Line data Source code
1 : pub(crate) mod analysis;
2 : pub(crate) mod compaction;
3 : pub mod delete;
4 : pub(crate) mod detach_ancestor;
5 : mod eviction_task;
6 : pub(crate) mod handle;
7 : mod init;
8 : pub mod layer_manager;
9 : pub(crate) mod logical_size;
10 : pub mod offload;
11 : pub mod span;
12 : pub mod uninit;
13 : mod walreceiver;
14 :
15 : use anyhow::{anyhow, bail, ensure, Context, Result};
16 : use arc_swap::ArcSwap;
17 : use bytes::Bytes;
18 : use camino::Utf8Path;
19 : use chrono::{DateTime, Utc};
20 : use enumset::EnumSet;
21 : use fail::fail_point;
22 : use handle::ShardTimelineId;
23 : use offload::OffloadError;
24 : use once_cell::sync::Lazy;
25 : use pageserver_api::{
26 : config::tenant_conf_defaults::DEFAULT_COMPACTION_THRESHOLD,
27 : key::{
28 : KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
29 : NON_INHERITED_SPARSE_RANGE,
30 : },
31 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
32 : models::{
33 : CompactionAlgorithm, CompactionAlgorithmSettings, DownloadRemoteLayersTaskInfo,
34 : DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy, InMemoryLayerInfo, LayerMapInfo,
35 : LsnLease, TimelineState,
36 : },
37 : reltag::BlockNumber,
38 : shard::{ShardIdentity, ShardNumber, TenantShardId},
39 : };
40 : use rand::Rng;
41 : use serde_with::serde_as;
42 : use storage_broker::BrokerClientChannel;
43 : use tokio::{
44 : runtime::Handle,
45 : sync::{oneshot, watch},
46 : };
47 : use tokio_util::sync::CancellationToken;
48 : use tracing::*;
49 : use utils::{
50 : fs_ext, pausable_failpoint,
51 : sync::gate::{Gate, GateGuard},
52 : };
53 : use wal_decoder::serialized_batch::SerializedValueBatch;
54 :
55 : use std::sync::atomic::Ordering as AtomicOrdering;
56 : use std::sync::{Arc, Mutex, RwLock, Weak};
57 : use std::time::{Duration, Instant, SystemTime};
58 : use std::{
59 : array,
60 : collections::{BTreeMap, HashMap, HashSet},
61 : sync::atomic::AtomicU64,
62 : };
63 : use std::{cmp::min, ops::ControlFlow};
64 : use std::{
65 : collections::btree_map::Entry,
66 : ops::{Deref, Range},
67 : };
68 : use std::{pin::pin, sync::OnceLock};
69 :
70 : use crate::{
71 : aux_file::AuxFileSizeEstimator,
72 : tenant::{
73 : config::AttachmentMode,
74 : layer_map::{LayerMap, SearchResult},
75 : metadata::TimelineMetadata,
76 : storage_layer::{inmemory_layer::IndexEntry, PersistentLayerDesc},
77 : },
78 : walingest::WalLagCooldown,
79 : walredo,
80 : };
81 : use crate::{
82 : context::{DownloadBehavior, RequestContext},
83 : disk_usage_eviction_task::DiskUsageEvictionInfo,
84 : pgdatadir_mapping::CollectKeySpaceError,
85 : };
86 : use crate::{
87 : disk_usage_eviction_task::finite_f32,
88 : tenant::storage_layer::{
89 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
90 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
91 : ValuesReconstructState,
92 : },
93 : };
94 : use crate::{
95 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
96 : };
97 : use crate::{
98 : l0_flush::{self, L0FlushGlobalState},
99 : metrics::GetKind,
100 : };
101 : use crate::{
102 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
103 : };
104 : use crate::{
105 : pgdatadir_mapping::DirectoryKind,
106 : virtual_file::{MaybeFatalIo, VirtualFile},
107 : };
108 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
109 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
110 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
111 :
112 : use crate::config::PageServerConf;
113 : use crate::keyspace::{KeyPartitioning, KeySpace};
114 : use crate::metrics::TimelineMetrics;
115 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
116 : use crate::tenant::config::TenantConfOpt;
117 : use pageserver_api::reltag::RelTag;
118 : use pageserver_api::shard::ShardIndex;
119 :
120 : use postgres_connection::PgConnectionConfig;
121 : use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
122 : use utils::{
123 : completion,
124 : generation::Generation,
125 : id::TimelineId,
126 : lsn::{AtomicLsn, Lsn, RecordLsn},
127 : seqwait::SeqWait,
128 : simple_rcu::{Rcu, RcuReadGuard},
129 : };
130 :
131 : use crate::task_mgr;
132 : use crate::task_mgr::TaskKind;
133 : use crate::tenant::gc_result::GcResult;
134 : use crate::ZERO_PAGE;
135 : use pageserver_api::key::Key;
136 :
137 : use self::delete::DeleteTimelineFlow;
138 : pub(super) use self::eviction_task::EvictionTaskTenantState;
139 : use self::eviction_task::EvictionTaskTimelineState;
140 : use self::layer_manager::LayerManager;
141 : use self::logical_size::LogicalSize;
142 : use self::walreceiver::{WalReceiver, WalReceiverConf};
143 :
144 : use super::{
145 : config::TenantConf, storage_layer::LayerVisibilityHint, upload_queue::NotInitialized,
146 : MaybeOffloaded,
147 : };
148 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
149 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
150 : use super::{
151 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
152 : storage_layer::ReadableLayer,
153 : };
154 : use super::{
155 : secondary::heatmap::{HeatMapLayer, HeatMapTimeline},
156 : GcError,
157 : };
158 :
159 : #[cfg(test)]
160 : use pageserver_api::value::Value;
161 :
162 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
163 : pub(crate) enum FlushLoopState {
164 : NotStarted,
165 : Running {
166 : #[cfg(test)]
167 : expect_initdb_optimization: bool,
168 : #[cfg(test)]
169 : initdb_optimization_count: usize,
170 : },
171 : Exited,
172 : }
173 :
174 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
175 : pub enum ImageLayerCreationMode {
176 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
177 : Try,
178 : /// Force creating the image layers if possible. For now, no image layers will be created
179 : /// for metadata keys. Used in compaction code path with force flag enabled.
180 : Force,
181 : /// Initial ingestion of the data, and no data should be dropped in this function. This
182 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
183 : /// code path.
184 : Initial,
185 : }
186 :
187 : impl std::fmt::Display for ImageLayerCreationMode {
188 716 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
189 716 : write!(f, "{:?}", self)
190 716 : }
191 : }
192 :
193 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
194 : /// Can be removed after all refactors are done.
195 28 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
196 28 : drop(rlock)
197 28 : }
198 :
199 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
200 : /// Can be removed after all refactors are done.
201 744 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
202 744 : drop(rlock)
203 744 : }
204 :
205 : /// The outward-facing resources required to build a Timeline
206 : pub struct TimelineResources {
207 : pub remote_client: RemoteTimelineClient,
208 : pub timeline_get_throttle:
209 : Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
210 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
211 : }
212 :
213 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
214 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
215 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
216 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
217 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
218 : pub(crate) struct RelSizeCache {
219 : pub(crate) complete_as_of: Lsn,
220 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
221 : }
222 :
223 : pub struct Timeline {
224 : pub(crate) conf: &'static PageServerConf,
225 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
226 :
227 : myself: Weak<Self>,
228 :
229 : pub(crate) tenant_shard_id: TenantShardId,
230 : pub timeline_id: TimelineId,
231 :
232 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
233 : /// Never changes for the lifetime of this [`Timeline`] object.
234 : ///
235 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
236 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
237 : pub(crate) generation: Generation,
238 :
239 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
240 : /// to shards, and is constant through the lifetime of this Timeline.
241 : shard_identity: ShardIdentity,
242 :
243 : pub pg_version: u32,
244 :
245 : /// The tuple has two elements.
246 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
247 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
248 : ///
249 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
250 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
251 : ///
252 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
253 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
254 : /// `PersistentLayerDesc`'s.
255 : ///
256 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
257 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
258 : /// runtime, e.g., during page reconstruction.
259 : ///
260 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
261 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
262 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
263 :
264 : last_freeze_at: AtomicLsn,
265 : // Atomic would be more appropriate here.
266 : last_freeze_ts: RwLock<Instant>,
267 :
268 : pub(crate) standby_horizon: AtomicLsn,
269 :
270 : // WAL redo manager. `None` only for broken tenants.
271 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
272 :
273 : /// Remote storage client.
274 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
275 : pub remote_client: Arc<RemoteTimelineClient>,
276 :
277 : // What page versions do we hold in the repository? If we get a
278 : // request > last_record_lsn, we need to wait until we receive all
279 : // the WAL up to the request. The SeqWait provides functions for
280 : // that. TODO: If we get a request for an old LSN, such that the
281 : // versions have already been garbage collected away, we should
282 : // throw an error, but we don't track that currently.
283 : //
284 : // last_record_lsn.load().last points to the end of last processed WAL record.
285 : //
286 : // We also remember the starting point of the previous record in
287 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
288 : // first WAL record when the node is started up. But here, we just
289 : // keep track of it.
290 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
291 :
292 : // All WAL records have been processed and stored durably on files on
293 : // local disk, up to this LSN. On crash and restart, we need to re-process
294 : // the WAL starting from this point.
295 : //
296 : // Some later WAL records might have been processed and also flushed to disk
297 : // already, so don't be surprised to see some, but there's no guarantee on
298 : // them yet.
299 : disk_consistent_lsn: AtomicLsn,
300 :
301 : // Parent timeline that this timeline was branched from, and the LSN
302 : // of the branch point.
303 : ancestor_timeline: Option<Arc<Timeline>>,
304 : ancestor_lsn: Lsn,
305 :
306 : pub(super) metrics: TimelineMetrics,
307 :
308 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
309 : // in `crate::page_service` writes these metrics.
310 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
311 :
312 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
313 :
314 : /// Ensures layers aren't frozen by checkpointer between
315 : /// [`Timeline::get_layer_for_write`] and layer reads.
316 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
317 : /// Must always be acquired before the layer map/individual layer lock
318 : /// to avoid deadlock.
319 : ///
320 : /// The state is cleared upon freezing.
321 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
322 :
323 : /// Used to avoid multiple `flush_loop` tasks running
324 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
325 :
326 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
327 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
328 : /// The flush cycle counter is sent back on the layer_flush_done channel when
329 : /// the flush finishes. You can use that to wait for the flush to finish.
330 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
331 : /// read by whoever sends an update
332 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
333 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
334 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
335 :
336 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
337 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
338 :
339 : // List of child timelines and their branch points. This is needed to avoid
340 : // garbage collecting data that is still needed by the child timelines.
341 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
342 :
343 : // It may change across major versions so for simplicity
344 : // keep it after running initdb for a timeline.
345 : // It is needed in checks when we want to error on some operations
346 : // when they are requested for pre-initdb lsn.
347 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
348 : // though let's keep them both for better error visibility.
349 : pub initdb_lsn: Lsn,
350 :
351 : /// When did we last calculate the partitioning? Make it pub to test cases.
352 : pub(super) partitioning: tokio::sync::Mutex<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
353 :
354 : /// Configuration: how often should the partitioning be recalculated.
355 : repartition_threshold: u64,
356 :
357 : last_image_layer_creation_check_at: AtomicLsn,
358 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
359 :
360 : /// Current logical size of the "datadir", at the last LSN.
361 : current_logical_size: LogicalSize,
362 :
363 : /// Information about the last processed message by the WAL receiver,
364 : /// or None if WAL receiver has not received anything for this timeline
365 : /// yet.
366 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
367 : pub walreceiver: Mutex<Option<WalReceiver>>,
368 :
369 : /// Relation size cache
370 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
371 :
372 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
373 :
374 : state: watch::Sender<TimelineState>,
375 :
376 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
377 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
378 : pub delete_progress: TimelineDeleteProgress,
379 :
380 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
381 :
382 : /// Load or creation time information about the disk_consistent_lsn and when the loading
383 : /// happened. Used for consumption metrics.
384 : pub(crate) loaded_at: (Lsn, SystemTime),
385 :
386 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
387 : pub(crate) gate: Gate,
388 :
389 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
390 : /// to the timeline should drop out when this token fires.
391 : pub(crate) cancel: CancellationToken,
392 :
393 : /// Make sure we only have one running compaction at a time in tests.
394 : ///
395 : /// Must only be taken in two places:
396 : /// - [`Timeline::compact`] (this file)
397 : /// - [`delete::delete_local_timeline_directory`]
398 : ///
399 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
400 : compaction_lock: tokio::sync::Mutex<()>,
401 :
402 : /// Make sure we only have one running gc at a time.
403 : ///
404 : /// Must only be taken in two places:
405 : /// - [`Timeline::gc`] (this file)
406 : /// - [`delete::delete_local_timeline_directory`]
407 : ///
408 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
409 : gc_lock: tokio::sync::Mutex<()>,
410 :
411 : /// Cloned from [`super::Tenant::timeline_get_throttle`] on construction.
412 : timeline_get_throttle:
413 : Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
414 :
415 : /// Size estimator for aux file v2
416 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
417 :
418 : /// Some test cases directly place keys into the timeline without actually modifying the directory
419 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
420 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
421 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
422 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
423 : #[cfg(test)]
424 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
425 :
426 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
427 :
428 : pub(crate) handles: handle::PerTimelineState<crate::page_service::TenantManagerTypes>,
429 :
430 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
431 :
432 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
433 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
434 : }
435 :
436 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
437 :
438 : pub struct WalReceiverInfo {
439 : pub wal_source_connconf: PgConnectionConfig,
440 : pub last_received_msg_lsn: Lsn,
441 : pub last_received_msg_ts: u128,
442 : }
443 :
444 : /// Information about how much history needs to be retained, needed by
445 : /// Garbage Collection.
446 : #[derive(Default)]
447 : pub(crate) struct GcInfo {
448 : /// Specific LSNs that are needed.
449 : ///
450 : /// Currently, this includes all points where child branches have
451 : /// been forked off from. In the future, could also include
452 : /// explicit user-defined snapshot points.
453 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
454 :
455 : /// The cutoff coordinates, which are combined by selecting the minimum.
456 : pub(crate) cutoffs: GcCutoffs,
457 :
458 : /// Leases granted to particular LSNs.
459 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
460 :
461 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
462 : pub(crate) within_ancestor_pitr: bool,
463 : }
464 :
465 : impl GcInfo {
466 228 : pub(crate) fn min_cutoff(&self) -> Lsn {
467 228 : self.cutoffs.select_min()
468 228 : }
469 :
470 232 : pub(super) fn insert_child(
471 232 : &mut self,
472 232 : child_id: TimelineId,
473 232 : child_lsn: Lsn,
474 232 : is_offloaded: MaybeOffloaded,
475 232 : ) {
476 232 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
477 232 : self.retain_lsns.sort_by_key(|i| i.0);
478 232 : }
479 :
480 4 : pub(super) fn remove_child_maybe_offloaded(
481 4 : &mut self,
482 4 : child_id: TimelineId,
483 4 : maybe_offloaded: MaybeOffloaded,
484 4 : ) -> bool {
485 4 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
486 4 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
487 4 : let mut removed = false;
488 6 : self.retain_lsns.retain(|i| {
489 6 : if removed {
490 2 : return true;
491 4 : }
492 4 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
493 4 : removed |= remove;
494 4 : !remove
495 6 : });
496 4 : removed
497 4 : }
498 :
499 4 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
500 4 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
501 4 : }
502 :
503 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
504 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
505 0 : }
506 : }
507 :
508 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
509 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
510 : /// between time-based and space-based retention for observability and consumption metrics purposes.
511 : #[derive(Debug, Clone)]
512 : pub(crate) struct GcCutoffs {
513 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
514 : /// history we must keep to retain a specified number of bytes of WAL.
515 : pub(crate) space: Lsn,
516 :
517 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
518 : /// history we must keep to enable reading back at least the PITR interval duration.
519 : pub(crate) time: Lsn,
520 : }
521 :
522 : impl Default for GcCutoffs {
523 418 : fn default() -> Self {
524 418 : Self {
525 418 : space: Lsn::INVALID,
526 418 : time: Lsn::INVALID,
527 418 : }
528 418 : }
529 : }
530 :
531 : impl GcCutoffs {
532 268 : fn select_min(&self) -> Lsn {
533 268 : std::cmp::min(self.space, self.time)
534 268 : }
535 : }
536 :
537 : pub(crate) struct TimelineVisitOutcome {
538 : completed_keyspace: KeySpace,
539 : image_covered_keyspace: KeySpace,
540 : }
541 :
542 : /// An error happened in a get() operation.
543 2 : #[derive(thiserror::Error, Debug)]
544 : pub(crate) enum PageReconstructError {
545 : #[error(transparent)]
546 : Other(anyhow::Error),
547 :
548 : #[error("Ancestor LSN wait error: {0}")]
549 : AncestorLsnTimeout(WaitLsnError),
550 :
551 : #[error("timeline shutting down")]
552 : Cancelled,
553 :
554 : /// An error happened replaying WAL records
555 : #[error(transparent)]
556 : WalRedo(anyhow::Error),
557 :
558 : #[error("{0}")]
559 : MissingKey(MissingKeyError),
560 : }
561 :
562 : impl From<anyhow::Error> for PageReconstructError {
563 0 : fn from(value: anyhow::Error) -> Self {
564 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
565 0 : match value.downcast::<PageReconstructError>() {
566 0 : Ok(pre) => pre,
567 0 : Err(other) => PageReconstructError::Other(other),
568 : }
569 0 : }
570 : }
571 :
572 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
573 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
574 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
575 0 : }
576 : }
577 :
578 : impl From<layer_manager::Shutdown> for PageReconstructError {
579 0 : fn from(_: layer_manager::Shutdown) -> Self {
580 0 : PageReconstructError::Cancelled
581 0 : }
582 : }
583 :
584 : impl GetVectoredError {
585 : #[cfg(test)]
586 6 : pub(crate) fn is_missing_key_error(&self) -> bool {
587 6 : matches!(self, Self::MissingKey(_))
588 6 : }
589 : }
590 :
591 : impl From<layer_manager::Shutdown> for GetVectoredError {
592 0 : fn from(_: layer_manager::Shutdown) -> Self {
593 0 : GetVectoredError::Cancelled
594 0 : }
595 : }
596 :
597 : #[derive(thiserror::Error)]
598 : pub struct MissingKeyError {
599 : key: Key,
600 : shard: ShardNumber,
601 : cont_lsn: Lsn,
602 : request_lsn: Lsn,
603 : ancestor_lsn: Option<Lsn>,
604 : backtrace: Option<std::backtrace::Backtrace>,
605 : }
606 :
607 : impl std::fmt::Debug for MissingKeyError {
608 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
609 0 : write!(f, "{}", self)
610 0 : }
611 : }
612 :
613 : impl std::fmt::Display for MissingKeyError {
614 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
615 0 : write!(
616 0 : f,
617 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
618 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
619 0 : )?;
620 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
621 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
622 0 : }
623 :
624 0 : if let Some(ref backtrace) = self.backtrace {
625 0 : write!(f, "\n{}", backtrace)?;
626 0 : }
627 :
628 0 : Ok(())
629 0 : }
630 : }
631 :
632 : impl PageReconstructError {
633 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
634 0 : pub(crate) fn is_stopping(&self) -> bool {
635 : use PageReconstructError::*;
636 0 : match self {
637 0 : Cancelled => true,
638 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
639 : }
640 0 : }
641 : }
642 :
643 0 : #[derive(thiserror::Error, Debug)]
644 : pub(crate) enum CreateImageLayersError {
645 : #[error("timeline shutting down")]
646 : Cancelled,
647 :
648 : #[error("read failed")]
649 : GetVectoredError(#[source] GetVectoredError),
650 :
651 : #[error("reconstruction failed")]
652 : PageReconstructError(#[source] PageReconstructError),
653 :
654 : #[error(transparent)]
655 : Other(#[from] anyhow::Error),
656 : }
657 :
658 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
659 0 : fn from(_: layer_manager::Shutdown) -> Self {
660 0 : CreateImageLayersError::Cancelled
661 0 : }
662 : }
663 :
664 0 : #[derive(thiserror::Error, Debug, Clone)]
665 : pub(crate) enum FlushLayerError {
666 : /// Timeline cancellation token was cancelled
667 : #[error("timeline shutting down")]
668 : Cancelled,
669 :
670 : /// We tried to flush a layer while the Timeline is in an unexpected state
671 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
672 : NotRunning(FlushLoopState),
673 :
674 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
675 : // loop via a watch channel, where we can only borrow it.
676 : #[error("create image layers (shared)")]
677 : CreateImageLayersError(Arc<CreateImageLayersError>),
678 :
679 : #[error("other (shared)")]
680 : Other(#[from] Arc<anyhow::Error>),
681 : }
682 :
683 : impl FlushLayerError {
684 : // When crossing from generic anyhow errors to this error type, we explicitly check
685 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
686 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
687 0 : let cancelled = timeline.cancel.is_cancelled()
688 : // The upload queue might have been shut down before the official cancellation of the timeline.
689 0 : || err
690 0 : .downcast_ref::<NotInitialized>()
691 0 : .map(NotInitialized::is_stopping)
692 0 : .unwrap_or_default();
693 0 : if cancelled {
694 0 : Self::Cancelled
695 : } else {
696 0 : Self::Other(Arc::new(err))
697 : }
698 0 : }
699 : }
700 :
701 : impl From<layer_manager::Shutdown> for FlushLayerError {
702 0 : fn from(_: layer_manager::Shutdown) -> Self {
703 0 : FlushLayerError::Cancelled
704 0 : }
705 : }
706 :
707 0 : #[derive(thiserror::Error, Debug)]
708 : pub(crate) enum GetVectoredError {
709 : #[error("timeline shutting down")]
710 : Cancelled,
711 :
712 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
713 : Oversized(u64),
714 :
715 : #[error("requested at invalid LSN: {0}")]
716 : InvalidLsn(Lsn),
717 :
718 : #[error("requested key not found: {0}")]
719 : MissingKey(MissingKeyError),
720 :
721 : #[error("ancestry walk")]
722 : GetReadyAncestorError(#[source] GetReadyAncestorError),
723 :
724 : #[error(transparent)]
725 : Other(#[from] anyhow::Error),
726 : }
727 :
728 : impl From<GetReadyAncestorError> for GetVectoredError {
729 2 : fn from(value: GetReadyAncestorError) -> Self {
730 : use GetReadyAncestorError::*;
731 2 : match value {
732 0 : Cancelled => GetVectoredError::Cancelled,
733 : AncestorLsnTimeout(_) | BadState { .. } => {
734 2 : GetVectoredError::GetReadyAncestorError(value)
735 : }
736 : }
737 2 : }
738 : }
739 :
740 2 : #[derive(thiserror::Error, Debug)]
741 : pub(crate) enum GetReadyAncestorError {
742 : #[error("ancestor LSN wait error")]
743 : AncestorLsnTimeout(#[from] WaitLsnError),
744 :
745 : #[error("bad state on timeline {timeline_id}: {state:?}")]
746 : BadState {
747 : timeline_id: TimelineId,
748 : state: TimelineState,
749 : },
750 :
751 : #[error("cancelled")]
752 : Cancelled,
753 : }
754 :
755 : #[derive(Clone, Copy)]
756 : pub enum LogicalSizeCalculationCause {
757 : Initial,
758 : ConsumptionMetricsSyntheticSize,
759 : EvictionTaskImitation,
760 : TenantSizeHandler,
761 : }
762 :
763 : pub enum GetLogicalSizePriority {
764 : User,
765 : Background,
766 : }
767 :
768 0 : #[derive(enumset::EnumSetType)]
769 : pub(crate) enum CompactFlags {
770 : ForceRepartition,
771 : ForceImageLayerCreation,
772 : ForceL0Compaction,
773 : EnhancedGcBottomMostCompaction,
774 : DryRun,
775 : }
776 :
777 : impl std::fmt::Debug for Timeline {
778 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
779 0 : write!(f, "Timeline<{}>", self.timeline_id)
780 0 : }
781 : }
782 :
783 0 : #[derive(thiserror::Error, Debug)]
784 : pub(crate) enum WaitLsnError {
785 : // Called on a timeline which is shutting down
786 : #[error("Shutdown")]
787 : Shutdown,
788 :
789 : // Called on an timeline not in active state or shutting down
790 : #[error("Bad timeline state: {0:?}")]
791 : BadState(TimelineState),
792 :
793 : // Timeout expired while waiting for LSN to catch up with goal.
794 : #[error("{0}")]
795 : Timeout(String),
796 : }
797 :
798 : // The impls below achieve cancellation mapping for errors.
799 : // Perhaps there's a way of achieving this with less cruft.
