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 : #[serde_with::serde_as]
778 0 : #[derive(Debug, Clone, serde::Deserialize)]
779 : pub(crate) struct CompactRange {
780 : #[serde_as(as = "serde_with::DisplayFromStr")]
781 : pub start: Key,
782 : #[serde_as(as = "serde_with::DisplayFromStr")]
783 : pub end: Key,
784 : }
785 :
786 : #[derive(Clone, Default)]
787 : pub(crate) struct CompactOptions {
788 : pub flags: EnumSet<CompactFlags>,
789 : pub compact_range: Option<CompactRange>,
790 : }
791 :
792 : impl std::fmt::Debug for Timeline {
793 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
794 0 : write!(f, "Timeline<{}>", self.timeline_id)
795 0 : }
796 : }
797 :
798 0 : #[derive(thiserror::Error, Debug)]
799 : pub(crate) enum WaitLsnError {
800 : // Called on a timeline which is shutting down
801 : #[error("Shutdown")]
802 : Shutdown,
803 :
804 : // Called on an timeline not in active state or shutting down
805 : #[error("Bad timeline state: {0:?}")]
806 : BadState(TimelineState),
807 :
808 : // Timeout expired while waiting for LSN to catch up with goal.
809 : #[error("{0}")]
810 : Timeout(String),
811 : }
812 :
813 : // The impls below achieve cancellation mapping for errors.
814 : // Perhaps there's a way of achieving this with less cruft.
815 :
816 : impl From<CreateImageLayersError> for CompactionError {
817 0 : fn from(e: CreateImageLayersError) -> Self {
818 0 : match e {
819 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
820 0 : CreateImageLayersError::Other(e) => {
821 0 : CompactionError::Other(e.context("create image layers"))
822 : }
823 0 : _ => CompactionError::Other(e.into()),
824 : }
825 0 : }
826 : }
827 :
828 : impl From<CreateImageLayersError> for FlushLayerError {
829 0 : fn from(e: CreateImageLayersError) -> Self {
830 0 : match e {
831 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
832 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
833 : }
834 0 : }
835 : }
836 :
837 : impl From<PageReconstructError> for CreateImageLayersError {
838 0 : fn from(e: PageReconstructError) -> Self {
839 0 : match e {
840 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
841 0 : _ => CreateImageLayersError::PageReconstructError(e),
842 : }
843 0 : }
844 : }
845 :
846 : impl From<GetVectoredError> for CreateImageLayersError {
847 0 : fn from(e: GetVectoredError) -> Self {
848 0 : match e {
849 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
850 0 : _ => CreateImageLayersError::GetVectoredError(e),
851 : }
852 0 : }
853 : }
854 :
855 : impl From<GetVectoredError> for PageReconstructError {
856 6 : fn from(e: GetVectoredError) -> Self {
857 6 : match e {
858 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
859 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
860 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
861 4 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
862 2 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
863 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
864 : }
865 6 : }
866 : }
867 :
868 : impl From<GetReadyAncestorError> for PageReconstructError {
869 2 : fn from(e: GetReadyAncestorError) -> Self {
870 : use GetReadyAncestorError::*;
871 2 : match e {
872 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
873 2 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
874 0 : Cancelled => PageReconstructError::Cancelled,
875 : }
876 2 : }
877 : }
878 :
879 : pub(crate) enum WaitLsnWaiter<'a> {
880 : Timeline(&'a Timeline),
881 : Tenant,
882 : PageService,
883 : }
884 :
885 : /// Argument to [`Timeline::shutdown`].
886 : #[derive(Debug, Clone, Copy)]
887 : pub(crate) enum ShutdownMode {
888 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
889 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
890 : ///
891 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
892 : /// the call to [`Timeline::shutdown`].
893 : FreezeAndFlush,
894 : /// Only flush the layers to the remote storage without freezing any open layers. This is the
895 : /// mode used by ancestor detach and any other operations that reloads a tenant but not increasing
896 : /// the generation number.
897 : Flush,
898 : /// Shut down immediately, without waiting for any open layers to flush.
899 : Hard,
900 : }
901 :
902 : struct ImageLayerCreationOutcome {
903 : image: Option<ResidentLayer>,
904 : next_start_key: Key,
905 : }
906 :
907 : /// Public interface functions
908 : impl Timeline {
909 : /// Get the LSN where this branch was created
910 4 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
911 4 : self.ancestor_lsn
912 4 : }
913 :
914 : /// Get the ancestor's timeline id
915 12 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
916 12 : self.ancestor_timeline
917 12 : .as_ref()
918 12 : .map(|ancestor| ancestor.timeline_id)
919 12 : }
920 :
921 : /// Get the ancestor timeline
922 2 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
923 2 : self.ancestor_timeline.as_ref()
924 2 : }
925 :
926 : /// Get the bytes written since the PITR cutoff on this branch, and
927 : /// whether this branch's ancestor_lsn is within its parent's PITR.
928 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
929 0 : let gc_info = self.gc_info.read().unwrap();
930 0 : let history = self
931 0 : .get_last_record_lsn()
932 0 : .checked_sub(gc_info.cutoffs.time)
933 0 : .unwrap_or(Lsn(0))
934 0 : .0;
935 0 : (history, gc_info.within_ancestor_pitr)
936 0 : }
937 :
938 : /// Lock and get timeline's GC cutoff
939 248 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
940 248 : self.latest_gc_cutoff_lsn.read()
941 248 : }
942 :
943 : /// Look up given page version.
944 : ///
945 : /// If a remote layer file is needed, it is downloaded as part of this
946 : /// call.
947 : ///
948 : /// This method enforces [`Self::timeline_get_throttle`] internally.
949 : ///
950 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
951 : /// abstraction above this needs to store suitable metadata to track what
952 : /// data exists with what keys, in separate metadata entries. If a
953 : /// non-existent key is requested, we may incorrectly return a value from
954 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
955 : /// non-existing key.
956 : ///
957 : /// # Cancel-Safety
958 : ///
959 : /// This method is cancellation-safe.
960 : #[inline(always)]
961 606758 : pub(crate) async fn get(
962 606758 : &self,
963 606758 : key: Key,
964 606758 : lsn: Lsn,
965 606758 : ctx: &RequestContext,
966 606758 : ) -> Result<Bytes, PageReconstructError> {
967 606758 : if !lsn.is_valid() {
968 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
969 606758 : }
970 606758 :
971 606758 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
972 606758 : // already checked the key against the shard_identity when looking up the Timeline from
973 606758 : // page_service.
974 606758 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
975 :
976 606758 : self.timeline_get_throttle.throttle(ctx, 1).await;
977 :
978 606758 : let keyspace = KeySpace {
979 606758 : ranges: vec![key..key.next()],
980 606758 : };
981 606758 :
982 606758 : // Initialise the reconstruct state for the key with the cache
983 606758 : // entry returned above.
984 606758 : let mut reconstruct_state = ValuesReconstructState::new();
985 :
986 606758 : let vectored_res = self
987 606758 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
988 183036 : .await;
989 :
990 606758 : let key_value = vectored_res?.pop_first();
991 606752 : match key_value {
992 606740 : Some((got_key, value)) => {
993 606740 : if got_key != key {
994 0 : error!(
995 0 : "Expected {}, but singular vectored get returned {}",
996 : key, got_key
997 : );
998 0 : Err(PageReconstructError::Other(anyhow!(
999 0 : "Singular vectored get returned wrong key"
1000 0 : )))
1001 : } else {
1002 606740 : value
1003 : }
1004 : }
1005 12 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1006 12 : key,
1007 12 : shard: self.shard_identity.get_shard_number(&key),
1008 12 : cont_lsn: Lsn(0),
1009 12 : request_lsn: lsn,
1010 12 : ancestor_lsn: None,
1011 12 : backtrace: None,
1012 12 : })),
1013 : }
1014 606758 : }
1015 :
1016 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1017 : pub(crate) const VEC_GET_LAYERS_VISITED_WARN_THRESH: f64 = 512.0;
1018 :
1019 : /// Look up multiple page versions at a given LSN
1020 : ///
1021 : /// This naive implementation will be replaced with a more efficient one
1022 : /// which actually vectorizes the read path.
1023 19528 : pub(crate) async fn get_vectored(
1024 19528 : &self,
1025 19528 : keyspace: KeySpace,
1026 19528 : lsn: Lsn,
1027 19528 : ctx: &RequestContext,
1028 19528 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1029 19528 : if !lsn.is_valid() {
1030 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1031 19528 : }
1032 19528 :
1033 19528 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1034 19528 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1035 0 : return Err(GetVectoredError::Oversized(key_count));
1036 19528 : }
1037 :
1038 39056 : for range in &keyspace.ranges {
1039 19528 : let mut key = range.start;
1040 39262 : while key != range.end {
1041 19734 : assert!(!self.shard_identity.is_key_disposable(&key));
1042 19734 : key = key.next();
1043 : }
1044 : }
1045 :
1046 19528 : trace!(
1047 0 : "get vectored request for {:?}@{} from task kind {:?}",
1048 0 : keyspace,
1049 0 : lsn,
1050 0 : ctx.task_kind(),
1051 : );
1052 :
1053 19528 : let start = crate::metrics::GET_VECTORED_LATENCY
1054 19528 : .for_task_kind(ctx.task_kind())
1055 19528 : .map(|metric| (metric, Instant::now()));
1056 :
1057 : // start counting after throttle so that throttle time
1058 : // is always less than observation time
1059 19528 : let throttled = self
1060 19528 : .timeline_get_throttle
1061 19528 : .throttle(ctx, key_count as usize)
1062 0 : .await;
1063 :
1064 19528 : let res = self
1065 19528 : .get_vectored_impl(
1066 19528 : keyspace.clone(),
1067 19528 : lsn,
1068 19528 : &mut ValuesReconstructState::new(),
1069 19528 : ctx,
1070 19528 : )
1071 7037 : .await;
1072 :
1073 19528 : if let Some((metric, start)) = start {
1074 0 : let elapsed = start.elapsed();
1075 0 : let ex_throttled = if let Some(throttled) = throttled {
1076 0 : elapsed.checked_sub(throttled)
1077 : } else {
1078 0 : Some(elapsed)
1079 : };
1080 :
1081 0 : if let Some(ex_throttled) = ex_throttled {
1082 0 : metric.observe(ex_throttled.as_secs_f64());
1083 0 : } else {
1084 0 : use utils::rate_limit::RateLimit;
1085 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1086 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
1087 0 : let mut rate_limit = LOGGED.lock().unwrap();
1088 0 : rate_limit.call(|| {
1089 0 : warn!("error deducting time spent throttled; this message is logged at a global rate limit");
1090 0 : });
1091 0 : }
1092 19528 : }
1093 :
1094 19528 : res
1095 19528 : }
1096 :
1097 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1098 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1099 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1100 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1101 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1102 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1103 : /// the scan operation will not cause OOM in the future.
1104 12 : pub(crate) async fn scan(
1105 12 : &self,
1106 12 : keyspace: KeySpace,
1107 12 : lsn: Lsn,
1108 12 : ctx: &RequestContext,
1109 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1110 12 : if !lsn.is_valid() {
1111 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1112 12 : }
1113 12 :
1114 12 : trace!(
1115 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1116 0 : keyspace,
1117 0 : lsn,
1118 0 : ctx.task_kind()
1119 : );
1120 :
1121 : // We should generalize this into Keyspace::contains in the future.
1122 24 : for range in &keyspace.ranges {
1123 12 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1124 12 : || range.end.field1 > METADATA_KEY_END_PREFIX
1125 : {
1126 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1127 0 : "only metadata keyspace can be scanned"
1128 0 : )));
1129 12 : }
1130 : }
1131 :
1132 12 : let start = crate::metrics::SCAN_LATENCY
1133 12 : .for_task_kind(ctx.task_kind())
1134 12 : .map(ScanLatencyOngoingRecording::start_recording);
1135 :
1136 : // start counting after throttle so that throttle time
1137 : // is always less than observation time
1138 12 : let throttled = self
1139 12 : .timeline_get_throttle
1140 12 : // assume scan = 1 quota for now until we find a better way to process this
1141 12 : .throttle(ctx, 1)
1142 0 : .await;
1143 :
1144 12 : let vectored_res = self
1145 12 : .get_vectored_impl(
1146 12 : keyspace.clone(),
1147 12 : lsn,
1148 12 : &mut ValuesReconstructState::default(),
1149 12 : ctx,
1150 12 : )
1151 0 : .await;
1152 :
1153 12 : if let Some(recording) = start {
1154 0 : recording.observe(throttled);
1155 12 : }
1156 :
1157 12 : vectored_res
1158 12 : }
1159 :
1160 626600 : pub(super) async fn get_vectored_impl(
1161 626600 : &self,
1162 626600 : keyspace: KeySpace,
1163 626600 : lsn: Lsn,
1164 626600 : reconstruct_state: &mut ValuesReconstructState,
1165 626600 : ctx: &RequestContext,
1166 626600 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1167 626600 : let get_kind = if keyspace.total_raw_size() == 1 {
1168 626164 : GetKind::Singular
1169 : } else {
1170 436 : GetKind::Vectored
1171 : };
1172 :
1173 626600 : let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
1174 626600 : .for_get_kind(get_kind)
1175 626600 : .start_timer();
1176 626600 : self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1177 192258 : .await?;
1178 626584 : get_data_timer.stop_and_record();
1179 626584 :
1180 626584 : let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
1181 626584 : .for_get_kind(get_kind)
1182 626584 : .start_timer();
1183 626584 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
1184 626584 : let layers_visited = reconstruct_state.get_layers_visited();
1185 :
1186 666910 : for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
1187 666910 : match res {
1188 0 : Err(err) => {
1189 0 : results.insert(key, Err(err));
1190 0 : }
1191 666910 : Ok(state) => {
1192 666910 : let state = ValueReconstructState::from(state);
1193 :
1194 666910 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
1195 666910 : results.insert(key, reconstruct_res);
1196 : }
1197 : }
1198 : }
1199 626584 : reconstruct_timer.stop_and_record();
1200 626584 :
1201 626584 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1202 626584 : // when they're missing. Instead they are omitted from the resulting btree
1203 626584 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1204 626584 : // to avoid infinite results.
1205 626584 : if !results.is_empty() {
1206 626378 : let avg = layers_visited as f64 / results.len() as f64;
1207 626378 : if avg >= Self::VEC_GET_LAYERS_VISITED_WARN_THRESH {
1208 0 : use utils::rate_limit::RateLimit;
1209 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1210 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1211 0 : let mut rate_limit = LOGGED.lock().unwrap();
1212 0 : rate_limit.call(|| {
1213 0 : tracing::info!(
1214 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1215 0 : lsn = %lsn,
1216 0 : "Vectored read for {} visited {} layers on average per key and {} in total. {}/{} pages were returned",
1217 0 : keyspace, avg, layers_visited, results.len(), keyspace.total_raw_size());
1218 0 : });
1219 626378 : }
1220 :
1221 : // Note that this is an approximation. Tracking the exact number of layers visited
1222 : // per key requires virtually unbounded memory usage and is inefficient
1223 : // (i.e. segment tree tracking each range queried from a layer)
1224 626378 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED.observe(avg);
1225 206 : }
1226 :
1227 626584 : Ok(results)
1228 626600 : }
1229 :
1230 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1231 274290 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1232 274290 : self.last_record_lsn.load().last
1233 274290 : }
1234 :
1235 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1236 0 : self.last_record_lsn.load().prev
1237 0 : }
1238 :
1239 : /// Atomically get both last and prev.
1240 228 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1241 228 : self.last_record_lsn.load()
1242 228 : }
1243 :
1244 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1245 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1246 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1247 0 : self.last_record_lsn.status_receiver()
1248 0 : }
1249 :
1250 414 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1251 414 : self.disk_consistent_lsn.load()
1252 414 : }
1253 :
1254 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1255 : /// not validated with control plane yet.
1256 : /// See [`Self::get_remote_consistent_lsn_visible`].
1257 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1258 0 : self.remote_client.remote_consistent_lsn_projected()
1259 0 : }
1260 :
1261 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1262 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1263 : /// generation validation in the deletion queue.
