lsn(0x10),
11280 4 : DEFAULT_PG_VERSION,
11281 4 : &ctx,
11282 4 : vec![], // in-memory layers
11283 4 : vec![
11284 4 : // delta1/2/4 only contain a single key but multiple updates
11285 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11286 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11287 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11288 4 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11289 4 : ], // delta layers
11290 4 : vec![(Lsn(0x10), img_layer)], // image layers
11291 4 : Lsn(0x50),
11292 4 : )
11293 4 : .await?;
11294 4 : {
11295 4 : tline
11296 4 : .applied_gc_cutoff_lsn
11297 4 : .lock_for_write()
11298 4 : .store_and_unlock(Lsn(0x30))
11299 4 : .wait()
11300 4 : .await;
11301 4 : // Update GC info
11302 4 : let mut guard = tline.gc_info.write().unwrap();
11303 4 : *guard = GcInfo {
11304 4 : retain_lsns: vec![
11305 4 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11306 4 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11307 4 : ],
11308 4 : cutoffs: GcCutoffs {
11309 4 : time: Lsn(0x30),
11310 4 : space: Lsn(0x30),
11311 4 : },
11312 4 : leases: Default::default(),
11313 4 : within_ancestor_pitr: false,
11314 4 : };
11315 4 : }
11316 4 :
11317 4 : let expected_result = [
11318 4 : Bytes::from_static(b"value 0@0x10"),
11319 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11320 4 : Bytes::from_static(b"value 2@0x10"),
11321 4 : Bytes::from_static(b"value 3@0x10"),
11322 4 : Bytes::from_static(b"value 4@0x10"),
11323 4 : Bytes::from_static(b"value 5@0x10"),
11324 4 : Bytes::from_static(b"value 6@0x10"),
11325 4 : Bytes::from_static(b"value 7@0x10"),
11326 4 : Bytes::from_static(b"value 8@0x10@0x48"),
11327 4 : Bytes::from_static(b"value 9@0x10@0x48"),
11328 4 : ];
11329 4 :
11330 4 : let expected_result_at_gc_horizon = [
11331 4 : Bytes::from_static(b"value 0@0x10"),
11332 4 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11333 4 : Bytes::from_static(b"value 2@0x10"),
11334 4 : Bytes::from_static(b"value 3@0x10"),
11335 4 : Bytes::from_static(b"value 4@0x10"),
11336 4 : Bytes::from_static(b"value 5@0x10"),
11337 4 : Bytes::from_static(b"value 6@0x10"),
11338 4 : Bytes::from_static(b"value 7@0x10"),
11339 4 : Bytes::from_static(b"value 8@0x10"),
11340 4 : Bytes::from_static(b"value 9@0x10"),
11341 4 : ];
11342 4 :
11343 4 : let expected_result_at_lsn_20 = [
11344 4 : Bytes::from_static(b"value 0@0x10"),
11345 4 : Bytes::from_static(b"value 1@0x10@0x20"),
11346 4 : Bytes::from_static(b"value 2@0x10"),
11347 4 : Bytes::from_static(b"value 3@0x10"),
11348 4 : Bytes::from_static(b"value 4@0x10"),
11349 4 : Bytes::from_static(b"value 5@0x10"),
11350 4 : Bytes::from_static(b"value 6@0x10"),
11351 4 : Bytes::from_static(b"value 7@0x10"),
11352 4 : Bytes::from_static(b"value 8@0x10"),
11353 4 : Bytes::from_static(b"value 9@0x10"),
11354 4 : ];
11355 4 :
11356 4 : let expected_result_at_lsn_10 = [
11357 4 : Bytes::from_static(b"value 0@0x10"),
11358 4 : Bytes::from_static(b"value 1@0x10"),
11359 4 : Bytes::from_static(b"value 2@0x10"),
11360 4 : Bytes::from_static(b"value 3@0x10"),
11361 4 : Bytes::from_static(b"value 4@0x10"),
11362 4 : Bytes::from_static(b"value 5@0x10"),
11363 4 : Bytes::from_static(b"value 6@0x10"),
11364 4 : Bytes::from_static(b"value 7@0x10"),
11365 4 : Bytes::from_static(b"value 8@0x10"),
11366 4 : Bytes::from_static(b"value 9@0x10"),
11367 4 : ];
11368 4 :
11369 20 : let verify_result = || async {
11370 20 : let gc_horizon = {
11371 20 : let gc_info = tline.gc_info.read().unwrap();
11372 20 : gc_info.cutoffs.time
11373 4 : };
11374 220 : for idx in 0..10 {
11375 200 : assert_eq!(
11376 200 : tline
11377 200 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
11378 200 : .await
11379 200 : .unwrap(),
11380 200 : &expected_result[idx]
11381 4 : );
11382 200 : assert_eq!(
11383 200 : tline
11384 200 : .get(get_key(idx as u32), gc_horizon, &ctx)
11385 200 : .await
11386 200 : .unwrap(),
11387 200 : &expected_result_at_gc_horizon[idx]
11388 4 : );
11389 200 : assert_eq!(
11390 200 : tline
11391 200 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
11392 200 : .await
11393 200 : .unwrap(),
11394 200 : &expected_result_at_lsn_20[idx]
11395 4 : );
11396 200 : assert_eq!(
11397 200 : tline
11398 200 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
11399 200 : .await
11400 200 : .unwrap(),
11401 200 : &expected_result_at_lsn_10[idx]
11402 4 : );
11403 4 : }
11404 40 : };
11405 4 :
11406 4 : verify_result().await;
11407 4 :
11408 4 : let cancel = CancellationToken::new();
11409 4 :
11410 4 : tline
11411 4 : .compact_with_gc(
11412 4 : &cancel,
11413 4 : CompactOptions {
11414 4 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11415 4 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
11416 4 : ..Default::default()
11417 4 : },
11418 4 : &ctx,
11419 4 : )
11420 4 : .await
11421 4 : .unwrap();
11422 4 : verify_result().await;
11423 4 :
11424 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11425 4 : check_layer_map_key_eq(
11426 4 : all_layers,
11427 4 : vec![
11428 4 : // The original image layer, not compacted
11429 4 : PersistentLayerKey {
11430 4 : key_range: get_key(0)..get_key(10),
11431 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11432 4 : is_delta: false,
11433 4 : },
11434 4 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
11435 4 : // the layer 0x28-0x30 into one.