800 :
801 : impl From<CreateImageLayersError> for CompactionError {
802 0 : fn from(e: CreateImageLayersError) -> Self {
803 0 : match e {
804 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
805 0 : CreateImageLayersError::Other(e) => {
806 0 : CompactionError::Other(e.context("create image layers"))
807 : }
808 0 : _ => CompactionError::Other(e.into()),
809 : }
810 0 : }
811 : }
812 :
813 : impl From<CreateImageLayersError> for FlushLayerError {
814 0 : fn from(e: CreateImageLayersError) -> Self {
815 0 : match e {
816 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
817 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
818 : }
819 0 : }
820 : }
821 :
822 : impl From<PageReconstructError> for CreateImageLayersError {
823 0 : fn from(e: PageReconstructError) -> Self {
824 0 : match e {
825 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
826 0 : _ => CreateImageLayersError::PageReconstructError(e),
827 : }
828 0 : }
829 : }
830 :
831 : impl From<GetVectoredError> for CreateImageLayersError {
832 0 : fn from(e: GetVectoredError) -> Self {
833 0 : match e {
834 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
835 0 : _ => CreateImageLayersError::GetVectoredError(e),
836 : }
837 0 : }
838 : }
839 :
840 : impl From<GetVectoredError> for PageReconstructError {
841 6 : fn from(e: GetVectoredError) -> Self {
842 6 : match e {
843 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
844 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
845 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
846 4 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
847 2 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
848 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
849 : }
850 6 : }
851 : }
852 :
853 : impl From<GetReadyAncestorError> for PageReconstructError {
854 2 : fn from(e: GetReadyAncestorError) -> Self {
855 : use GetReadyAncestorError::*;
856 2 : match e {
857 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
858 2 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
859 0 : Cancelled => PageReconstructError::Cancelled,
860 : }
861 2 : }
862 : }
863 :
864 : pub(crate) enum WaitLsnWaiter<'a> {
865 : Timeline(&'a Timeline),
866 : Tenant,
867 : PageService,
868 : }
869 :
870 : /// Argument to [`Timeline::shutdown`].
871 : #[derive(Debug, Clone, Copy)]
872 : pub(crate) enum ShutdownMode {
873 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
874 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
875 : ///
876 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
877 : /// the call to [`Timeline::shutdown`].
878 : FreezeAndFlush,
879 : /// Only flush the layers to the remote storage without freezing any open layers. This is the
880 : /// mode used by ancestor detach and any other operations that reloads a tenant but not increasing
881 : /// the generation number.
882 : Flush,
883 : /// Shut down immediately, without waiting for any open layers to flush.
884 : Hard,
885 : }
886 :
887 : struct ImageLayerCreationOutcome {
888 : image: Option<ResidentLayer>,
889 : next_start_key: Key,
890 : }
891 :
892 : /// Public interface functions
893 : impl Timeline {
894 : /// Get the LSN where this branch was created
895 4 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
896 4 : self.ancestor_lsn
897 4 : }
898 :
899 : /// Get the ancestor's timeline id
900 12 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
901 12 : self.ancestor_timeline
902 12 : .as_ref()
903 12 : .map(|ancestor| ancestor.timeline_id)
904 12 : }
905 :
906 : /// Get the ancestor timeline
907 2 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
908 2 : self.ancestor_timeline.as_ref()
909 2 : }
910 :
911 : /// Get the bytes written since the PITR cutoff on this branch, and
912 : /// whether this branch's ancestor_lsn is within its parent's PITR.
913 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
914 0 : let gc_info = self.gc_info.read().unwrap();
915 0 : let history = self
916 0 : .get_last_record_lsn()
917 0 : .checked_sub(gc_info.cutoffs.time)
918 0 : .unwrap_or(Lsn(0))
919 0 : .0;
920 0 : (history, gc_info.within_ancestor_pitr)
921 0 : }
922 :
923 : /// Lock and get timeline's GC cutoff
924 248 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
925 248 : self.latest_gc_cutoff_lsn.read()
926 248 : }
927 :
928 : /// Look up given page version.
929 : ///
930 : /// If a remote layer file is needed, it is downloaded as part of this
931 : /// call.
932 : ///
933 : /// This method enforces [`Self::timeline_get_throttle`] internally.
934 : ///
935 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
936 : /// abstraction above this needs to store suitable metadata to track what
937 : /// data exists with what keys, in separate metadata entries. If a
938 : /// non-existent key is requested, we may incorrectly return a value from
939 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
940 : /// non-existing key.
941 : ///
942 : /// # Cancel-Safety
943 : ///
944 : /// This method is cancellation-safe.
945 : #[inline(always)]
946 606728 : pub(crate) async fn get(
947 606728 : &self,
948 606728 : key: Key,
949 606728 : lsn: Lsn,
950 606728 : ctx: &RequestContext,
951 606728 : ) -> Result<Bytes, PageReconstructError> {
952 606728 : if !lsn.is_valid() {
953 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
954 606728 : }
955 606728 :
956 606728 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
957 606728 : // already checked the key against the shard_identity when looking up the Timeline from
958 606728 : // page_service.
959 606728 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
960 :
961 606728 : self.timeline_get_throttle.throttle(ctx, 1).await;
962 :
963 606728 : let keyspace = KeySpace {
964 606728 : ranges: vec![key..key.next()],
965 606728 : };
966 606728 :
967 606728 : // Initialise the reconstruct state for the key with the cache
968 606728 : // entry returned above.
969 606728 : let mut reconstruct_state = ValuesReconstructState::new();
970 :
971 606728 : let vectored_res = self
972 606728 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
973 183045 : .await;
974 :
975 606728 : let key_value = vectored_res?.pop_first();
976 606722 : match key_value {
977 606710 : Some((got_key, value)) => {
978 606710 : if got_key != key {
979 0 : error!(
980 0 : "Expected {}, but singular vectored get returned {}",
981 : key, got_key
982 : );
983 0 : Err(PageReconstructError::Other(anyhow!(
984 0 : "Singular vectored get returned wrong key"
985 0 : )))
986 : } else {
987 606710 : value
988 : }
989 : }
990 12 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
991 12 : key,
992 12 : shard: self.shard_identity.get_shard_number(&key),
993 12 : cont_lsn: Lsn(0),
994 12 : request_lsn: lsn,
995 12 : ancestor_lsn: None,
996 12 : backtrace: None,
997 12 : })),
998 : }
999 606728 : }
1000 :
1001 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1002 : pub(crate) const VEC_GET_LAYERS_VISITED_WARN_THRESH: f64 = 512.0;
1003 :
1004 : /// Look up multiple page versions at a given LSN
1005 : ///
1006 : /// This naive implementation will be replaced with a more efficient one
1007 : /// which actually vectorizes the read path.
1008 19528 : pub(crate) async fn get_vectored(
1009 19528 : &self,
1010 19528 : keyspace: KeySpace,
1011 19528 : lsn: Lsn,
1012 19528 : ctx: &RequestContext,
1013 19528 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1014 19528 : if !lsn.is_valid() {
1015 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1016 19528 : }
1017 19528 :
1018 19528 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1019 19528 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1020 0 : return Err(GetVectoredError::Oversized(key_count));
1021 19528 : }
1022 :
1023 39056 : for range in &keyspace.ranges {
1024 19528 : let mut key = range.start;
1025 39262 : while key != range.end {
1026 19734 : assert!(!self.shard_identity.is_key_disposable(&key));
1027 19734 : key = key.next();
1028 : }
1029 : }
1030 :
1031 19528 : trace!(
1032 0 : "get vectored request for {:?}@{} from task kind {:?}",
1033 0 : keyspace,
1034 0 : lsn,
1035 0 : ctx.task_kind(),
1036 : );
1037 :
1038 19528 : let start = crate::metrics::GET_VECTORED_LATENCY
1039 19528 : .for_task_kind(ctx.task_kind())
1040 19528 : .map(|metric| (metric, Instant::now()));
1041 :
1042 : // start counting after throttle so that throttle time
1043 : // is always less than observation time
1044 19528 : let throttled = self
1045 19528 : .timeline_get_throttle
1046 19528 : .throttle(ctx, key_count as usize)
1047 0 : .await;
1048 :
1049 19528 : let res = self
1050 19528 : .get_vectored_impl(
1051 19528 : keyspace.clone(),
1052 19528 : lsn,
1053 19528 : &mut ValuesReconstructState::new(),
1054 19528 : ctx,
1055 19528 : )
1056 7038 : .await;
1057 :
1058 19528 : if let Some((metric, start)) = start {
1059 0 : let elapsed = start.elapsed();
1060 0 : let ex_throttled = if let Some(throttled) = throttled {
1061 0 : elapsed.checked_sub(throttled)
1062 : } else {
1063 0 : Some(elapsed)
1064 : };
1065 :
1066 0 : if let Some(ex_throttled) = ex_throttled {
1067 0 : metric.observe(ex_throttled.as_secs_f64());
1068 0 : } else {
1069 0 : use utils::rate_limit::RateLimit;
1070 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1071 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
1072 0 : let mut rate_limit = LOGGED.lock().unwrap();
1073 0 : rate_limit.call(|| {
1074 0 : warn!("error deducting time spent throttled; this message is logged at a global rate limit");
1075 0 : });
1076 0 : }
1077 19528 : }
1078 :
1079 19528 : res
1080 19528 : }
1081 :
1082 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1083 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1084 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1085 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1086 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1087 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1088 : /// the scan operation will not cause OOM in the future.
1089 12 : pub(crate) async fn scan(
1090 12 : &self,
1091 12 : keyspace: KeySpace,
1092 12 : lsn: Lsn,
1093 12 : ctx: &RequestContext,
1094 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1095 12 : if !lsn.is_valid() {
1096 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1097 12 : }
1098 12 :
1099 12 : trace!(
1100 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1101 0 : keyspace,
1102 0 : lsn,
1103 0 : ctx.task_kind()
1104 : );
1105 :
1106 : // We should generalize this into Keyspace::contains in the future.
1107 24 : for range in &keyspace.ranges {
1108 12 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1109 12 : || range.end.field1 > METADATA_KEY_END_PREFIX
1110 : {
1111 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1112 0 : "only metadata keyspace can be scanned"
1113 0 : )));
1114 12 : }
1115 : }
1116 :
1117 12 : let start = crate::metrics::SCAN_LATENCY
1118 12 : .for_task_kind(ctx.task_kind())
1119 12 : .map(ScanLatencyOngoingRecording::start_recording);
1120 :
1121 : // start counting after throttle so that throttle time
1122 : // is always less than observation time
1123 12 : let throttled = self
1124 12 : .timeline_get_throttle
1125 12 : // assume scan = 1 quota for now until we find a better way to process this
1126 12 : .throttle(ctx, 1)
1127 0 : .await;
1128 :
1129 12 : let vectored_res = self
1130 12 : .get_vectored_impl(
1131 12 : keyspace.clone(),
1132 12 : lsn,
1133 12 : &mut ValuesReconstructState::default(),
1134 12 : ctx,
1135 12 : )
1136 0 : .await;
1137 :
1138 12 : if let Some(recording) = start {
1139 0 : recording.observe(throttled);
1140 12 : }
1141 :
1142 12 : vectored_res
1143 12 : }
1144 :
1145 626570 : pub(super) async fn get_vectored_impl(
1146 626570 : &self,
1147 626570 : keyspace: KeySpace,
1148 626570 : lsn: Lsn,
1149 626570 : reconstruct_state: &mut ValuesReconstructState,
1150 626570 : ctx: &RequestContext,
1151 626570 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1152 626570 : let get_kind = if keyspace.total_raw_size() == 1 {
1153 626134 : GetKind::Singular
1154 : } else {
1155 436 : GetKind::Vectored
1156 : };
1157 :
1158 626570 : let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
1159 626570 : .for_get_kind(get_kind)
1160 626570 : .start_timer();
1161 626570 : self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1162 192204 : .await?;
1163 626554 : get_data_timer.stop_and_record();
1164 626554 :
1165 626554 : let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
1166 626554 : .for_get_kind(get_kind)
1167 626554 : .start_timer();
1168 626554 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
1169 626554 : let layers_visited = reconstruct_state.get_layers_visited();
1170 :
1171 666880 : for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
1172 666880 : match res {
1173 0 : Err(err) => {
1174 0 : results.insert(key, Err(err));
1175 0 : }
1176 666880 : Ok(state) => {
1177 666880 : let state = ValueReconstructState::from(state);
1178 :
1179 666880 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
1180 666880 : results.insert(key, reconstruct_res);
1181 : }
1182 : }
1183 : }
1184 626554 : reconstruct_timer.stop_and_record();
1185 626554 :
1186 626554 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1187 626554 : // when they're missing. Instead they are omitted from the resulting btree
1188 626554 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1189 626554 : // to avoid infinite results.
1190 626554 : if !results.is_empty() {
1191 626348 : let avg = layers_visited as f64 / results.len() as f64;
1192 626348 : if avg >= Self::VEC_GET_LAYERS_VISITED_WARN_THRESH {
1193 0 : use utils::rate_limit::RateLimit;
1194 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1195 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1196 0 : let mut rate_limit = LOGGED.lock().unwrap();
1197 0 : rate_limit.call(|| {
1198 0 : tracing::info!(
1199 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1200 0 : lsn = %lsn,
1201 0 : "Vectored read for {} visited {} layers on average per key and {} in total. {}/{} pages were returned",
1202 0 : keyspace, avg, layers_visited, results.len(), keyspace.total_raw_size());
1203 0 : });
1204 626348 : }
1205 :
1206 : // Note that this is an approximation. Tracking the exact number of layers visited
1207 : // per key requires virtually unbounded memory usage and is inefficient
1208 : // (i.e. segment tree tracking each range queried from a layer)
1209 626348 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED.observe(avg);
1210 206 : }
1211 :
1212 626554 : Ok(results)
1213 626570 : }
1214 :
1215 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1216 274290 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1217 274290 : self.last_record_lsn.load().last
1218 274290 : }
1219 :
1220 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1221 0 : self.last_record_lsn.load().prev
1222 0 : }
1223 :
1224 : /// Atomically get both last and prev.
1225 228 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1226 228 : self.last_record_lsn.load()
1227 228 : }
1228 :
1229 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1230 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1231 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1232 0 : self.last_record_lsn.status_receiver()
1233 0 : }
1234 :
1235 414 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1236 414 : self.disk_consistent_lsn.load()
1237 414 : }
1238 :
1239 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1240 : /// not validated with control plane yet.
1241 : /// See [`Self::get_remote_consistent_lsn_visible`].
1242 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1243 0 : self.remote_client.remote_consistent_lsn_projected()
1244 0 : }
1245 :
1246 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1247 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1248 : /// generation validation in the deletion queue.
1249 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1250 0 : self.remote_client.remote_consistent_lsn_visible()
1251 0 : }
1252 :
1253 : /// The sum of the file size of all historic layers in the layer map.
1254 : /// This method makes no distinction between local and remote layers.
1255 : /// Hence, the result **does not represent local filesystem usage**.
1256 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1257 0 : let guard = self.layers.read().await;
1258 0 : guard.layer_size_sum()
1259 0 : }
1260 :
1261 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1262 0 : self.metrics.resident_physical_size_get()
1263 0 : }
1264 :
1265 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1266 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1267 0 : }
1268 :
1269 : ///
1270 : /// Wait until WAL has been received and processed up to this LSN.
1271 : ///
1272 : /// You should call this before any of the other get_* or list_* functions. Calling
1273 : /// those functions with an LSN that has been processed yet is an error.
1274 : ///
1275 224854 : pub(crate) async fn wait_lsn(
1276 224854 : &self,
1277 224854 : lsn: Lsn,
1278 224854 : who_is_waiting: WaitLsnWaiter<'_>,
1279 224854 : ctx: &RequestContext, /* Prepare for use by cancellation */
1280 224854 : ) -> Result<(), WaitLsnError> {
1281 224854 : let state = self.current_state();
1282 224854 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1283 0 : return Err(WaitLsnError::Shutdown);
1284 224854 : } else if !matches!(state, TimelineState::Active) {
1285 0 : return Err(WaitLsnError::BadState(state));
1286 224854 : }
1287 224854 :
1288 224854 : if cfg!(debug_assertions) {
1289 224854 : match ctx.task_kind() {
1290 : TaskKind::WalReceiverManager
1291 : | TaskKind::WalReceiverConnectionHandler
1292 : | TaskKind::WalReceiverConnectionPoller => {
1293 0 : let is_myself = match who_is_waiting {
1294 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1295 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1296 : };
1297 0 : if is_myself {
1298 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1299 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1300 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1301 0 : }
1302 0 : } else {
1303 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1304 0 : // our walreceiver task can make progress independent of theirs
1305 0 : }
1306 : }
1307 224854 : _ => {}
1308 : }
1309 0 : }
1310 :
1311 224854 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1312 224854 :
1313 224854 : match self
1314 224854 : .last_record_lsn
1315 224854 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1316 0 : .await
1317 : {
1318 224854 : Ok(()) => Ok(()),
1319 0 : Err(e) => {
1320 : use utils::seqwait::SeqWaitError::*;
1321 0 : match e {
1322 0 : Shutdown => Err(WaitLsnError::Shutdown),
1323 : Timeout => {
1324 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1325 0 : drop(_timer);
1326 0 : let walreceiver_status = self.walreceiver_status();
1327 0 : Err(WaitLsnError::Timeout(format!(
1328 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1329 0 : lsn,
1330 0 : self.get_last_record_lsn(),
1331 0 : self.get_disk_consistent_lsn(),
1332 0 : walreceiver_status,
1333 0 : )))
1334 : }
1335 : }
1336 : }
1337 : }
1338 224854 : }
1339 :
1340 0 : pub(crate) fn walreceiver_status(&self) -> String {
1341 0 : match &*self.walreceiver.lock().unwrap() {
1342 0 : None => "stopping or stopped".to_string(),
1343 0 : Some(walreceiver) => match walreceiver.status() {
1344 0 : Some(status) => status.to_human_readable_string(),
1345 0 : None => "Not active".to_string(),
1346 : },
1347 : }
1348 0 : }
1349 :
1350 : /// Check that it is valid to request operations with that lsn.
1351 232 : pub(crate) fn check_lsn_is_in_scope(
1352 232 : &self,
1353 232 : lsn: Lsn,
1354 232 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1355 232 : ) -> anyhow::Result<()> {
1356 232 : ensure!(
1357 232 : lsn >= **latest_gc_cutoff_lsn,
1358 4 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1359 4 : lsn,
1360 4 : **latest_gc_cutoff_lsn,
1361 : );
1362 228 : Ok(())
1363 232 : }
1364 :
1365 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1366 10 : pub(crate) fn init_lsn_lease(
1367 10 : &self,
1368 10 : lsn: Lsn,
1369 10 : length: Duration,
1370 10 : ctx: &RequestContext,
1371 10 : ) -> anyhow::Result<LsnLease> {
1372 10 : self.make_lsn_lease(lsn, length, true, ctx)
1373 10 : }
1374 :
1375 : /// Renews a lease at a particular LSN. The requested LSN is not validated against the `latest_gc_cutoff_lsn` when we are in the grace period.
1376 4 : pub(crate) fn renew_lsn_lease(
1377 4 : &self,
1378 4 : lsn: Lsn,
1379 4 : length: Duration,
1380 4 : ctx: &RequestContext,
1381 4 : ) -> anyhow::Result<LsnLease> {
1382 4 : self.make_lsn_lease(lsn, length, false, ctx)
1383 4 : }
1384 :
1385 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1386 : ///
1387 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1388 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1389 : ///
1390 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1391 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1392 14 : fn make_lsn_lease(
1393 14 : &self,
1394 14 : lsn: Lsn,
1395 14 : length: Duration,
1396 14 : init: bool,
1397 14 : _ctx: &RequestContext,
1398 14 : ) -> anyhow::Result<LsnLease> {
1399 12 : let lease = {
1400 : // Normalize the requested LSN to be aligned, and move to the first record
1401 : // if it points to the beginning of the page (header).
1402 14 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1403 14 :
1404 14 : let mut gc_info = self.gc_info.write().unwrap();
1405 14 :
1406 14 : let valid_until = SystemTime::now() + length;
1407 14 :
1408 14 : let entry = gc_info.leases.entry(lsn);
1409 14 :
1410 14 : match entry {
1411 6 : Entry::Occupied(mut occupied) => {
1412 6 : let existing_lease = occupied.get_mut();
1413 6 : if valid_until > existing_lease.valid_until {
1414 2 : existing_lease.valid_until = valid_until;
1415 2 : let dt: DateTime<Utc> = valid_until.into();
1416 2 : info!("lease extended to {}", dt);
1417 : } else {
1418 4 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1419 4 : info!("existing lease covers greater length, valid until {}", dt);
1420 : }
1421 :
1422 6 : existing_lease.clone()
1423 : }
1424 8 : Entry::Vacant(vacant) => {
1425 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
1426 : // not blocked by the lsn lease deadline.
1427 8 : let validate = {
1428 8 : let conf = self.tenant_conf.load();
1429 8 : conf.location.attach_mode == AttachmentMode::Single
1430 8 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1431 : };
1432 :
1433 8 : if init || validate {
1434 8 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1435 8 : if lsn < *latest_gc_cutoff_lsn {
1436 2 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1437 6 : }
1438 0 : }
1439 :
1440 6 : let dt: DateTime<Utc> = valid_until.into();
1441 6 : info!("lease created, valid until {}", dt);
1442 6 : vacant.insert(LsnLease { valid_until }).clone()
1443 : }
1444 : }
1445 : };
1446 :
1447 12 : Ok(lease)
1448 14 : }
1449 :
1450 : /// Flush to disk all data that was written with the put_* functions
1451 1092 : #[instrument(skip(self), fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%self.timeline_id))]
1452 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1453 : self.freeze_and_flush0().await
1454 : }
1455 :
1456 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1457 : // polluting the span hierarchy.
1458 1092 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1459 1092 : let token = {
1460 : // Freeze the current open in-memory layer. It will be written to disk on next
1461 : // iteration.
1462 1092 : let mut g = self.write_lock.lock().await;
1463 :
1464 1092 : let to_lsn = self.get_last_record_lsn();
1465 1092 : self.freeze_inmem_layer_at(to_lsn, &mut g).await?
1466 : };
1467 1092 : self.wait_flush_completion(token).await
1468 1092 : }
1469 :
1470 : // Check if an open ephemeral layer should be closed: this provides
1471 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1472 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1473 : // ephemeral layer bytes has been breached.
1474 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1475 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1476 : // If the write lock is held, there is an active wal receiver: rolling open layers
1477 : // is their responsibility while they hold this lock.
1478 0 : return;
1479 : };
1480 :
1481 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1482 : // time, and this was missed.
1483 : // if write_guard.is_none() { return; }
1484 :
1485 0 : let Ok(layers_guard) = self.layers.try_read() else {
1486 : // Don't block if the layer lock is busy
1487 0 : return;
1488 : };
1489 :
1490 0 : let Ok(lm) = layers_guard.layer_map() else {
1491 0 : return;
1492 : };
1493 :
1494 0 : let Some(open_layer) = &lm.open_layer else {
1495 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1496 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1497 : // that didn't result in writes to this shard.
1498 :
1499 : // Must not hold the layers lock while waiting for a flush.
1500 0 : drop(layers_guard);
1501 0 :
1502 0 : let last_record_lsn = self.get_last_record_lsn();
1503 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1504 0 : if last_record_lsn > disk_consistent_lsn {
1505 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1506 : // we are a sharded tenant and have skipped some WAL
1507 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1508 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1509 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1510 : // without any data ingested (yet) doesn't write a remote index as soon as it
1511 : // sees its LSN advance: we only do this if we've been layer-less
1512 : // for some time.
1513 0 : tracing::debug!(
1514 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1515 : disk_consistent_lsn,
1516 : last_record_lsn
1517 : );
1518 :
1519 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1520 : // We know there is no open layer, so we can request freezing without actually
1521 : // freezing anything. This is true even if we have dropped the layers_guard, we
1522 : // still hold the write_guard.
1523 0 : let _ = async {
1524 0 : let token = self
1525 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1526 0 : .await?;
1527 0 : self.wait_flush_completion(token).await
1528 0 : }
1529 0 : .await;
1530 0 : }
1531 0 : }
1532 :
1533 0 : return;
1534 : };
1535 :
1536 0 : let Some(current_size) = open_layer.try_len() else {
1537 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1538 : // read lock to get size should always succeed.
1539 0 : tracing::warn!("Lock conflict while reading size of open layer");
1540 0 : return;
1541 : };
1542 :
1543 0 : let current_lsn = self.get_last_record_lsn();
1544 :
1545 0 : let checkpoint_distance_override = open_layer.tick().await;
1546 :
1547 0 : if let Some(size_override) = checkpoint_distance_override {
1548 0 : if current_size > size_override {
1549 : // This is not harmful, but it only happens in relatively rare cases where
1550 : // time-based checkpoints are not happening fast enough to keep the amount of
1551 : // ephemeral data within configured limits. It's a sign of stress on the system.
1552 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1553 0 : }
1554 0 : }
1555 :
1556 0 : let checkpoint_distance =
1557 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1558 0 :
1559 0 : if self.should_roll(
1560 0 : current_size,
1561 0 : current_size,
1562 0 : checkpoint_distance,
1563 0 : self.get_last_record_lsn(),
1564 0 : self.last_freeze_at.load(),
1565 0 : open_layer.get_opened_at(),
1566 0 : ) {
1567 0 : match open_layer.info() {
1568 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1569 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1570 0 : // happens asynchronously in the background.
1571 0 : tracing::debug!(
1572 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1573 : );
1574 : }
1575 : InMemoryLayerInfo::Open { .. } => {
1576 : // Upgrade to a write lock and freeze the layer
1577 0 : drop(layers_guard);
1578 0 : let res = self
1579 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1580 0 : .await;
1581 :
1582 0 : if let Err(e) = res {
1583 0 : tracing::info!(
1584 0 : "failed to flush frozen layer after background freeze: {e:#}"
1585 : );
1586 0 : }
1587 : }
1588 : }
1589 0 : }
1590 0 : }
1591 :
1592 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1593 : ///
1594 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1595 : /// child timelines that are not offloaded yet.
1596 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1597 0 : if self.remote_client.is_archived() != Some(true) {
1598 0 : return (false, "the timeline is not archived");
1599 0 : }
1600 0 : if !self.remote_client.no_pending_work() {
1601 : // if the remote client is still processing some work, we can't offload
1602 0 : return (false, "the upload queue is not drained yet");
1603 0 : }
1604 0 :
1605 0 : (true, "ok")
1606 0 : }
1607 :
1608 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1609 : /// compaction tasks.