1264 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1265 0 : self.remote_client.remote_consistent_lsn_visible()
1266 0 : }
1267 :
1268 : /// The sum of the file size of all historic layers in the layer map.
1269 : /// This method makes no distinction between local and remote layers.
1270 : /// Hence, the result **does not represent local filesystem usage**.
1271 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1272 0 : let guard = self.layers.read().await;
1273 0 : guard.layer_size_sum()
1274 0 : }
1275 :
1276 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1277 0 : self.metrics.resident_physical_size_get()
1278 0 : }
1279 :
1280 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1281 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1282 0 : }
1283 :
1284 : ///
1285 : /// Wait until WAL has been received and processed up to this LSN.
1286 : ///
1287 : /// You should call this before any of the other get_* or list_* functions. Calling
1288 : /// those functions with an LSN that has been processed yet is an error.
1289 : ///
1290 225596 : pub(crate) async fn wait_lsn(
1291 225596 : &self,
1292 225596 : lsn: Lsn,
1293 225596 : who_is_waiting: WaitLsnWaiter<'_>,
1294 225596 : ctx: &RequestContext, /* Prepare for use by cancellation */
1295 225596 : ) -> Result<(), WaitLsnError> {
1296 225596 : let state = self.current_state();
1297 225596 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1298 0 : return Err(WaitLsnError::Shutdown);
1299 225596 : } else if !matches!(state, TimelineState::Active) {
1300 0 : return Err(WaitLsnError::BadState(state));
1301 225596 : }
1302 225596 :
1303 225596 : if cfg!(debug_assertions) {
1304 225596 : match ctx.task_kind() {
1305 : TaskKind::WalReceiverManager
1306 : | TaskKind::WalReceiverConnectionHandler
1307 : | TaskKind::WalReceiverConnectionPoller => {
1308 0 : let is_myself = match who_is_waiting {
1309 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1310 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1311 : };
1312 0 : if is_myself {
1313 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1314 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1315 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1316 0 : }
1317 0 : } else {
1318 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1319 0 : // our walreceiver task can make progress independent of theirs
1320 0 : }
1321 : }
1322 225596 : _ => {}
1323 : }
1324 0 : }
1325 :
1326 225596 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1327 225596 :
1328 225596 : match self
1329 225596 : .last_record_lsn
1330 225596 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1331 0 : .await
1332 : {
1333 225596 : Ok(()) => Ok(()),
1334 0 : Err(e) => {
1335 : use utils::seqwait::SeqWaitError::*;
1336 0 : match e {
1337 0 : Shutdown => Err(WaitLsnError::Shutdown),
1338 : Timeout => {
1339 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1340 0 : drop(_timer);
1341 0 : let walreceiver_status = self.walreceiver_status();
1342 0 : Err(WaitLsnError::Timeout(format!(
1343 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1344 0 : lsn,
1345 0 : self.get_last_record_lsn(),
1346 0 : self.get_disk_consistent_lsn(),
1347 0 : walreceiver_status,
1348 0 : )))
1349 : }
1350 : }
1351 : }
1352 : }
1353 225596 : }
1354 :
1355 0 : pub(crate) fn walreceiver_status(&self) -> String {
1356 0 : match &*self.walreceiver.lock().unwrap() {
1357 0 : None => "stopping or stopped".to_string(),
1358 0 : Some(walreceiver) => match walreceiver.status() {
1359 0 : Some(status) => status.to_human_readable_string(),
1360 0 : None => "Not active".to_string(),
1361 : },
1362 : }
1363 0 : }
1364 :
1365 : /// Check that it is valid to request operations with that lsn.
1366 232 : pub(crate) fn check_lsn_is_in_scope(
1367 232 : &self,
1368 232 : lsn: Lsn,
1369 232 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1370 232 : ) -> anyhow::Result<()> {
1371 232 : ensure!(
1372 232 : lsn >= **latest_gc_cutoff_lsn,
1373 4 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1374 4 : lsn,
1375 4 : **latest_gc_cutoff_lsn,
1376 : );
1377 228 : Ok(())
1378 232 : }
1379 :
1380 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1381 10 : pub(crate) fn init_lsn_lease(
1382 10 : &self,
1383 10 : lsn: Lsn,
1384 10 : length: Duration,
1385 10 : ctx: &RequestContext,
1386 10 : ) -> anyhow::Result<LsnLease> {
1387 10 : self.make_lsn_lease(lsn, length, true, ctx)
1388 10 : }
1389 :
1390 : /// 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.
1391 4 : pub(crate) fn renew_lsn_lease(
1392 4 : &self,
1393 4 : lsn: Lsn,
1394 4 : length: Duration,
1395 4 : ctx: &RequestContext,
1396 4 : ) -> anyhow::Result<LsnLease> {
1397 4 : self.make_lsn_lease(lsn, length, false, ctx)
1398 4 : }
1399 :
1400 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1401 : ///
1402 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1403 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1404 : ///
1405 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1406 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1407 14 : fn make_lsn_lease(
1408 14 : &self,
1409 14 : lsn: Lsn,
1410 14 : length: Duration,
1411 14 : init: bool,
1412 14 : _ctx: &RequestContext,
1413 14 : ) -> anyhow::Result<LsnLease> {
1414 12 : let lease = {
1415 : // Normalize the requested LSN to be aligned, and move to the first record
1416 : // if it points to the beginning of the page (header).
1417 14 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1418 14 :
1419 14 : let mut gc_info = self.gc_info.write().unwrap();
1420 14 :
1421 14 : let valid_until = SystemTime::now() + length;
1422 14 :
1423 14 : let entry = gc_info.leases.entry(lsn);
1424 14 :
1425 14 : match entry {
1426 6 : Entry::Occupied(mut occupied) => {
1427 6 : let existing_lease = occupied.get_mut();
1428 6 : if valid_until > existing_lease.valid_until {
1429 2 : existing_lease.valid_until = valid_until;
1430 2 : let dt: DateTime<Utc> = valid_until.into();
1431 2 : info!("lease extended to {}", dt);
1432 : } else {
1433 4 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1434 4 : info!("existing lease covers greater length, valid until {}", dt);
1435 : }
1436 :
1437 6 : existing_lease.clone()
1438 : }
1439 8 : Entry::Vacant(vacant) => {
1440 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
1441 : // not blocked by the lsn lease deadline.
1442 8 : let validate = {
1443 8 : let conf = self.tenant_conf.load();
1444 8 : conf.location.attach_mode == AttachmentMode::Single
1445 8 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1446 : };
1447 :
1448 8 : if init || validate {
1449 8 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1450 8 : if lsn < *latest_gc_cutoff_lsn {
1451 2 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1452 6 : }
1453 0 : }
1454 :
1455 6 : let dt: DateTime<Utc> = valid_until.into();
1456 6 : info!("lease created, valid until {}", dt);
1457 6 : vacant.insert(LsnLease { valid_until }).clone()
1458 : }
1459 : }
1460 : };
1461 :
1462 12 : Ok(lease)
1463 14 : }
1464 :
1465 : /// Flush to disk all data that was written with the put_* functions
1466 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))]
1467 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1468 : self.freeze_and_flush0().await
1469 : }
1470 :
1471 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1472 : // polluting the span hierarchy.
1473 1092 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1474 1092 : let token = {
1475 : // Freeze the current open in-memory layer. It will be written to disk on next
1476 : // iteration.
1477 1092 : let mut g = self.write_lock.lock().await;
1478 :
1479 1092 : let to_lsn = self.get_last_record_lsn();
1480 1092 : self.freeze_inmem_layer_at(to_lsn, &mut g).await?
1481 : };
1482 1092 : self.wait_flush_completion(token).await
1483 1092 : }
1484 :
1485 : // Check if an open ephemeral layer should be closed: this provides
1486 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1487 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1488 : // ephemeral layer bytes has been breached.
1489 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1490 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1491 : // If the write lock is held, there is an active wal receiver: rolling open layers
1492 : // is their responsibility while they hold this lock.
1493 0 : return;
1494 : };
1495 :
1496 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1497 : // time, and this was missed.
1498 : // if write_guard.is_none() { return; }
1499 :
1500 0 : let Ok(layers_guard) = self.layers.try_read() else {
1501 : // Don't block if the layer lock is busy
1502 0 : return;
1503 : };
1504 :
1505 0 : let Ok(lm) = layers_guard.layer_map() else {
1506 0 : return;
1507 : };
1508 :
1509 0 : let Some(open_layer) = &lm.open_layer else {
1510 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1511 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1512 : // that didn't result in writes to this shard.
1513 :
1514 : // Must not hold the layers lock while waiting for a flush.
1515 0 : drop(layers_guard);
1516 0 :
1517 0 : let last_record_lsn = self.get_last_record_lsn();
1518 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1519 0 : if last_record_lsn > disk_consistent_lsn {
1520 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1521 : // we are a sharded tenant and have skipped some WAL
1522 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1523 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1524 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1525 : // without any data ingested (yet) doesn't write a remote index as soon as it
1526 : // sees its LSN advance: we only do this if we've been layer-less
1527 : // for some time.
1528 0 : tracing::debug!(
1529 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1530 : disk_consistent_lsn,
1531 : last_record_lsn
1532 : );
1533 :
1534 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1535 : // We know there is no open layer, so we can request freezing without actually
1536 : // freezing anything. This is true even if we have dropped the layers_guard, we
1537 : // still hold the write_guard.
1538 0 : let _ = async {
1539 0 : let token = self
1540 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1541 0 : .await?;
1542 0 : self.wait_flush_completion(token).await
1543 0 : }
1544 0 : .await;
1545 0 : }
1546 0 : }
1547 :
1548 0 : return;
1549 : };
1550 :
1551 0 : let Some(current_size) = open_layer.try_len() else {
1552 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1553 : // read lock to get size should always succeed.
1554 0 : tracing::warn!("Lock conflict while reading size of open layer");
1555 0 : return;
1556 : };
1557 :
1558 0 : let current_lsn = self.get_last_record_lsn();
1559 :
1560 0 : let checkpoint_distance_override = open_layer.tick().await;
1561 :
1562 0 : if let Some(size_override) = checkpoint_distance_override {
1563 0 : if current_size > size_override {
1564 : // This is not harmful, but it only happens in relatively rare cases where
1565 : // time-based checkpoints are not happening fast enough to keep the amount of
1566 : // ephemeral data within configured limits. It's a sign of stress on the system.
1567 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1568 0 : }
1569 0 : }
1570 :
1571 0 : let checkpoint_distance =
1572 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1573 0 :
1574 0 : if self.should_roll(
1575 0 : current_size,
1576 0 : current_size,
1577 0 : checkpoint_distance,
1578 0 : self.get_last_record_lsn(),
1579 0 : self.last_freeze_at.load(),
1580 0 : open_layer.get_opened_at(),
1581 0 : ) {
1582 0 : match open_layer.info() {
1583 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1584 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1585 0 : // happens asynchronously in the background.
1586 0 : tracing::debug!(
1587 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1588 : );
1589 : }
1590 : InMemoryLayerInfo::Open { .. } => {
1591 : // Upgrade to a write lock and freeze the layer
1592 0 : drop(layers_guard);
1593 0 : let res = self
1594 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1595 0 : .await;
1596 :
1597 0 : if let Err(e) = res {
1598 0 : tracing::info!(
1599 0 : "failed to flush frozen layer after background freeze: {e:#}"
1600 : );
1601 0 : }
1602 : }
1603 : }
1604 0 : }
1605 0 : }
1606 :
1607 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1608 : ///
1609 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1610 : /// child timelines that are not offloaded yet.
1611 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1612 0 : if self.remote_client.is_archived() != Some(true) {
1613 0 : return (false, "the timeline is not archived");
1614 0 : }
1615 0 : if !self.remote_client.no_pending_work() {
1616 : // if the remote client is still processing some work, we can't offload
1617 0 : return (false, "the upload queue is not drained yet");
1618 0 : }
1619 0 :
1620 0 : (true, "ok")
1621 0 : }
1622 :
1623 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1624 : /// compaction tasks.
1625 364 : pub(crate) async fn compact(
1626 364 : self: &Arc<Self>,
1627 364 : cancel: &CancellationToken,
1628 364 : flags: EnumSet<CompactFlags>,
1629 364 : ctx: &RequestContext,
1630 364 : ) -> Result<bool, CompactionError> {
1631 364 : self.compact_with_options(
1632 364 : cancel,
1633 364 : CompactOptions {
1634 364 : flags,
1635 364 : compact_range: None,
1636 364 : },
1637 364 : ctx,
1638 364 : )
1639 37158 : .await
1640 364 : }
1641 :
1642 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1643 : /// compaction tasks.
1644 364 : pub(crate) async fn compact_with_options(
1645 364 : self: &Arc<Self>,
1646 364 : cancel: &CancellationToken,
1647 364 : options: CompactOptions,
1648 364 : ctx: &RequestContext,
1649 364 : ) -> Result<bool, CompactionError> {
1650 364 : // most likely the cancellation token is from background task, but in tests it could be the
1651 364 : // request task as well.
1652 364 :
1653 364 : let prepare = async move {
1654 364 : let guard = self.compaction_lock.lock().await;
1655 :
1656 364 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1657 364 : BackgroundLoopKind::Compaction,
1658 364 : ctx,
1659 364 : )
1660 0 : .await;
1661 :
1662 364 : (guard, permit)
1663 364 : };
1664 :
1665 : // this wait probably never needs any "long time spent" logging, because we already nag if
1666 : // compaction task goes over it's period (20s) which is quite often in production.
1667 364 : let (_guard, _permit) = tokio::select! {
1668 364 : tuple = prepare => { tuple },
1669 364 : _ = self.cancel.cancelled() => return Ok(false),
1670 364 : _ = cancel.cancelled() => return Ok(false),
1671 : };
1672 :
1673 364 : let last_record_lsn = self.get_last_record_lsn();
1674 364 :
1675 364 : // Last record Lsn could be zero in case the timeline was just created
1676 364 : if !last_record_lsn.is_valid() {
1677 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1678 0 : return Ok(false);
1679 364 : }
1680 364 :
1681 364 : match self.get_compaction_algorithm_settings().kind {
1682 : CompactionAlgorithm::Tiered => {
1683 0 : self.compact_tiered(cancel, ctx).await?;
1684 0 : Ok(false)
1685 : }
1686 37158 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1687 : }
1688 364 : }
1689 :
1690 : /// Mutate the timeline with a [`TimelineWriter`].
1691 5133162 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1692 5133162 : TimelineWriter {
1693 5133162 : tl: self,
1694 5133162 : write_guard: self.write_lock.lock().await,
1695 : }
1696 5133162 : }
1697 :
1698 0 : pub(crate) fn activate(
1699 0 : self: &Arc<Self>,
1700 0 : parent: Arc<crate::tenant::Tenant>,
1701 0 : broker_client: BrokerClientChannel,
1702 0 : background_jobs_can_start: Option<&completion::Barrier>,
1703 0 : ctx: &RequestContext,
1704 0 : ) {
1705 0 : if self.tenant_shard_id.is_shard_zero() {
1706 0 : // Logical size is only maintained accurately on shard zero.
1707 0 : self.spawn_initial_logical_size_computation_task(ctx);
1708 0 : }
1709 0 : self.launch_wal_receiver(ctx, broker_client);
1710 0 : self.set_state(TimelineState::Active);
1711 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1712 0 : }
1713 :
1714 : /// After this function returns, there are no timeline-scoped tasks are left running.
1715 : ///
1716 : /// The preferred pattern for is:
1717 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1718 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1719 : /// go the extra mile and keep track of JoinHandles
1720 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1721 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1722 : ///
1723 : /// For legacy reasons, we still have multiple tasks spawned using
1724 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1725 : /// We refer to these as "timeline-scoped task_mgr tasks".
1726 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1727 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1728 : /// or [`task_mgr::shutdown_watcher`].
1729 : /// We want to gradually convert the code base away from these.