11436 4 : PersistentLayerKey {
11437 4 : key_range: get_key(1)..get_key(2),
11438 4 : lsn_range: Lsn(0x20)..Lsn(0x30),
11439 4 : is_delta: true,
11440 4 : },
11441 4 : // Above the upper bound and untouched
11442 4 : PersistentLayerKey {
11443 4 : key_range: get_key(1)..get_key(2),
11444 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11445 4 : is_delta: true,
11446 4 : },
11447 4 : // This layer is untouched
11448 4 : PersistentLayerKey {
11449 4 : key_range: get_key(8)..get_key(10),
11450 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11451 4 : is_delta: true,
11452 4 : },
11453 4 : ],
11454 4 : );
11455 4 :
11456 4 : tline
11457 4 : .compact_with_gc(
11458 4 : &cancel,
11459 4 : CompactOptions {
11460 4 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
11461 4 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
11462 4 : ..Default::default()
11463 4 : },
11464 4 : &ctx,
11465 4 : )
11466 4 : .await
11467 4 : .unwrap();
11468 4 : verify_result().await;
11469 4 :
11470 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11471 4 : check_layer_map_key_eq(
11472 4 : all_layers,
11473 4 : vec![
11474 4 : // The original image layer, not compacted
11475 4 : PersistentLayerKey {
11476 4 : key_range: get_key(0)..get_key(10),
11477 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11478 4 : is_delta: false,
11479 4 : },
11480 4 : // Not in the compaction key range, uncompacted
11481 4 : PersistentLayerKey {
11482 4 : key_range: get_key(1)..get_key(2),
11483 4 : lsn_range: Lsn(0x20)..Lsn(0x30),
11484 4 : is_delta: true,
11485 4 : },
11486 4 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
11487 4 : PersistentLayerKey {
11488 4 : key_range: get_key(1)..get_key(2),
11489 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11490 4 : is_delta: true,
11491 4 : },
11492 4 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
11493 4 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
11494 4 : // becomes 0x50.
11495 4 : PersistentLayerKey {
11496 4 : key_range: get_key(8)..get_key(10),
11497 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11498 4 : is_delta: true,
11499 4 : },
11500 4 : ],
11501 4 : );
11502 4 :
11503 4 : // compact again
11504 4 : tline
11505 4 : .compact_with_gc(
11506 4 : &cancel,
11507 4 : CompactOptions {
11508 4 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
11509 4 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
11510 4 : ..Default::default()
11511 4 : },
11512 4 : &ctx,
11513 4 : )
11514 4 : .await
11515 4 : .unwrap();
11516 4 : verify_result().await;
11517 4 :
11518 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11519 4 : check_layer_map_key_eq(
11520 4 : all_layers,
11521 4 : vec![
11522 4 : // The original image layer, not compacted
11523 4 : PersistentLayerKey {
11524 4 : key_range: get_key(0)..get_key(10),
11525 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11526 4 : is_delta: false,
11527 4 : },
11528 4 : // The range gets compacted
11529 4 : PersistentLayerKey {
11530 4 : key_range: get_key(1)..get_key(2),
11531 4 : lsn_range: Lsn(0x20)..Lsn(0x50),
11532 4 : is_delta: true,
11533 4 : },
11534 4 : // Not touched during this iteration of compaction
11535 4 : PersistentLayerKey {
11536 4 : key_range: get_key(8)..get_key(10),
11537 4 : lsn_range: Lsn(0x30)..Lsn(0x50),
11538 4 : is_delta: true,
11539 4 : },
11540 4 : ],
11541 4 : );
11542 4 :
11543 4 : // final full compaction
11544 4 : tline
11545 4 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
11546 4 : .await
11547 4 : .unwrap();
11548 4 : verify_result().await;
11549 4 :
11550 4 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11551 4 : check_layer_map_key_eq(
11552 4 : all_layers,
11553 4 : vec![
11554 4 : // The compacted image layer (full key range)
11555 4 : PersistentLayerKey {
11556 4 : key_range: Key::MIN..Key::MAX,
11557 4 : lsn_range: Lsn(0x10)..Lsn(0x11),
11558 4 : is_delta: false,
11559 4 : },
11560 4 : // All other data in the delta layer
11561 4 : PersistentLayerKey {
11562 4 : key_range: get_key(1)..get_key(10),
11563 4 : lsn_range: Lsn(0x10)..Lsn(0x50),
11564 4 : is_delta: true,
11565 4 : },
11566 4 : ],
11567 4 : );
11568 4 :
11569 4 : Ok(())
11570 4 : }
11571 : }
|