1610 364 : pub(crate) async fn compact(
1611 364 : self: &Arc<Self>,
1612 364 : cancel: &CancellationToken,
1613 364 : flags: EnumSet<CompactFlags>,
1614 364 : ctx: &RequestContext,
1615 364 : ) -> Result<bool, CompactionError> {
1616 364 : // most likely the cancellation token is from background task, but in tests it could be the
1617 364 : // request task as well.
1618 364 :
1619 364 : let prepare = async move {
1620 364 : let guard = self.compaction_lock.lock().await;
1621 :
1622 364 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1623 364 : BackgroundLoopKind::Compaction,
1624 364 : ctx,
1625 364 : )
1626 0 : .await;
1627 :
1628 364 : (guard, permit)
1629 364 : };
1630 :
1631 : // this wait probably never needs any "long time spent" logging, because we already nag if
1632 : // compaction task goes over it's period (20s) which is quite often in production.
1633 364 : let (_guard, _permit) = tokio::select! {
1634 364 : tuple = prepare => { tuple },
1635 364 : _ = self.cancel.cancelled() => return Ok(false),
1636 364 : _ = cancel.cancelled() => return Ok(false),
1637 : };
1638 :
1639 364 : let last_record_lsn = self.get_last_record_lsn();
1640 364 :
1641 364 : // Last record Lsn could be zero in case the timeline was just created
1642 364 : if !last_record_lsn.is_valid() {
1643 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1644 0 : return Ok(false);
1645 364 : }
1646 364 :
1647 364 : match self.get_compaction_algorithm_settings().kind {
1648 : CompactionAlgorithm::Tiered => {
1649 0 : self.compact_tiered(cancel, ctx).await?;
1650 0 : Ok(false)
1651 : }
1652 37131 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, flags, ctx).await,
1653 : }
1654 364 : }
1655 :
1656 : /// Mutate the timeline with a [`TimelineWriter`].
1657 5133162 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1658 5133162 : TimelineWriter {
1659 5133162 : tl: self,
1660 5133162 : write_guard: self.write_lock.lock().await,
1661 : }
1662 5133162 : }
1663 :
1664 0 : pub(crate) fn activate(
1665 0 : self: &Arc<Self>,
1666 0 : parent: Arc<crate::tenant::Tenant>,
1667 0 : broker_client: BrokerClientChannel,
1668 0 : background_jobs_can_start: Option<&completion::Barrier>,
1669 0 : ctx: &RequestContext,
1670 0 : ) {
1671 0 : if self.tenant_shard_id.is_shard_zero() {
1672 0 : // Logical size is only maintained accurately on shard zero.
1673 0 : self.spawn_initial_logical_size_computation_task(ctx);
1674 0 : }
1675 0 : self.launch_wal_receiver(ctx, broker_client);
1676 0 : self.set_state(TimelineState::Active);
1677 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1678 0 : }
1679 :
1680 : /// After this function returns, there are no timeline-scoped tasks are left running.
1681 : ///
1682 : /// The preferred pattern for is:
1683 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1684 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1685 : /// go the extra mile and keep track of JoinHandles
1686 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1687 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1688 : ///
1689 : /// For legacy reasons, we still have multiple tasks spawned using
1690 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1691 : /// We refer to these as "timeline-scoped task_mgr tasks".
1692 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1693 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1694 : /// or [`task_mgr::shutdown_watcher`].
1695 : /// We want to gradually convert the code base away from these.
1696 : ///
1697 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1698 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1699 : /// ones that aren't mentioned here):
1700 : /// - [`TaskKind::TimelineDeletionWorker`]
1701 : /// - NB: also used for tenant deletion
1702 : /// - [`TaskKind::RemoteUploadTask`]`
1703 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1704 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1705 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1706 : /// - [`TaskKind::Eviction`]
1707 : /// - [`TaskKind::LayerFlushTask`]
1708 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1709 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1710 10 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1711 10 : debug_assert_current_span_has_tenant_and_timeline_id();
1712 10 :
1713 10 : // Regardless of whether we're going to try_freeze_and_flush
1714 10 : // or not, stop ingesting any more data. Walreceiver only provides
1715 10 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1716 10 : // So, only after the self.gate.close() below will we know for sure that
1717 10 : // no walreceiver tasks are left.
1718 10 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1719 10 : // data during the call to `self.freeze_and_flush()` below.
1720 10 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1721 10 : // which is what we'd need to properly model early shutdown of the walreceiver
1722 10 : // task sub-tree before the other Timeline task sub-trees.
1723 10 : let walreceiver = self.walreceiver.lock().unwrap().take();
1724 10 : tracing::debug!(
1725 0 : is_some = walreceiver.is_some(),
1726 0 : "Waiting for WalReceiverManager..."
1727 : );
1728 10 : if let Some(walreceiver) = walreceiver {
1729 0 : walreceiver.cancel();
1730 10 : }
1731 : // ... and inform any waiters for newer LSNs that there won't be any.
1732 10 : self.last_record_lsn.shutdown();
1733 10 :
1734 10 : if let ShutdownMode::FreezeAndFlush = mode {
1735 6 : if let Some((open, frozen)) = self
1736 6 : .layers
1737 6 : .read()
1738 0 : .await
1739 6 : .layer_map()
1740 6 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1741 6 : .ok()
1742 6 : .filter(|(open, frozen)| *open || *frozen > 0)
1743 : {
1744 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1745 6 : } else {
1746 6 : // this is double-shutdown, ignore it
1747 6 : }
1748 :
1749 : // we shut down walreceiver above, so, we won't add anything more
1750 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1751 : // to reach the disk & upload queue, then shut the upload queue and
1752 : // wait for it to drain.
1753 6 : match self.freeze_and_flush().await {
1754 : Ok(_) => {
1755 : // drain the upload queue
1756 : // if we did not wait for completion here, it might be our shutdown process
1757 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1758 : // be spawned.
1759 : //
1760 : // what is problematic is the shutting down of RemoteTimelineClient, because
1761 : // obviously it does not make sense to stop while we wait for it, but what
1762 : // about corner cases like s3 suddenly hanging up?
1763 6 : self.remote_client.shutdown().await;
1764 : }
1765 : Err(FlushLayerError::Cancelled) => {
1766 : // this is likely the second shutdown, ignore silently.
1767 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1768 0 : debug_assert!(self.cancel.is_cancelled());
1769 : }
1770 0 : Err(e) => {
1771 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1772 0 : // we have some extra WAL replay to do next time the timeline starts.
1773 0 : warn!("failed to freeze and flush: {e:#}");
1774 : }
1775 : }
1776 :
1777 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
1778 : // we also do a final check here to ensure that the queue is empty.
1779 6 : if !self.remote_client.no_pending_work() {
1780 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1781 6 : }
1782 4 : }
1783 :
1784 10 : if let ShutdownMode::Flush = mode {
1785 : // drain the upload queue
1786 2 : self.remote_client.shutdown().await;
1787 2 : if !self.remote_client.no_pending_work() {
1788 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1789 2 : }
1790 8 : }
1791 :
1792 : // Signal any subscribers to our cancellation token to drop out
1793 10 : tracing::debug!("Cancelling CancellationToken");
1794 10 : self.cancel.cancel();
1795 10 :
1796 10 : // Ensure Prevent new page service requests from starting.
1797 10 : self.handles.shutdown();
1798 10 :
1799 10 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1800 10 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1801 10 : self.remote_client.stop();
1802 10 :
1803 10 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1804 10 : // to shut down the upload queue tasks.
1805 10 : // TODO: fix that, task management should be encapsulated inside remote_client.
1806 10 : task_mgr::shutdown_tasks(
1807 10 : Some(TaskKind::RemoteUploadTask),
1808 10 : Some(self.tenant_shard_id),
1809 10 : Some(self.timeline_id),
1810 10 : )
1811 0 : .await;
1812 :
1813 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
1814 10 : tracing::debug!("Waiting for tasks...");
1815 10 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1816 :
1817 : {
1818 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
1819 : // open.
1820 10 : let mut write_guard = self.write_lock.lock().await;
1821 10 : self.layers.write().await.shutdown(&mut write_guard);
1822 10 : }
1823 10 :
1824 10 : // Finally wait until any gate-holders are complete.
1825 10 : //
1826 10 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1827 10 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1828 10 : self.gate.close().await;
1829 :
1830 10 : self.metrics.shutdown();
1831 10 : }
1832 :
1833 420 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1834 420 : match (self.current_state(), new_state) {
1835 420 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1836 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1837 : }
1838 0 : (st, TimelineState::Loading) => {
1839 0 : error!("ignoring transition from {st:?} into Loading state");
1840 : }
1841 0 : (TimelineState::Broken { .. }, new_state) => {
1842 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1843 : }
1844 : (TimelineState::Stopping, TimelineState::Active) => {
1845 0 : error!("Not activating a Stopping timeline");
1846 : }
1847 418 : (_, new_state) => {
1848 418 : self.state.send_replace(new_state);
1849 418 : }
1850 : }
1851 420 : }
1852 :
1853 2 : pub(crate) fn set_broken(&self, reason: String) {
1854 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1855 2 : let broken_state = TimelineState::Broken {
1856 2 : reason,
1857 2 : backtrace: backtrace_str,
1858 2 : };
1859 2 : self.set_state(broken_state);
1860 2 :
1861 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1862 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1863 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1864 2 : self.cancel.cancel();
1865 2 : }
1866 :
1867 225864 : pub(crate) fn current_state(&self) -> TimelineState {
1868 225864 : self.state.borrow().clone()
1869 225864 : }
1870 :
1871 6 : pub(crate) fn is_broken(&self) -> bool {
1872 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1873 6 : }
1874 :
1875 222 : pub(crate) fn is_active(&self) -> bool {
1876 222 : self.current_state() == TimelineState::Active
1877 222 : }
1878 :
1879 2 : pub(crate) fn is_archived(&self) -> Option<bool> {
1880 2 : self.remote_client.is_archived()
1881 2 : }
1882 :
1883 368 : pub(crate) fn is_stopping(&self) -> bool {
1884 368 : self.current_state() == TimelineState::Stopping
1885 368 : }
1886 :
1887 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1888 0 : self.state.subscribe()
1889 0 : }
1890 :
1891 224856 : pub(crate) async fn wait_to_become_active(
1892 224856 : &self,
1893 224856 : _ctx: &RequestContext, // Prepare for use by cancellation
1894 224856 : ) -> Result<(), TimelineState> {
1895 224856 : let mut receiver = self.state.subscribe();
1896 : loop {
1897 224856 : let current_state = receiver.borrow().clone();
1898 224856 : match current_state {
1899 : TimelineState::Loading => {
1900 0 : receiver
1901 0 : .changed()
1902 0 : .await
1903 0 : .expect("holding a reference to self");
1904 : }
1905 : TimelineState::Active { .. } => {
1906 224854 : return Ok(());
1907 : }
1908 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1909 : // There's no chance the timeline can transition back into ::Active
1910 2 : return Err(current_state);
1911 : }
1912 : }
1913 : }
1914 224856 : }
1915 :
1916 0 : pub(crate) async fn layer_map_info(
1917 0 : &self,
1918 0 : reset: LayerAccessStatsReset,
1919 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
1920 0 : let guard = self.layers.read().await;
1921 0 : let layer_map = guard.layer_map()?;
1922 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
1923 0 : if let Some(open_layer) = &layer_map.open_layer {
1924 0 : in_memory_layers.push(open_layer.info());
1925 0 : }
1926 0 : for frozen_layer in &layer_map.frozen_layers {
1927 0 : in_memory_layers.push(frozen_layer.info());
1928 0 : }
1929 :
1930 0 : let historic_layers = layer_map
1931 0 : .iter_historic_layers()
1932 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
1933 0 : .collect();
1934 0 :
1935 0 : Ok(LayerMapInfo {
1936 0 : in_memory_layers,
1937 0 : historic_layers,
1938 0 : })
1939 0 : }
1940 :
1941 0 : #[instrument(skip_all, fields(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id))]
1942 : pub(crate) async fn download_layer(
1943 : &self,
1944 : layer_file_name: &LayerName,
1945 : ) -> anyhow::Result<Option<bool>> {
1946 : let Some(layer) = self.find_layer(layer_file_name).await? else {
1947 : return Ok(None);
1948 : };
1949 :
1950 : layer.download().await?;
1951 :
1952 : Ok(Some(true))
1953 : }
1954 :
1955 : /// Evict just one layer.
1956 : ///
1957 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
1958 0 : pub(crate) async fn evict_layer(
1959 0 : &self,
1960 0 : layer_file_name: &LayerName,
1961 0 : ) -> anyhow::Result<Option<bool>> {
1962 0 : let _gate = self
1963 0 : .gate
1964 0 : .enter()
1965 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
1966 :
1967 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
1968 0 : return Ok(None);
1969 : };
1970 :
1971 : // curl has this by default
1972 0 : let timeout = std::time::Duration::from_secs(120);
1973 0 :
1974 0 : match local_layer.evict_and_wait(timeout).await {
1975 0 : Ok(()) => Ok(Some(true)),
1976 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
1977 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
1978 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
1979 : }
1980 0 : }
1981 :
1982 4803010 : fn should_roll(
1983 4803010 : &self,
1984 4803010 : layer_size: u64,
1985 4803010 : projected_layer_size: u64,
1986 4803010 : checkpoint_distance: u64,
1987 4803010 : projected_lsn: Lsn,
1988 4803010 : last_freeze_at: Lsn,
1989 4803010 : opened_at: Instant,
1990 4803010 : ) -> bool {
1991 4803010 : let distance = projected_lsn.widening_sub(last_freeze_at);
1992 4803010 :
1993 4803010 : // Rolling the open layer can be triggered by:
1994 4803010 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
1995 4803010 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
1996 4803010 : // account for how writes are distributed across shards: we expect each node to consume
1997 4803010 : // 1/count of the LSN on average.
1998 4803010 : // 2. The size of the currently open layer.
1999 4803010 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2000 4803010 : // up and suspend activity.
2001 4803010 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2002 0 : info!(
2003 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2004 : projected_lsn, layer_size, distance
2005 : );
2006 :
2007 0 : true
2008 4803010 : } else if projected_layer_size >= checkpoint_distance {
2009 : // NB: this check is relied upon by:
2010 80 : let _ = IndexEntry::validate_checkpoint_distance;
2011 80 : info!(
2012 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2013 : projected_lsn, layer_size, projected_layer_size
2014 : );
2015 :
2016 80 : true
2017 4802930 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2018 0 : info!(
2019 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2020 0 : projected_lsn,
2021 0 : layer_size,
2022 0 : opened_at.elapsed()
2023 : );
2024 :
2025 0 : true
2026 : } else {
2027 4802930 : false
2028 : }
2029 4803010 : }
2030 : }
2031 :
2032 : /// Number of times we will compute partition within a checkpoint distance.
2033 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2034 :
2035 : // Private functions
2036 : impl Timeline {
2037 12 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2038 12 : let tenant_conf = self.tenant_conf.load();
2039 12 : tenant_conf
2040 12 : .tenant_conf
2041 12 : .lsn_lease_length
2042 12 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2043 12 : }
2044 :
2045 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2046 0 : let tenant_conf = self.tenant_conf.load();
2047 0 : tenant_conf
2048 0 : .tenant_conf
2049 0 : .lsn_lease_length_for_ts
2050 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2051 0 : }
2052 :
2053 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2054 0 : let tenant_conf = self.tenant_conf.load();
2055 0 : tenant_conf
2056 0 : .tenant_conf
2057 0 : .lazy_slru_download
2058 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2059 0 : }
2060 :
2061 4804616 : fn get_checkpoint_distance(&self) -> u64 {
2062 4804616 : let tenant_conf = self.tenant_conf.load();
2063 4804616 : tenant_conf
2064 4804616 : .tenant_conf
2065 4804616 : .checkpoint_distance
2066 4804616 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2067 4804616 : }
2068 :
2069 4802930 : fn get_checkpoint_timeout(&self) -> Duration {
2070 4802930 : let tenant_conf = self.tenant_conf.load();
2071 4802930 : tenant_conf
2072 4802930 : .tenant_conf
2073 4802930 : .checkpoint_timeout
2074 4802930 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2075 4802930 : }
2076 :
2077 618 : fn get_compaction_target_size(&self) -> u64 {
2078 618 : let tenant_conf = self.tenant_conf.load();
2079 618 : tenant_conf
2080 618 : .tenant_conf
2081 618 : .compaction_target_size
2082 618 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2083 618 : }
2084 :
2085 1536 : fn get_compaction_threshold(&self) -> usize {
2086 1536 : let tenant_conf = self.tenant_conf.load();
2087 1536 : tenant_conf
2088 1536 : .tenant_conf
2089 1536 : .compaction_threshold
2090 1536 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2091 1536 : }
2092 :
2093 14 : fn get_image_creation_threshold(&self) -> usize {
2094 14 : let tenant_conf = self.tenant_conf.load();
2095 14 : tenant_conf
2096 14 : .tenant_conf
2097 14 : .image_creation_threshold
2098 14 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2099 14 : }
2100 :
2101 364 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2102 364 : let tenant_conf = &self.tenant_conf.load();
2103 364 : tenant_conf
2104 364 : .tenant_conf
2105 364 : .compaction_algorithm
2106 364 : .as_ref()
2107 364 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2108 364 : .clone()
2109 364 : }
2110 :
2111 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2112 0 : let tenant_conf = self.tenant_conf.load();
2113 0 : tenant_conf
2114 0 : .tenant_conf
2115 0 : .eviction_policy
2116 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2117 0 : }
2118 :
2119 418 : fn get_evictions_low_residence_duration_metric_threshold(
2120 418 : tenant_conf: &TenantConfOpt,
2121 418 : default_tenant_conf: &TenantConf,
2122 418 : ) -> Duration {
2123 418 : tenant_conf
2124 418 : .evictions_low_residence_duration_metric_threshold
2125 418 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2126 418 : }
2127 :
2128 716 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2129 716 : let tenant_conf = self.tenant_conf.load();
2130 716 : tenant_conf
2131 716 : .tenant_conf
2132 716 : .image_layer_creation_check_threshold
2133 716 : .unwrap_or(
2134 716 : self.conf
2135 716 : .default_tenant_conf
2136 716 : .image_layer_creation_check_threshold,
2137 716 : )
2138 716 : }
2139 :
2140 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2141 0 : // NB: Most tenant conf options are read by background loops, so,
2142 0 : // changes will automatically be picked up.
2143 0 :
2144 0 : // The threshold is embedded in the metric. So, we need to update it.
2145 0 : {
2146 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2147 0 : new_conf,
2148 0 : &self.conf.default_tenant_conf,
2149 0 : );
2150 0 :
2151 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2152 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2153 0 :
2154 0 : let timeline_id_str = self.timeline_id.to_string();
2155 0 : self.metrics
2156 0 : .evictions_with_low_residence_duration
2157 0 : .write()
2158 0 : .unwrap()
2159 0 : .change_threshold(
2160 0 : &tenant_id_str,
2161 0 : &shard_id_str,
2162 0 : &timeline_id_str,
2163 0 : new_threshold,
2164 0 : );
2165 0 : }
2166 0 : }
2167 :
2168 : /// Open a Timeline handle.
2169 : ///
2170 : /// Loads the metadata for the timeline into memory, but not the layer map.
2171 : #[allow(clippy::too_many_arguments)]
2172 418 : pub(super) fn new(
2173 418 : conf: &'static PageServerConf,
2174 418 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2175 418 : metadata: &TimelineMetadata,
2176 418 : ancestor: Option<Arc<Timeline>>,
2177 418 : timeline_id: TimelineId,
2178 418 : tenant_shard_id: TenantShardId,
2179 418 : generation: Generation,
2180 418 : shard_identity: ShardIdentity,
2181 418 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2182 418 : resources: TimelineResources,
2183 418 : pg_version: u32,
2184 418 : state: TimelineState,
2185 418 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2186 418 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2187 418 : cancel: CancellationToken,
2188 418 : ) -> Arc<Self> {
2189 418 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2190 418 : let (state, _) = watch::channel(state);
2191 418 :
2192 418 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2193 418 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2194 418 :
2195 418 : let evictions_low_residence_duration_metric_threshold = {
2196 418 : let loaded_tenant_conf = tenant_conf.load();
2197 418 : Self::get_evictions_low_residence_duration_metric_threshold(
2198 418 : &loaded_tenant_conf.tenant_conf,
2199 418 : &conf.default_tenant_conf,
2200 418 : )
2201 : };
2202 :
2203 418 : if let Some(ancestor) = &ancestor {
2204 230 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2205 230 : // If we construct an explicit timeline object, it's obviously not offloaded
2206 230 : let is_offloaded = MaybeOffloaded::No;
2207 230 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2208 230 : }
2209 :
2210 418 : Arc::new_cyclic(|myself| {
2211 418 : let metrics = TimelineMetrics::new(
2212 418 : &tenant_shard_id,
2213 418 : &timeline_id,
2214 418 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2215 418 : "mtime",
2216 418 : evictions_low_residence_duration_metric_threshold,
2217 418 : ),
2218 418 : );
2219 418 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2220 :
2221 418 : let mut result = Timeline {
2222 418 : conf,
2223 418 : tenant_conf,
2224 418 : myself: myself.clone(),
2225 418 : timeline_id,
2226 418 : tenant_shard_id,
2227 418 : generation,
2228 418 : shard_identity,
2229 418 : pg_version,
2230 418 : layers: Default::default(),
2231 418 :
2232 418 : walredo_mgr,
2233 418 : walreceiver: Mutex::new(None),
2234 418 :
2235 418 : remote_client: Arc::new(resources.remote_client),
2236 418 :
2237 418 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2238 418 : last_record_lsn: SeqWait::new(RecordLsn {
2239 418 : last: disk_consistent_lsn,
2240 418 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2241 418 : }),
2242 418 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2243 418 :
2244 418 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2245 418 : last_freeze_ts: RwLock::new(Instant::now()),
2246 418 :
2247 418 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2248 418 :
2249 418 : ancestor_timeline: ancestor,
2250 418 : ancestor_lsn: metadata.ancestor_lsn(),
2251 418 :
2252 418 : metrics,
2253 418 :
2254 418 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2255 418 : &tenant_shard_id,
2256 418 : &timeline_id,
2257 418 : ),
2258 418 :
2259 2926 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2260 418 :
2261 418 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2262 418 :
2263 418 : layer_flush_start_tx,
2264 418 : layer_flush_done_tx,
2265 418 :
2266 418 : write_lock: tokio::sync::Mutex::new(None),
2267 418 :
2268 418 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2269 418 :
2270 418 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2271 418 : initdb_lsn: metadata.initdb_lsn(),
2272 418 :
2273 418 : current_logical_size: if disk_consistent_lsn.is_valid() {
2274 : // we're creating timeline data with some layer files existing locally,
2275 : // need to recalculate timeline's logical size based on data in the layers.
2276 234 : LogicalSize::deferred_initial(disk_consistent_lsn)
2277 : } else {
2278 : // we're creating timeline data without any layers existing locally,
2279 : // initial logical size is 0.
2280 184 : LogicalSize::empty_initial()
2281 : },
2282 418 : partitioning: tokio::sync::Mutex::new((
2283 418 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2284 418 : Lsn(0),
2285 418 : )),
2286 418 : repartition_threshold: 0,
2287 418 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2288 418 : last_image_layer_creation_check_instant: Mutex::new(None),
2289 418 :
2290 418 : last_received_wal: Mutex::new(None),
2291 418 : rel_size_cache: RwLock::new(RelSizeCache {
2292 418 : complete_as_of: disk_consistent_lsn,
2293 418 : map: HashMap::new(),
2294 418 : }),
2295 418 :
2296 418 : download_all_remote_layers_task_info: RwLock::new(None),
2297 418 :
2298 418 : state,
2299 418 :
2300 418 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2301 418 : EvictionTaskTimelineState::default(),
2302 418 : ),
2303 418 : delete_progress: TimelineDeleteProgress::default(),
2304 418 :
2305 418 : cancel,
2306 418 : gate: Gate::default(),
2307 418 :
2308 418 : compaction_lock: tokio::sync::Mutex::default(),
2309 418 : gc_lock: tokio::sync::Mutex::default(),
2310 418 :
2311 418 : standby_horizon: AtomicLsn::new(0),
2312 418 :
2313 418 : timeline_get_throttle: resources.timeline_get_throttle,
2314 418 :
2315 418 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2316 418 :
2317 418 : #[cfg(test)]
2318 418 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2319 418 :
2320 418 : l0_flush_global_state: resources.l0_flush_global_state,
2321 418 :
2322 418 : handles: Default::default(),
2323 418 :
2324 418 : attach_wal_lag_cooldown,
2325 418 :
2326 418 : create_idempotency,
2327 418 : };
2328 418 :
2329 418 : result.repartition_threshold =
2330 418 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2331 418 :
2332 418 : result
2333 418 : .metrics
2334 418 : .last_record_gauge
2335 418 : .set(disk_consistent_lsn.0 as i64);
2336 418 : result
2337 418 : })
2338 418 : }
2339 :
2340 588 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2341 588 : let Ok(guard) = self.gate.enter() else {
2342 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2343 0 : return;
2344 : };
2345 588 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2346 588 : match *flush_loop_state {
2347 412 : FlushLoopState::NotStarted => (),
2348 : FlushLoopState::Running { .. } => {
2349 176 : info!(
2350 0 : "skipping attempt to start flush_loop twice {}/{}",
2351 0 : self.tenant_shard_id, self.timeline_id
2352 : );
2353 176 : return;
2354 : }
2355 : FlushLoopState::Exited => {
2356 0 : warn!(
2357 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2358 0 : self.tenant_shard_id, self.timeline_id
2359 : );
2360 0 : return;
2361 : }
2362 : }
2363 :
2364 412 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2365 412 : let self_clone = Arc::clone(self);
2366 412 :
2367 412 : debug!("spawning flush loop");
2368 412 : *flush_loop_state = FlushLoopState::Running {
2369 412 : #[cfg(test)]
2370 412 : expect_initdb_optimization: false,
2371 412 : #[cfg(test)]
2372 412 : initdb_optimization_count: 0,
2373 412 : };
2374 412 : task_mgr::spawn(
2375 412 : task_mgr::BACKGROUND_RUNTIME.handle(),
2376 412 : task_mgr::TaskKind::LayerFlushTask,
2377 412 : self.tenant_shard_id,
2378 412 : Some(self.timeline_id),
2379 412 : "layer flush task",
2380 412 : async move {
2381 412 : let _guard = guard;
2382 412 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2383 18044 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2384 10 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2385 10 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2386 10 : *flush_loop_state = FlushLoopState::Exited;
2387 10 : Ok(())
2388 10 : }
2389 412 : .instrument(info_span!(parent: None, "layer flush task", tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id))
2390 : );
2391 588 : }
2392 :
2393 : /// Creates and starts the wal receiver.