1730 : ///
1731 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1732 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1733 : /// ones that aren't mentioned here):
1734 : /// - [`TaskKind::TimelineDeletionWorker`]
1735 : /// - NB: also used for tenant deletion
1736 : /// - [`TaskKind::RemoteUploadTask`]`
1737 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1738 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1739 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1740 : /// - [`TaskKind::Eviction`]
1741 : /// - [`TaskKind::LayerFlushTask`]
1742 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1743 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1744 10 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1745 10 : debug_assert_current_span_has_tenant_and_timeline_id();
1746 10 :
1747 10 : // Regardless of whether we're going to try_freeze_and_flush
1748 10 : // or not, stop ingesting any more data. Walreceiver only provides
1749 10 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1750 10 : // So, only after the self.gate.close() below will we know for sure that
1751 10 : // no walreceiver tasks are left.
1752 10 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1753 10 : // data during the call to `self.freeze_and_flush()` below.
1754 10 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1755 10 : // which is what we'd need to properly model early shutdown of the walreceiver
1756 10 : // task sub-tree before the other Timeline task sub-trees.
1757 10 : let walreceiver = self.walreceiver.lock().unwrap().take();
1758 10 : tracing::debug!(
1759 0 : is_some = walreceiver.is_some(),
1760 0 : "Waiting for WalReceiverManager..."
1761 : );
1762 10 : if let Some(walreceiver) = walreceiver {
1763 0 : walreceiver.cancel();
1764 10 : }
1765 : // ... and inform any waiters for newer LSNs that there won't be any.
1766 10 : self.last_record_lsn.shutdown();
1767 10 :
1768 10 : if let ShutdownMode::FreezeAndFlush = mode {
1769 6 : if let Some((open, frozen)) = self
1770 6 : .layers
1771 6 : .read()
1772 0 : .await
1773 6 : .layer_map()
1774 6 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1775 6 : .ok()
1776 6 : .filter(|(open, frozen)| *open || *frozen > 0)
1777 : {
1778 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1779 6 : } else {
1780 6 : // this is double-shutdown, ignore it
1781 6 : }
1782 :
1783 : // we shut down walreceiver above, so, we won't add anything more
1784 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1785 : // to reach the disk & upload queue, then shut the upload queue and
1786 : // wait for it to drain.
1787 6 : match self.freeze_and_flush().await {
1788 : Ok(_) => {
1789 : // drain the upload queue
1790 : // if we did not wait for completion here, it might be our shutdown process
1791 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1792 : // be spawned.
1793 : //
1794 : // what is problematic is the shutting down of RemoteTimelineClient, because
1795 : // obviously it does not make sense to stop while we wait for it, but what
1796 : // about corner cases like s3 suddenly hanging up?
1797 6 : self.remote_client.shutdown().await;
1798 : }
1799 : Err(FlushLayerError::Cancelled) => {
1800 : // this is likely the second shutdown, ignore silently.
1801 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1802 0 : debug_assert!(self.cancel.is_cancelled());
1803 : }
1804 0 : Err(e) => {
1805 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1806 0 : // we have some extra WAL replay to do next time the timeline starts.
1807 0 : warn!("failed to freeze and flush: {e:#}");
1808 : }
1809 : }
1810 :
1811 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
1812 : // we also do a final check here to ensure that the queue is empty.
1813 6 : if !self.remote_client.no_pending_work() {
1814 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1815 6 : }
1816 4 : }
1817 :
1818 10 : if let ShutdownMode::Flush = mode {
1819 : // drain the upload queue
1820 2 : self.remote_client.shutdown().await;
1821 2 : if !self.remote_client.no_pending_work() {
1822 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1823 2 : }
1824 8 : }
1825 :
1826 : // Signal any subscribers to our cancellation token to drop out
1827 10 : tracing::debug!("Cancelling CancellationToken");
1828 10 : self.cancel.cancel();
1829 10 :
1830 10 : // Ensure Prevent new page service requests from starting.
1831 10 : self.handles.shutdown();
1832 10 :
1833 10 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1834 10 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1835 10 : self.remote_client.stop();
1836 10 :
1837 10 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1838 10 : // to shut down the upload queue tasks.
1839 10 : // TODO: fix that, task management should be encapsulated inside remote_client.
1840 10 : task_mgr::shutdown_tasks(
1841 10 : Some(TaskKind::RemoteUploadTask),
1842 10 : Some(self.tenant_shard_id),
1843 10 : Some(self.timeline_id),
1844 10 : )
1845 0 : .await;
1846 :
1847 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
1848 10 : tracing::debug!("Waiting for tasks...");
1849 10 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1850 :
1851 : {
1852 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
1853 : // open.
1854 10 : let mut write_guard = self.write_lock.lock().await;
1855 10 : self.layers.write().await.shutdown(&mut write_guard);
1856 10 : }
1857 10 :
1858 10 : // Finally wait until any gate-holders are complete.
1859 10 : //
1860 10 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1861 10 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1862 10 : self.gate.close().await;
1863 :
1864 10 : self.metrics.shutdown();
1865 10 : }
1866 :
1867 420 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1868 420 : match (self.current_state(), new_state) {
1869 420 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1870 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1871 : }
1872 0 : (st, TimelineState::Loading) => {
1873 0 : error!("ignoring transition from {st:?} into Loading state");
1874 : }
1875 0 : (TimelineState::Broken { .. }, new_state) => {
1876 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1877 : }
1878 : (TimelineState::Stopping, TimelineState::Active) => {
1879 0 : error!("Not activating a Stopping timeline");
1880 : }
1881 418 : (_, new_state) => {
1882 418 : self.state.send_replace(new_state);
1883 418 : }
1884 : }
1885 420 : }
1886 :
1887 2 : pub(crate) fn set_broken(&self, reason: String) {
1888 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1889 2 : let broken_state = TimelineState::Broken {
1890 2 : reason,
1891 2 : backtrace: backtrace_str,
1892 2 : };
1893 2 : self.set_state(broken_state);
1894 2 :
1895 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1896 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1897 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1898 2 : self.cancel.cancel();
1899 2 : }
1900 :
1901 226606 : pub(crate) fn current_state(&self) -> TimelineState {
1902 226606 : self.state.borrow().clone()
1903 226606 : }
1904 :
1905 6 : pub(crate) fn is_broken(&self) -> bool {
1906 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1907 6 : }
1908 :
1909 222 : pub(crate) fn is_active(&self) -> bool {
1910 222 : self.current_state() == TimelineState::Active
1911 222 : }
1912 :
1913 2 : pub(crate) fn is_archived(&self) -> Option<bool> {
1914 2 : self.remote_client.is_archived()
1915 2 : }
1916 :
1917 368 : pub(crate) fn is_stopping(&self) -> bool {
1918 368 : self.current_state() == TimelineState::Stopping
1919 368 : }
1920 :
1921 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1922 0 : self.state.subscribe()
1923 0 : }
1924 :
1925 225598 : pub(crate) async fn wait_to_become_active(
1926 225598 : &self,
1927 225598 : _ctx: &RequestContext, // Prepare for use by cancellation
1928 225598 : ) -> Result<(), TimelineState> {
1929 225598 : let mut receiver = self.state.subscribe();
1930 : loop {
1931 225598 : let current_state = receiver.borrow().clone();
1932 225598 : match current_state {
1933 : TimelineState::Loading => {
1934 0 : receiver
1935 0 : .changed()
1936 0 : .await
1937 0 : .expect("holding a reference to self");
1938 : }
1939 : TimelineState::Active { .. } => {
1940 225596 : return Ok(());
1941 : }
1942 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1943 : // There's no chance the timeline can transition back into ::Active
1944 2 : return Err(current_state);
1945 : }
1946 : }
1947 : }
1948 225598 : }
1949 :
1950 0 : pub(crate) async fn layer_map_info(
1951 0 : &self,
1952 0 : reset: LayerAccessStatsReset,
1953 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
1954 0 : let guard = self.layers.read().await;
1955 0 : let layer_map = guard.layer_map()?;
1956 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
1957 0 : if let Some(open_layer) = &layer_map.open_layer {
1958 0 : in_memory_layers.push(open_layer.info());
1959 0 : }
1960 0 : for frozen_layer in &layer_map.frozen_layers {
1961 0 : in_memory_layers.push(frozen_layer.info());
1962 0 : }
1963 :
1964 0 : let historic_layers = layer_map
1965 0 : .iter_historic_layers()
1966 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
1967 0 : .collect();
1968 0 :
1969 0 : Ok(LayerMapInfo {
1970 0 : in_memory_layers,
1971 0 : historic_layers,
1972 0 : })
1973 0 : }
1974 :
1975 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))]
1976 : pub(crate) async fn download_layer(
1977 : &self,
1978 : layer_file_name: &LayerName,
1979 : ) -> anyhow::Result<Option<bool>> {
1980 : let Some(layer) = self.find_layer(layer_file_name).await? else {
1981 : return Ok(None);
1982 : };
1983 :
1984 : layer.download().await?;
1985 :
1986 : Ok(Some(true))
1987 : }
1988 :
1989 : /// Evict just one layer.
1990 : ///
1991 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
1992 0 : pub(crate) async fn evict_layer(
1993 0 : &self,
1994 0 : layer_file_name: &LayerName,
1995 0 : ) -> anyhow::Result<Option<bool>> {
1996 0 : let _gate = self
1997 0 : .gate
1998 0 : .enter()
1999 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2000 :
2001 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2002 0 : return Ok(None);
2003 : };
2004 :
2005 : // curl has this by default
2006 0 : let timeout = std::time::Duration::from_secs(120);
2007 0 :
2008 0 : match local_layer.evict_and_wait(timeout).await {
2009 0 : Ok(()) => Ok(Some(true)),
2010 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2011 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2012 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2013 : }
2014 0 : }
2015 :
2016 4803010 : fn should_roll(
2017 4803010 : &self,
2018 4803010 : layer_size: u64,
2019 4803010 : projected_layer_size: u64,
2020 4803010 : checkpoint_distance: u64,
2021 4803010 : projected_lsn: Lsn,
2022 4803010 : last_freeze_at: Lsn,
2023 4803010 : opened_at: Instant,
2024 4803010 : ) -> bool {
2025 4803010 : let distance = projected_lsn.widening_sub(last_freeze_at);
2026 4803010 :
2027 4803010 : // Rolling the open layer can be triggered by:
2028 4803010 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2029 4803010 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2030 4803010 : // account for how writes are distributed across shards: we expect each node to consume
2031 4803010 : // 1/count of the LSN on average.
2032 4803010 : // 2. The size of the currently open layer.
2033 4803010 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2034 4803010 : // up and suspend activity.
2035 4803010 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2036 0 : info!(
2037 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2038 : projected_lsn, layer_size, distance
2039 : );
2040 :
2041 0 : true
2042 4803010 : } else if projected_layer_size >= checkpoint_distance {
2043 : // NB: this check is relied upon by:
2044 80 : let _ = IndexEntry::validate_checkpoint_distance;
2045 80 : info!(
2046 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2047 : projected_lsn, layer_size, projected_layer_size
2048 : );
2049 :
2050 80 : true
2051 4802930 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2052 0 : info!(
2053 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2054 0 : projected_lsn,
2055 0 : layer_size,
2056 0 : opened_at.elapsed()
2057 : );
2058 :
2059 0 : true
2060 : } else {
2061 4802930 : false
2062 : }
2063 4803010 : }
2064 : }
2065 :
2066 : /// Number of times we will compute partition within a checkpoint distance.
2067 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2068 :
2069 : // Private functions
2070 : impl Timeline {
2071 12 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2072 12 : let tenant_conf = self.tenant_conf.load();
2073 12 : tenant_conf
2074 12 : .tenant_conf
2075 12 : .lsn_lease_length
2076 12 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2077 12 : }
2078 :
2079 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2080 0 : let tenant_conf = self.tenant_conf.load();
2081 0 : tenant_conf
2082 0 : .tenant_conf
2083 0 : .lsn_lease_length_for_ts
2084 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2085 0 : }
2086 :
2087 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2088 0 : let tenant_conf = self.tenant_conf.load();
2089 0 : tenant_conf
2090 0 : .tenant_conf
2091 0 : .lazy_slru_download
2092 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2093 0 : }
2094 :
2095 4804616 : fn get_checkpoint_distance(&self) -> u64 {
2096 4804616 : let tenant_conf = self.tenant_conf.load();
2097 4804616 : tenant_conf
2098 4804616 : .tenant_conf
2099 4804616 : .checkpoint_distance
2100 4804616 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2101 4804616 : }
2102 :
2103 4802930 : fn get_checkpoint_timeout(&self) -> Duration {
2104 4802930 : let tenant_conf = self.tenant_conf.load();
2105 4802930 : tenant_conf
2106 4802930 : .tenant_conf
2107 4802930 : .checkpoint_timeout
2108 4802930 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2109 4802930 : }
2110 :
2111 618 : fn get_compaction_target_size(&self) -> u64 {
2112 618 : let tenant_conf = self.tenant_conf.load();
2113 618 : tenant_conf
2114 618 : .tenant_conf
2115 618 : .compaction_target_size
2116 618 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2117 618 : }
2118 :
2119 1536 : fn get_compaction_threshold(&self) -> usize {
2120 1536 : let tenant_conf = self.tenant_conf.load();
2121 1536 : tenant_conf
2122 1536 : .tenant_conf
2123 1536 : .compaction_threshold
2124 1536 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2125 1536 : }
2126 :
2127 14 : fn get_image_creation_threshold(&self) -> usize {
2128 14 : let tenant_conf = self.tenant_conf.load();
2129 14 : tenant_conf
2130 14 : .tenant_conf
2131 14 : .image_creation_threshold
2132 14 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2133 14 : }
2134 :
2135 364 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2136 364 : let tenant_conf = &self.tenant_conf.load();
2137 364 : tenant_conf
2138 364 : .tenant_conf
2139 364 : .compaction_algorithm
2140 364 : .as_ref()
2141 364 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2142 364 : .clone()
2143 364 : }
2144 :
2145 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2146 0 : let tenant_conf = self.tenant_conf.load();
2147 0 : tenant_conf
2148 0 : .tenant_conf
2149 0 : .eviction_policy
2150 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2151 0 : }
2152 :
2153 418 : fn get_evictions_low_residence_duration_metric_threshold(
2154 418 : tenant_conf: &TenantConfOpt,
2155 418 : default_tenant_conf: &TenantConf,
2156 418 : ) -> Duration {
2157 418 : tenant_conf
2158 418 : .evictions_low_residence_duration_metric_threshold
2159 418 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2160 418 : }
2161 :
2162 716 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2163 716 : let tenant_conf = self.tenant_conf.load();
2164 716 : tenant_conf
2165 716 : .tenant_conf
2166 716 : .image_layer_creation_check_threshold
2167 716 : .unwrap_or(
2168 716 : self.conf
2169 716 : .default_tenant_conf
2170 716 : .image_layer_creation_check_threshold,
2171 716 : )
2172 716 : }
2173 :
2174 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2175 0 : // NB: Most tenant conf options are read by background loops, so,
2176 0 : // changes will automatically be picked up.
2177 0 :
2178 0 : // The threshold is embedded in the metric. So, we need to update it.
2179 0 : {
2180 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2181 0 : new_conf,
2182 0 : &self.conf.default_tenant_conf,
2183 0 : );
2184 0 :
2185 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2186 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2187 0 :
2188 0 : let timeline_id_str = self.timeline_id.to_string();
2189 0 : self.metrics
2190 0 : .evictions_with_low_residence_duration
2191 0 : .write()
2192 0 : .unwrap()
2193 0 : .change_threshold(
2194 0 : &tenant_id_str,
2195 0 : &shard_id_str,
2196 0 : &timeline_id_str,
2197 0 : new_threshold,
2198 0 : );
2199 0 : }
2200 0 : }
2201 :
2202 : /// Open a Timeline handle.
2203 : ///
2204 : /// Loads the metadata for the timeline into memory, but not the layer map.