2394 : ///
2395 : /// This function is expected to be called at most once per Timeline's lifecycle
2396 : /// when the timeline is activated.
2397 0 : fn launch_wal_receiver(
2398 0 : self: &Arc<Self>,
2399 0 : ctx: &RequestContext,
2400 0 : broker_client: BrokerClientChannel,
2401 0 : ) {
2402 0 : info!(
2403 0 : "launching WAL receiver for timeline {} of tenant {}",
2404 0 : self.timeline_id, self.tenant_shard_id
2405 : );
2406 :
2407 0 : let tenant_conf = self.tenant_conf.load();
2408 0 : let wal_connect_timeout = tenant_conf
2409 0 : .tenant_conf
2410 0 : .walreceiver_connect_timeout
2411 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2412 0 : let lagging_wal_timeout = tenant_conf
2413 0 : .tenant_conf
2414 0 : .lagging_wal_timeout
2415 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2416 0 : let max_lsn_wal_lag = tenant_conf
2417 0 : .tenant_conf
2418 0 : .max_lsn_wal_lag
2419 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2420 0 :
2421 0 : let mut guard = self.walreceiver.lock().unwrap();
2422 0 : assert!(
2423 0 : guard.is_none(),
2424 0 : "multiple launches / re-launches of WAL receiver are not supported"
2425 : );
2426 0 : *guard = Some(WalReceiver::start(
2427 0 : Arc::clone(self),
2428 0 : WalReceiverConf {
2429 0 : wal_connect_timeout,
2430 0 : lagging_wal_timeout,
2431 0 : max_lsn_wal_lag,
2432 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2433 0 : availability_zone: self.conf.availability_zone.clone(),
2434 0 : ingest_batch_size: self.conf.ingest_batch_size,
2435 0 : },
2436 0 : broker_client,
2437 0 : ctx,
2438 0 : ));
2439 0 : }
2440 :
2441 : /// Initialize with an empty layer map. Used when creating a new timeline.
2442 412 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2443 412 : let mut layers = self.layers.try_write().expect(
2444 412 : "in the context where we call this function, no other task has access to the object",
2445 412 : );
2446 412 : layers
2447 412 : .open_mut()
2448 412 : .expect("in this context the LayerManager must still be open")
2449 412 : .initialize_empty(Lsn(start_lsn.0));
2450 412 : }
2451 :
2452 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2453 : /// files.
2454 6 : pub(super) async fn load_layer_map(
2455 6 : &self,
2456 6 : disk_consistent_lsn: Lsn,
2457 6 : index_part: IndexPart,
2458 6 : ) -> anyhow::Result<()> {
2459 : use init::{Decision::*, Discovered, DismissedLayer};
2460 : use LayerName::*;
2461 :
2462 6 : let mut guard = self.layers.write().await;
2463 :
2464 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2465 6 :
2466 6 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2467 6 : // structs representing all files on disk
2468 6 : let timeline_path = self
2469 6 : .conf
2470 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2471 6 : let conf = self.conf;
2472 6 : let span = tracing::Span::current();
2473 6 :
2474 6 : // Copy to move into the task we're about to spawn
2475 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2476 :
2477 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2478 6 : move || {
2479 6 : let _g = span.entered();
2480 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2481 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2482 6 : let mut unrecognized_files = Vec::new();
2483 6 :
2484 6 : let mut path = timeline_path;
2485 :
2486 22 : for discovered in discovered {
2487 16 : let (name, kind) = match discovered {
2488 16 : Discovered::Layer(layer_file_name, local_metadata) => {
2489 16 : discovered_layers.push((layer_file_name, local_metadata));
2490 16 : continue;
2491 : }
2492 0 : Discovered::IgnoredBackup(path) => {
2493 0 : std::fs::remove_file(path)
2494 0 : .or_else(fs_ext::ignore_not_found)
2495 0 : .fatal_err("Removing .old file");
2496 0 : continue;
2497 : }
2498 0 : Discovered::Unknown(file_name) => {
2499 0 : // we will later error if there are any
2500 0 : unrecognized_files.push(file_name);
2501 0 : continue;
2502 : }
2503 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2504 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2505 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2506 : };
2507 0 : path.push(Utf8Path::new(&name));
2508 0 : init::cleanup(&path, kind)?;
2509 0 : path.pop();
2510 : }
2511 :
2512 6 : if !unrecognized_files.is_empty() {
2513 : // assume that if there are any there are many many.
2514 0 : let n = unrecognized_files.len();
2515 0 : let first = &unrecognized_files[..n.min(10)];
2516 0 : anyhow::bail!(
2517 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2518 0 : );
2519 6 : }
2520 6 :
2521 6 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2522 6 :
2523 6 : let mut loaded_layers = Vec::new();
2524 6 : let mut needs_cleanup = Vec::new();
2525 6 : let mut total_physical_size = 0;
2526 :
2527 22 : for (name, decision) in decided {
2528 16 : let decision = match decision {
2529 16 : Ok(decision) => decision,
2530 0 : Err(DismissedLayer::Future { local }) => {
2531 0 : if let Some(local) = local {
2532 0 : init::cleanup_future_layer(
2533 0 : &local.local_path,
2534 0 : &name,
2535 0 : disk_consistent_lsn,
2536 0 : )?;
2537 0 : }
2538 0 : needs_cleanup.push(name);
2539 0 : continue;
2540 : }
2541 0 : Err(DismissedLayer::LocalOnly(local)) => {
2542 0 : init::cleanup_local_only_file(&name, &local)?;
2543 : // this file never existed remotely, we will have to do rework
2544 0 : continue;
2545 : }
2546 0 : Err(DismissedLayer::BadMetadata(local)) => {
2547 0 : init::cleanup_local_file_for_remote(&local)?;
2548 : // this file never existed remotely, we will have to do rework
2549 0 : continue;
2550 : }
2551 : };
2552 :
2553 16 : match &name {
2554 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2555 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2556 : }
2557 :
2558 16 : tracing::debug!(layer=%name, ?decision, "applied");
2559 :
2560 16 : let layer = match decision {
2561 16 : Resident { local, remote } => {
2562 16 : total_physical_size += local.file_size;
2563 16 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2564 16 : .drop_eviction_guard()
2565 : }
2566 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2567 : };
2568 :
2569 16 : loaded_layers.push(layer);
2570 : }
2571 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2572 6 : }
2573 6 : })
2574 3 : .await
2575 6 : .map_err(anyhow::Error::new)
2576 6 : .and_then(|x| x)?;
2577 :
2578 6 : let num_layers = loaded_layers.len();
2579 6 :
2580 6 : guard
2581 6 : .open_mut()
2582 6 : .expect("layermanager must be open during init")
2583 6 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2584 6 :
2585 6 : self.remote_client
2586 6 : .schedule_layer_file_deletion(&needs_cleanup)?;
2587 6 : self.remote_client
2588 6 : .schedule_index_upload_for_file_changes()?;
2589 : // This barrier orders above DELETEs before any later operations.
2590 : // This is critical because code executing after the barrier might
2591 : // create again objects with the same key that we just scheduled for deletion.
2592 : // For example, if we just scheduled deletion of an image layer "from the future",
2593 : // later compaction might run again and re-create the same image layer.
2594 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2595 : // "same" here means same key range and LSN.
2596 : //
2597 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2598 : // the upload queue may execute the PUT first, then the DELETE.
2599 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2600 : //
2601 : // 1. a future image layer is created and uploaded
2602 : // 2. ps restart
2603 : // 3. the future layer from (1) is deleted during load layer map
2604 : // 4. image layer is re-created and uploaded
2605 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2606 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2607 : //
2608 : // See https://github.com/neondatabase/neon/issues/5878
2609 : //
2610 : // NB: generation numbers naturally protect against this because they disambiguate
2611 : // (1) and (4)
2612 6 : self.remote_client.schedule_barrier()?;
2613 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2614 : // on retry.
2615 :
2616 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2617 6 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2618 6 : self.update_layer_visibility().await?;
2619 :
2620 6 : info!(
2621 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2622 : num_layers, disk_consistent_lsn, total_physical_size
2623 : );
2624 :
2625 6 : timer.stop_and_record();
2626 6 : Ok(())
2627 6 : }
2628 :
2629 : /// Retrieve current logical size of the timeline.
2630 : ///
2631 : /// The size could be lagging behind the actual number, in case
2632 : /// the initial size calculation has not been run (gets triggered on the first size access).
2633 : ///
2634 : /// return size and boolean flag that shows if the size is exact
2635 0 : pub(crate) fn get_current_logical_size(
2636 0 : self: &Arc<Self>,
2637 0 : priority: GetLogicalSizePriority,
2638 0 : ctx: &RequestContext,
2639 0 : ) -> logical_size::CurrentLogicalSize {
2640 0 : if !self.tenant_shard_id.is_shard_zero() {
2641 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2642 : // when HTTP API is serving a GET for timeline zero, return zero
2643 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2644 0 : }
2645 0 :
2646 0 : let current_size = self.current_logical_size.current_size();
2647 0 : debug!("Current size: {current_size:?}");
2648 :
2649 0 : match (current_size.accuracy(), priority) {
2650 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2651 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2652 0 : // background task will eventually deliver an exact value, we're in no rush
2653 0 : }
2654 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2655 : // background task is not ready, but user is asking for it now;
2656 : // => make the background task skip the line
2657 : // (The alternative would be to calculate the size here, but,
2658 : // it can actually take a long time if the user has a lot of rels.
2659 : // And we'll inevitable need it again; So, let the background task do the work.)
2660 0 : match self
2661 0 : .current_logical_size
2662 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2663 0 : .get()
2664 : {
2665 0 : Some(cancel) => cancel.cancel(),
2666 : None => {
2667 0 : let state = self.current_state();
2668 0 : if matches!(
2669 0 : state,
2670 : TimelineState::Broken { .. } | TimelineState::Stopping
2671 0 : ) {
2672 0 :
2673 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2674 0 : // Don't make noise.
2675 0 : } else {
2676 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2677 0 : debug_assert!(false);
2678 : }
2679 : }
2680 : };
2681 : }
2682 : }
2683 :
2684 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2685 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2686 0 : let first = self
2687 0 : .current_logical_size
2688 0 : .did_return_approximate_to_walreceiver
2689 0 : .compare_exchange(
2690 0 : false,
2691 0 : true,
2692 0 : AtomicOrdering::Relaxed,
2693 0 : AtomicOrdering::Relaxed,
2694 0 : )
2695 0 : .is_ok();
2696 0 : if first {
2697 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2698 0 : }
2699 0 : }
2700 0 : }
2701 :
2702 0 : current_size
2703 0 : }
2704 :
2705 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2706 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2707 : // nothing to do for freshly created timelines;
2708 0 : assert_eq!(
2709 0 : self.current_logical_size.current_size().accuracy(),
2710 0 : logical_size::Accuracy::Exact,
2711 0 : );
2712 0 : self.current_logical_size.initialized.add_permits(1);
2713 0 : return;
2714 : };
2715 :
2716 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2717 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2718 0 : self.current_logical_size
2719 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2720 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2721 0 :
2722 0 : let self_clone = Arc::clone(self);
2723 0 : let background_ctx = ctx.detached_child(
2724 0 : TaskKind::InitialLogicalSizeCalculation,
2725 0 : DownloadBehavior::Download,
2726 0 : );
2727 0 : task_mgr::spawn(
2728 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2729 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2730 0 : self.tenant_shard_id,
2731 0 : Some(self.timeline_id),
2732 0 : "initial size calculation",
2733 : // NB: don't log errors here, task_mgr will do that.
2734 0 : async move {
2735 0 : let cancel = task_mgr::shutdown_token();
2736 0 : self_clone
2737 0 : .initial_logical_size_calculation_task(
2738 0 : initial_part_end,
2739 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2740 0 : cancel,
2741 0 : background_ctx,
2742 0 : )
2743 0 : .await;
2744 0 : Ok(())
2745 0 : }
2746 0 : .instrument(info_span!(parent: None, "initial_size_calculation", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%self.timeline_id)),
2747 : );
2748 0 : }
2749 :
2750 0 : async fn initial_logical_size_calculation_task(
2751 0 : self: Arc<Self>,
2752 0 : initial_part_end: Lsn,
2753 0 : skip_concurrency_limiter: CancellationToken,
2754 0 : cancel: CancellationToken,
2755 0 : background_ctx: RequestContext,
2756 0 : ) {
2757 0 : scopeguard::defer! {
2758 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2759 0 : self.current_logical_size.initialized.add_permits(1);
2760 0 : }
2761 0 :
2762 0 : let try_once = |attempt: usize| {
2763 0 : let background_ctx = &background_ctx;
2764 0 : let self_ref = &self;
2765 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2766 0 : async move {
2767 0 : let cancel = task_mgr::shutdown_token();
2768 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2769 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2770 0 : background_ctx,
2771 0 : );
2772 :
2773 : use crate::metrics::initial_logical_size::StartCircumstances;
2774 0 : let (_maybe_permit, circumstances) = tokio::select! {
2775 0 : permit = wait_for_permit => {
2776 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2777 : }
2778 0 : _ = self_ref.cancel.cancelled() => {
2779 0 : return Err(CalculateLogicalSizeError::Cancelled);
2780 : }
2781 0 : _ = cancel.cancelled() => {
2782 0 : return Err(CalculateLogicalSizeError::Cancelled);
2783 : },
2784 0 : () = skip_concurrency_limiter.cancelled() => {
2785 : // Some action that is part of a end user interaction requested logical size
2786 : // => break out of the rate limit
2787 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2788 : // but then again what happens if they cancel; also, we should just be using
2789 : // one runtime across the entire process, so, let's leave this for now.
2790 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2791 : }
2792 : };
2793 :
2794 0 : let metrics_guard = if attempt == 1 {
2795 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2796 : } else {
2797 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2798 : };
2799 :
2800 0 : let calculated_size = self_ref
2801 0 : .logical_size_calculation_task(
2802 0 : initial_part_end,
2803 0 : LogicalSizeCalculationCause::Initial,
2804 0 : background_ctx,
2805 0 : )
2806 0 : .await?;
2807 :
2808 0 : self_ref
2809 0 : .trigger_aux_file_size_computation(initial_part_end, background_ctx)
2810 0 : .await?;
2811 :
2812 : // TODO: add aux file size to logical size
2813 :
2814 0 : Ok((calculated_size, metrics_guard))
2815 0 : }
2816 0 : };
2817 :
2818 0 : let retrying = async {
2819 0 : let mut attempt = 0;
2820 : loop {
2821 0 : attempt += 1;
2822 0 :
2823 0 : match try_once(attempt).await {
2824 0 : Ok(res) => return ControlFlow::Continue(res),
2825 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
2826 : Err(
2827 0 : e @ (CalculateLogicalSizeError::Decode(_)
2828 0 : | CalculateLogicalSizeError::PageRead(_)),
2829 0 : ) => {
2830 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2831 : // exponential back-off doesn't make sense at these long intervals;
2832 : // use fixed retry interval with generous jitter instead
2833 0 : let sleep_duration = Duration::from_secs(
2834 0 : u64::try_from(
2835 0 : // 1hour base
2836 0 : (60_i64 * 60_i64)
2837 0 : // 10min jitter
2838 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2839 0 : )
2840 0 : .expect("10min < 1hour"),
2841 0 : );
2842 0 : tokio::time::sleep(sleep_duration).await;
2843 : }
2844 : }
2845 : }
2846 0 : };
2847 :
2848 0 : let (calculated_size, metrics_guard) = tokio::select! {
2849 0 : res = retrying => {
2850 0 : match res {
2851 0 : ControlFlow::Continue(calculated_size) => calculated_size,
2852 0 : ControlFlow::Break(()) => return,
2853 : }
2854 : }
2855 0 : _ = cancel.cancelled() => {
2856 0 : return;
2857 : }
2858 : };
2859 :
2860 : // we cannot query current_logical_size.current_size() to know the current
2861 : // *negative* value, only truncated to u64.
2862 0 : let added = self
2863 0 : .current_logical_size
2864 0 : .size_added_after_initial
2865 0 : .load(AtomicOrdering::Relaxed);
2866 0 :
2867 0 : let sum = calculated_size.saturating_add_signed(added);
2868 0 :
2869 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2870 0 : self.metrics.current_logical_size_gauge.set(sum);
2871 0 :
2872 0 : self.current_logical_size
2873 0 : .initial_logical_size
2874 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
2875 0 : .ok()
2876 0 : .expect("only this task sets it");
2877 0 : }
2878 :
2879 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
2880 0 : self: &Arc<Self>,
2881 0 : lsn: Lsn,
2882 0 : cause: LogicalSizeCalculationCause,
2883 0 : ctx: RequestContext,
2884 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
2885 0 : let (sender, receiver) = oneshot::channel();
2886 0 : let self_clone = Arc::clone(self);
2887 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
2888 0 : // we should stop the size calculation work and return an error.
2889 0 : // That would require restructuring this function's API to
2890 0 : // return the result directly, instead of a Receiver for the result.
2891 0 : let ctx = ctx.detached_child(
2892 0 : TaskKind::OndemandLogicalSizeCalculation,
2893 0 : DownloadBehavior::Download,
2894 0 : );
2895 0 : task_mgr::spawn(
2896 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2897 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
2898 0 : self.tenant_shard_id,
2899 0 : Some(self.timeline_id),
2900 0 : "ondemand logical size calculation",
2901 0 : async move {
2902 0 : let res = self_clone
2903 0 : .logical_size_calculation_task(lsn, cause, &ctx)
2904 0 : .await;
2905 0 : let _ = sender.send(res).ok();
2906 0 : Ok(()) // Receiver is responsible for handling errors
2907 0 : }
2908 0 : .in_current_span(),
2909 0 : );
2910 0 : receiver
2911 0 : }
2912 :
2913 : /// # Cancel-Safety
2914 : ///
2915 : /// This method is cancellation-safe.
2916 0 : #[instrument(skip_all)]
2917 : async fn logical_size_calculation_task(
2918 : self: &Arc<Self>,
2919 : lsn: Lsn,
2920 : cause: LogicalSizeCalculationCause,
2921 : ctx: &RequestContext,
2922 : ) -> Result<u64, CalculateLogicalSizeError> {
2923 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
2924 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
2925 : // accurate relation sizes, and they do not emit consumption metrics.
2926 : debug_assert!(self.tenant_shard_id.is_shard_zero());
2927 :
2928 : let guard = self
2929 : .gate
2930 : .enter()
2931 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
2932 :
2933 : let self_calculation = Arc::clone(self);
2934 :
2935 0 : let mut calculation = pin!(async {
2936 0 : let ctx = ctx.attached_child();
2937 0 : self_calculation
2938 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
2939 0 : .await
2940 0 : });
2941 :
2942 : tokio::select! {
2943 : res = &mut calculation => { res }
2944 : _ = self.cancel.cancelled() => {
2945 : debug!("cancelling logical size calculation for timeline shutdown");
2946 : calculation.await
2947 : }
2948 : }
2949 : }
2950 :
2951 : /// Calculate the logical size of the database at the latest LSN.
2952 : ///
2953 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
2954 : /// especially if we need to download remote layers.
2955 : ///
2956 : /// # Cancel-Safety
2957 : ///
2958 : /// This method is cancellation-safe.
2959 0 : async fn calculate_logical_size(
2960 0 : &self,
2961 0 : up_to_lsn: Lsn,
2962 0 : cause: LogicalSizeCalculationCause,
2963 0 : _guard: &GateGuard,
2964 0 : ctx: &RequestContext,
2965 0 : ) -> Result<u64, CalculateLogicalSizeError> {
2966 0 : info!(
2967 0 : "Calculating logical size for timeline {} at {}",
2968 : self.timeline_id, up_to_lsn
2969 : );
2970 :
2971 0 : pausable_failpoint!("timeline-calculate-logical-size-pause");
2972 :
2973 : // See if we've already done the work for initial size calculation.
2974 : // This is a short-cut for timelines that are mostly unused.
2975 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
2976 0 : return Ok(size);
2977 0 : }
2978 0 : let storage_time_metrics = match cause {
2979 : LogicalSizeCalculationCause::Initial
2980 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
2981 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
2982 : LogicalSizeCalculationCause::EvictionTaskImitation => {
2983 0 : &self.metrics.imitate_logical_size_histo
2984 : }
2985 : };
2986 0 : let timer = storage_time_metrics.start_timer();
2987 0 : let logical_size = self
2988 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
2989 0 : .await?;
2990 0 : debug!("calculated logical size: {logical_size}");
2991 0 : timer.stop_and_record();
2992 0 : Ok(logical_size)
2993 0 : }
2994 :
2995 : /// Update current logical size, adding `delta' to the old value.
2996 270570 : fn update_current_logical_size(&self, delta: i64) {
2997 270570 : let logical_size = &self.current_logical_size;
2998 270570 : logical_size.increment_size(delta);
2999 270570 :
3000 270570 : // Also set the value in the prometheus gauge. Note that
3001 270570 : // there is a race condition here: if this is is called by two
3002 270570 : // threads concurrently, the prometheus gauge might be set to
3003 270570 : // one value while current_logical_size is set to the
3004 270570 : // other.
3005 270570 : match logical_size.current_size() {
3006 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
3007 270570 : .metrics
3008 270570 : .current_logical_size_gauge
3009 270570 : .set(new_current_size.into()),
3010 0 : CurrentLogicalSize::Approximate(_) => {
3011 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3012 0 : // forth between the initial size calculation task.
3013 0 : }
3014 : }
3015 270570 : }
3016 :
3017 2834 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3018 2834 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3019 2834 : let aux_metric =
3020 2834 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3021 2834 :
3022 2834 : let sum_of_entries = self
3023 2834 : .directory_metrics
3024 2834 : .iter()
3025 19838 : .map(|v| v.load(AtomicOrdering::Relaxed))
3026 2834 : .sum();
3027 : // Set a high general threshold and a lower threshold for the auxiliary files,
3028 : // as we can have large numbers of relations in the db directory.
3029 : const SUM_THRESHOLD: u64 = 5000;
3030 : const AUX_THRESHOLD: u64 = 1000;
3031 2834 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3032 0 : self.metrics
3033 0 : .directory_entries_count_gauge
3034 0 : .set(sum_of_entries);
3035 2834 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3036 0 : metric.set(sum_of_entries);
3037 2834 : }
3038 2834 : }
3039 :
3040 0 : async fn find_layer(
3041 0 : &self,
3042 0 : layer_name: &LayerName,
3043 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3044 0 : let guard = self.layers.read().await;
3045 0 : let layer = guard
3046 0 : .layer_map()?
3047 0 : .iter_historic_layers()
3048 0 : .find(|l| &l.layer_name() == layer_name)
3049 0 : .map(|found| guard.get_from_desc(&found));
3050 0 : Ok(layer)
3051 0 : }
3052 :
3053 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3054 : /// indicating which layers are currently on-disk on the primary.
3055 : ///
3056 : /// None is returned if the Timeline is in a state where uploading a heatmap
3057 : /// doesn't make sense, such as shutting down or initializing. The caller
3058 : /// should treat this as a cue to simply skip doing any heatmap uploading
3059 : /// for this timeline.
3060 2 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3061 2 : if !self.is_active() {
3062 0 : return None;
3063 2 : }
3064 :
3065 2 : let guard = self.layers.read().await;
3066 :
3067 10 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3068 10 : match layer.visibility() {
3069 : LayerVisibilityHint::Visible => {
3070 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3071 8 : let last_activity_ts = layer.latest_activity();
3072 8 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
3073 : }
3074 : LayerVisibilityHint::Covered => {
3075 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3076 2 : None
3077 : }
3078 : }
3079 10 : });
3080 2 :
3081 2 : let mut layers = resident.collect::<Vec<_>>();
3082 2 :
3083 2 : // Sort layers in order of which to download first. For a large set of layers to download, we
3084 2 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3085 2 : // or hours later:
3086 2 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3087 2 : // only exist for a few minutes before being compacted into L1s.
3088 2 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3089 2 : // the layer is likely to be covered by an image layer during compaction.