2205 : #[allow(clippy::too_many_arguments)]
2206 418 : pub(super) fn new(
2207 418 : conf: &'static PageServerConf,
2208 418 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2209 418 : metadata: &TimelineMetadata,
2210 418 : ancestor: Option<Arc<Timeline>>,
2211 418 : timeline_id: TimelineId,
2212 418 : tenant_shard_id: TenantShardId,
2213 418 : generation: Generation,
2214 418 : shard_identity: ShardIdentity,
2215 418 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2216 418 : resources: TimelineResources,
2217 418 : pg_version: u32,
2218 418 : state: TimelineState,
2219 418 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2220 418 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2221 418 : cancel: CancellationToken,
2222 418 : ) -> Arc<Self> {
2223 418 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2224 418 : let (state, _) = watch::channel(state);
2225 418 :
2226 418 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2227 418 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2228 418 :
2229 418 : let evictions_low_residence_duration_metric_threshold = {
2230 418 : let loaded_tenant_conf = tenant_conf.load();
2231 418 : Self::get_evictions_low_residence_duration_metric_threshold(
2232 418 : &loaded_tenant_conf.tenant_conf,
2233 418 : &conf.default_tenant_conf,
2234 418 : )
2235 : };
2236 :
2237 418 : if let Some(ancestor) = &ancestor {
2238 230 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2239 230 : // If we construct an explicit timeline object, it's obviously not offloaded
2240 230 : let is_offloaded = MaybeOffloaded::No;
2241 230 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2242 230 : }
2243 :
2244 418 : Arc::new_cyclic(|myself| {
2245 418 : let metrics = TimelineMetrics::new(
2246 418 : &tenant_shard_id,
2247 418 : &timeline_id,
2248 418 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2249 418 : "mtime",
2250 418 : evictions_low_residence_duration_metric_threshold,
2251 418 : ),
2252 418 : );
2253 418 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2254 :
2255 418 : let mut result = Timeline {
2256 418 : conf,
2257 418 : tenant_conf,
2258 418 : myself: myself.clone(),
2259 418 : timeline_id,
2260 418 : tenant_shard_id,
2261 418 : generation,
2262 418 : shard_identity,
2263 418 : pg_version,
2264 418 : layers: Default::default(),
2265 418 :
2266 418 : walredo_mgr,
2267 418 : walreceiver: Mutex::new(None),
2268 418 :
2269 418 : remote_client: Arc::new(resources.remote_client),
2270 418 :
2271 418 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2272 418 : last_record_lsn: SeqWait::new(RecordLsn {
2273 418 : last: disk_consistent_lsn,
2274 418 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2275 418 : }),
2276 418 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2277 418 :
2278 418 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2279 418 : last_freeze_ts: RwLock::new(Instant::now()),
2280 418 :
2281 418 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2282 418 :
2283 418 : ancestor_timeline: ancestor,
2284 418 : ancestor_lsn: metadata.ancestor_lsn(),
2285 418 :
2286 418 : metrics,
2287 418 :
2288 418 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2289 418 : &tenant_shard_id,
2290 418 : &timeline_id,
2291 418 : ),
2292 418 :
2293 2926 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2294 418 :
2295 418 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2296 418 :
2297 418 : layer_flush_start_tx,
2298 418 : layer_flush_done_tx,
2299 418 :
2300 418 : write_lock: tokio::sync::Mutex::new(None),
2301 418 :
2302 418 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2303 418 :
2304 418 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2305 418 : initdb_lsn: metadata.initdb_lsn(),
2306 418 :
2307 418 : current_logical_size: if disk_consistent_lsn.is_valid() {
2308 : // we're creating timeline data with some layer files existing locally,
2309 : // need to recalculate timeline's logical size based on data in the layers.
2310 234 : LogicalSize::deferred_initial(disk_consistent_lsn)
2311 : } else {
2312 : // we're creating timeline data without any layers existing locally,
2313 : // initial logical size is 0.
2314 184 : LogicalSize::empty_initial()
2315 : },
2316 418 : partitioning: tokio::sync::Mutex::new((
2317 418 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2318 418 : Lsn(0),
2319 418 : )),
2320 418 : repartition_threshold: 0,
2321 418 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2322 418 : last_image_layer_creation_check_instant: Mutex::new(None),
2323 418 :
2324 418 : last_received_wal: Mutex::new(None),
2325 418 : rel_size_cache: RwLock::new(RelSizeCache {
2326 418 : complete_as_of: disk_consistent_lsn,
2327 418 : map: HashMap::new(),
2328 418 : }),
2329 418 :
2330 418 : download_all_remote_layers_task_info: RwLock::new(None),
2331 418 :
2332 418 : state,
2333 418 :
2334 418 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2335 418 : EvictionTaskTimelineState::default(),
2336 418 : ),
2337 418 : delete_progress: TimelineDeleteProgress::default(),
2338 418 :
2339 418 : cancel,
2340 418 : gate: Gate::default(),
2341 418 :
2342 418 : compaction_lock: tokio::sync::Mutex::default(),
2343 418 : gc_lock: tokio::sync::Mutex::default(),
2344 418 :
2345 418 : standby_horizon: AtomicLsn::new(0),
2346 418 :
2347 418 : timeline_get_throttle: resources.timeline_get_throttle,
2348 418 :
2349 418 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2350 418 :
2351 418 : #[cfg(test)]
2352 418 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2353 418 :
2354 418 : l0_flush_global_state: resources.l0_flush_global_state,
2355 418 :
2356 418 : handles: Default::default(),
2357 418 :
2358 418 : attach_wal_lag_cooldown,
2359 418 :
2360 418 : create_idempotency,
2361 418 : };
2362 418 :
2363 418 : result.repartition_threshold =
2364 418 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2365 418 :
2366 418 : result
2367 418 : .metrics
2368 418 : .last_record_gauge
2369 418 : .set(disk_consistent_lsn.0 as i64);
2370 418 : result
2371 418 : })
2372 418 : }
2373 :
2374 588 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2375 588 : let Ok(guard) = self.gate.enter() else {
2376 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2377 0 : return;
2378 : };
2379 588 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2380 588 : match *flush_loop_state {
2381 412 : FlushLoopState::NotStarted => (),
2382 : FlushLoopState::Running { .. } => {
2383 176 : info!(
2384 0 : "skipping attempt to start flush_loop twice {}/{}",
2385 0 : self.tenant_shard_id, self.timeline_id
2386 : );
2387 176 : return;
2388 : }
2389 : FlushLoopState::Exited => {
2390 0 : warn!(
2391 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2392 0 : self.tenant_shard_id, self.timeline_id
2393 : );
2394 0 : return;
2395 : }
2396 : }
2397 :
2398 412 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2399 412 : let self_clone = Arc::clone(self);
2400 412 :
2401 412 : debug!("spawning flush loop");
2402 412 : *flush_loop_state = FlushLoopState::Running {
2403 412 : #[cfg(test)]
2404 412 : expect_initdb_optimization: false,
2405 412 : #[cfg(test)]
2406 412 : initdb_optimization_count: 0,
2407 412 : };
2408 412 : task_mgr::spawn(
2409 412 : task_mgr::BACKGROUND_RUNTIME.handle(),
2410 412 : task_mgr::TaskKind::LayerFlushTask,
2411 412 : self.tenant_shard_id,
2412 412 : Some(self.timeline_id),
2413 412 : "layer flush task",
2414 412 : async move {
2415 412 : let _guard = guard;
2416 412 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2417 18032 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2418 10 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2419 10 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2420 10 : *flush_loop_state = FlushLoopState::Exited;
2421 10 : Ok(())
2422 10 : }
2423 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))
2424 : );
2425 588 : }
2426 :
2427 : /// Creates and starts the wal receiver.
2428 : ///
2429 : /// This function is expected to be called at most once per Timeline's lifecycle
2430 : /// when the timeline is activated.
2431 0 : fn launch_wal_receiver(
2432 0 : self: &Arc<Self>,
2433 0 : ctx: &RequestContext,
2434 0 : broker_client: BrokerClientChannel,
2435 0 : ) {
2436 0 : info!(
2437 0 : "launching WAL receiver for timeline {} of tenant {}",
2438 0 : self.timeline_id, self.tenant_shard_id
2439 : );
2440 :
2441 0 : let tenant_conf = self.tenant_conf.load();
2442 0 : let wal_connect_timeout = tenant_conf
2443 0 : .tenant_conf
2444 0 : .walreceiver_connect_timeout
2445 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2446 0 : let lagging_wal_timeout = tenant_conf
2447 0 : .tenant_conf
2448 0 : .lagging_wal_timeout
2449 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2450 0 : let max_lsn_wal_lag = tenant_conf
2451 0 : .tenant_conf
2452 0 : .max_lsn_wal_lag
2453 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2454 0 :
2455 0 : let mut guard = self.walreceiver.lock().unwrap();
2456 0 : assert!(
2457 0 : guard.is_none(),
2458 0 : "multiple launches / re-launches of WAL receiver are not supported"
2459 : );
2460 0 : *guard = Some(WalReceiver::start(
2461 0 : Arc::clone(self),
2462 0 : WalReceiverConf {
2463 0 : wal_connect_timeout,
2464 0 : lagging_wal_timeout,
2465 0 : max_lsn_wal_lag,
2466 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2467 0 : availability_zone: self.conf.availability_zone.clone(),
2468 0 : ingest_batch_size: self.conf.ingest_batch_size,
2469 0 : },
2470 0 : broker_client,
2471 0 : ctx,
2472 0 : ));
2473 0 : }
2474 :
2475 : /// Initialize with an empty layer map. Used when creating a new timeline.
2476 412 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2477 412 : let mut layers = self.layers.try_write().expect(
2478 412 : "in the context where we call this function, no other task has access to the object",
2479 412 : );
2480 412 : layers
2481 412 : .open_mut()
2482 412 : .expect("in this context the LayerManager must still be open")
2483 412 : .initialize_empty(Lsn(start_lsn.0));
2484 412 : }
2485 :
2486 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2487 : /// files.
2488 6 : pub(super) async fn load_layer_map(
2489 6 : &self,
2490 6 : disk_consistent_lsn: Lsn,
2491 6 : index_part: IndexPart,
2492 6 : ) -> anyhow::Result<()> {
2493 : use init::{Decision::*, Discovered, DismissedLayer};
2494 : use LayerName::*;
2495 :
2496 6 : let mut guard = self.layers.write().await;
2497 :
2498 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2499 6 :
2500 6 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2501 6 : // structs representing all files on disk
2502 6 : let timeline_path = self
2503 6 : .conf
2504 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2505 6 : let conf = self.conf;
2506 6 : let span = tracing::Span::current();
2507 6 :
2508 6 : // Copy to move into the task we're about to spawn
2509 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2510 :
2511 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2512 6 : move || {
2513 6 : let _g = span.entered();
2514 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2515 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2516 6 : let mut unrecognized_files = Vec::new();
2517 6 :
2518 6 : let mut path = timeline_path;
2519 :
2520 22 : for discovered in discovered {
2521 16 : let (name, kind) = match discovered {
2522 16 : Discovered::Layer(layer_file_name, local_metadata) => {
2523 16 : discovered_layers.push((layer_file_name, local_metadata));
2524 16 : continue;
2525 : }
2526 0 : Discovered::IgnoredBackup(path) => {
2527 0 : std::fs::remove_file(path)
2528 0 : .or_else(fs_ext::ignore_not_found)
2529 0 : .fatal_err("Removing .old file");
2530 0 : continue;
2531 : }
2532 0 : Discovered::Unknown(file_name) => {
2533 0 : // we will later error if there are any
2534 0 : unrecognized_files.push(file_name);
2535 0 : continue;
2536 : }
2537 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2538 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2539 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2540 : };
2541 0 : path.push(Utf8Path::new(&name));
2542 0 : init::cleanup(&path, kind)?;
2543 0 : path.pop();
2544 : }
2545 :
2546 6 : if !unrecognized_files.is_empty() {
2547 : // assume that if there are any there are many many.
2548 0 : let n = unrecognized_files.len();
2549 0 : let first = &unrecognized_files[..n.min(10)];
2550 0 : anyhow::bail!(
2551 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2552 0 : );
2553 6 : }
2554 6 :
2555 6 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2556 6 :
2557 6 : let mut loaded_layers = Vec::new();
2558 6 : let mut needs_cleanup = Vec::new();
2559 6 : let mut total_physical_size = 0;
2560 :
2561 22 : for (name, decision) in decided {
2562 16 : let decision = match decision {
2563 16 : Ok(decision) => decision,
2564 0 : Err(DismissedLayer::Future { local }) => {
2565 0 : if let Some(local) = local {
2566 0 : init::cleanup_future_layer(
2567 0 : &local.local_path,
2568 0 : &name,
2569 0 : disk_consistent_lsn,
2570 0 : )?;
2571 0 : }
2572 0 : needs_cleanup.push(name);
2573 0 : continue;
2574 : }
2575 0 : Err(DismissedLayer::LocalOnly(local)) => {
2576 0 : init::cleanup_local_only_file(&name, &local)?;
2577 : // this file never existed remotely, we will have to do rework
2578 0 : continue;
2579 : }
2580 0 : Err(DismissedLayer::BadMetadata(local)) => {
2581 0 : init::cleanup_local_file_for_remote(&local)?;
2582 : // this file never existed remotely, we will have to do rework
2583 0 : continue;
2584 : }
2585 : };
2586 :
2587 16 : match &name {
2588 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2589 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2590 : }
2591 :
2592 16 : tracing::debug!(layer=%name, ?decision, "applied");
2593 :
2594 16 : let layer = match decision {
2595 16 : Resident { local, remote } => {
2596 16 : total_physical_size += local.file_size;
2597 16 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2598 16 : .drop_eviction_guard()
2599 : }
2600 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2601 : };
2602 :
2603 16 : loaded_layers.push(layer);
2604 : }
2605 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2606 6 : }
2607 6 : })
2608 4 : .await
2609 6 : .map_err(anyhow::Error::new)
2610 6 : .and_then(|x| x)?;
2611 :
2612 6 : let num_layers = loaded_layers.len();
2613 6 :
2614 6 : guard
2615 6 : .open_mut()
2616 6 : .expect("layermanager must be open during init")
2617 6 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2618 6 :
2619 6 : self.remote_client
2620 6 : .schedule_layer_file_deletion(&needs_cleanup)?;
2621 6 : self.remote_client
2622 6 : .schedule_index_upload_for_file_changes()?;
2623 : // This barrier orders above DELETEs before any later operations.
2624 : // This is critical because code executing after the barrier might
2625 : // create again objects with the same key that we just scheduled for deletion.
2626 : // For example, if we just scheduled deletion of an image layer "from the future",
2627 : // later compaction might run again and re-create the same image layer.
2628 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2629 : // "same" here means same key range and LSN.
2630 : //
2631 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2632 : // the upload queue may execute the PUT first, then the DELETE.
2633 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2634 : //
2635 : // 1. a future image layer is created and uploaded
2636 : // 2. ps restart
2637 : // 3. the future layer from (1) is deleted during load layer map
2638 : // 4. image layer is re-created and uploaded
2639 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2640 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2641 : //
2642 : // See https://github.com/neondatabase/neon/issues/5878
2643 : //
2644 : // NB: generation numbers naturally protect against this because they disambiguate
2645 : // (1) and (4)
2646 6 : self.remote_client.schedule_barrier()?;
2647 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2648 : // on retry.
2649 :
2650 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2651 6 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2652 6 : self.update_layer_visibility().await?;
2653 :
2654 6 : info!(
2655 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2656 : num_layers, disk_consistent_lsn, total_physical_size
2657 : );
2658 :
2659 6 : timer.stop_and_record();
2660 6 : Ok(())
2661 6 : }
2662 :
2663 : /// Retrieve current logical size of the timeline.
2664 : ///
2665 : /// The size could be lagging behind the actual number, in case
2666 : /// the initial size calculation has not been run (gets triggered on the first size access).