3090 20 : layers.sort_by_key(|(desc, _meta, _atime)| {
3091 20 : std::cmp::Reverse((
3092 20 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3093 20 : desc.lsn_range.end,
3094 20 : ))
3095 20 : });
3096 2 :
3097 2 : let layers = layers
3098 2 : .into_iter()
3099 8 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3100 2 : .collect();
3101 2 :
3102 2 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3103 2 : }
3104 :
3105 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3106 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3107 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3108 0 : // branchpoint in the value in IndexPart::lineage
3109 0 : self.ancestor_lsn == lsn
3110 0 : || (self.ancestor_lsn == Lsn::INVALID
3111 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3112 0 : }
3113 : }
3114 :
3115 : impl Timeline {
3116 : #[allow(clippy::doc_lazy_continuation)]
3117 : /// Get the data needed to reconstruct all keys in the provided keyspace
3118 : ///
3119 : /// The algorithm is as follows:
3120 : /// 1. While some keys are still not done and there's a timeline to visit:
3121 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3122 : /// 2.1: Build the fringe for the current keyspace
3123 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3124 : /// intersects
3125 : /// 2.3. Pop the timeline from the fringe
3126 : /// 2.4. If the fringe is empty, go back to 1
3127 626570 : async fn get_vectored_reconstruct_data(
3128 626570 : &self,
3129 626570 : mut keyspace: KeySpace,
3130 626570 : request_lsn: Lsn,
3131 626570 : reconstruct_state: &mut ValuesReconstructState,
3132 626570 : ctx: &RequestContext,
3133 626570 : ) -> Result<(), GetVectoredError> {
3134 626570 : let mut timeline_owned: Arc<Timeline>;
3135 626570 : let mut timeline = self;
3136 626570 :
3137 626570 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3138 :
3139 626568 : let missing_keyspace = loop {
3140 851424 : if self.cancel.is_cancelled() {
3141 0 : return Err(GetVectoredError::Cancelled);
3142 851424 : }
3143 :
3144 : let TimelineVisitOutcome {
3145 851424 : completed_keyspace: completed,
3146 851424 : image_covered_keyspace,
3147 851424 : } = Self::get_vectored_reconstruct_data_timeline(
3148 851424 : timeline,
3149 851424 : keyspace.clone(),
3150 851424 : cont_lsn,
3151 851424 : reconstruct_state,
3152 851424 : &self.cancel,
3153 851424 : ctx,
3154 851424 : )
3155 192204 : .await?;
3156 :
3157 851424 : keyspace.remove_overlapping_with(&completed);
3158 851424 :
3159 851424 : // Do not descend into the ancestor timeline for aux files.
3160 851424 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3161 851424 : // stalling compaction.
3162 851424 : keyspace.remove_overlapping_with(&KeySpace {
3163 851424 : ranges: vec![NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE],
3164 851424 : });
3165 851424 :
3166 851424 : // Keyspace is fully retrieved
3167 851424 : if keyspace.is_empty() {
3168 626554 : break None;
3169 224870 : }
3170 :
3171 224870 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3172 : // Not fully retrieved but no ancestor timeline.
3173 14 : break Some(keyspace);
3174 : };
3175 :
3176 : // Now we see if there are keys covered by the image layer but does not exist in the
3177 : // image layer, which means that the key does not exist.
3178 :
3179 : // The block below will stop the vectored search if any of the keys encountered an image layer
3180 : // which did not contain a snapshot for said key. Since we have already removed all completed
3181 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3182 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3183 : // and stop the search as a result of that.
3184 224856 : let removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3185 224856 : if !removed.is_empty() {
3186 0 : break Some(removed);
3187 224856 : }
3188 224856 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3189 224856 : // keyspace.
3190 224856 :
3191 224856 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3192 224856 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3193 224856 : timeline_owned = timeline
3194 224856 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3195 2 : .await?;
3196 224854 : timeline = &*timeline_owned;
3197 : };
3198 :
3199 626568 : if let Some(missing_keyspace) = missing_keyspace {
3200 14 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3201 14 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3202 14 : shard: self
3203 14 : .shard_identity
3204 14 : .get_shard_number(&missing_keyspace.start().unwrap()),
3205 14 : cont_lsn,
3206 14 : request_lsn,
3207 14 : ancestor_lsn: Some(timeline.ancestor_lsn),
3208 14 : backtrace: None,
3209 14 : }));
3210 626554 : }
3211 626554 :
3212 626554 : Ok(())
3213 626570 : }
3214 :
3215 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3216 : ///
3217 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3218 : /// the current keyspace. The current keyspace of the search at any given timeline
3219 : /// is the original keyspace minus all the keys that have been completed minus
3220 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3221 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3222 : ///
3223 : /// This is basically a depth-first search visitor implementation where a vertex
3224 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3225 : ///
3226 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3227 : /// and get all the required reconstruct data from the layer in one go.
3228 : ///
3229 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3230 : /// decides how to deal with these two keyspaces.
3231 851424 : async fn get_vectored_reconstruct_data_timeline(
3232 851424 : timeline: &Timeline,
3233 851424 : keyspace: KeySpace,
3234 851424 : mut cont_lsn: Lsn,
3235 851424 : reconstruct_state: &mut ValuesReconstructState,
3236 851424 : cancel: &CancellationToken,
3237 851424 : ctx: &RequestContext,
3238 851424 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3239 851424 : let mut unmapped_keyspace = keyspace.clone();
3240 851424 : let mut fringe = LayerFringe::new();
3241 851424 :
3242 851424 : let mut completed_keyspace = KeySpace::default();
3243 851424 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3244 :
3245 : loop {
3246 1696854 : if cancel.is_cancelled() {
3247 0 : return Err(GetVectoredError::Cancelled);
3248 1696854 : }
3249 1696854 :
3250 1696854 : let (keys_done_last_step, keys_with_image_coverage) =
3251 1696854 : reconstruct_state.consume_done_keys();
3252 1696854 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3253 1696854 : completed_keyspace.merge(&keys_done_last_step);
3254 1696854 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3255 21808 : unmapped_keyspace
3256 21808 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3257 21808 : image_covered_keyspace.add_range(keys_with_image_coverage);
3258 1675046 : }
3259 :
3260 : // Do not descent any further if the last layer we visited
3261 : // completed all keys in the keyspace it inspected. This is not
3262 : // required for correctness, but avoids visiting extra layers
3263 : // which turns out to be a perf bottleneck in some cases.
3264 1696854 : if !unmapped_keyspace.is_empty() {
3265 1072544 : let guard = timeline.layers.read().await;
3266 1072544 : let layers = guard.layer_map()?;
3267 :
3268 1072544 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3269 912921 : let start_lsn = l.get_lsn_range().start;
3270 912921 : cont_lsn > start_lsn
3271 1072544 : });
3272 1072544 :
3273 1072544 : match in_memory_layer {
3274 606316 : Some(l) => {
3275 606316 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3276 606316 : fringe.update(
3277 606316 : ReadableLayer::InMemoryLayer(l),
3278 606316 : unmapped_keyspace.clone(),
3279 606316 : lsn_range,
3280 606316 : );
3281 606316 : }
3282 : None => {
3283 466250 : for range in unmapped_keyspace.ranges.iter() {
3284 466250 : let results = layers.range_search(range.clone(), cont_lsn);
3285 466250 :
3286 466250 : results
3287 466250 : .found
3288 466250 : .into_iter()
3289 466250 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3290 239128 : (
3291 239128 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3292 239128 : keyspace_accum.to_keyspace(),
3293 239128 : lsn_floor..cont_lsn,
3294 239128 : )
3295 466250 : })
3296 466250 : .for_each(|(layer, keyspace, lsn_range)| {
3297 239128 : fringe.update(layer, keyspace, lsn_range)
3298 466250 : });
3299 466250 : }
3300 : }
3301 : }
3302 :
3303 : // It's safe to drop the layer map lock after planning the next round of reads.
3304 : // The fringe keeps readable handles for the layers which are safe to read even
3305 : // if layers were compacted or flushed.
3306 : //
3307 : // The more interesting consideration is: "Why is the read algorithm still correct
3308 : // if the layer map changes while it is operating?". Doing a vectored read on a
3309 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3310 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3311 : // range at *a particular point in time*. It is fine for the answer to be different
3312 : // at two different time points.
3313 1072544 : drop(guard);
3314 624310 : }
3315 :
3316 1696854 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3317 845430 : let next_cont_lsn = lsn_range.start;
3318 845430 : layer_to_read
3319 845430 : .get_values_reconstruct_data(
3320 845430 : keyspace_to_read.clone(),
3321 845430 : lsn_range,
3322 845430 : reconstruct_state,
3323 845430 : ctx,
3324 845430 : )
3325 183072 : .await?;
3326 :
3327 845430 : unmapped_keyspace = keyspace_to_read;
3328 845430 : cont_lsn = next_cont_lsn;
3329 845430 :
3330 845430 : reconstruct_state.on_layer_visited(&layer_to_read);
3331 : } else {
3332 851424 : break;
3333 851424 : }
3334 851424 : }
3335 851424 :
3336 851424 : Ok(TimelineVisitOutcome {
3337 851424 : completed_keyspace,
3338 851424 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3339 851424 : })
3340 851424 : }
3341 :
3342 224856 : async fn get_ready_ancestor_timeline(
3343 224856 : &self,
3344 224856 : ancestor: &Arc<Timeline>,
3345 224856 : ctx: &RequestContext,
3346 224856 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3347 224856 : // It's possible that the ancestor timeline isn't active yet, or
3348 224856 : // is active but hasn't yet caught up to the branch point. Wait
3349 224856 : // for it.
3350 224856 : //
3351 224856 : // This cannot happen while the pageserver is running normally,
3352 224856 : // because you cannot create a branch from a point that isn't
3353 224856 : // present in the pageserver yet. However, we don't wait for the
3354 224856 : // branch point to be uploaded to cloud storage before creating
3355 224856 : // a branch. I.e., the branch LSN need not be remote consistent
3356 224856 : // for the branching operation to succeed.
3357 224856 : //
3358 224856 : // Hence, if we try to load a tenant in such a state where
3359 224856 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3360 224856 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3361 224856 : // then we will need to wait for the ancestor timeline to
3362 224856 : // re-stream WAL up to branch_lsn before we access it.
3363 224856 : //
3364 224856 : // How can a tenant get in such a state?
3365 224856 : // - ungraceful pageserver process exit
3366 224856 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3367 224856 : //
3368 224856 : // NB: this could be avoided by requiring
3369 224856 : // branch_lsn >= remote_consistent_lsn
3370 224856 : // during branch creation.
3371 224856 : match ancestor.wait_to_become_active(ctx).await {
3372 224854 : Ok(()) => {}
3373 : Err(TimelineState::Stopping) => {
3374 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3375 0 : return Err(GetReadyAncestorError::Cancelled);
3376 : }
3377 2 : Err(state) => {
3378 2 : return Err(GetReadyAncestorError::BadState {
3379 2 : timeline_id: ancestor.timeline_id,
3380 2 : state,
3381 2 : });
3382 : }
3383 : }
3384 224854 : ancestor
3385 224854 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3386 0 : .await
3387 224854 : .map_err(|e| match e {
3388 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3389 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3390 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3391 0 : timeline_id: ancestor.timeline_id,
3392 0 : state,
3393 0 : },
3394 224854 : })?;
3395 :
3396 224854 : Ok(ancestor.clone())
3397 224856 : }
3398 :
3399 151304 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3400 151304 : &self.shard_identity
3401 151304 : }
3402 :
3403 : #[inline(always)]
3404 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3405 0 : ShardTimelineId {
3406 0 : shard_index: ShardIndex {
3407 0 : shard_number: self.shard_identity.number,
3408 0 : shard_count: self.shard_identity.count,
3409 0 : },
3410 0 : timeline_id: self.timeline_id,
3411 0 : }
3412 0 : }
3413 :
3414 : /// Returns a non-frozen open in-memory layer for ingestion.
3415 : ///
3416 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3417 : /// this function without holding the mutex.
3418 1268 : async fn get_layer_for_write(
3419 1268 : &self,
3420 1268 : lsn: Lsn,
3421 1268 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3422 1268 : ctx: &RequestContext,
3423 1268 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3424 1268 : let mut guard = self.layers.write().await;
3425 1268 : let gate_guard = self.gate.enter().context("enter gate for inmem layer")?;
3426 :
3427 1268 : let last_record_lsn = self.get_last_record_lsn();
3428 1268 : ensure!(
3429 1268 : lsn > last_record_lsn,
3430 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3431 : lsn,
3432 : last_record_lsn,
3433 : );
3434 :
3435 1268 : let layer = guard
3436 1268 : .open_mut()?
3437 1268 : .get_layer_for_write(
3438 1268 : lsn,
3439 1268 : self.conf,
3440 1268 : self.timeline_id,
3441 1268 : self.tenant_shard_id,
3442 1268 : gate_guard,
3443 1268 : ctx,
3444 1268 : )
3445 719 : .await?;
3446 1268 : Ok(layer)
3447 1268 : }
3448 :
3449 5279064 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3450 5279064 : assert!(new_lsn.is_aligned());
3451 :
3452 5279064 : self.metrics.last_record_gauge.set(new_lsn.0 as i64);
3453 5279064 : self.last_record_lsn.advance(new_lsn);
3454 5279064 : }
3455 :
3456 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3457 : ///
3458 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3459 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3460 1172 : async fn freeze_inmem_layer_at(
3461 1172 : &self,
3462 1172 : at: Lsn,
3463 1172 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3464 1172 : ) -> Result<u64, FlushLayerError> {
3465 1172 : let frozen = {
3466 1172 : let mut guard = self.layers.write().await;
3467 1172 : guard
3468 1172 : .open_mut()?
3469 1172 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock)
3470 1 : .await
3471 : };
3472 :
3473 1172 : if frozen {
3474 1144 : let now = Instant::now();
3475 1144 : *(self.last_freeze_ts.write().unwrap()) = now;
3476 1144 : }
3477 :
3478 : // Increment the flush cycle counter and wake up the flush task.
3479 : // Remember the new value, so that when we listen for the flush
3480 : // to finish, we know when the flush that we initiated has
3481 : // finished, instead of some other flush that was started earlier.
3482 1172 : let mut my_flush_request = 0;
3483 1172 :
3484 1172 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3485 1172 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3486 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3487 1172 : }
3488 1172 :
3489 1172 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3490 1172 : my_flush_request = *counter + 1;
3491 1172 : *counter = my_flush_request;
3492 1172 : *lsn = std::cmp::max(at, *lsn);
3493 1172 : });
3494 1172 :
3495 1172 : assert_ne!(my_flush_request, 0);
3496 :
3497 1172 : Ok(my_flush_request)
3498 1172 : }
3499 :
3500 : /// Layer flusher task's main loop.
3501 412 : async fn flush_loop(
3502 412 : self: &Arc<Self>,
3503 412 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3504 412 : ctx: &RequestContext,
3505 412 : ) {
3506 412 : info!("started flush loop");
3507 : loop {
3508 1546 : tokio::select! {
3509 1546 : _ = self.cancel.cancelled() => {
3510 10 : info!("shutting down layer flush task due to Timeline::cancel");
3511 10 : break;
3512 : },
3513 1546 : _ = layer_flush_start_rx.changed() => {}
3514 1134 : }
3515 1134 : trace!("waking up");
3516 1134 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3517 1134 :
3518 1134 : // The highest LSN to which we flushed in the loop over frozen layers
3519 1134 : let mut flushed_to_lsn = Lsn(0);
3520 :
3521 1134 : let result = loop {
3522 2278 : if self.cancel.is_cancelled() {
3523 0 : info!("dropping out of flush loop for timeline shutdown");
3524 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3525 : // anyone waiting on that will respect self.cancel as well: they will stop
3526 : // waiting at the same time we as drop out of this loop.
3527 0 : return;
3528 2278 : }
3529 2278 :
3530 2278 : let timer = self.metrics.flush_time_histo.start_timer();
3531 :
3532 : let num_frozen_layers;
3533 : let frozen_layer_total_size;
3534 2278 : let layer_to_flush = {
3535 2278 : let guard = self.layers.read().await;
3536 2278 : let Ok(lm) = guard.layer_map() else {
3537 0 : info!("dropping out of flush loop for timeline shutdown");
3538 0 : return;
3539 : };
3540 2278 : num_frozen_layers = lm.frozen_layers.len();
3541 2278 : frozen_layer_total_size = lm
3542 2278 : .frozen_layers
3543 2278 : .iter()
3544 2278 : .map(|l| l.estimated_in_mem_size())
3545 2278 : .sum::<u64>();
3546 2278 : lm.frozen_layers.front().cloned()
3547 : // drop 'layers' lock to allow concurrent reads and writes
3548 : };
3549 2278 : let Some(layer_to_flush) = layer_to_flush else {
3550 1134 : break Ok(());
3551 : };
3552 1144 : if num_frozen_layers
3553 1144 : > std::cmp::max(
3554 1144 : self.get_compaction_threshold(),
3555 1144 : DEFAULT_COMPACTION_THRESHOLD,
3556 1144 : )
3557 0 : && frozen_layer_total_size >= /* 128 MB */ 128000000
3558 : {
3559 0 : tracing::warn!(
3560 0 : "too many frozen layers: {num_frozen_layers} layers with estimated in-mem size of {frozen_layer_total_size} bytes",
3561 : );
3562 1144 : }
3563 17084 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3564 1144 : Ok(this_layer_to_lsn) => {
3565 1144 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3566 1144 : }
3567 : Err(FlushLayerError::Cancelled) => {
3568 0 : info!("dropping out of flush loop for timeline shutdown");
3569 0 : return;
3570 : }
3571 0 : err @ Err(
3572 0 : FlushLayerError::NotRunning(_)
3573 0 : | FlushLayerError::Other(_)
3574 0 : | FlushLayerError::CreateImageLayersError(_),
3575 0 : ) => {
3576 0 : error!("could not flush frozen layer: {err:?}");
3577 0 : break err.map(|_| ());
3578 : }
3579 : }
3580 1144 : timer.stop_and_record();
3581 : };
3582 :
3583 : // Unsharded tenants should never advance their LSN beyond the end of the
3584 : // highest layer they write: such gaps between layer data and the frozen LSN
3585 : // are only legal on sharded tenants.
3586 1134 : debug_assert!(
3587 1134 : self.shard_identity.count.count() > 1
3588 1134 : || flushed_to_lsn >= frozen_to_lsn
3589 68 : || !flushed_to_lsn.is_valid()
3590 : );
3591 :
3592 1134 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3593 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3594 : // to us via layer_flush_start_rx, then advance it here.
3595 : //
3596 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3597 : // never encounter a gap in the wal.
3598 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3599 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3600 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3601 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3602 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3603 0 : }
3604 0 : }
3605 1134 : }
3606 :
3607 : // Notify any listeners that we're done
3608 1134 : let _ = self
3609 1134 : .layer_flush_done_tx
3610 1134 : .send_replace((flush_counter, result));
3611 : }
3612 10 : }
3613 :
3614 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
3615 1092 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3616 1092 : let mut rx = self.layer_flush_done_tx.subscribe();
3617 : loop {
3618 : {
3619 2177 : let (last_result_counter, last_result) = &*rx.borrow();
3620 2177 : if *last_result_counter >= request {
3621 1092 : if let Err(err) = last_result {
3622 : // We already logged the original error in
3623 : // flush_loop. We cannot propagate it to the caller
3624 : // here, because it might not be Cloneable
3625 0 : return Err(err.clone());
3626 : } else {
3627 1092 : return Ok(());
3628 : }
3629 1085 : }
3630 1085 : }
3631 1085 : trace!("waiting for flush to complete");
3632 1085 : tokio::select! {
3633 1085 : rx_e = rx.changed() => {
3634 1085 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3635 : },
3636 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3637 : // the notification from [`flush_loop`] that it completed.
3638 1085 : _ = self.cancel.cancelled() => {
3639 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3640 0 : return Ok(())
3641 : }
3642 : };
3643 1085 : trace!("done")
3644 : }
3645 1092 : }
3646 :
3647 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3648 : ///
3649 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3650 1144 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3651 : async fn flush_frozen_layer(
3652 : self: &Arc<Self>,
3653 : frozen_layer: Arc<InMemoryLayer>,
3654 : ctx: &RequestContext,
3655 : ) -> Result<Lsn, FlushLayerError> {
3656 : debug_assert_current_span_has_tenant_and_timeline_id();
3657 :
3658 : // As a special case, when we have just imported an image into the repository,
3659 : // instead of writing out a L0 delta layer, we directly write out image layer
3660 : // files instead. This is possible as long as *all* the data imported into the
3661 : // repository have the same LSN.
3662 : let lsn_range = frozen_layer.get_lsn_range();
3663 :
3664 : // Whether to directly create image layers for this flush, or flush them as delta layers
3665 : let create_image_layer =
3666 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3667 :
3668 : #[cfg(test)]
3669 : {
3670 : match &mut *self.flush_loop_state.lock().unwrap() {
3671 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3672 : panic!("flush loop not running")
3673 : }
3674 : FlushLoopState::Running {
3675 : expect_initdb_optimization,
3676 : initdb_optimization_count,
3677 : ..
3678 : } => {
3679 : if create_image_layer {
3680 : *initdb_optimization_count += 1;
3681 : } else {
3682 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3683 : }
3684 : }
3685 : }
3686 : }
3687 :
3688 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3689 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3690 : // require downloading anything during initial import.
3691 : let ((rel_partition, metadata_partition), _lsn) = self
3692 : .repartition(
3693 : self.initdb_lsn,
3694 : self.get_compaction_target_size(),
3695 : EnumSet::empty(),
3696 : ctx,
3697 : )
3698 : .await
3699 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
3700 :
3701 : if self.cancel.is_cancelled() {
3702 : return Err(FlushLayerError::Cancelled);
3703 : }
3704 :
3705 : let mut layers_to_upload = Vec::new();
3706 : layers_to_upload.extend(
3707 : self.create_image_layers(
3708 : &rel_partition,
3709 : self.initdb_lsn,
3710 : ImageLayerCreationMode::Initial,
3711 : ctx,
3712 : )
3713 : .await?,
3714 : );
3715 : if !metadata_partition.parts.is_empty() {
3716 : assert_eq!(
3717 : metadata_partition.parts.len(),
3718 : 1,
3719 : "currently sparse keyspace should only contain a single metadata keyspace"
3720 : );
3721 : layers_to_upload.extend(
3722 : self.create_image_layers(
3723 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
3724 : // every single key within the keyspace, and therefore, it's safe to force converting it
3725 : // into a dense keyspace before calling this function.
3726 : &metadata_partition.into_dense(),
3727 : self.initdb_lsn,
3728 : ImageLayerCreationMode::Initial,
3729 : ctx,
3730 : )
3731 : .await?,
3732 : );
3733 : }
3734 :
3735 : (layers_to_upload, None)
3736 : } else {
3737 : // Normal case, write out a L0 delta layer file.
3738 : // `create_delta_layer` will not modify the layer map.
3739 : // We will remove frozen layer and add delta layer in one atomic operation later.
3740 : let Some(layer) = self
3741 : .create_delta_layer(&frozen_layer, None, ctx)
3742 : .await
3743 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
3744 : else {
3745 : panic!("delta layer cannot be empty if no filter is applied");
3746 : };
3747 : (
3748 : // FIXME: even though we have a single image and single delta layer assumption
3749 : // we push them to vec
3750 : vec![layer.clone()],
3751 : Some(layer),
3752 : )
3753 : };
3754 :
3755 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
3756 :
3757 : if self.cancel.is_cancelled() {
3758 : return Err(FlushLayerError::Cancelled);
3759 : }
3760 :
3761 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
3762 :
3763 : // The new on-disk layers are now in the layer map. We can remove the
3764 : // in-memory layer from the map now. The flushed layer is stored in
3765 : // the mapping in `create_delta_layer`.
3766 : {
3767 : let mut guard = self.layers.write().await;
3768 :
3769 : guard.open_mut()?.finish_flush_l0_layer(
3770 : delta_layer_to_add.as_ref(),
3771 : &frozen_layer,
3772 : &self.metrics,
3773 : );
3774 :
3775 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
3776 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
3777 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
3778 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
3779 : }
3780 : // release lock on 'layers'
3781 : };
3782 :
3783 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
3784 : // This makes us refuse ingest until the new layers have been persisted to the remote.
3785 : self.remote_client
3786 : .wait_completion()
3787 : .await
3788 0 : .map_err(|e| match e {
3789 : WaitCompletionError::UploadQueueShutDownOrStopped
3790 : | WaitCompletionError::NotInitialized(
3791 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
3792 0 : ) => FlushLayerError::Cancelled,
3793 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
3794 0 : FlushLayerError::Other(anyhow!(e).into())
3795 : }
3796 0 : })?;
3797 :
3798 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
3799 : // a compaction can delete the file and then it won't be available for uploads any more.
3800 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
3801 : // race situation.
3802 : // See https://github.com/neondatabase/neon/issues/4526
3803 : pausable_failpoint!("flush-frozen-pausable");
3804 :
3805 : // This failpoint is used by another test case `test_pageserver_recovery`.
3806 : fail_point!("flush-frozen-exit");
3807 :
3808 : Ok(Lsn(lsn_range.end.0 - 1))
3809 : }
3810 :
3811 : /// Return true if the value changed
3812 : ///
3813 : /// This function must only be used from the layer flush task.