2667 : ///
2668 : /// return size and boolean flag that shows if the size is exact
2669 0 : pub(crate) fn get_current_logical_size(
2670 0 : self: &Arc<Self>,
2671 0 : priority: GetLogicalSizePriority,
2672 0 : ctx: &RequestContext,
2673 0 : ) -> logical_size::CurrentLogicalSize {
2674 0 : if !self.tenant_shard_id.is_shard_zero() {
2675 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2676 : // when HTTP API is serving a GET for timeline zero, return zero
2677 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2678 0 : }
2679 0 :
2680 0 : let current_size = self.current_logical_size.current_size();
2681 0 : debug!("Current size: {current_size:?}");
2682 :
2683 0 : match (current_size.accuracy(), priority) {
2684 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2685 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2686 0 : // background task will eventually deliver an exact value, we're in no rush
2687 0 : }
2688 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2689 : // background task is not ready, but user is asking for it now;
2690 : // => make the background task skip the line
2691 : // (The alternative would be to calculate the size here, but,
2692 : // it can actually take a long time if the user has a lot of rels.
2693 : // And we'll inevitable need it again; So, let the background task do the work.)
2694 0 : match self
2695 0 : .current_logical_size
2696 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2697 0 : .get()
2698 : {
2699 0 : Some(cancel) => cancel.cancel(),
2700 : None => {
2701 0 : let state = self.current_state();
2702 0 : if matches!(
2703 0 : state,
2704 : TimelineState::Broken { .. } | TimelineState::Stopping
2705 0 : ) {
2706 0 :
2707 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2708 0 : // Don't make noise.
2709 0 : } else {
2710 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2711 0 : debug_assert!(false);
2712 : }
2713 : }
2714 : };
2715 : }
2716 : }
2717 :
2718 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2719 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2720 0 : let first = self
2721 0 : .current_logical_size
2722 0 : .did_return_approximate_to_walreceiver
2723 0 : .compare_exchange(
2724 0 : false,
2725 0 : true,
2726 0 : AtomicOrdering::Relaxed,
2727 0 : AtomicOrdering::Relaxed,
2728 0 : )
2729 0 : .is_ok();
2730 0 : if first {
2731 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2732 0 : }
2733 0 : }
2734 0 : }
2735 :
2736 0 : current_size
2737 0 : }
2738 :
2739 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2740 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2741 : // nothing to do for freshly created timelines;
2742 0 : assert_eq!(
2743 0 : self.current_logical_size.current_size().accuracy(),
2744 0 : logical_size::Accuracy::Exact,
2745 0 : );
2746 0 : self.current_logical_size.initialized.add_permits(1);
2747 0 : return;
2748 : };
2749 :
2750 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2751 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2752 0 : self.current_logical_size
2753 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2754 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2755 0 :
2756 0 : let self_clone = Arc::clone(self);
2757 0 : let background_ctx = ctx.detached_child(
2758 0 : TaskKind::InitialLogicalSizeCalculation,
2759 0 : DownloadBehavior::Download,
2760 0 : );
2761 0 : task_mgr::spawn(
2762 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2763 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2764 0 : self.tenant_shard_id,
2765 0 : Some(self.timeline_id),
2766 0 : "initial size calculation",
2767 : // NB: don't log errors here, task_mgr will do that.
2768 0 : async move {
2769 0 : let cancel = task_mgr::shutdown_token();
2770 0 : self_clone
2771 0 : .initial_logical_size_calculation_task(
2772 0 : initial_part_end,
2773 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2774 0 : cancel,
2775 0 : background_ctx,
2776 0 : )
2777 0 : .await;
2778 0 : Ok(())
2779 0 : }
2780 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)),
2781 : );
2782 0 : }
2783 :
2784 0 : async fn initial_logical_size_calculation_task(
2785 0 : self: Arc<Self>,
2786 0 : initial_part_end: Lsn,
2787 0 : skip_concurrency_limiter: CancellationToken,
2788 0 : cancel: CancellationToken,
2789 0 : background_ctx: RequestContext,
2790 0 : ) {
2791 0 : scopeguard::defer! {
2792 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2793 0 : self.current_logical_size.initialized.add_permits(1);
2794 0 : }
2795 0 :
2796 0 : let try_once = |attempt: usize| {
2797 0 : let background_ctx = &background_ctx;
2798 0 : let self_ref = &self;
2799 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2800 0 : async move {
2801 0 : let cancel = task_mgr::shutdown_token();
2802 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2803 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2804 0 : background_ctx,
2805 0 : );
2806 :
2807 : use crate::metrics::initial_logical_size::StartCircumstances;
2808 0 : let (_maybe_permit, circumstances) = tokio::select! {
2809 0 : permit = wait_for_permit => {
2810 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2811 : }
2812 0 : _ = self_ref.cancel.cancelled() => {
2813 0 : return Err(CalculateLogicalSizeError::Cancelled);
2814 : }
2815 0 : _ = cancel.cancelled() => {
2816 0 : return Err(CalculateLogicalSizeError::Cancelled);
2817 : },
2818 0 : () = skip_concurrency_limiter.cancelled() => {
2819 : // Some action that is part of a end user interaction requested logical size
2820 : // => break out of the rate limit
2821 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2822 : // but then again what happens if they cancel; also, we should just be using
2823 : // one runtime across the entire process, so, let's leave this for now.
2824 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2825 : }
2826 : };
2827 :
2828 0 : let metrics_guard = if attempt == 1 {
2829 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2830 : } else {
2831 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2832 : };
2833 :
2834 0 : let calculated_size = self_ref
2835 0 : .logical_size_calculation_task(
2836 0 : initial_part_end,
2837 0 : LogicalSizeCalculationCause::Initial,
2838 0 : background_ctx,
2839 0 : )
2840 0 : .await?;
2841 :
2842 0 : self_ref
2843 0 : .trigger_aux_file_size_computation(initial_part_end, background_ctx)
2844 0 : .await?;
2845 :
2846 : // TODO: add aux file size to logical size
2847 :
2848 0 : Ok((calculated_size, metrics_guard))
2849 0 : }
2850 0 : };
2851 :
2852 0 : let retrying = async {
2853 0 : let mut attempt = 0;
2854 : loop {
2855 0 : attempt += 1;
2856 0 :
2857 0 : match try_once(attempt).await {
2858 0 : Ok(res) => return ControlFlow::Continue(res),
2859 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
2860 : Err(
2861 0 : e @ (CalculateLogicalSizeError::Decode(_)
2862 0 : | CalculateLogicalSizeError::PageRead(_)),
2863 0 : ) => {
2864 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2865 : // exponential back-off doesn't make sense at these long intervals;
2866 : // use fixed retry interval with generous jitter instead
2867 0 : let sleep_duration = Duration::from_secs(
2868 0 : u64::try_from(
2869 0 : // 1hour base
2870 0 : (60_i64 * 60_i64)
2871 0 : // 10min jitter
2872 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2873 0 : )
2874 0 : .expect("10min < 1hour"),
2875 0 : );
2876 0 : tokio::time::sleep(sleep_duration).await;
2877 : }
2878 : }
2879 : }
2880 0 : };
2881 :
2882 0 : let (calculated_size, metrics_guard) = tokio::select! {
2883 0 : res = retrying => {
2884 0 : match res {
2885 0 : ControlFlow::Continue(calculated_size) => calculated_size,
2886 0 : ControlFlow::Break(()) => return,
2887 : }
2888 : }
2889 0 : _ = cancel.cancelled() => {
2890 0 : return;
2891 : }
2892 : };
2893 :
2894 : // we cannot query current_logical_size.current_size() to know the current
2895 : // *negative* value, only truncated to u64.
2896 0 : let added = self
2897 0 : .current_logical_size
2898 0 : .size_added_after_initial
2899 0 : .load(AtomicOrdering::Relaxed);
2900 0 :
2901 0 : let sum = calculated_size.saturating_add_signed(added);
2902 0 :
2903 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2904 0 : self.metrics.current_logical_size_gauge.set(sum);
2905 0 :
2906 0 : self.current_logical_size
2907 0 : .initial_logical_size
2908 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
2909 0 : .ok()
2910 0 : .expect("only this task sets it");
2911 0 : }
2912 :
2913 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
2914 0 : self: &Arc<Self>,
2915 0 : lsn: Lsn,
2916 0 : cause: LogicalSizeCalculationCause,
2917 0 : ctx: RequestContext,
2918 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
2919 0 : let (sender, receiver) = oneshot::channel();
2920 0 : let self_clone = Arc::clone(self);
2921 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
2922 0 : // we should stop the size calculation work and return an error.
2923 0 : // That would require restructuring this function's API to
2924 0 : // return the result directly, instead of a Receiver for the result.
2925 0 : let ctx = ctx.detached_child(
2926 0 : TaskKind::OndemandLogicalSizeCalculation,
2927 0 : DownloadBehavior::Download,
2928 0 : );
2929 0 : task_mgr::spawn(
2930 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2931 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
2932 0 : self.tenant_shard_id,
2933 0 : Some(self.timeline_id),
2934 0 : "ondemand logical size calculation",
2935 0 : async move {
2936 0 : let res = self_clone
2937 0 : .logical_size_calculation_task(lsn, cause, &ctx)
2938 0 : .await;
2939 0 : let _ = sender.send(res).ok();
2940 0 : Ok(()) // Receiver is responsible for handling errors
2941 0 : }
2942 0 : .in_current_span(),
2943 0 : );
2944 0 : receiver
2945 0 : }
2946 :
2947 : /// # Cancel-Safety
2948 : ///
2949 : /// This method is cancellation-safe.
2950 0 : #[instrument(skip_all)]
2951 : async fn logical_size_calculation_task(
2952 : self: &Arc<Self>,
2953 : lsn: Lsn,
2954 : cause: LogicalSizeCalculationCause,
2955 : ctx: &RequestContext,
2956 : ) -> Result<u64, CalculateLogicalSizeError> {
2957 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
2958 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
2959 : // accurate relation sizes, and they do not emit consumption metrics.
2960 : debug_assert!(self.tenant_shard_id.is_shard_zero());
2961 :
2962 : let guard = self
2963 : .gate
2964 : .enter()
2965 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
2966 :
2967 : let self_calculation = Arc::clone(self);
2968 :
2969 0 : let mut calculation = pin!(async {
2970 0 : let ctx = ctx.attached_child();
2971 0 : self_calculation
2972 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
2973 0 : .await
2974 0 : });
2975 :
2976 : tokio::select! {
2977 : res = &mut calculation => { res }
2978 : _ = self.cancel.cancelled() => {
2979 : debug!("cancelling logical size calculation for timeline shutdown");
2980 : calculation.await
2981 : }
2982 : }
2983 : }
2984 :
2985 : /// Calculate the logical size of the database at the latest LSN.
2986 : ///
2987 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
2988 : /// especially if we need to download remote layers.
2989 : ///
2990 : /// # Cancel-Safety
2991 : ///
2992 : /// This method is cancellation-safe.
2993 0 : async fn calculate_logical_size(
2994 0 : &self,
2995 0 : up_to_lsn: Lsn,
2996 0 : cause: LogicalSizeCalculationCause,
2997 0 : _guard: &GateGuard,
2998 0 : ctx: &RequestContext,
2999 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3000 0 : info!(
3001 0 : "Calculating logical size for timeline {} at {}",
3002 : self.timeline_id, up_to_lsn
3003 : );
3004 :
3005 0 : pausable_failpoint!("timeline-calculate-logical-size-pause");
3006 :
3007 : // See if we've already done the work for initial size calculation.
3008 : // This is a short-cut for timelines that are mostly unused.
3009 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3010 0 : return Ok(size);
3011 0 : }
3012 0 : let storage_time_metrics = match cause {
3013 : LogicalSizeCalculationCause::Initial
3014 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3015 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3016 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3017 0 : &self.metrics.imitate_logical_size_histo
3018 : }
3019 : };
3020 0 : let timer = storage_time_metrics.start_timer();
3021 0 : let logical_size = self
3022 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3023 0 : .await?;
3024 0 : debug!("calculated logical size: {logical_size}");
3025 0 : timer.stop_and_record();
3026 0 : Ok(logical_size)
3027 0 : }
3028 :
3029 : /// Update current logical size, adding `delta' to the old value.
3030 270570 : fn update_current_logical_size(&self, delta: i64) {
3031 270570 : let logical_size = &self.current_logical_size;
3032 270570 : logical_size.increment_size(delta);
3033 270570 :
3034 270570 : // Also set the value in the prometheus gauge. Note that
3035 270570 : // there is a race condition here: if this is is called by two
3036 270570 : // threads concurrently, the prometheus gauge might be set to
3037 270570 : // one value while current_logical_size is set to the
3038 270570 : // other.
3039 270570 : match logical_size.current_size() {
3040 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
3041 270570 : .metrics
3042 270570 : .current_logical_size_gauge
3043 270570 : .set(new_current_size.into()),
3044 0 : CurrentLogicalSize::Approximate(_) => {
3045 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3046 0 : // forth between the initial size calculation task.
3047 0 : }
3048 : }
3049 270570 : }
3050 :
3051 2834 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3052 2834 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3053 2834 : let aux_metric =
3054 2834 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3055 2834 :
3056 2834 : let sum_of_entries = self
3057 2834 : .directory_metrics
3058 2834 : .iter()
3059 19838 : .map(|v| v.load(AtomicOrdering::Relaxed))
3060 2834 : .sum();
3061 : // Set a high general threshold and a lower threshold for the auxiliary files,
3062 : // as we can have large numbers of relations in the db directory.
3063 : const SUM_THRESHOLD: u64 = 5000;
3064 : const AUX_THRESHOLD: u64 = 1000;
3065 2834 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3066 0 : self.metrics
3067 0 : .directory_entries_count_gauge
3068 0 : .set(sum_of_entries);
3069 2834 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3070 0 : metric.set(sum_of_entries);
3071 2834 : }
3072 2834 : }
3073 :
3074 0 : async fn find_layer(
3075 0 : &self,
3076 0 : layer_name: &LayerName,
3077 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3078 0 : let guard = self.layers.read().await;
3079 0 : let layer = guard
3080 0 : .layer_map()?
3081 0 : .iter_historic_layers()
3082 0 : .find(|l| &l.layer_name() == layer_name)
3083 0 : .map(|found| guard.get_from_desc(&found));
3084 0 : Ok(layer)
3085 0 : }
3086 :
3087 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3088 : /// indicating which layers are currently on-disk on the primary.
3089 : ///
3090 : /// None is returned if the Timeline is in a state where uploading a heatmap
3091 : /// doesn't make sense, such as shutting down or initializing. The caller
3092 : /// should treat this as a cue to simply skip doing any heatmap uploading
3093 : /// for this timeline.
3094 2 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3095 2 : if !self.is_active() {
3096 0 : return None;
3097 2 : }
3098 :
3099 2 : let guard = self.layers.read().await;
3100 :
3101 10 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3102 10 : match layer.visibility() {
3103 : LayerVisibilityHint::Visible => {
3104 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3105 8 : let last_activity_ts = layer.latest_activity();
3106 8 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
3107 : }
3108 : LayerVisibilityHint::Covered => {
3109 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3110 2 : None
3111 : }
3112 : }
3113 10 : });
3114 2 :
3115 2 : let mut layers = resident.collect::<Vec<_>>();
3116 2 :
3117 2 : // Sort layers in order of which to download first. For a large set of layers to download, we
3118 2 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3119 2 : // or hours later:
3120 2 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3121 2 : // only exist for a few minutes before being compacted into L1s.
3122 2 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3123 2 : // the layer is likely to be covered by an image layer during compaction.
3124 24 : layers.sort_by_key(|(desc, _meta, _atime)| {
3125 24 : std::cmp::Reverse((
3126 24 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3127 24 : desc.lsn_range.end,
3128 24 : ))
3129 24 : });
3130 2 :
3131 2 : let layers = layers
3132 2 : .into_iter()
3133 8 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3134 2 : .collect();
3135 2 :
3136 2 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3137 2 : }
3138 :
3139 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3140 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3141 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3142 0 : // branchpoint in the value in IndexPart::lineage
3143 0 : self.ancestor_lsn == lsn
3144 0 : || (self.ancestor_lsn == Lsn::INVALID
3145 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3146 0 : }
3147 : }
3148 :
3149 : impl Timeline {
3150 : #[allow(clippy::doc_lazy_continuation)]
3151 : /// Get the data needed to reconstruct all keys in the provided keyspace
3152 : ///
3153 : /// The algorithm is as follows:
3154 : /// 1. While some keys are still not done and there's a timeline to visit:
3155 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3156 : /// 2.1: Build the fringe for the current keyspace
3157 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3158 : /// intersects
3159 : /// 2.3. Pop the timeline from the fringe
3160 : /// 2.4. If the fringe is empty, go back to 1
3161 626600 : async fn get_vectored_reconstruct_data(
3162 626600 : &self,
3163 626600 : mut keyspace: KeySpace,
3164 626600 : request_lsn: Lsn,
3165 626600 : reconstruct_state: &mut ValuesReconstructState,
3166 626600 : ctx: &RequestContext,
3167 626600 : ) -> Result<(), GetVectoredError> {
3168 626600 : let mut timeline_owned: Arc<Timeline>;
3169 626600 : let mut timeline = self;
3170 626600 :
3171 626600 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3172 :
3173 626598 : let missing_keyspace = loop {
3174 852196 : if self.cancel.is_cancelled() {
3175 0 : return Err(GetVectoredError::Cancelled);
3176 852196 : }
3177 :
3178 : let TimelineVisitOutcome {
3179 852196 : completed_keyspace: completed,
3180 852196 : image_covered_keyspace,
3181 852196 : } = Self::get_vectored_reconstruct_data_timeline(
3182 852196 : timeline,
3183 852196 : keyspace.clone(),
3184 852196 : cont_lsn,
3185 852196 : reconstruct_state,
3186 852196 : &self.cancel,
3187 852196 : ctx,
3188 852196 : )
3189 192258 : .await?;
3190 :
3191 852196 : keyspace.remove_overlapping_with(&completed);
3192 852196 :
3193 852196 : // Do not descend into the ancestor timeline for aux files.
3194 852196 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3195 852196 : // stalling compaction.
3196 852196 : keyspace.remove_overlapping_with(&KeySpace {
3197 852196 : ranges: vec![NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE],
3198 852196 : });
3199 852196 :
3200 852196 : // Keyspace is fully retrieved
3201 852196 : if keyspace.is_empty() {
3202 626584 : break None;
3203 225612 : }
3204 :
3205 225612 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3206 : // Not fully retrieved but no ancestor timeline.
3207 14 : break Some(keyspace);
3208 : };
3209 :
3210 : // Now we see if there are keys covered by the image layer but does not exist in the
3211 : // image layer, which means that the key does not exist.
3212 :
3213 : // The block below will stop the vectored search if any of the keys encountered an image layer
3214 : // which did not contain a snapshot for said key. Since we have already removed all completed
3215 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3216 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3217 : // and stop the search as a result of that.
3218 225598 : let removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3219 225598 : if !removed.is_empty() {
3220 0 : break Some(removed);
3221 225598 : }
3222 225598 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3223 225598 : // keyspace.
3224 225598 :
3225 225598 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3226 225598 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3227 225598 : timeline_owned = timeline
3228 225598 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3229 2 : .await?;
3230 225596 : timeline = &*timeline_owned;
3231 : };
3232 :
3233 626598 : if let Some(missing_keyspace) = missing_keyspace {
3234 14 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3235 14 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3236 14 : shard: self
3237 14 : .shard_identity
3238 14 : .get_shard_number(&missing_keyspace.start().unwrap()),
3239 14 : cont_lsn,
3240 14 : request_lsn,
3241 14 : ancestor_lsn: Some(timeline.ancestor_lsn),
3242 14 : backtrace: None,
3243 14 : }));
3244 626584 : }
3245 626584 :
3246 626584 : Ok(())
3247 626600 : }
3248 :
3249 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3250 : ///
3251 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3252 : /// the current keyspace. The current keyspace of the search at any given timeline
3253 : /// is the original keyspace minus all the keys that have been completed minus
3254 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3255 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3256 : ///
3257 : /// This is basically a depth-first search visitor implementation where a vertex
3258 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3259 : ///
3260 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3261 : /// and get all the required reconstruct data from the layer in one go.
3262 : ///
3263 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3264 : /// decides how to deal with these two keyspaces.
3265 852196 : async fn get_vectored_reconstruct_data_timeline(
3266 852196 : timeline: &Timeline,
3267 852196 : keyspace: KeySpace,
3268 852196 : mut cont_lsn: Lsn,
3269 852196 : reconstruct_state: &mut ValuesReconstructState,
3270 852196 : cancel: &CancellationToken,
3271 852196 : ctx: &RequestContext,
3272 852196 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3273 852196 : let mut unmapped_keyspace = keyspace.clone();
3274 852196 : let mut fringe = LayerFringe::new();
3275 852196 :
3276 852196 : let mut completed_keyspace = KeySpace::default();
3277 852196 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3278 :
3279 : loop {
3280 1698287 : if cancel.is_cancelled() {
3281 0 : return Err(GetVectoredError::Cancelled);
3282 1698287 : }
3283 1698287 :
3284 1698287 : let (keys_done_last_step, keys_with_image_coverage) =
3285 1698287 : reconstruct_state.consume_done_keys();
3286 1698287 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3287 1698287 : completed_keyspace.merge(&keys_done_last_step);
3288 1698287 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3289 21808 : unmapped_keyspace
3290 21808 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3291 21808 : image_covered_keyspace.add_range(keys_with_image_coverage);
3292 1676479 : }
3293 :
3294 : // Do not descent any further if the last layer we visited
3295 : // completed all keys in the keyspace it inspected. This is not
3296 : // required for correctness, but avoids visiting extra layers
3297 : // which turns out to be a perf bottleneck in some cases.
3298 1698287 : if !unmapped_keyspace.is_empty() {
3299 1073947 : let guard = timeline.layers.read().await;
3300 1073947 : let layers = guard.layer_map()?;
3301 :
3302 1073947 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3303 913505 : let start_lsn = l.get_lsn_range().start;
3304 913505 : cont_lsn > start_lsn
3305 1073947 : });
3306 1073947 :
3307 1073947 : match in_memory_layer {
3308 606325 : Some(l) => {
3309 606325 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3310 606325 : fringe.update(
3311 606325 : ReadableLayer::InMemoryLayer(l),
3312 606325 : unmapped_keyspace.clone(),
3313 606325 : lsn_range,
3314 606325 : );
3315 606325 : }
3316 : None => {
3317 467644 : for range in unmapped_keyspace.ranges.iter() {
3318 467644 : let results = layers.range_search(range.clone(), cont_lsn);
3319 467644 :
3320 467644 : results
3321 467644 : .found
3322 467644 : .into_iter()
3323 467644 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3324 239780 : (
3325 239780 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3326 239780 : keyspace_accum.to_keyspace(),
3327 239780 : lsn_floor..cont_lsn,
3328 239780 : )
3329 467644 : })
3330 467644 : .for_each(|(layer, keyspace, lsn_range)| {
3331 239780 : fringe.update(layer, keyspace, lsn_range)
3332 467644 : });
3333 467644 : }
3334 : }
3335 : }
3336 :
3337 : // It's safe to drop the layer map lock after planning the next round of reads.
3338 : // The fringe keeps readable handles for the layers which are safe to read even
3339 : // if layers were compacted or flushed.
3340 : //
3341 : // The more interesting consideration is: "Why is the read algorithm still correct
3342 : // if the layer map changes while it is operating?". Doing a vectored read on a
3343 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3344 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3345 : // range at *a particular point in time*. It is fine for the answer to be different
3346 : // at two different time points.
3347 1073947 : drop(guard);
3348 624340 : }
3349 :
3350 1698287 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3351 846091 : let next_cont_lsn = lsn_range.start;
3352 846091 : layer_to_read
3353 846091 : .get_values_reconstruct_data(
3354 846091 : keyspace_to_read.clone(),
3355 846091 : lsn_range,
3356 846091 : reconstruct_state,
3357 846091 : ctx,
3358 846091 : )
3359 183044 : .await?;
3360 :
3361 846091 : unmapped_keyspace = keyspace_to_read;
3362 846091 : cont_lsn = next_cont_lsn;
3363 846091 :
3364 846091 : reconstruct_state.on_layer_visited(&layer_to_read);
3365 : } else {
3366 852196 : break;
3367 852196 : }
3368 852196 : }
3369 852196 :
3370 852196 : Ok(TimelineVisitOutcome {
3371 852196 : completed_keyspace,
3372 852196 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3373 852196 : })
3374 852196 : }
3375 :
3376 225598 : async fn get_ready_ancestor_timeline(
3377 225598 : &self,
3378 225598 : ancestor: &Arc<Timeline>,
3379 225598 : ctx: &RequestContext,
3380 225598 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3381 225598 : // It's possible that the ancestor timeline isn't active yet, or
3382 225598 : // is active but hasn't yet caught up to the branch point. Wait
3383 225598 : // for it.
3384 225598 : //
3385 225598 : // This cannot happen while the pageserver is running normally,
3386 225598 : // because you cannot create a branch from a point that isn't
3387 225598 : // present in the pageserver yet. However, we don't wait for the
3388 225598 : // branch point to be uploaded to cloud storage before creating
3389 225598 : // a branch. I.e., the branch LSN need not be remote consistent
3390 225598 : // for the branching operation to succeed.
3391 225598 : //
3392 225598 : // Hence, if we try to load a tenant in such a state where
3393 225598 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3394 225598 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3395 225598 : // then we will need to wait for the ancestor timeline to
3396 225598 : // re-stream WAL up to branch_lsn before we access it.
3397 225598 : //
3398 225598 : // How can a tenant get in such a state?
3399 225598 : // - ungraceful pageserver process exit
3400 225598 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3401 225598 : //
3402 225598 : // NB: this could be avoided by requiring
3403 225598 : // branch_lsn >= remote_consistent_lsn
3404 225598 : // during branch creation.
3405 225598 : match ancestor.wait_to_become_active(ctx).await {
3406 225596 : Ok(()) => {}
3407 : Err(TimelineState::Stopping) => {
3408 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3409 0 : return Err(GetReadyAncestorError::Cancelled);
3410 : }
3411 2 : Err(state) => {
3412 2 : return Err(GetReadyAncestorError::BadState {
3413 2 : timeline_id: ancestor.timeline_id,
3414 2 : state,
3415 2 : });
3416 : }
3417 : }
3418 225596 : ancestor
3419 225596 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3420 0 : .await
3421 225596 : .map_err(|e| match e {
3422 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3423 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3424 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3425 0 : timeline_id: ancestor.timeline_id,
3426 0 : state,
3427 0 : },
3428 225596 : })?;
3429 :
3430 225596 : Ok(ancestor.clone())
3431 225598 : }
3432 :
3433 151304 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3434 151304 : &self.shard_identity
3435 151304 : }
3436 :
3437 : #[inline(always)]
3438 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3439 0 : ShardTimelineId {
3440 0 : shard_index: ShardIndex {
3441 0 : shard_number: self.shard_identity.number,
3442 0 : shard_count: self.shard_identity.count,
3443 0 : },
3444 0 : timeline_id: self.timeline_id,
3445 0 : }
3446 0 : }
3447 :
3448 : /// Returns a non-frozen open in-memory layer for ingestion.
3449 : ///
3450 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3451 : /// this function without holding the mutex.
3452 1268 : async fn get_layer_for_write(
3453 1268 : &self,
3454 1268 : lsn: Lsn,
3455 1268 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3456 1268 : ctx: &RequestContext,
3457 1268 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3458 1268 : let mut guard = self.layers.write().await;
3459 1268 : let gate_guard = self.gate.enter().context("enter gate for inmem layer")?;
3460 :
3461 1268 : let last_record_lsn = self.get_last_record_lsn();
3462 1268 : ensure!(
3463 1268 : lsn > last_record_lsn,
3464 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3465 : lsn,
3466 : last_record_lsn,
3467 : );
3468 :
3469 1268 : let layer = guard
3470 1268 : .open_mut()?
3471 1268 : .get_layer_for_write(
3472 1268 : lsn,
3473 1268 : self.conf,
3474 1268 : self.timeline_id,
3475 1268 : self.tenant_shard_id,
3476 1268 : gate_guard,
3477 1268 : ctx,
3478 1268 : )
3479 719 : .await?;
3480 1268 : Ok(layer)
3481 1268 : }
3482 :
3483 5279064 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3484 5279064 : assert!(new_lsn.is_aligned());
3485 :
3486 5279064 : self.metrics.last_record_gauge.set(new_lsn.0 as i64);
3487 5279064 : self.last_record_lsn.advance(new_lsn);
3488 5279064 : }
3489 :
3490 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3491 : ///
3492 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3493 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3494 1172 : async fn freeze_inmem_layer_at(
3495 1172 : &self,
3496 1172 : at: Lsn,
3497 1172 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3498 1172 : ) -> Result<u64, FlushLayerError> {
3499 1172 : let frozen = {
3500 1172 : let mut guard = self.layers.write().await;
3501 1172 : guard
3502 1172 : .open_mut()?
3503 1172 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock)
3504 1 : .await
3505 : };
3506 :
3507 1172 : if frozen {
3508 1144 : let now = Instant::now();
3509 1144 : *(self.last_freeze_ts.write().unwrap()) = now;
3510 1144 : }
3511 :
3512 : // Increment the flush cycle counter and wake up the flush task.
3513 : // Remember the new value, so that when we listen for the flush
3514 : // to finish, we know when the flush that we initiated has
3515 : // finished, instead of some other flush that was started earlier.
3516 1172 : let mut my_flush_request = 0;
3517 1172 :
3518 1172 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3519 1172 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3520 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3521 1172 : }
3522 1172 :
3523 1172 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3524 1172 : my_flush_request = *counter + 1;
3525 1172 : *counter = my_flush_request;
3526 1172 : *lsn = std::cmp::max(at, *lsn);
3527 1172 : });
3528 1172 :
3529 1172 : assert_ne!(my_flush_request, 0);
3530 :
3531 1172 : Ok(my_flush_request)
3532 1172 : }
3533 :
3534 : /// Layer flusher task's main loop.
3535 412 : async fn flush_loop(
3536 412 : self: &Arc<Self>,
3537 412 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3538 412 : ctx: &RequestContext,
3539 412 : ) {
3540 412 : info!("started flush loop");
3541 : loop {
3542 1547 : tokio::select! {
3543 1547 : _ = self.cancel.cancelled() => {
3544 10 : info!("shutting down layer flush task due to Timeline::cancel");
3545 10 : break;
3546 : },
3547 1547 : _ = layer_flush_start_rx.changed() => {}
3548 1135 : }
3549 1135 : trace!("waking up");
3550 1135 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3551 1135 :
3552 1135 : // The highest LSN to which we flushed in the loop over frozen layers
3553 1135 : let mut flushed_to_lsn = Lsn(0);
3554 :
3555 1135 : let result = loop {
3556 2279 : if self.cancel.is_cancelled() {
3557 0 : info!("dropping out of flush loop for timeline shutdown");
3558 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3559 : // anyone waiting on that will respect self.cancel as well: they will stop
3560 : // waiting at the same time we as drop out of this loop.