3814 1144 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
3815 1144 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
3816 1144 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
3817 1144 : new_value != old_value
3818 1144 : }
3819 :
3820 : /// Update metadata file
3821 1146 : fn schedule_uploads(
3822 1146 : &self,
3823 1146 : disk_consistent_lsn: Lsn,
3824 1146 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
3825 1146 : ) -> anyhow::Result<()> {
3826 1146 : // We can only save a valid 'prev_record_lsn' value on disk if we
3827 1146 : // flushed *all* in-memory changes to disk. We only track
3828 1146 : // 'prev_record_lsn' in memory for the latest processed record, so we
3829 1146 : // don't remember what the correct value that corresponds to some old
3830 1146 : // LSN is. But if we flush everything, then the value corresponding
3831 1146 : // current 'last_record_lsn' is correct and we can store it on disk.
3832 1146 : let RecordLsn {
3833 1146 : last: last_record_lsn,
3834 1146 : prev: prev_record_lsn,
3835 1146 : } = self.last_record_lsn.load();
3836 1146 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
3837 1066 : Some(prev_record_lsn)
3838 : } else {
3839 80 : None
3840 : };
3841 :
3842 1146 : let update = crate::tenant::metadata::MetadataUpdate::new(
3843 1146 : disk_consistent_lsn,
3844 1146 : ondisk_prev_record_lsn,
3845 1146 : *self.latest_gc_cutoff_lsn.read(),
3846 1146 : );
3847 1146 :
3848 1146 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
3849 0 : "{}",
3850 0 : x.unwrap()
3851 1146 : ));
3852 :
3853 2302 : for layer in layers_to_upload {
3854 1156 : self.remote_client.schedule_layer_file_upload(layer)?;
3855 : }
3856 1146 : self.remote_client
3857 1146 : .schedule_index_upload_for_metadata_update(&update)?;
3858 :
3859 1146 : Ok(())
3860 1146 : }
3861 :
3862 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
3863 0 : self.remote_client
3864 0 : .preserve_initdb_archive(
3865 0 : &self.tenant_shard_id.tenant_id,
3866 0 : &self.timeline_id,
3867 0 : &self.cancel,
3868 0 : )
3869 0 : .await
3870 0 : }
3871 :
3872 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
3873 : // in layer map immediately. The caller is responsible to put it into the layer map.
3874 968 : async fn create_delta_layer(
3875 968 : self: &Arc<Self>,
3876 968 : frozen_layer: &Arc<InMemoryLayer>,
3877 968 : key_range: Option<Range<Key>>,
3878 968 : ctx: &RequestContext,
3879 968 : ) -> anyhow::Result<Option<ResidentLayer>> {
3880 968 : let self_clone = Arc::clone(self);
3881 968 : let frozen_layer = Arc::clone(frozen_layer);
3882 968 : let ctx = ctx.attached_child();
3883 968 : let work = async move {
3884 968 : let Some((desc, path)) = frozen_layer
3885 968 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
3886 10250 : .await?
3887 : else {
3888 0 : return Ok(None);
3889 : };
3890 968 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
3891 :
3892 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
3893 : // We just need to fsync the directory in which these inodes are linked,
3894 : // which we know to be the timeline directory.
3895 : //
3896 : // We use fatal_err() below because the after write_to_disk returns with success,
3897 : // the in-memory state of the filesystem already has the layer file in its final place,
3898 : // and subsequent pageserver code could think it's durable while it really isn't.
3899 968 : let timeline_dir = VirtualFile::open(
3900 968 : &self_clone
3901 968 : .conf
3902 968 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
3903 968 : &ctx,
3904 968 : )
3905 486 : .await
3906 968 : .fatal_err("VirtualFile::open for timeline dir fsync");
3907 968 : timeline_dir
3908 968 : .sync_all()
3909 484 : .await
3910 968 : .fatal_err("VirtualFile::sync_all timeline dir");
3911 968 : anyhow::Ok(Some(new_delta))
3912 968 : };
3913 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
3914 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
3915 : use crate::virtual_file::io_engine::IoEngine;
3916 968 : match crate::virtual_file::io_engine::get() {
3917 0 : IoEngine::NotSet => panic!("io engine not set"),
3918 : IoEngine::StdFs => {
3919 484 : let span = tracing::info_span!("blocking");
3920 484 : tokio::task::spawn_blocking({
3921 484 : move || Handle::current().block_on(work.instrument(span))
3922 484 : })
3923 484 : .await
3924 484 : .context("spawn_blocking")
3925 484 : .and_then(|x| x)
3926 : }
3927 : #[cfg(target_os = "linux")]
3928 11215 : IoEngine::TokioEpollUring => work.await,
3929 : }
3930 968 : }
3931 :
3932 540 : async fn repartition(
3933 540 : &self,
3934 540 : lsn: Lsn,
3935 540 : partition_size: u64,
3936 540 : flags: EnumSet<CompactFlags>,
3937 540 : ctx: &RequestContext,
3938 540 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
3939 540 : let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
3940 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
3941 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
3942 : // and hence before the compaction task starts.
3943 0 : return Err(CompactionError::Other(anyhow!(
3944 0 : "repartition() called concurrently, this should not happen"
3945 0 : )));
3946 : };
3947 540 : let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
3948 540 : if lsn < *partition_lsn {
3949 0 : return Err(CompactionError::Other(anyhow!(
3950 0 : "repartition() called with LSN going backwards, this should not happen"
3951 0 : )));
3952 540 : }
3953 540 :
3954 540 : let distance = lsn.0 - partition_lsn.0;
3955 540 : if *partition_lsn != Lsn(0)
3956 262 : && distance <= self.repartition_threshold
3957 262 : && !flags.contains(CompactFlags::ForceRepartition)
3958 : {
3959 248 : debug!(
3960 : distance,
3961 : threshold = self.repartition_threshold,
3962 0 : "no repartitioning needed"
3963 : );
3964 248 : return Ok((
3965 248 : (dense_partition.clone(), sparse_partition.clone()),
3966 248 : *partition_lsn,
3967 248 : ));
3968 292 : }
3969 :
3970 15831 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
3971 292 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
3972 292 : let sparse_partitioning = SparseKeyPartitioning {
3973 292 : parts: vec![sparse_ks],
3974 292 : }; // no partitioning for metadata keys for now
3975 292 : *partitioning_guard = ((dense_partitioning, sparse_partitioning), lsn);
3976 292 :
3977 292 : Ok((partitioning_guard.0.clone(), partitioning_guard.1))
3978 540 : }
3979 :
3980 : // Is it time to create a new image layer for the given partition?
3981 14 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
3982 14 : let threshold = self.get_image_creation_threshold();
3983 :
3984 14 : let guard = self.layers.read().await;
3985 14 : let Ok(layers) = guard.layer_map() else {
3986 0 : return false;
3987 : };
3988 :
3989 14 : let mut max_deltas = 0;
3990 28 : for part_range in &partition.ranges {
3991 14 : let image_coverage = layers.image_coverage(part_range, lsn);
3992 28 : for (img_range, last_img) in image_coverage {
3993 14 : let img_lsn = if let Some(last_img) = last_img {
3994 0 : last_img.get_lsn_range().end
3995 : } else {
3996 14 : Lsn(0)
3997 : };
3998 : // Let's consider an example:
3999 : //
4000 : // delta layer with LSN range 71-81
4001 : // delta layer with LSN range 81-91
4002 : // delta layer with LSN range 91-101
4003 : // image layer at LSN 100
4004 : //
4005 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4006 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4007 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4008 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4009 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4010 14 : if img_lsn < lsn {
4011 14 : let num_deltas =
4012 14 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4013 14 :
4014 14 : max_deltas = max_deltas.max(num_deltas);
4015 14 : if num_deltas >= threshold {
4016 0 : debug!(
4017 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4018 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4019 : );
4020 0 : return true;
4021 14 : }
4022 0 : }
4023 : }
4024 : }
4025 :
4026 14 : debug!(
4027 : max_deltas,
4028 0 : "none of the partitioned ranges had >= {threshold} deltas"
4029 : );
4030 14 : false
4031 14 : }
4032 :
4033 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4034 : /// so that at most one image layer will be produced from this function.
4035 202 : async fn create_image_layer_for_rel_blocks(
4036 202 : self: &Arc<Self>,
4037 202 : partition: &KeySpace,
4038 202 : mut image_layer_writer: ImageLayerWriter,
4039 202 : lsn: Lsn,
4040 202 : ctx: &RequestContext,
4041 202 : img_range: Range<Key>,
4042 202 : start: Key,
4043 202 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4044 202 : let mut wrote_keys = false;
4045 202 :
4046 202 : let mut key_request_accum = KeySpaceAccum::new();
4047 1344 : for range in &partition.ranges {
4048 1142 : let mut key = range.start;
4049 2472 : while key < range.end {
4050 : // Decide whether to retain this key: usually we do, but sharded tenants may
4051 : // need to drop keys that don't belong to them. If we retain the key, add it
4052 : // to `key_request_accum` for later issuing a vectored get
4053 1330 : if self.shard_identity.is_key_disposable(&key) {
4054 0 : debug!(
4055 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4056 0 : key,
4057 0 : self.shard_identity.get_shard_number(&key)
4058 : );
4059 1330 : } else {
4060 1330 : key_request_accum.add_key(key);
4061 1330 : }
4062 :
4063 1330 : let last_key_in_range = key.next() == range.end;
4064 1330 : key = key.next();
4065 1330 :
4066 1330 : // Maybe flush `key_rest_accum`
4067 1330 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4068 1330 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4069 : {
4070 1142 : let results = self
4071 1142 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
4072 50 : .await?;
4073 :
4074 1142 : if self.cancel.is_cancelled() {
4075 0 : return Err(CreateImageLayersError::Cancelled);
4076 1142 : }
4077 :
4078 2472 : for (img_key, img) in results {
4079 1330 : let img = match img {
4080 1330 : Ok(img) => img,
4081 0 : Err(err) => {
4082 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4083 0 : // page without losing any actual user data. That seems better
4084 0 : // than failing repeatedly and getting stuck.
4085 0 : //
4086 0 : // We had a bug at one point, where we truncated the FSM and VM
4087 0 : // in the pageserver, but the Postgres didn't know about that
4088 0 : // and continued to generate incremental WAL records for pages
4089 0 : // that didn't exist in the pageserver. Trying to replay those
4090 0 : // WAL records failed to find the previous image of the page.
4091 0 : // This special case allows us to recover from that situation.
4092 0 : // See https://github.com/neondatabase/neon/issues/2601.
4093 0 : //
4094 0 : // Unfortunately we cannot do this for the main fork, or for
4095 0 : // any metadata keys, keys, as that would lead to actual data
4096 0 : // loss.
4097 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4098 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4099 0 : ZERO_PAGE.clone()
4100 : } else {
4101 0 : return Err(CreateImageLayersError::from(err));
4102 : }
4103 : }
4104 : };
4105 :
4106 : // Write all the keys we just read into our new image layer.
4107 1457 : image_layer_writer.put_image(img_key, img, ctx).await?;
4108 1330 : wrote_keys = true;
4109 : }
4110 188 : }
4111 : }
4112 : }
4113 :
4114 202 : if wrote_keys {
4115 : // Normal path: we have written some data into the new image layer for this
4116 : // partition, so flush it to disk.
4117 404 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4118 202 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4119 202 : info!("created image layer for rel {}", image_layer.local_path());
4120 202 : Ok(ImageLayerCreationOutcome {
4121 202 : image: Some(image_layer),
4122 202 : next_start_key: img_range.end,
4123 202 : })
4124 : } else {
4125 : // Special case: the image layer may be empty if this is a sharded tenant and the
4126 : // partition does not cover any keys owned by this shard. In this case, to ensure
4127 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4128 : // layer we write will cover the key range that we just scanned.
4129 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4130 0 : Ok(ImageLayerCreationOutcome {
4131 0 : image: None,
4132 0 : next_start_key: start,
4133 0 : })
4134 : }
4135 202 : }
4136 :
4137 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4138 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4139 : /// would not be too large to fit in a single image layer.
4140 : #[allow(clippy::too_many_arguments)]
4141 192 : async fn create_image_layer_for_metadata_keys(
4142 192 : self: &Arc<Self>,
4143 192 : partition: &KeySpace,
4144 192 : mut image_layer_writer: ImageLayerWriter,
4145 192 : lsn: Lsn,
4146 192 : ctx: &RequestContext,
4147 192 : img_range: Range<Key>,
4148 192 : mode: ImageLayerCreationMode,
4149 192 : start: Key,
4150 192 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4151 192 : // Metadata keys image layer creation.
4152 192 : let mut reconstruct_state = ValuesReconstructState::default();
4153 192 : let begin = Instant::now();
4154 192 : let data = self
4155 192 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4156 1025 : .await?;
4157 192 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4158 192 : let mut new_data = BTreeMap::new();
4159 192 : let mut total_kb_retrieved = 0;
4160 192 : let mut total_keys_retrieved = 0;
4161 10204 : for (k, v) in data {
4162 10012 : let v = v?;
4163 10012 : total_kb_retrieved += KEY_SIZE + v.len();
4164 10012 : total_keys_retrieved += 1;
4165 10012 : new_data.insert(k, v);
4166 : }
4167 192 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4168 192 : };
4169 192 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4170 192 : let elapsed = begin.elapsed();
4171 192 :
4172 192 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4173 192 : info!(
4174 0 : "metadata key compaction: trigger_generation={trigger_generation}, delta_files_accessed={delta_files_accessed}, total_kb_retrieved={total_kb_retrieved}, total_keys_retrieved={total_keys_retrieved}, read_time={}s", elapsed.as_secs_f64()
4175 : );
4176 :
4177 192 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4178 2 : return Ok(ImageLayerCreationOutcome {
4179 2 : image: None,
4180 2 : next_start_key: img_range.end,
4181 2 : });
4182 190 : }
4183 190 : if self.cancel.is_cancelled() {
4184 0 : return Err(CreateImageLayersError::Cancelled);
4185 190 : }
4186 190 : let mut wrote_any_image = false;
4187 10202 : for (k, v) in data {
4188 10012 : if v.is_empty() {
4189 : // the key has been deleted, it does not need an image
4190 : // in metadata keyspace, an empty image == tombstone
4191 8 : continue;
4192 10004 : }
4193 10004 : wrote_any_image = true;
4194 10004 :
4195 10004 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4196 10004 :
4197 10004 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4198 10004 : // on the normal data path either.
4199 10160 : image_layer_writer.put_image(k, v, ctx).await?;
4200 : }
4201 :
4202 190 : if wrote_any_image {
4203 : // Normal path: we have written some data into the new image layer for this
4204 : // partition, so flush it to disk.
4205 25 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4206 12 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4207 12 : info!(
4208 0 : "created image layer for metadata {}",
4209 0 : image_layer.local_path()
4210 : );
4211 12 : Ok(ImageLayerCreationOutcome {
4212 12 : image: Some(image_layer),
4213 12 : next_start_key: img_range.end,
4214 12 : })
4215 : } else {
4216 : // Special case: the image layer may be empty if this is a sharded tenant and the
4217 : // partition does not cover any keys owned by this shard. In this case, to ensure
4218 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4219 : // layer we write will cover the key range that we just scanned.
4220 178 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4221 178 : Ok(ImageLayerCreationOutcome {
4222 178 : image: None,
4223 178 : next_start_key: start,
4224 178 : })
4225 : }
4226 192 : }
4227 :
4228 : /// Predicate function which indicates whether we should check if new image layers
4229 : /// are required. Since checking if new image layers are required is expensive in
4230 : /// terms of CPU, we only do it in the following cases:
4231 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4232 : /// 2. If enough time has passed since the last check:
4233 : /// 1. For large tenants, we wish to perform the check more often since they
4234 : /// suffer from the lack of image layers
4235 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4236 716 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4237 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4238 :
4239 716 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4240 716 : let distance = lsn
4241 716 : .checked_sub(last_checks_at)
4242 716 : .expect("Attempt to compact with LSN going backwards");
4243 716 : let min_distance =
4244 716 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4245 716 :
4246 716 : let distance_based_decision = distance.0 >= min_distance;
4247 716 :
4248 716 : let mut time_based_decision = false;
4249 716 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4250 716 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4251 614 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4252 : {
4253 0 : self.get_checkpoint_timeout()
4254 : } else {
4255 614 : Duration::from_secs(3600 * 48)
4256 : };
4257 :
4258 614 : time_based_decision = match *last_check_instant {
4259 438 : Some(last_check) => {
4260 438 : let elapsed = last_check.elapsed();
4261 438 : elapsed >= check_required_after
4262 : }
4263 176 : None => true,
4264 : };
4265 102 : }
4266 :
4267 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4268 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4269 : // check.
4270 716 : let decision = distance_based_decision || time_based_decision;
4271 :
4272 716 : if decision {
4273 178 : self.last_image_layer_creation_check_at.store(lsn);
4274 178 : *last_check_instant = Some(Instant::now());
4275 538 : }
4276 :
4277 716 : decision
4278 716 : }
4279 :
4280 716 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4281 : async fn create_image_layers(
4282 : self: &Arc<Timeline>,
4283 : partitioning: &KeyPartitioning,
4284 : lsn: Lsn,
4285 : mode: ImageLayerCreationMode,
4286 : ctx: &RequestContext,
4287 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4288 : let timer = self.metrics.create_images_time_histo.start_timer();
4289 : let mut image_layers = Vec::new();
4290 :
4291 : // We need to avoid holes between generated image layers.
4292 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4293 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4294 : //
4295 : // How such hole between partitions can appear?
4296 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4297 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4298 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4299 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4300 : let mut start = Key::MIN;
4301 :
4302 : let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
4303 :
4304 : for partition in partitioning.parts.iter() {
4305 : if self.cancel.is_cancelled() {
4306 : return Err(CreateImageLayersError::Cancelled);
4307 : }
4308 :
4309 : let img_range = start..partition.ranges.last().unwrap().end;
4310 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4311 : if compact_metadata {
4312 : for range in &partition.ranges {
4313 : assert!(
4314 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4315 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4316 : "metadata keys must be partitioned separately"
4317 : );
4318 : }
4319 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4320 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4321 : // might mess up with evictions.
4322 : start = img_range.end;
4323 : continue;
4324 : }
4325 : // For initial and force modes, we always generate image layers for metadata keys.
4326 : } else if let ImageLayerCreationMode::Try = mode {
4327 : // check_for_image_layers = false -> skip
4328 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4329 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4330 : start = img_range.end;
4331 : continue;
4332 : }
4333 : }
4334 : if let ImageLayerCreationMode::Force = mode {
4335 : // When forced to create image layers, we might try and create them where they already
4336 : // exist. This mode is only used in tests/debug.
4337 : let layers = self.layers.read().await;
4338 : if layers.contains_key(&PersistentLayerKey {
4339 : key_range: img_range.clone(),
4340 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4341 : is_delta: false,
4342 : }) {
4343 : tracing::info!(
4344 : "Skipping image layer at {lsn} {}..{}, already exists",
4345 : img_range.start,
4346 : img_range.end
4347 : );
4348 : start = img_range.end;
4349 : continue;
4350 : }
4351 : }
4352 :
4353 : let image_layer_writer = ImageLayerWriter::new(
4354 : self.conf,
4355 : self.timeline_id,
4356 : self.tenant_shard_id,
4357 : &img_range,
4358 : lsn,
4359 : ctx,
4360 : )
4361 : .await?;
4362 :
4363 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4364 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4365 0 : "failpoint image-layer-writer-fail-before-finish"
4366 0 : )))
4367 0 : });
4368 :
4369 : if !compact_metadata {
4370 : let ImageLayerCreationOutcome {
4371 : image,
4372 : next_start_key,
4373 : } = self
4374 : .create_image_layer_for_rel_blocks(
4375 : partition,
4376 : image_layer_writer,
4377 : lsn,
4378 : ctx,
4379 : img_range,
4380 : start,
4381 : )
4382 : .await?;
4383 :
4384 : start = next_start_key;
4385 : image_layers.extend(image);
4386 : } else {
4387 : let ImageLayerCreationOutcome {
4388 : image,
4389 : next_start_key,
4390 : } = self
4391 : .create_image_layer_for_metadata_keys(
4392 : partition,
4393 : image_layer_writer,
4394 : lsn,
4395 : ctx,
4396 : img_range,
4397 : mode,
4398 : start,
4399 : )
4400 : .await?;
4401 : start = next_start_key;
4402 : image_layers.extend(image);
4403 : }
4404 : }
4405 :
4406 : let mut guard = self.layers.write().await;
4407 :
4408 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4409 : // now they are being scheduled outside of write lock; current way is inconsistent with
4410 : // compaction lock order.
4411 : guard
4412 : .open_mut()?
4413 : .track_new_image_layers(&image_layers, &self.metrics);
4414 : drop_wlock(guard);
4415 : timer.stop_and_record();
4416 :
4417 : // Creating image layers may have caused some previously visible layers to be covered
4418 : if !image_layers.is_empty() {
4419 : self.update_layer_visibility().await?;
4420 : }
4421 :
4422 : Ok(image_layers)
4423 : }
4424 :
4425 : /// Wait until the background initial logical size calculation is complete, or
4426 : /// this Timeline is shut down. Calling this function will cause the initial
4427 : /// logical size calculation to skip waiting for the background jobs barrier.
4428 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4429 0 : if !self.shard_identity.is_shard_zero() {
4430 : // We don't populate logical size on shard >0: skip waiting for it.
4431 0 : return;
4432 0 : }
4433 0 :
4434 0 : if self.remote_client.is_deleting() {
4435 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4436 0 : return;
4437 0 : }
4438 0 :
4439 0 : if self.current_logical_size.current_size().is_exact() {
4440 : // root timelines are initialized with exact count, but never start the background
4441 : // calculation
4442 0 : return;
4443 0 : }
4444 :
4445 0 : if let Some(await_bg_cancel) = self
4446 0 : .current_logical_size
4447 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4448 0 : .get()
4449 0 : {
4450 0 : await_bg_cancel.cancel();
4451 0 : } else {
4452 : // We should not wait if we were not able to explicitly instruct
4453 : // the logical size cancellation to skip the concurrency limit semaphore.
4454 : // TODO: this is an unexpected case. We should restructure so that it
4455 : // can't happen.
4456 0 : tracing::warn!(
4457 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4458 : );
4459 0 : debug_assert!(false);
4460 : }
4461 :
4462 0 : tokio::select!(
4463 0 : _ = self.current_logical_size.initialized.acquire() => {},
4464 0 : _ = self.cancel.cancelled() => {}
4465 : )
4466 0 : }
4467 :
4468 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4469 : /// Timelines layers up to the ancestor_lsn.
4470 : ///
4471 : /// Requires a timeline that:
4472 : /// - has an ancestor to detach from
4473 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4474 : /// a technical requirement
4475 : ///
4476 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4477 : /// polled again until completion.
4478 : ///
4479 : /// During the operation all timelines sharing the data with this timeline will be reparented
4480 : /// from our ancestor to be branches of this timeline.
4481 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4482 0 : self: &Arc<Timeline>,
4483 0 : tenant: &crate::tenant::Tenant,
4484 0 : options: detach_ancestor::Options,
4485 0 : ctx: &RequestContext,
4486 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
4487 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4488 0 : }
4489 :
4490 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
4491 : /// reparents any reparentable children of previous ancestor.
4492 : ///
4493 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
4494 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
4495 : /// successfully, tenant must be reloaded.
4496 : ///
4497 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
4498 : /// resetting the tenant.
4499 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
4500 0 : self: &Arc<Timeline>,
4501 0 : tenant: &crate::tenant::Tenant,
4502 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4503 0 : ctx: &RequestContext,
4504 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
4505 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
4506 0 : }
4507 :
4508 : /// Final step which unblocks the GC.
4509 : ///
4510 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
4511 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4512 0 : self: &Arc<Timeline>,
4513 0 : tenant: &crate::tenant::Tenant,
4514 0 : attempt: detach_ancestor::Attempt,
4515 0 : ctx: &RequestContext,
4516 0 : ) -> Result<(), detach_ancestor::Error> {
4517 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
4518 0 : }
4519 : }
4520 :
4521 : impl Drop for Timeline {
4522 10 : fn drop(&mut self) {
4523 10 : if let Some(ancestor) = &self.ancestor_timeline {
4524 : // This lock should never be poisoned, but in case it is we do a .map() instead of
4525 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
4526 4 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
4527 4 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
4528 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
4529 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
4530 4 : }
4531 0 : }
4532 6 : }
4533 10 : }
4534 : }
4535 :
4536 : /// Top-level failure to compact.
4537 0 : #[derive(Debug, thiserror::Error)]
4538 : pub(crate) enum CompactionError {
4539 : #[error("The timeline or pageserver is shutting down")]
4540 : ShuttingDown,
4541 : /// Compaction tried to offload a timeline and failed
4542 : #[error("Failed to offload timeline: {0}")]
4543 : Offload(OffloadError),
4544 : /// Compaction cannot be done right now; page reconstruction and so on.