3561 0 : return;
3562 2279 : }
3563 2279 :
3564 2279 : let timer = self.metrics.flush_time_histo.start_timer();
3565 :
3566 : let num_frozen_layers;
3567 : let frozen_layer_total_size;
3568 2279 : let layer_to_flush = {
3569 2279 : let guard = self.layers.read().await;
3570 2279 : let Ok(lm) = guard.layer_map() else {
3571 0 : info!("dropping out of flush loop for timeline shutdown");
3572 0 : return;
3573 : };
3574 2279 : num_frozen_layers = lm.frozen_layers.len();
3575 2279 : frozen_layer_total_size = lm
3576 2279 : .frozen_layers
3577 2279 : .iter()
3578 2279 : .map(|l| l.estimated_in_mem_size())
3579 2279 : .sum::<u64>();
3580 2279 : lm.frozen_layers.front().cloned()
3581 : // drop 'layers' lock to allow concurrent reads and writes
3582 : };
3583 2279 : let Some(layer_to_flush) = layer_to_flush else {
3584 1135 : break Ok(());
3585 : };
3586 1144 : if num_frozen_layers
3587 1144 : > std::cmp::max(
3588 1144 : self.get_compaction_threshold(),
3589 1144 : DEFAULT_COMPACTION_THRESHOLD,
3590 1144 : )
3591 0 : && frozen_layer_total_size >= /* 128 MB */ 128000000
3592 : {
3593 0 : tracing::warn!(
3594 0 : "too many frozen layers: {num_frozen_layers} layers with estimated in-mem size of {frozen_layer_total_size} bytes",
3595 : );
3596 1144 : }
3597 17070 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3598 1144 : Ok(this_layer_to_lsn) => {
3599 1144 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3600 1144 : }
3601 : Err(FlushLayerError::Cancelled) => {
3602 0 : info!("dropping out of flush loop for timeline shutdown");
3603 0 : return;
3604 : }
3605 0 : err @ Err(
3606 0 : FlushLayerError::NotRunning(_)
3607 0 : | FlushLayerError::Other(_)
3608 0 : | FlushLayerError::CreateImageLayersError(_),
3609 0 : ) => {
3610 0 : error!("could not flush frozen layer: {err:?}");
3611 0 : break err.map(|_| ());
3612 : }
3613 : }
3614 1144 : timer.stop_and_record();
3615 : };
3616 :
3617 : // Unsharded tenants should never advance their LSN beyond the end of the
3618 : // highest layer they write: such gaps between layer data and the frozen LSN
3619 : // are only legal on sharded tenants.
3620 1135 : debug_assert!(
3621 1135 : self.shard_identity.count.count() > 1
3622 1135 : || flushed_to_lsn >= frozen_to_lsn
3623 66 : || !flushed_to_lsn.is_valid()
3624 : );
3625 :
3626 1135 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3627 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3628 : // to us via layer_flush_start_rx, then advance it here.
3629 : //
3630 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3631 : // never encounter a gap in the wal.
3632 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3633 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3634 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3635 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3636 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3637 0 : }
3638 0 : }
3639 1135 : }
3640 :
3641 : // Notify any listeners that we're done
3642 1135 : let _ = self
3643 1135 : .layer_flush_done_tx
3644 1135 : .send_replace((flush_counter, result));
3645 : }
3646 10 : }
3647 :
3648 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
3649 1092 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3650 1092 : let mut rx = self.layer_flush_done_tx.subscribe();
3651 : loop {
3652 : {
3653 2179 : let (last_result_counter, last_result) = &*rx.borrow();
3654 2179 : if *last_result_counter >= request {
3655 1092 : if let Err(err) = last_result {
3656 : // We already logged the original error in
3657 : // flush_loop. We cannot propagate it to the caller
3658 : // here, because it might not be Cloneable
3659 0 : return Err(err.clone());
3660 : } else {
3661 1092 : return Ok(());
3662 : }
3663 1087 : }
3664 1087 : }
3665 1087 : trace!("waiting for flush to complete");
3666 1087 : tokio::select! {
3667 1087 : rx_e = rx.changed() => {
3668 1087 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3669 : },
3670 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3671 : // the notification from [`flush_loop`] that it completed.
3672 1087 : _ = self.cancel.cancelled() => {
3673 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3674 0 : return Ok(())
3675 : }
3676 : };
3677 1087 : trace!("done")
3678 : }
3679 1092 : }
3680 :
3681 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3682 : ///
3683 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3684 1144 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3685 : async fn flush_frozen_layer(
3686 : self: &Arc<Self>,
3687 : frozen_layer: Arc<InMemoryLayer>,
3688 : ctx: &RequestContext,
3689 : ) -> Result<Lsn, FlushLayerError> {
3690 : debug_assert_current_span_has_tenant_and_timeline_id();
3691 :
3692 : // As a special case, when we have just imported an image into the repository,
3693 : // instead of writing out a L0 delta layer, we directly write out image layer
3694 : // files instead. This is possible as long as *all* the data imported into the
3695 : // repository have the same LSN.
3696 : let lsn_range = frozen_layer.get_lsn_range();
3697 :
3698 : // Whether to directly create image layers for this flush, or flush them as delta layers
3699 : let create_image_layer =
3700 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3701 :
3702 : #[cfg(test)]
3703 : {
3704 : match &mut *self.flush_loop_state.lock().unwrap() {
3705 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3706 : panic!("flush loop not running")
3707 : }
3708 : FlushLoopState::Running {
3709 : expect_initdb_optimization,
3710 : initdb_optimization_count,
3711 : ..
3712 : } => {
3713 : if create_image_layer {
3714 : *initdb_optimization_count += 1;
3715 : } else {
3716 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3717 : }
3718 : }
3719 : }
3720 : }
3721 :
3722 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3723 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3724 : // require downloading anything during initial import.
3725 : let ((rel_partition, metadata_partition), _lsn) = self
3726 : .repartition(
3727 : self.initdb_lsn,
3728 : self.get_compaction_target_size(),
3729 : EnumSet::empty(),
3730 : ctx,
3731 : )
3732 : .await
3733 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
3734 :
3735 : if self.cancel.is_cancelled() {
3736 : return Err(FlushLayerError::Cancelled);
3737 : }
3738 :
3739 : let mut layers_to_upload = Vec::new();
3740 : layers_to_upload.extend(
3741 : self.create_image_layers(
3742 : &rel_partition,
3743 : self.initdb_lsn,
3744 : ImageLayerCreationMode::Initial,
3745 : ctx,
3746 : )
3747 : .await?,
3748 : );
3749 : if !metadata_partition.parts.is_empty() {
3750 : assert_eq!(
3751 : metadata_partition.parts.len(),
3752 : 1,
3753 : "currently sparse keyspace should only contain a single metadata keyspace"
3754 : );
3755 : layers_to_upload.extend(
3756 : self.create_image_layers(
3757 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
3758 : // every single key within the keyspace, and therefore, it's safe to force converting it
3759 : // into a dense keyspace before calling this function.
3760 : &metadata_partition.into_dense(),
3761 : self.initdb_lsn,
3762 : ImageLayerCreationMode::Initial,
3763 : ctx,
3764 : )
3765 : .await?,
3766 : );
3767 : }
3768 :
3769 : (layers_to_upload, None)
3770 : } else {
3771 : // Normal case, write out a L0 delta layer file.
3772 : // `create_delta_layer` will not modify the layer map.
3773 : // We will remove frozen layer and add delta layer in one atomic operation later.
3774 : let Some(layer) = self
3775 : .create_delta_layer(&frozen_layer, None, ctx)
3776 : .await
3777 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
3778 : else {
3779 : panic!("delta layer cannot be empty if no filter is applied");
3780 : };
3781 : (
3782 : // FIXME: even though we have a single image and single delta layer assumption
3783 : // we push them to vec
3784 : vec![layer.clone()],
3785 : Some(layer),
3786 : )
3787 : };
3788 :
3789 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
3790 :
3791 : if self.cancel.is_cancelled() {
3792 : return Err(FlushLayerError::Cancelled);
3793 : }
3794 :
3795 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
3796 :
3797 : // The new on-disk layers are now in the layer map. We can remove the
3798 : // in-memory layer from the map now. The flushed layer is stored in
3799 : // the mapping in `create_delta_layer`.
3800 : {
3801 : let mut guard = self.layers.write().await;
3802 :
3803 : guard.open_mut()?.finish_flush_l0_layer(
3804 : delta_layer_to_add.as_ref(),
3805 : &frozen_layer,
3806 : &self.metrics,
3807 : );
3808 :
3809 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
3810 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
3811 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
3812 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
3813 : }
3814 : // release lock on 'layers'
3815 : };
3816 :
3817 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
3818 : // This makes us refuse ingest until the new layers have been persisted to the remote.
3819 : self.remote_client
3820 : .wait_completion()
3821 : .await
3822 0 : .map_err(|e| match e {
3823 : WaitCompletionError::UploadQueueShutDownOrStopped
3824 : | WaitCompletionError::NotInitialized(
3825 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
3826 0 : ) => FlushLayerError::Cancelled,
3827 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
3828 0 : FlushLayerError::Other(anyhow!(e).into())
3829 : }
3830 0 : })?;
3831 :
3832 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
3833 : // a compaction can delete the file and then it won't be available for uploads any more.
3834 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
3835 : // race situation.
3836 : // See https://github.com/neondatabase/neon/issues/4526
3837 : pausable_failpoint!("flush-frozen-pausable");
3838 :
3839 : // This failpoint is used by another test case `test_pageserver_recovery`.
3840 : fail_point!("flush-frozen-exit");
3841 :
3842 : Ok(Lsn(lsn_range.end.0 - 1))
3843 : }
3844 :
3845 : /// Return true if the value changed
3846 : ///
3847 : /// This function must only be used from the layer flush task.
3848 1144 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
3849 1144 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
3850 1144 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
3851 1144 : new_value != old_value
3852 1144 : }
3853 :
3854 : /// Update metadata file
3855 1146 : fn schedule_uploads(
3856 1146 : &self,
3857 1146 : disk_consistent_lsn: Lsn,
3858 1146 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
3859 1146 : ) -> anyhow::Result<()> {
3860 1146 : // We can only save a valid 'prev_record_lsn' value on disk if we
3861 1146 : // flushed *all* in-memory changes to disk. We only track
3862 1146 : // 'prev_record_lsn' in memory for the latest processed record, so we
3863 1146 : // don't remember what the correct value that corresponds to some old
3864 1146 : // LSN is. But if we flush everything, then the value corresponding
3865 1146 : // current 'last_record_lsn' is correct and we can store it on disk.
3866 1146 : let RecordLsn {
3867 1146 : last: last_record_lsn,
3868 1146 : prev: prev_record_lsn,
3869 1146 : } = self.last_record_lsn.load();
3870 1146 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
3871 1069 : Some(prev_record_lsn)
3872 : } else {
3873 77 : None
3874 : };
3875 :
3876 1146 : let update = crate::tenant::metadata::MetadataUpdate::new(
3877 1146 : disk_consistent_lsn,
3878 1146 : ondisk_prev_record_lsn,
3879 1146 : *self.latest_gc_cutoff_lsn.read(),
3880 1146 : );
3881 1146 :
3882 1146 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
3883 0 : "{}",
3884 0 : x.unwrap()
3885 1146 : ));
3886 :
3887 2302 : for layer in layers_to_upload {
3888 1156 : self.remote_client.schedule_layer_file_upload(layer)?;
3889 : }
3890 1146 : self.remote_client
3891 1146 : .schedule_index_upload_for_metadata_update(&update)?;
3892 :
3893 1146 : Ok(())
3894 1146 : }
3895 :
3896 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
3897 0 : self.remote_client
3898 0 : .preserve_initdb_archive(
3899 0 : &self.tenant_shard_id.tenant_id,
3900 0 : &self.timeline_id,
3901 0 : &self.cancel,
3902 0 : )
3903 0 : .await
3904 0 : }
3905 :
3906 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
3907 : // in layer map immediately. The caller is responsible to put it into the layer map.
3908 968 : async fn create_delta_layer(
3909 968 : self: &Arc<Self>,
3910 968 : frozen_layer: &Arc<InMemoryLayer>,
3911 968 : key_range: Option<Range<Key>>,
3912 968 : ctx: &RequestContext,
3913 968 : ) -> anyhow::Result<Option<ResidentLayer>> {
3914 968 : let self_clone = Arc::clone(self);
3915 968 : let frozen_layer = Arc::clone(frozen_layer);
3916 968 : let ctx = ctx.attached_child();
3917 968 : let work = async move {
3918 968 : let Some((desc, path)) = frozen_layer
3919 968 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
3920 10250 : .await?
3921 : else {
3922 0 : return Ok(None);
3923 : };
3924 968 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
3925 :
3926 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
3927 : // We just need to fsync the directory in which these inodes are linked,
3928 : // which we know to be the timeline directory.
3929 : //
3930 : // We use fatal_err() below because the after write_to_disk returns with success,
3931 : // the in-memory state of the filesystem already has the layer file in its final place,
3932 : // and subsequent pageserver code could think it's durable while it really isn't.
3933 968 : let timeline_dir = VirtualFile::open(
3934 968 : &self_clone
3935 968 : .conf
3936 968 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
3937 968 : &ctx,
3938 968 : )
3939 490 : .await
3940 968 : .fatal_err("VirtualFile::open for timeline dir fsync");
3941 968 : timeline_dir
3942 968 : .sync_all()
3943 484 : .await
3944 968 : .fatal_err("VirtualFile::sync_all timeline dir");
3945 968 : anyhow::Ok(Some(new_delta))
3946 968 : };
3947 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
3948 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
3949 : use crate::virtual_file::io_engine::IoEngine;
3950 968 : match crate::virtual_file::io_engine::get() {
3951 0 : IoEngine::NotSet => panic!("io engine not set"),
3952 : IoEngine::StdFs => {
3953 484 : let span = tracing::info_span!("blocking");
3954 484 : tokio::task::spawn_blocking({
3955 484 : move || Handle::current().block_on(work.instrument(span))
3956 484 : })
3957 484 : .await
3958 484 : .context("spawn_blocking")
3959 484 : .and_then(|x| x)
3960 : }
3961 : #[cfg(target_os = "linux")]
3962 11219 : IoEngine::TokioEpollUring => work.await,
3963 : }
3964 968 : }
3965 :
3966 540 : async fn repartition(
3967 540 : &self,
3968 540 : lsn: Lsn,
3969 540 : partition_size: u64,
3970 540 : flags: EnumSet<CompactFlags>,
3971 540 : ctx: &RequestContext,
3972 540 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
3973 540 : let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
3974 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
3975 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
3976 : // and hence before the compaction task starts.
3977 0 : return Err(CompactionError::Other(anyhow!(
3978 0 : "repartition() called concurrently, this should not happen"
3979 0 : )));
3980 : };
3981 540 : let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
3982 540 : if lsn < *partition_lsn {
3983 0 : return Err(CompactionError::Other(anyhow!(
3984 0 : "repartition() called with LSN going backwards, this should not happen"
3985 0 : )));
3986 540 : }
3987 540 :
3988 540 : let distance = lsn.0 - partition_lsn.0;
3989 540 : if *partition_lsn != Lsn(0)
3990 262 : && distance <= self.repartition_threshold
3991 262 : && !flags.contains(CompactFlags::ForceRepartition)
3992 : {
3993 248 : debug!(
3994 : distance,
3995 : threshold = self.repartition_threshold,
3996 0 : "no repartitioning needed"
3997 : );
3998 248 : return Ok((
3999 248 : (dense_partition.clone(), sparse_partition.clone()),
4000 248 : *partition_lsn,
4001 248 : ));
4002 292 : }
4003 :
4004 15812 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4005 292 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4006 292 : let sparse_partitioning = SparseKeyPartitioning {
4007 292 : parts: vec![sparse_ks],
4008 292 : }; // no partitioning for metadata keys for now
4009 292 : *partitioning_guard = ((dense_partitioning, sparse_partitioning), lsn);
4010 292 :
4011 292 : Ok((partitioning_guard.0.clone(), partitioning_guard.1))
4012 540 : }
4013 :
4014 : // Is it time to create a new image layer for the given partition?