4545 : #[error(transparent)]
4546 : Other(anyhow::Error),
4547 : }
4548 :
4549 : impl From<OffloadError> for CompactionError {
4550 0 : fn from(e: OffloadError) -> Self {
4551 0 : match e {
4552 0 : OffloadError::Cancelled => Self::ShuttingDown,
4553 0 : _ => Self::Offload(e),
4554 : }
4555 0 : }
4556 : }
4557 :
4558 : impl CompactionError {
4559 0 : pub fn is_cancelled(&self) -> bool {
4560 0 : matches!(self, CompactionError::ShuttingDown)
4561 0 : }
4562 : }
4563 :
4564 : impl From<CollectKeySpaceError> for CompactionError {
4565 0 : fn from(err: CollectKeySpaceError) -> Self {
4566 0 : match err {
4567 : CollectKeySpaceError::Cancelled
4568 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4569 0 : CompactionError::ShuttingDown
4570 : }
4571 0 : e => CompactionError::Other(e.into()),
4572 : }
4573 0 : }
4574 : }
4575 :
4576 : impl From<super::upload_queue::NotInitialized> for CompactionError {
4577 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
4578 0 : match value {
4579 : super::upload_queue::NotInitialized::Uninitialized => {
4580 0 : CompactionError::Other(anyhow::anyhow!(value))
4581 : }
4582 : super::upload_queue::NotInitialized::ShuttingDown
4583 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
4584 : }
4585 0 : }
4586 : }
4587 :
4588 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
4589 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
4590 0 : match e {
4591 : super::storage_layer::layer::DownloadError::TimelineShutdown
4592 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
4593 0 : CompactionError::ShuttingDown
4594 : }
4595 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
4596 : | super::storage_layer::layer::DownloadError::DownloadRequired
4597 : | super::storage_layer::layer::DownloadError::NotFile(_)
4598 : | super::storage_layer::layer::DownloadError::DownloadFailed
4599 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
4600 0 : CompactionError::Other(anyhow::anyhow!(e))
4601 : }
4602 : #[cfg(test)]
4603 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
4604 0 : CompactionError::Other(anyhow::anyhow!(e))
4605 : }
4606 : }
4607 0 : }
4608 : }
4609 :
4610 : impl From<layer_manager::Shutdown> for CompactionError {
4611 0 : fn from(_: layer_manager::Shutdown) -> Self {
4612 0 : CompactionError::ShuttingDown
4613 0 : }
4614 : }
4615 :
4616 : #[serde_as]
4617 196 : #[derive(serde::Serialize)]
4618 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4619 :
4620 : #[derive(Default)]
4621 : enum DurationRecorder {
4622 : #[default]
4623 : NotStarted,
4624 : Recorded(RecordedDuration, tokio::time::Instant),
4625 : }
4626 :
4627 : impl DurationRecorder {
4628 504 : fn till_now(&self) -> DurationRecorder {
4629 504 : match self {
4630 : DurationRecorder::NotStarted => {
4631 0 : panic!("must only call on recorded measurements")
4632 : }
4633 504 : DurationRecorder::Recorded(_, ended) => {
4634 504 : let now = tokio::time::Instant::now();
4635 504 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4636 504 : }
4637 504 : }
4638 504 : }
4639 196 : fn into_recorded(self) -> Option<RecordedDuration> {
4640 196 : match self {
4641 0 : DurationRecorder::NotStarted => None,
4642 196 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4643 : }
4644 196 : }
4645 : }
4646 :
4647 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
4648 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
4649 : /// the layer descriptor requires the user to provide the ranges, which should cover all
4650 : /// keys specified in the `data` field.
4651 : #[cfg(test)]
4652 : #[derive(Clone)]
4653 : pub struct DeltaLayerTestDesc {
4654 : pub lsn_range: Range<Lsn>,
4655 : pub key_range: Range<Key>,
4656 : pub data: Vec<(Key, Lsn, Value)>,
4657 : }
4658 :
4659 : #[cfg(test)]
4660 : impl DeltaLayerTestDesc {
4661 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
4662 2 : Self {
4663 2 : lsn_range,
4664 2 : key_range,
4665 2 : data,
4666 2 : }
4667 2 : }
4668 :
4669 72 : pub fn new_with_inferred_key_range(
4670 72 : lsn_range: Range<Lsn>,
4671 72 : data: Vec<(Key, Lsn, Value)>,
4672 72 : ) -> Self {
4673 196 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
4674 196 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
4675 72 : Self {
4676 72 : key_range: (*key_min)..(key_max.next()),
4677 72 : lsn_range,
4678 72 : data,
4679 72 : }
4680 72 : }
4681 :
4682 10 : pub(crate) fn layer_name(&self) -> LayerName {
4683 10 : LayerName::Delta(super::storage_layer::DeltaLayerName {
4684 10 : key_range: self.key_range.clone(),
4685 10 : lsn_range: self.lsn_range.clone(),
4686 10 : })
4687 10 : }
4688 : }
4689 :
4690 : impl Timeline {
4691 28 : async fn finish_compact_batch(
4692 28 : self: &Arc<Self>,
4693 28 : new_deltas: &[ResidentLayer],
4694 28 : new_images: &[ResidentLayer],
4695 28 : layers_to_remove: &[Layer],
4696 28 : ) -> Result<(), CompactionError> {
4697 28 : let mut guard = tokio::select! {
4698 28 : guard = self.layers.write() => guard,
4699 28 : _ = self.cancel.cancelled() => {
4700 0 : return Err(CompactionError::ShuttingDown);
4701 : }
4702 : };
4703 :
4704 28 : let mut duplicated_layers = HashSet::new();
4705 28 :
4706 28 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
4707 :
4708 336 : for l in new_deltas {
4709 308 : if guard.contains(l.as_ref()) {
4710 : // expected in tests
4711 0 : tracing::error!(layer=%l, "duplicated L1 layer");
4712 :
4713 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
4714 : // `guard` on self.layers. as of writing this, there are no error returns except
4715 : // for compact_level0_phase1 creating an L0, which does not happen in practice
4716 : // because we have not implemented L0 => L0 compaction.
4717 0 : duplicated_layers.insert(l.layer_desc().key());
4718 308 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
4719 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
4720 308 : } else {
4721 308 : insert_layers.push(l.clone());
4722 308 : }
4723 : }
4724 :
4725 : // only remove those inputs which were not outputs
4726 28 : let remove_layers: Vec<Layer> = layers_to_remove
4727 28 : .iter()
4728 402 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
4729 28 : .cloned()
4730 28 : .collect();
4731 28 :
4732 28 : if !new_images.is_empty() {
4733 0 : guard
4734 0 : .open_mut()?
4735 0 : .track_new_image_layers(new_images, &self.metrics);
4736 28 : }
4737 :
4738 28 : guard
4739 28 : .open_mut()?
4740 28 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
4741 28 :
4742 28 : self.remote_client
4743 28 : .schedule_compaction_update(&remove_layers, new_deltas)?;
4744 :
4745 28 : drop_wlock(guard);
4746 28 :
4747 28 : Ok(())
4748 28 : }
4749 :
4750 0 : async fn rewrite_layers(
4751 0 : self: &Arc<Self>,
4752 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
4753 0 : mut drop_layers: Vec<Layer>,
4754 0 : ) -> Result<(), CompactionError> {
4755 0 : let mut guard = self.layers.write().await;
4756 :
4757 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
4758 : // to avoid double-removing, and avoid rewriting something that was removed.
4759 0 : replace_layers.retain(|(l, _)| guard.contains(l));
4760 0 : drop_layers.retain(|l| guard.contains(l));
4761 0 :
4762 0 : guard
4763 0 : .open_mut()?
4764 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
4765 0 :
4766 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
4767 0 :
4768 0 : self.remote_client
4769 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
4770 :
4771 0 : Ok(())
4772 0 : }
4773 :
4774 : /// Schedules the uploads of the given image layers
4775 364 : fn upload_new_image_layers(
4776 364 : self: &Arc<Self>,
4777 364 : new_images: impl IntoIterator<Item = ResidentLayer>,
4778 364 : ) -> Result<(), super::upload_queue::NotInitialized> {
4779 390 : for layer in new_images {
4780 26 : self.remote_client.schedule_layer_file_upload(layer)?;
4781 : }
4782 : // should any new image layer been created, not uploading index_part will
4783 : // result in a mismatch between remote_physical_size and layermap calculated
4784 : // size, which will fail some tests, but should not be an issue otherwise.
4785 364 : self.remote_client
4786 364 : .schedule_index_upload_for_file_changes()?;
4787 364 : Ok(())
4788 364 : }
4789 :
4790 : /// Find the Lsns above which layer files need to be retained on
4791 : /// garbage collection.
4792 : ///
4793 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
4794 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
4795 : /// the space-based retention.
4796 : ///
4797 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
4798 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
4799 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
4800 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
4801 : /// in the response as the GC cutoff point for the timeline.
4802 4 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
4803 : pub(super) async fn find_gc_cutoffs(
4804 : &self,
4805 : space_cutoff: Lsn,
4806 : pitr: Duration,
4807 : cancel: &CancellationToken,
4808 : ctx: &RequestContext,
4809 : ) -> Result<GcCutoffs, PageReconstructError> {
4810 : let _timer = self
4811 : .metrics
4812 : .find_gc_cutoffs_histo
4813 : .start_timer()
4814 : .record_on_drop();
4815 :
4816 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
4817 :
4818 : if cfg!(test) {
4819 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
4820 : if pitr == Duration::ZERO {
4821 : return Ok(GcCutoffs {
4822 : time: self.get_last_record_lsn(),
4823 : space: space_cutoff,
4824 : });
4825 : }
4826 : }
4827 :
4828 : // Calculate a time-based limit on how much to retain:
4829 : // - if PITR interval is set, then this is our cutoff.
4830 : // - if PITR interval is not set, then we do a lookup
4831 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
4832 : let time_cutoff = {
4833 : let now = SystemTime::now();
4834 : let time_range = if pitr == Duration::ZERO {
4835 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
4836 : } else {
4837 : pitr
4838 : };
4839 :
4840 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
4841 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
4842 : let timestamp = to_pg_timestamp(time_cutoff);
4843 :
4844 : match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
4845 : LsnForTimestamp::Present(lsn) => Some(lsn),
4846 : LsnForTimestamp::Future(lsn) => {
4847 : // The timestamp is in the future. That sounds impossible,
4848 : // but what it really means is that there hasn't been
4849 : // any commits since the cutoff timestamp.
4850 : //
4851 : // In this case we should use the LSN of the most recent commit,
4852 : // which is implicitly the last LSN in the log.
4853 : debug!("future({})", lsn);
4854 : Some(self.get_last_record_lsn())
4855 : }
4856 : LsnForTimestamp::Past(lsn) => {
4857 : debug!("past({})", lsn);
4858 : None
4859 : }
4860 : LsnForTimestamp::NoData(lsn) => {
4861 : debug!("nodata({})", lsn);
4862 : None
4863 : }
4864 : }
4865 : };
4866 :
4867 : Ok(match (pitr, time_cutoff) {
4868 : (Duration::ZERO, Some(time_cutoff)) => {
4869 : // PITR is not set. Retain the size-based limit, or the default time retention,
4870 : // whichever requires less data.
4871 : GcCutoffs {
4872 : time: self.get_last_record_lsn(),
4873 : space: std::cmp::max(time_cutoff, space_cutoff),
4874 : }
4875 : }
4876 : (Duration::ZERO, None) => {
4877 : // PITR is not set, and time lookup failed
4878 : GcCutoffs {
4879 : time: self.get_last_record_lsn(),
4880 : space: space_cutoff,
4881 : }
4882 : }
4883 : (_, None) => {
4884 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
4885 : // cannot advance beyond what was already GC'd, and respect space-based retention
4886 : GcCutoffs {
4887 : time: *self.get_latest_gc_cutoff_lsn(),
4888 : space: space_cutoff,
4889 : }
4890 : }
4891 : (_, Some(time_cutoff)) => {
4892 : // PITR interval is set and we looked up timestamp successfully. Ignore
4893 : // size based retention and make time cutoff authoritative
4894 : GcCutoffs {
4895 : time: time_cutoff,
4896 : space: time_cutoff,
4897 : }
4898 : }
4899 : })
4900 : }
4901 :
4902 : /// Garbage collect layer files on a timeline that are no longer needed.
4903 : ///
4904 : /// Currently, we don't make any attempt at removing unneeded page versions
4905 : /// within a layer file. We can only remove the whole file if it's fully
4906 : /// obsolete.
4907 4 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
4908 : // this is most likely the background tasks, but it might be the spawned task from
4909 : // immediate_gc
4910 4 : let _g = tokio::select! {
4911 4 : guard = self.gc_lock.lock() => guard,
4912 4 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
4913 : };
4914 4 : let timer = self.metrics.garbage_collect_histo.start_timer();
4915 4 :
4916 4 : fail_point!("before-timeline-gc");
4917 4 :
4918 4 : // Is the timeline being deleted?
4919 4 : if self.is_stopping() {
4920 0 : return Err(GcError::TimelineCancelled);
4921 4 : }
4922 4 :
4923 4 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
4924 4 : let gc_info = self.gc_info.read().unwrap();
4925 4 :
4926 4 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
4927 4 : let time_cutoff = gc_info.cutoffs.time;
4928 4 : let retain_lsns = gc_info
4929 4 : .retain_lsns
4930 4 : .iter()
4931 4 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
4932 4 : .collect();
4933 4 :
4934 4 : // Gets the maximum LSN that holds the valid lease.
4935 4 : //
4936 4 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
4937 4 : // Here, we do not check for stale leases again.
4938 4 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
4939 4 :
4940 4 : (
4941 4 : space_cutoff,
4942 4 : time_cutoff,
4943 4 : retain_lsns,
4944 4 : max_lsn_with_valid_lease,
4945 4 : )
4946 4 : };
4947 4 :
4948 4 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
4949 4 : let standby_horizon = self.standby_horizon.load();
4950 4 : // Hold GC for the standby, but as a safety guard do it only within some
4951 4 : // reasonable lag.
4952 4 : if standby_horizon != Lsn::INVALID {
4953 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
4954 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
4955 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
4956 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
4957 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
4958 : } else {
4959 0 : warn!(
4960 0 : "standby is lagging for more than {}MB, not holding gc for it",
4961 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
4962 : )
4963 : }
4964 0 : }
4965 4 : }
4966 :
4967 : // Reset standby horizon to ignore it if it is not updated till next GC.
4968 : // It is an easy way to unset it when standby disappears without adding
4969 : // more conf options.
4970 4 : self.standby_horizon.store(Lsn::INVALID);
4971 4 : self.metrics
4972 4 : .standby_horizon_gauge
4973 4 : .set(Lsn::INVALID.0 as i64);
4974 :
4975 4 : let res = self
4976 4 : .gc_timeline(
4977 4 : space_cutoff,
4978 4 : time_cutoff,
4979 4 : retain_lsns,
4980 4 : max_lsn_with_valid_lease,
4981 4 : new_gc_cutoff,
4982 4 : )
4983 4 : .instrument(
4984 4 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
4985 : )
4986 0 : .await?;
4987 :
4988 : // only record successes
4989 4 : timer.stop_and_record();
4990 4 :
4991 4 : Ok(res)
4992 4 : }
4993 :
4994 4 : async fn gc_timeline(
4995 4 : &self,
4996 4 : space_cutoff: Lsn,
4997 4 : time_cutoff: Lsn,
4998 4 : retain_lsns: Vec<Lsn>,
4999 4 : max_lsn_with_valid_lease: Option<Lsn>,
5000 4 : new_gc_cutoff: Lsn,
5001 4 : ) -> Result<GcResult, GcError> {
5002 4 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5003 4 :
5004 4 : let now = SystemTime::now();
5005 4 : let mut result: GcResult = GcResult::default();
5006 4 :
5007 4 : // Nothing to GC. Return early.
5008 4 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
5009 4 : if latest_gc_cutoff >= new_gc_cutoff {
5010 0 : info!(
5011 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5012 : );
5013 0 : return Ok(result);
5014 4 : }
5015 :
5016 : // We need to ensure that no one tries to read page versions or create
5017 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5018 : // for details. This will block until the old value is no longer in use.
5019 : //
5020 : // The GC cutoff should only ever move forwards.
5021 4 : let waitlist = {
5022 4 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
5023 4 : if *write_guard > new_gc_cutoff {
5024 0 : return Err(GcError::BadLsn {
5025 0 : why: format!(
5026 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5027 0 : *write_guard, new_gc_cutoff
5028 0 : ),
5029 0 : });
5030 4 : }
5031 4 :
5032 4 : write_guard.store_and_unlock(new_gc_cutoff)
5033 4 : };
5034 4 : waitlist.wait().await;
5035 :
5036 4 : info!("GC starting");
5037 :
5038 4 : debug!("retain_lsns: {:?}", retain_lsns);
5039 :
5040 4 : let mut layers_to_remove = Vec::new();
5041 :
5042 : // Scan all layers in the timeline (remote or on-disk).
5043 : //
5044 : // Garbage collect the layer if all conditions are satisfied:
5045 : // 1. it is older than cutoff LSN;
5046 : // 2. it is older than PITR interval;
5047 : // 3. it doesn't need to be retained for 'retain_lsns';
5048 : // 4. it does not need to be kept for LSNs holding valid leases.
5049 : // 5. newer on-disk image layers cover the layer's whole key range
5050 : //
5051 : // TODO holding a write lock is too agressive and avoidable
5052 4 : let mut guard = self.layers.write().await;
5053 4 : let layers = guard.layer_map()?;
5054 24 : 'outer: for l in layers.iter_historic_layers() {
5055 24 : result.layers_total += 1;
5056 24 :
5057 24 : // 1. Is it newer than GC horizon cutoff point?
5058 24 : if l.get_lsn_range().end > space_cutoff {
5059 2 : info!(
5060 0 : "keeping {} because it's newer than space_cutoff {}",
5061 0 : l.layer_name(),
5062 : space_cutoff,
5063 : );
5064 2 : result.layers_needed_by_cutoff += 1;
5065 2 : continue 'outer;
5066 22 : }
5067 22 :
5068 22 : // 2. It is newer than PiTR cutoff point?
5069 22 : if l.get_lsn_range().end > time_cutoff {
5070 0 : info!(
5071 0 : "keeping {} because it's newer than time_cutoff {}",
5072 0 : l.layer_name(),
5073 : time_cutoff,
5074 : );
5075 0 : result.layers_needed_by_pitr += 1;
5076 0 : continue 'outer;
5077 22 : }
5078 :
5079 : // 3. Is it needed by a child branch?
5080 : // NOTE With that we would keep data that
5081 : // might be referenced by child branches forever.
5082 : // We can track this in child timeline GC and delete parent layers when
5083 : // they are no longer needed. This might be complicated with long inheritance chains.
5084 : //
5085 : // TODO Vec is not a great choice for `retain_lsns`
5086 22 : for retain_lsn in &retain_lsns {
5087 : // start_lsn is inclusive
5088 0 : if &l.get_lsn_range().start <= retain_lsn {
5089 0 : info!(
5090 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5091 0 : l.layer_name(),
5092 0 : retain_lsn,
5093 0 : l.is_incremental(),
5094 : );
5095 0 : result.layers_needed_by_branches += 1;
5096 0 : continue 'outer;
5097 0 : }
5098 : }
5099 :
5100 : // 4. Is there a valid lease that requires us to keep this layer?
5101 22 : if let Some(lsn) = &max_lsn_with_valid_lease {
5102 : // keep if layer start <= any of the lease
5103 18 : if &l.get_lsn_range().start <= lsn {
5104 14 : info!(
5105 0 : "keeping {} because there is a valid lease preventing GC at {}",
5106 0 : l.layer_name(),
5107 : lsn,
5108 : );
5109 14 : result.layers_needed_by_leases += 1;
5110 14 : continue 'outer;
5111 4 : }
5112 4 : }
5113 :
5114 : // 5. Is there a later on-disk layer for this relation?
5115 : //
5116 : // The end-LSN is exclusive, while disk_consistent_lsn is
5117 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5118 : // OK for a delta layer to have end LSN 101, but if the end LSN
5119 : // is 102, then it might not have been fully flushed to disk
5120 : // before crash.
5121 : //
5122 : // For example, imagine that the following layers exist:
5123 : //
5124 : // 1000 - image (A)
5125 : // 1000-2000 - delta (B)
5126 : // 2000 - image (C)
5127 : // 2000-3000 - delta (D)
5128 : // 3000 - image (E)
5129 : //
5130 : // If GC horizon is at 2500, we can remove layers A and B, but
5131 : // we cannot remove C, even though it's older than 2500, because
5132 : // the delta layer 2000-3000 depends on it.
5133 8 : if !layers
5134 8 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5135 : {
5136 6 : info!("keeping {} because it is the latest layer", l.layer_name());
5137 6 : result.layers_not_updated += 1;
5138 6 : continue 'outer;
5139 2 : }
5140 2 :
5141 2 : // We didn't find any reason to keep this file, so remove it.
5142 2 : info!(
5143 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5144 0 : l.layer_name(),
5145 0 : l.is_incremental(),
5146 : );
5147 2 : layers_to_remove.push(l);
5148 : }
5149 :
5150 4 : if !layers_to_remove.is_empty() {
5151 : // Persist the new GC cutoff value before we actually remove anything.
5152 : // This unconditionally schedules also an index_part.json update, even though, we will
5153 : // be doing one a bit later with the unlinked gc'd layers.
5154 2 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5155 2 : self.schedule_uploads(disk_consistent_lsn, None)
5156 2 : .map_err(|e| {
5157 0 : if self.cancel.is_cancelled() {
5158 0 : GcError::TimelineCancelled
5159 : } else {
5160 0 : GcError::Remote(e)
5161 : }
5162 2 : })?;
5163 :
5164 2 : let gc_layers = layers_to_remove
5165 2 : .iter()
5166 2 : .map(|x| guard.get_from_desc(x))
5167 2 : .collect::<Vec<Layer>>();
5168 2 :
5169 2 : result.layers_removed = gc_layers.len() as u64;
5170 2 :
5171 2 : self.remote_client.schedule_gc_update(&gc_layers)?;
5172 :
5173 2 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5174 2 :
5175 2 : #[cfg(feature = "testing")]
5176 2 : {
5177 2 : result.doomed_layers = gc_layers;
5178 2 : }
5179 2 : }
5180 :
5181 4 : info!(
5182 0 : "GC completed removing {} layers, cutoff {}",
5183 : result.layers_removed, new_gc_cutoff
5184 : );
5185 :
5186 4 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5187 4 : Ok(result)
5188 4 : }
5189 :
5190 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5191 667406 : async fn reconstruct_value(
5192 667406 : &self,
5193 667406 : key: Key,
5194 667406 : request_lsn: Lsn,
5195 667406 : mut data: ValueReconstructState,
5196 667406 : ) -> Result<Bytes, PageReconstructError> {
5197 667406 : // Perform WAL redo if needed
5198 667406 : data.records.reverse();
5199 667406 :
5200 667406 : // If we have a page image, and no WAL, we're all set
5201 667406 : if data.records.is_empty() {
5202 666996 : if let Some((img_lsn, img)) = &data.img {
5203 666996 : trace!(
5204 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5205 : key,
5206 : img_lsn,
5207 : request_lsn,
5208 : );
5209 666996 : Ok(img.clone())
5210 : } else {
5211 0 : Err(PageReconstructError::from(anyhow!(
5212 0 : "base image for {key} at {request_lsn} not found"
5213 0 : )))
5214 : }
5215 : } else {
5216 : // We need to do WAL redo.
5217 : //
5218 : // If we don't have a base image, then the oldest WAL record better initialize
5219 : // the page
5220 410 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5221 0 : Err(PageReconstructError::from(anyhow!(
5222 0 : "Base image for {} at {} not found, but got {} WAL records",
5223 0 : key,
5224 0 : request_lsn,
5225 0 : data.records.len()
5226 0 : )))
5227 : } else {
5228 410 : if data.img.is_some() {
5229 344 : trace!(
5230 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5231 0 : data.records.len(),
5232 : key,
5233 : request_lsn
5234 : );
5235 : } else {
5236 66 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5237 : };
5238 410 : let res = self
5239 410 : .walredo_mgr
5240 410 : .as_ref()
5241 410 : .context("timeline has no walredo manager")
5242 410 : .map_err(PageReconstructError::WalRedo)?
5243 410 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5244 0 : .await;
5245 410 : let img = match res {
5246 410 : Ok(img) => img,
5247 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5248 0 : Err(walredo::Error::Other(e)) => {
5249 0 : return Err(PageReconstructError::WalRedo(
5250 0 : e.context("reconstruct a page image"),
5251 0 : ))
5252 : }
5253 : };
5254 410 : Ok(img)
5255 : }
5256 : }
5257 667406 : }
5258 :
5259 0 : pub(crate) async fn spawn_download_all_remote_layers(
5260 0 : self: Arc<Self>,
5261 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5262 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5263 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5264 :
5265 : // this is not really needed anymore; it has tests which really check the return value from
5266 : // http api. it would be better not to maintain this anymore.
5267 :
5268 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5269 0 : if let Some(st) = &*status_guard {
5270 0 : match &st.state {
5271 : DownloadRemoteLayersTaskState::Running => {
5272 0 : return Err(st.clone());
5273 : }
5274 : DownloadRemoteLayersTaskState::ShutDown
5275 0 : | DownloadRemoteLayersTaskState::Completed => {
5276 0 : *status_guard = None;
5277 0 : }
5278 : }
5279 0 : }
5280 :
5281 0 : let self_clone = Arc::clone(&self);
5282 0 : let task_id = task_mgr::spawn(
5283 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5284 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5285 0 : self.tenant_shard_id,
5286 0 : Some(self.timeline_id),
5287 0 : "download all remote layers task",
5288 0 : async move {
5289 0 : self_clone.download_all_remote_layers(request).await;
5290 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5291 0 : match &mut *status_guard {
5292 : None => {
5293 0 : warn!("tasks status is supposed to be Some(), since we are running");
5294 : }
5295 0 : Some(st) => {
5296 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5297 0 : if st.task_id != exp_task_id {
5298 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5299 0 : } else {
5300 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5301 0 : }
5302 : }
5303 : };
5304 0 : Ok(())
5305 0 : }
5306 0 : .instrument(info_span!(parent: None, "download_all_remote_layers", tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id))
5307 : );
5308 :
5309 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5310 0 : task_id: format!("{task_id}"),
5311 0 : state: DownloadRemoteLayersTaskState::Running,
5312 0 : total_layer_count: 0,
5313 0 : successful_download_count: 0,
5314 0 : failed_download_count: 0,
5315 0 : };
5316 0 : *status_guard = Some(initial_info.clone());
5317 0 :
5318 0 : Ok(initial_info)
5319 0 : }
5320 :
5321 0 : async fn download_all_remote_layers(
5322 0 : self: &Arc<Self>,
5323 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5324 0 : ) {
5325 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5326 :
5327 0 : let remaining = {
5328 0 : let guard = self.layers.read().await;
5329 0 : let Ok(lm) = guard.layer_map() else {
5330 : // technically here we could look into iterating accessible layers, but downloading
5331 : // all layers of a shutdown timeline makes no sense regardless.
5332 0 : tracing::info!("attempted to download all layers of shutdown timeline");
5333 0 : return;
5334 : };
5335 0 : lm.iter_historic_layers()
5336 0 : .map(|desc| guard.get_from_desc(&desc))
5337 0 : .collect::<Vec<_>>()
5338 0 : };
5339 0 : let total_layer_count = remaining.len();
5340 :
5341 : macro_rules! lock_status {
5342 : ($st:ident) => {
5343 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5344 : let st = st
5345 : .as_mut()
5346 : .expect("this function is only called after the task has been spawned");
5347 : assert_eq!(
5348 : st.task_id,
5349 : format!(
5350 : "{}",
5351 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5352 : )
5353 : );
5354 : let $st = st;
5355 : };
5356 : }
5357 :
5358 : {
5359 0 : lock_status!(st);
5360 0 : st.total_layer_count = total_layer_count as u64;
5361 0 : }
5362 0 :
5363 0 : let mut remaining = remaining.into_iter();
5364 0 : let mut have_remaining = true;
5365 0 : let mut js = tokio::task::JoinSet::new();
5366 0 :
5367 0 : let cancel = task_mgr::shutdown_token();
5368 0 :
5369 0 : let limit = request.max_concurrent_downloads;
5370 :
5371 : loop {
5372 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5373 0 : let Some(next) = remaining.next() else {
5374 0 : have_remaining = false;
5375 0 : break;
5376 : };
5377 :
5378 0 : let span = tracing::info_span!("download", layer = %next);
5379 :
5380 0 : js.spawn(
5381 0 : async move {
5382 0 : let res = next.download().await;
5383 0 : (next, res)
5384 0 : }
5385 0 : .instrument(span),
5386 0 : );
5387 0 : }
5388 :
5389 0 : while let Some(res) = js.join_next().await {
5390 0 : match res {
5391 : Ok((_, Ok(_))) => {
5392 0 : lock_status!(st);
5393 0 : st.successful_download_count += 1;
5394 : }
5395 0 : Ok((layer, Err(e))) => {
5396 0 : tracing::error!(%layer, "download failed: {e:#}");
5397 0 : lock_status!(st);
5398 0 : st.failed_download_count += 1;
5399 : }
5400 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5401 0 : Err(je) if je.is_panic() => {
5402 0 : lock_status!(st);
5403 0 : st.failed_download_count += 1;
5404 : }
5405 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5406 : }
5407 : }
5408 :
5409 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5410 0 : break;
5411 0 : }
5412 : }
5413 :
5414 : {
5415 0 : lock_status!(st);
5416 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5417 : }
5418 0 : }
5419 :
5420 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5421 0 : &self,
5422 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5423 0 : self.download_all_remote_layers_task_info
5424 0 : .read()
5425 0 : .unwrap()
5426 0 : .clone()
5427 0 : }
5428 : }
5429 :
5430 : impl Timeline {
5431 : /// Returns non-remote layers for eviction.
5432 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5433 0 : let guard = self.layers.read().await;
5434 0 : let mut max_layer_size: Option<u64> = None;
5435 0 :
5436 0 : let resident_layers = guard
5437 0 : .likely_resident_layers()
5438 0 : .map(|layer| {
5439 0 : let file_size = layer.layer_desc().file_size;
5440 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5441 0 :
5442 0 : let last_activity_ts = layer.latest_activity();
5443 0 :
5444 0 : EvictionCandidate {
5445 0 : layer: layer.to_owned().into(),
5446 0 : last_activity_ts,
5447 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5448 0 : visibility: layer.visibility(),
5449 0 : }
5450 0 : })
5451 0 : .collect();
5452 0 :
5453 0 : DiskUsageEvictionInfo {
5454 0 : max_layer_size,
5455 0 : resident_layers,
5456 0 : }
5457 0 : }
5458 :
5459 1730 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5460 1730 : ShardIndex {
5461 1730 : shard_number: self.tenant_shard_id.shard_number,
5462 1730 : shard_count: self.tenant_shard_id.shard_count,
5463 1730 : }
5464 1730 : }
5465 :
5466 : /// Persistently blocks gc for `Manual` reason.
5467 : ///
5468 : /// Returns true if no such block existed before, false otherwise.
5469 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
5470 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5471 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5472 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
5473 0 : }
5474 :
5475 : /// Persistently unblocks gc for `Manual` reason.
5476 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
5477 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5478 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5479 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
5480 0 : }
5481 :
5482 : #[cfg(test)]
5483 40 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5484 40 : self.last_record_lsn.advance(new_lsn);
5485 40 : }
5486 :
5487 : #[cfg(test)]
5488 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
5489 2 : self.disk_consistent_lsn.store(new_value);
5490 2 : }
5491 :
5492 : /// Force create an image layer and place it into the layer map.
5493 : ///
5494 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5495 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5496 : /// placed into the layer map in one run AND be validated.
5497 : #[cfg(test)]
5498 52 : pub(super) async fn force_create_image_layer(
5499 52 : self: &Arc<Timeline>,
5500 52 : lsn: Lsn,
5501 52 : mut images: Vec<(Key, Bytes)>,
5502 52 : check_start_lsn: Option<Lsn>,
5503 52 : ctx: &RequestContext,
5504 52 : ) -> anyhow::Result<()> {
5505 52 : let last_record_lsn = self.get_last_record_lsn();
5506 52 : assert!(
5507 52 : lsn <= last_record_lsn,
5508 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5509 : );
5510 52 : if let Some(check_start_lsn) = check_start_lsn {
5511 52 : assert!(lsn >= check_start_lsn);
5512 0 : }
5513 126 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5514 52 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5515 52 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5516 52 : let mut image_layer_writer = ImageLayerWriter::new(
5517 52 : self.conf,
5518 52 : self.timeline_id,
5519 52 : self.tenant_shard_id,
5520 52 : &(min_key..end_key),
5521 52 : lsn,
5522 52 : ctx,
5523 52 : )
5524 26 : .await?;
5525 230 : for (key, img) in images {
5526 178 : image_layer_writer.put_image(key, img, ctx).await?;
5527 : }
5528 104 : let (desc, path) = image_layer_writer.finish(ctx).await?;
5529 52 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5530 52 : info!("force created image layer {}", image_layer.local_path());
5531 : {
5532 52 : let mut guard = self.layers.write().await;
5533 52 : guard.open_mut().unwrap().force_insert_layer(image_layer);
5534 52 : }
5535 52 :
5536 52 : Ok(())
5537 52 : }
5538 :
5539 : /// Force create a delta layer and place it into the layer map.
5540 : ///
5541 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5542 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5543 : /// placed into the layer map in one run AND be validated.
5544 : #[cfg(test)]
5545 74 : pub(super) async fn force_create_delta_layer(
5546 74 : self: &Arc<Timeline>,
5547 74 : mut deltas: DeltaLayerTestDesc,
5548 74 : check_start_lsn: Option<Lsn>,
5549 74 : ctx: &RequestContext,
5550 74 : ) -> anyhow::Result<()> {
5551 74 : let last_record_lsn = self.get_last_record_lsn();
5552 74 : deltas
5553 74 : .data
5554 124 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5555 74 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5556 74 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5557 272 : for (_, lsn, _) in &deltas.data {
5558 198 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5559 : }
5560 74 : assert!(
5561 74 : deltas.lsn_range.end <= last_record_lsn,
5562 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
5563 : deltas.lsn_range.end,
5564 : last_record_lsn
5565 : );
5566 74 : if let Some(check_start_lsn) = check_start_lsn {
5567 74 : assert!(deltas.lsn_range.start >= check_start_lsn);
5568 0 : }
5569 74 : let mut delta_layer_writer = DeltaLayerWriter::new(
5570 74 : self.conf,
5571 74 : self.timeline_id,
5572 74 : self.tenant_shard_id,
5573 74 : deltas.key_range.start,
5574 74 : deltas.lsn_range,
5575 74 : ctx,
5576 74 : )
5577 37 : .await?;
5578 272 : for (key, lsn, val) in deltas.data {
5579 198 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5580 : }
5581 185 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
5582 74 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5583 74 : info!("force created delta layer {}", delta_layer.local_path());
5584 : {
5585 74 : let mut guard = self.layers.write().await;
5586 74 : guard.open_mut().unwrap().force_insert_layer(delta_layer);
5587 74 : }
5588 74 :
5589 74 : Ok(())
5590 74 : }
5591 :
5592 : /// Return all keys at the LSN in the image layers
5593 : #[cfg(test)]
5594 6 : pub(crate) async fn inspect_image_layers(
5595 6 : self: &Arc<Timeline>,
5596 6 : lsn: Lsn,
5597 6 : ctx: &RequestContext,
5598 6 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
5599 6 : let mut all_data = Vec::new();
5600 6 : let guard = self.layers.read().await;
5601 34 : for layer in guard.layer_map()?.iter_historic_layers() {
5602 34 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
5603 8 : let layer = guard.get_from_desc(&layer);
5604 8 : let mut reconstruct_data = ValuesReconstructState::default();
5605 8 : layer
5606 8 : .get_values_reconstruct_data(
5607 8 : KeySpace::single(Key::MIN..Key::MAX),
5608 8 : lsn..Lsn(lsn.0 + 1),
5609 8 : &mut reconstruct_data,
5610 8 : ctx,
5611 8 : )
5612 13 : .await?;
5613 74 : for (k, v) in reconstruct_data.keys {
5614 66 : all_data.push((k, v?.img.unwrap().1));
5615 : }
5616 26 : }
5617 : }
5618 6 : all_data.sort();
5619 6 : Ok(all_data)
5620 6 : }
5621 :
5622 : /// Get all historic layer descriptors in the layer map
5623 : #[cfg(test)]
5624 12 : pub(crate) async fn inspect_historic_layers(
5625 12 : self: &Arc<Timeline>,
5626 12 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
5627 12 : let mut layers = Vec::new();
5628 12 : let guard = self.layers.read().await;
5629 70 : for layer in guard.layer_map()?.iter_historic_layers() {
5630 70 : layers.push(layer.key());
5631 70 : }
5632 12 : Ok(layers)
5633 12 : }
5634 :
5635 : #[cfg(test)]
5636 10 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
5637 10 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
5638 10 : keyspace.merge(&ks);
5639 10 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
5640 10 : }
5641 : }
5642 :
5643 : /// Tracking writes ingestion does to a particular in-memory layer.
5644 : ///
5645 : /// Cleared upon freezing a layer.
5646 : pub(crate) struct TimelineWriterState {
5647 : open_layer: Arc<InMemoryLayer>,
5648 : current_size: u64,
5649 : // Previous Lsn which passed through
5650 : prev_lsn: Option<Lsn>,
5651 : // Largest Lsn which passed through the current writer
5652 : max_lsn: Option<Lsn>,
5653 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
5654 : cached_last_freeze_at: Lsn,
5655 : }
5656 :
5657 : impl TimelineWriterState {
5658 1268 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
5659 1268 : Self {
5660 1268 : open_layer,
5661 1268 : current_size,
5662 1268 : prev_lsn: None,
5663 1268 : max_lsn: None,
5664 1268 : cached_last_freeze_at: last_freeze_at,
5665 1268 : }
5666 1268 : }
5667 : }
5668 :
5669 : /// Various functions to mutate the timeline.
5670 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
5671 : // This is probably considered a bad practice in Rust and should be fixed eventually,
5672 : // but will cause large code changes.
5673 : pub(crate) struct TimelineWriter<'a> {
5674 : tl: &'a Timeline,
5675 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
5676 : }
5677 :
5678 : impl Deref for TimelineWriter<'_> {
5679 : type Target = Timeline;
5680 :
5681 4807192 : fn deref(&self) -> &Self::Target {
5682 4807192 : self.tl
5683 4807192 : }
5684 : }
5685 :
5686 : #[derive(PartialEq)]
5687 : enum OpenLayerAction {
5688 : Roll,
5689 : Open,
5690 : None,
5691 : }
5692 :
5693 : impl<'a> TimelineWriter<'a> {
5694 4804204 : async fn handle_open_layer_action(
5695 4804204 : &mut self,
5696 4804204 : at: Lsn,
5697 4804204 : action: OpenLayerAction,
5698 4804204 : ctx: &RequestContext,
5699 4804204 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
5700 4804204 : match action {
5701 : OpenLayerAction::Roll => {
5702 80 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
5703 80 : self.roll_layer(freeze_at).await?;
5704 80 : self.open_layer(at, ctx).await?;
5705 : }
5706 1188 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
5707 : OpenLayerAction::None => {
5708 4802936 : assert!(self.write_guard.is_some());
5709 : }
5710 : }
5711 :
5712 4804204 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
5713 4804204 : }
5714 :
5715 1268 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
5716 1268 : let layer = self
5717 1268 : .tl
5718 1268 : .get_layer_for_write(at, &self.write_guard, ctx)
5719 719 : .await?;
5720 1268 : let initial_size = layer.size().await?;
5721 :
5722 1268 : let last_freeze_at = self.last_freeze_at.load();
5723 1268 : self.write_guard.replace(TimelineWriterState::new(
5724 1268 : layer,
5725 1268 : initial_size,
5726 1268 : last_freeze_at,
5727 1268 : ));
5728 1268 :
5729 1268 : Ok(())
5730 1268 : }
5731 :
5732 80 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
5733 80 : let current_size = self.write_guard.as_ref().unwrap().current_size;
5734 80 :
5735 80 : // self.write_guard will be taken by the freezing
5736 80 : self.tl
5737 80 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
5738 4 : .await?;
5739 :
5740 80 : assert!(self.write_guard.is_none());
5741 :
5742 80 : if current_size >= self.get_checkpoint_distance() * 2 {
5743 0 : warn!("Flushed oversized open layer with size {}", current_size)
5744 80 : }
5745 :
5746 80 : Ok(())
5747 80 : }
5748 :
5749 4804204 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
5750 4804204 : let state = &*self.write_guard;
5751 4804204 : let Some(state) = &state else {
5752 1188 : return OpenLayerAction::Open;
5753 : };
5754 :
5755 : #[cfg(feature = "testing")]
5756 4803016 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
5757 : // this check and assertion are not really needed because
5758 : // LayerManager::try_freeze_in_memory_layer will always clear out the
5759 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
5760 : // is no TimelineWriterState.
5761 0 : assert!(
5762 0 : state.open_layer.end_lsn.get().is_some(),
5763 0 : "our open_layer must be outdated"
5764 : );
5765 :
5766 : // this would be a memory leak waiting to happen because the in-memory layer always has
5767 : // an index
5768 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
5769 4803016 : }
5770 4803016 :
5771 4803016 : if state.prev_lsn == Some(lsn) {
5772 : // Rolling mid LSN is not supported by [downstream code].
5773 : // Hence, only roll at LSN boundaries.
5774 : //
5775 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
5776 6 : return OpenLayerAction::None;
5777 4803010 : }
5778 4803010 :
5779 4803010 : if state.current_size == 0 {
5780 : // Don't roll empty layers
5781 0 : return OpenLayerAction::None;
5782 4803010 : }
5783 4803010 :
5784 4803010 : if self.tl.should_roll(
5785 4803010 : state.current_size,
5786 4803010 : state.current_size + new_value_size,
5787 4803010 : self.get_checkpoint_distance(),
5788 4803010 : lsn,
5789 4803010 : state.cached_last_freeze_at,
5790 4803010 : state.open_layer.get_opened_at(),
5791 4803010 : ) {
5792 80 : OpenLayerAction::Roll
5793 : } else {
5794 4802930 : OpenLayerAction::None
5795 : }
5796 4804204 : }
5797 :
5798 : /// Put a batch of keys at the specified Lsns.
5799 4804202 : pub(crate) async fn put_batch(
5800 4804202 : &mut self,
5801 4804202 : batch: SerializedValueBatch,
5802 4804202 : ctx: &RequestContext,
5803 4804202 : ) -> anyhow::Result<()> {
5804 4804202 : if batch.is_empty() {
5805 0 : return Ok(());
5806 4804202 : }
5807 4804202 :
5808 4804202 : let batch_max_lsn = batch.max_lsn;
5809 4804202 : let buf_size: u64 = batch.buffer_size() as u64;
5810 4804202 :
5811 4804202 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
5812 4804202 : let layer = self
5813 4804202 : .handle_open_layer_action(batch_max_lsn, action, ctx)
5814 723 : .await?;
5815 :
5816 4804202 : let res = layer.put_batch(batch, ctx).await;
5817 :
5818 4804202 : if res.is_ok() {
5819 4804202 : // Update the current size only when the entire write was ok.
5820 4804202 : // In case of failures, we may have had partial writes which
5821 4804202 : // render the size tracking out of sync. That's ok because
5822 4804202 : // the checkpoint distance should be significantly smaller
5823 4804202 : // than the S3 single shot upload limit of 5GiB.
5824 4804202 : let state = self.write_guard.as_mut().unwrap();
5825 4804202 :
5826 4804202 : state.current_size += buf_size;
5827 4804202 : state.prev_lsn = Some(batch_max_lsn);
5828 4804202 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
5829 4804202 : }
5830 :
5831 4804202 : res
5832 4804202 : }
5833 :
5834 : #[cfg(test)]
5835 : /// Test helper, for tests that would like to poke individual values without composing a batch
5836 4390154 : pub(crate) async fn put(
5837 4390154 : &mut self,
5838 4390154 : key: Key,
5839 4390154 : lsn: Lsn,
5840 4390154 : value: &Value,
5841 4390154 : ctx: &RequestContext,
5842 4390154 : ) -> anyhow::Result<()> {
5843 : use utils::bin_ser::BeSer;
5844 4390154 : if !key.is_valid_key_on_write_path() {
5845 0 : bail!(
5846 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
5847 0 : key
5848 0 : );
5849 4390154 : }
5850 4390154 : let val_ser_size = value.serialized_size().unwrap() as usize;
5851 4390154 : let batch = SerializedValueBatch::from_values(vec![(
5852 4390154 : key.to_compact(),
5853 4390154 : lsn,
5854 4390154 : val_ser_size,
5855 4390154 : value.clone(),
5856 4390154 : )]);
5857 4390154 :
5858 4390154 : self.put_batch(batch, ctx).await
5859 4390154 : }
5860 :
5861 2 : pub(crate) async fn delete_batch(
5862 2 : &mut self,
5863 2 : batch: &[(Range<Key>, Lsn)],
5864 2 : ctx: &RequestContext,
5865 2 : ) -> anyhow::Result<()> {
5866 2 : if let Some((_, lsn)) = batch.first() {
5867 2 : let action = self.get_open_layer_action(*lsn, 0);
5868 2 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
5869 2 : layer.put_tombstones(batch).await?;
5870 0 : }
5871 :
5872 2 : Ok(())
5873 2 : }
5874 :
5875 : /// Track the end of the latest digested WAL record.
5876 : /// Remember the (end of) last valid WAL record remembered in the timeline.
5877 : ///
5878 : /// Call this after you have finished writing all the WAL up to 'lsn'.
5879 : ///
5880 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
5881 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
5882 : /// the latest and can be read.
5883 5279064 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
5884 5279064 : self.tl.finish_write(new_lsn);
5885 5279064 : }
5886 :
5887 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
5888 270570 : self.tl.update_current_logical_size(delta)
5889 270570 : }
5890 : }
5891 :
5892 : // We need TimelineWriter to be send in upcoming conversion of
5893 : // Timeline::layers to tokio::sync::RwLock.
5894 : #[test]
5895 2 : fn is_send() {
5896 2 : fn _assert_send<T: Send>() {}
5897 2 : _assert_send::<TimelineWriter<'_>>();
5898 2 : }
5899 :
5900 : #[cfg(test)]
5901 : mod tests {
5902 : use pageserver_api::key::Key;
5903 : use pageserver_api::value::Value;
5904 : use utils::{id::TimelineId, lsn::Lsn};
5905 :
5906 : use crate::tenant::{
5907 : harness::{test_img, TenantHarness},
5908 : layer_map::LayerMap,
5909 : storage_layer::{Layer, LayerName},
5910 : timeline::{DeltaLayerTestDesc, EvictionError},
5911 : Timeline,
5912 : };
5913 :
5914 : #[tokio::test]
5915 2 : async fn test_heatmap_generation() {
5916 2 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
5917 2 :
5918 2 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
5919 2 : Lsn(0x10)..Lsn(0x20),
5920 2 : vec![(
5921 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
5922 2 : Lsn(0x11),
5923 2 : Value::Image(test_img("foo")),
5924 2 : )],
5925 2 : );
5926 2 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
5927 2 : Lsn(0x10)..Lsn(0x20),
5928 2 : vec![(
5929 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
5930 2 : Lsn(0x11),
5931 2 : Value::Image(test_img("foo")),
5932 2 : )],
5933 2 : );
5934 2 : let l0_delta = DeltaLayerTestDesc::new(
5935 2 : Lsn(0x20)..Lsn(0x30),
5936 2 : Key::from_hex("000000000000000000000000000000000000").unwrap()
5937 2 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
5938 2 : vec![(
5939 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
5940 2 : Lsn(0x25),
5941 2 : Value::Image(test_img("foo")),
5942 2 : )],
5943 2 : );
5944 2 : let delta_layers = vec![
5945 2 : covered_delta.clone(),
5946 2 : visible_delta.clone(),
5947 2 : l0_delta.clone(),
5948 2 : ];
5949 2 :
5950 2 : let image_layer = (
5951 2 : Lsn(0x40),
5952 2 : vec![(
5953 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
5954 2 : test_img("bar"),
5955 2 : )],
5956 2 : );
5957 2 : let image_layers = vec![image_layer];
5958 2 :
5959 20 : let (tenant, ctx) = harness.load().await;
5960 2 : let timeline = tenant
5961 2 : .create_test_timeline_with_layers(
5962 2 : TimelineId::generate(),
5963 2 : Lsn(0x10),
5964 2 : 14,
5965 2 : &ctx,
5966 2 : delta_layers,
5967 2 : image_layers,
5968 2 : Lsn(0x100),
5969 2 : )
5970 31 : .await
5971 2 : .unwrap();
5972 2 :
5973 2 : // Layer visibility is an input to heatmap generation, so refresh it first
5974 2 : timeline.update_layer_visibility().await.unwrap();
5975 2 :
5976 2 : let heatmap = timeline
5977 2 : .generate_heatmap()
5978 2 : .await
5979 2 : .expect("Infallible while timeline is not shut down");
5980 2 :
5981 2 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
5982 2 :
5983 2 : // L0 should come last
5984 2 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
5985 2 :
5986 2 : let mut last_lsn = Lsn::MAX;
5987 10 : for layer in heatmap.layers {
5988 2 : // Covered layer should be omitted
5989 8 : assert!(layer.name != covered_delta.layer_name());
5990 2 :
5991 8 : let layer_lsn = match &layer.name {
5992 4 : LayerName::Delta(d) => d.lsn_range.end,
5993 4 : LayerName::Image(i) => i.lsn,
5994 2 : };
5995 2 :
5996 2 : // Apart from L0s, newest Layers should come first
5997 8 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
5998 6 : assert!(layer_lsn <= last_lsn);
5999 6 : last_lsn = layer_lsn;
6000 2 : }
6001 2 : }
6002 2 : }
6003 :
6004 : #[tokio::test]
6005 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
6006 2 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6007 2 : .await
6008 2 : .unwrap();
6009 2 :
6010 20 : let (tenant, ctx) = harness.load().await;
6011 2 : let timeline = tenant
6012 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6013 6 : .await
6014 2 : .unwrap();
6015 2 :
6016 2 : let layer = find_some_layer(&timeline).await;
6017 2 : let layer = layer
6018 2 : .keep_resident()
6019 2 : .await
6020 2 : .expect("no download => no downloading errors")
6021 2 : .drop_eviction_guard();
6022 2 :
6023 2 : let forever = std::time::Duration::from_secs(120);
6024 2 :
6025 2 : let first = layer.evict_and_wait(forever);
6026 2 : let second = layer.evict_and_wait(forever);
6027 2 :
6028 2 : let (first, second) = tokio::join!(first, second);
6029 2 :
6030 2 : let res = layer.keep_resident().await;
6031 2 : assert!(res.is_none(), "{res:?}");
6032 2 :
6033 2 : match (first, second) {
6034 2 : (Ok(()), Ok(())) => {
6035 2 : // because there are no more timeline locks being taken on eviction path, we can
6036 2 : // witness all three outcomes here.
6037 2 : }
6038 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
6039 0 : // if one completes before the other, this is fine just as well.
6040 0 : }
6041 2 : other => unreachable!("unexpected {:?}", other),
6042 2 : }
6043 2 : }
6044 :
6045 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
6046 2 : let layers = timeline.layers.read().await;
6047 2 : let desc = layers
6048 2 : .layer_map()
6049 2 : .unwrap()
6050 2 : .iter_historic_layers()
6051 2 : .next()
6052 2 : .expect("must find one layer to evict");
6053 2 :
6054 2 : layers.get_from_desc(&desc)
6055 2 : }
6056 : }
|