4015 14 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4016 14 : let threshold = self.get_image_creation_threshold();
4017 :
4018 14 : let guard = self.layers.read().await;
4019 14 : let Ok(layers) = guard.layer_map() else {
4020 0 : return false;
4021 : };
4022 :
4023 14 : let mut max_deltas = 0;
4024 28 : for part_range in &partition.ranges {
4025 14 : let image_coverage = layers.image_coverage(part_range, lsn);
4026 28 : for (img_range, last_img) in image_coverage {
4027 14 : let img_lsn = if let Some(last_img) = last_img {
4028 0 : last_img.get_lsn_range().end
4029 : } else {
4030 14 : Lsn(0)
4031 : };
4032 : // Let's consider an example:
4033 : //
4034 : // delta layer with LSN range 71-81
4035 : // delta layer with LSN range 81-91
4036 : // delta layer with LSN range 91-101
4037 : // image layer at LSN 100
4038 : //
4039 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4040 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4041 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4042 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4043 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4044 14 : if img_lsn < lsn {
4045 14 : let num_deltas =
4046 14 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4047 14 :
4048 14 : max_deltas = max_deltas.max(num_deltas);
4049 14 : if num_deltas >= threshold {
4050 0 : debug!(
4051 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4052 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4053 : );
4054 0 : return true;
4055 14 : }
4056 0 : }
4057 : }
4058 : }
4059 :
4060 14 : debug!(
4061 : max_deltas,
4062 0 : "none of the partitioned ranges had >= {threshold} deltas"
4063 : );
4064 14 : false
4065 14 : }
4066 :
4067 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4068 : /// so that at most one image layer will be produced from this function.
4069 202 : async fn create_image_layer_for_rel_blocks(
4070 202 : self: &Arc<Self>,
4071 202 : partition: &KeySpace,
4072 202 : mut image_layer_writer: ImageLayerWriter,
4073 202 : lsn: Lsn,
4074 202 : ctx: &RequestContext,
4075 202 : img_range: Range<Key>,
4076 202 : start: Key,
4077 202 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4078 202 : let mut wrote_keys = false;
4079 202 :
4080 202 : let mut key_request_accum = KeySpaceAccum::new();
4081 1344 : for range in &partition.ranges {
4082 1142 : let mut key = range.start;
4083 2472 : while key < range.end {
4084 : // Decide whether to retain this key: usually we do, but sharded tenants may
4085 : // need to drop keys that don't belong to them. If we retain the key, add it
4086 : // to `key_request_accum` for later issuing a vectored get
4087 1330 : if self.shard_identity.is_key_disposable(&key) {
4088 0 : debug!(
4089 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4090 0 : key,
4091 0 : self.shard_identity.get_shard_number(&key)
4092 : );
4093 1330 : } else {
4094 1330 : key_request_accum.add_key(key);
4095 1330 : }
4096 :
4097 1330 : let last_key_in_range = key.next() == range.end;
4098 1330 : key = key.next();
4099 1330 :
4100 1330 : // Maybe flush `key_rest_accum`
4101 1330 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4102 1330 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4103 : {
4104 1142 : let results = self
4105 1142 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
4106 49 : .await?;
4107 :
4108 1142 : if self.cancel.is_cancelled() {
4109 0 : return Err(CreateImageLayersError::Cancelled);
4110 1142 : }
4111 :
4112 2472 : for (img_key, img) in results {
4113 1330 : let img = match img {
4114 1330 : Ok(img) => img,
4115 0 : Err(err) => {
4116 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4117 0 : // page without losing any actual user data. That seems better
4118 0 : // than failing repeatedly and getting stuck.
4119 0 : //
4120 0 : // We had a bug at one point, where we truncated the FSM and VM
4121 0 : // in the pageserver, but the Postgres didn't know about that
4122 0 : // and continued to generate incremental WAL records for pages
4123 0 : // that didn't exist in the pageserver. Trying to replay those
4124 0 : // WAL records failed to find the previous image of the page.
4125 0 : // This special case allows us to recover from that situation.
4126 0 : // See https://github.com/neondatabase/neon/issues/2601.
4127 0 : //
4128 0 : // Unfortunately we cannot do this for the main fork, or for
4129 0 : // any metadata keys, keys, as that would lead to actual data
4130 0 : // loss.
4131 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4132 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4133 0 : ZERO_PAGE.clone()
4134 : } else {
4135 0 : return Err(CreateImageLayersError::from(err));
4136 : }
4137 : }
4138 : };
4139 :
4140 : // Write all the keys we just read into our new image layer.
4141 1454 : image_layer_writer.put_image(img_key, img, ctx).await?;
4142 1330 : wrote_keys = true;
4143 : }
4144 188 : }
4145 : }
4146 : }
4147 :
4148 202 : if wrote_keys {
4149 : // Normal path: we have written some data into the new image layer for this
4150 : // partition, so flush it to disk.
4151 406 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4152 202 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4153 202 : info!("created image layer for rel {}", image_layer.local_path());
4154 202 : Ok(ImageLayerCreationOutcome {
4155 202 : image: Some(image_layer),
4156 202 : next_start_key: img_range.end,
4157 202 : })
4158 : } else {
4159 : // Special case: the image layer may be empty if this is a sharded tenant and the
4160 : // partition does not cover any keys owned by this shard. In this case, to ensure
4161 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4162 : // layer we write will cover the key range that we just scanned.
4163 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4164 0 : Ok(ImageLayerCreationOutcome {
4165 0 : image: None,
4166 0 : next_start_key: start,
4167 0 : })
4168 : }
4169 202 : }
4170 :
4171 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4172 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4173 : /// would not be too large to fit in a single image layer.
4174 : #[allow(clippy::too_many_arguments)]
4175 192 : async fn create_image_layer_for_metadata_keys(
4176 192 : self: &Arc<Self>,
4177 192 : partition: &KeySpace,
4178 192 : mut image_layer_writer: ImageLayerWriter,
4179 192 : lsn: Lsn,
4180 192 : ctx: &RequestContext,
4181 192 : img_range: Range<Key>,
4182 192 : mode: ImageLayerCreationMode,
4183 192 : start: Key,
4184 192 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4185 192 : // Metadata keys image layer creation.
4186 192 : let mut reconstruct_state = ValuesReconstructState::default();
4187 192 : let begin = Instant::now();
4188 192 : let data = self
4189 192 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4190 1070 : .await?;
4191 192 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4192 192 : let mut new_data = BTreeMap::new();
4193 192 : let mut total_kb_retrieved = 0;
4194 192 : let mut total_keys_retrieved = 0;
4195 10204 : for (k, v) in data {
4196 10012 : let v = v?;
4197 10012 : total_kb_retrieved += KEY_SIZE + v.len();
4198 10012 : total_keys_retrieved += 1;
4199 10012 : new_data.insert(k, v);
4200 : }
4201 192 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4202 192 : };
4203 192 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4204 192 : let elapsed = begin.elapsed();
4205 192 :
4206 192 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4207 192 : info!(
4208 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()
4209 : );
4210 :
4211 192 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4212 2 : return Ok(ImageLayerCreationOutcome {
4213 2 : image: None,
4214 2 : next_start_key: img_range.end,
4215 2 : });
4216 190 : }
4217 190 : if self.cancel.is_cancelled() {
4218 0 : return Err(CreateImageLayersError::Cancelled);
4219 190 : }
4220 190 : let mut wrote_any_image = false;
4221 10202 : for (k, v) in data {
4222 10012 : if v.is_empty() {
4223 : // the key has been deleted, it does not need an image
4224 : // in metadata keyspace, an empty image == tombstone
4225 8 : continue;
4226 10004 : }
4227 10004 : wrote_any_image = true;
4228 10004 :
4229 10004 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4230 10004 :
4231 10004 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4232 10004 : // on the normal data path either.
4233 10160 : image_layer_writer.put_image(k, v, ctx).await?;
4234 : }
4235 :
4236 190 : if wrote_any_image {
4237 : // Normal path: we have written some data into the new image layer for this
4238 : // partition, so flush it to disk.
4239 26 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4240 12 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4241 12 : info!(
4242 0 : "created image layer for metadata {}",
4243 0 : image_layer.local_path()
4244 : );
4245 12 : Ok(ImageLayerCreationOutcome {
4246 12 : image: Some(image_layer),
4247 12 : next_start_key: img_range.end,
4248 12 : })
4249 : } else {
4250 : // Special case: the image layer may be empty if this is a sharded tenant and the
4251 : // partition does not cover any keys owned by this shard. In this case, to ensure
4252 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4253 : // layer we write will cover the key range that we just scanned.
4254 178 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4255 178 : Ok(ImageLayerCreationOutcome {
4256 178 : image: None,
4257 178 : next_start_key: start,
4258 178 : })
4259 : }
4260 192 : }
4261 :
4262 : /// Predicate function which indicates whether we should check if new image layers
4263 : /// are required. Since checking if new image layers are required is expensive in
4264 : /// terms of CPU, we only do it in the following cases:
4265 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4266 : /// 2. If enough time has passed since the last check:
4267 : /// 1. For large tenants, we wish to perform the check more often since they
4268 : /// suffer from the lack of image layers
4269 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4270 716 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4271 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4272 :
4273 716 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4274 716 : let distance = lsn
4275 716 : .checked_sub(last_checks_at)
4276 716 : .expect("Attempt to compact with LSN going backwards");
4277 716 : let min_distance =
4278 716 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4279 716 :
4280 716 : let distance_based_decision = distance.0 >= min_distance;
4281 716 :
4282 716 : let mut time_based_decision = false;
4283 716 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4284 716 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4285 614 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4286 : {
4287 0 : self.get_checkpoint_timeout()
4288 : } else {
4289 614 : Duration::from_secs(3600 * 48)
4290 : };
4291 :
4292 614 : time_based_decision = match *last_check_instant {
4293 438 : Some(last_check) => {
4294 438 : let elapsed = last_check.elapsed();
4295 438 : elapsed >= check_required_after
4296 : }
4297 176 : None => true,
4298 : };
4299 102 : }
4300 :
4301 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4302 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4303 : // check.
4304 716 : let decision = distance_based_decision || time_based_decision;
4305 :
4306 716 : if decision {
4307 178 : self.last_image_layer_creation_check_at.store(lsn);
4308 178 : *last_check_instant = Some(Instant::now());
4309 538 : }
4310 :
4311 716 : decision
4312 716 : }
4313 :
4314 716 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4315 : async fn create_image_layers(
4316 : self: &Arc<Timeline>,
4317 : partitioning: &KeyPartitioning,
4318 : lsn: Lsn,
4319 : mode: ImageLayerCreationMode,
4320 : ctx: &RequestContext,
4321 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4322 : let timer = self.metrics.create_images_time_histo.start_timer();
4323 : let mut image_layers = Vec::new();
4324 :
4325 : // We need to avoid holes between generated image layers.
4326 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4327 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4328 : //
4329 : // How such hole between partitions can appear?
4330 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4331 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4332 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4333 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4334 : let mut start = Key::MIN;
4335 :
4336 : let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
4337 :
4338 : for partition in partitioning.parts.iter() {
4339 : if self.cancel.is_cancelled() {
4340 : return Err(CreateImageLayersError::Cancelled);
4341 : }
4342 :
4343 : let img_range = start..partition.ranges.last().unwrap().end;
4344 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4345 : if compact_metadata {
4346 : for range in &partition.ranges {
4347 : assert!(
4348 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4349 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4350 : "metadata keys must be partitioned separately"
4351 : );
4352 : }
4353 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4354 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4355 : // might mess up with evictions.
4356 : start = img_range.end;
4357 : continue;
4358 : }
4359 : // For initial and force modes, we always generate image layers for metadata keys.
4360 : } else if let ImageLayerCreationMode::Try = mode {
4361 : // check_for_image_layers = false -> skip
4362 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4363 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4364 : start = img_range.end;
4365 : continue;
4366 : }
4367 : }
4368 : if let ImageLayerCreationMode::Force = mode {
4369 : // When forced to create image layers, we might try and create them where they already
4370 : // exist. This mode is only used in tests/debug.
4371 : let layers = self.layers.read().await;
4372 : if layers.contains_key(&PersistentLayerKey {
4373 : key_range: img_range.clone(),
4374 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4375 : is_delta: false,
4376 : }) {
4377 : tracing::info!(
4378 : "Skipping image layer at {lsn} {}..{}, already exists",
4379 : img_range.start,
4380 : img_range.end
4381 : );
4382 : start = img_range.end;
4383 : continue;
4384 : }
4385 : }
4386 :
4387 : let image_layer_writer = ImageLayerWriter::new(
4388 : self.conf,
4389 : self.timeline_id,
4390 : self.tenant_shard_id,
4391 : &img_range,
4392 : lsn,
4393 : ctx,
4394 : )
4395 : .await?;
4396 :
4397 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4398 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4399 0 : "failpoint image-layer-writer-fail-before-finish"
4400 0 : )))
4401 0 : });
4402 :
4403 : if !compact_metadata {
4404 : let ImageLayerCreationOutcome {
4405 : image,
4406 : next_start_key,
4407 : } = self
4408 : .create_image_layer_for_rel_blocks(
4409 : partition,
4410 : image_layer_writer,
4411 : lsn,
4412 : ctx,
4413 : img_range,
4414 : start,
4415 : )
4416 : .await?;
4417 :
4418 : start = next_start_key;
4419 : image_layers.extend(image);
4420 : } else {
4421 : let ImageLayerCreationOutcome {
4422 : image,
4423 : next_start_key,
4424 : } = self
4425 : .create_image_layer_for_metadata_keys(
4426 : partition,
4427 : image_layer_writer,
4428 : lsn,
4429 : ctx,
4430 : img_range,
4431 : mode,
4432 : start,
4433 : )
4434 : .await?;
4435 : start = next_start_key;
4436 : image_layers.extend(image);
4437 : }
4438 : }
4439 :
4440 : let mut guard = self.layers.write().await;
4441 :
4442 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4443 : // now they are being scheduled outside of write lock; current way is inconsistent with
4444 : // compaction lock order.
4445 : guard
4446 : .open_mut()?
4447 : .track_new_image_layers(&image_layers, &self.metrics);
4448 : drop_wlock(guard);
4449 : timer.stop_and_record();
4450 :
4451 : // Creating image layers may have caused some previously visible layers to be covered
4452 : if !image_layers.is_empty() {
4453 : self.update_layer_visibility().await?;
4454 : }
4455 :
4456 : Ok(image_layers)
4457 : }
4458 :
4459 : /// Wait until the background initial logical size calculation is complete, or
4460 : /// this Timeline is shut down. Calling this function will cause the initial
4461 : /// logical size calculation to skip waiting for the background jobs barrier.
4462 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4463 0 : if !self.shard_identity.is_shard_zero() {
4464 : // We don't populate logical size on shard >0: skip waiting for it.
4465 0 : return;
4466 0 : }
4467 0 :
4468 0 : if self.remote_client.is_deleting() {
4469 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4470 0 : return;
4471 0 : }
4472 0 :
4473 0 : if self.current_logical_size.current_size().is_exact() {
4474 : // root timelines are initialized with exact count, but never start the background
4475 : // calculation
4476 0 : return;
4477 0 : }
4478 :
4479 0 : if let Some(await_bg_cancel) = self
4480 0 : .current_logical_size
4481 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4482 0 : .get()
4483 0 : {
4484 0 : await_bg_cancel.cancel();
4485 0 : } else {
4486 : // We should not wait if we were not able to explicitly instruct
4487 : // the logical size cancellation to skip the concurrency limit semaphore.
4488 : // TODO: this is an unexpected case. We should restructure so that it
4489 : // can't happen.
4490 0 : tracing::warn!(
4491 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4492 : );
4493 0 : debug_assert!(false);
4494 : }
4495 :
4496 0 : tokio::select!(
4497 0 : _ = self.current_logical_size.initialized.acquire() => {},
4498 0 : _ = self.cancel.cancelled() => {}
4499 : )
4500 0 : }
4501 :
4502 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4503 : /// Timelines layers up to the ancestor_lsn.
4504 : ///
4505 : /// Requires a timeline that:
4506 : /// - has an ancestor to detach from
4507 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4508 : /// a technical requirement
4509 : ///
4510 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4511 : /// polled again until completion.
4512 : ///
4513 : /// During the operation all timelines sharing the data with this timeline will be reparented
4514 : /// from our ancestor to be branches of this timeline.
4515 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4516 0 : self: &Arc<Timeline>,
4517 0 : tenant: &crate::tenant::Tenant,
4518 0 : options: detach_ancestor::Options,
4519 0 : ctx: &RequestContext,
4520 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
4521 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4522 0 : }
4523 :
4524 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
4525 : /// reparents any reparentable children of previous ancestor.
4526 : ///
4527 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
4528 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
4529 : /// successfully, tenant must be reloaded.
4530 : ///
4531 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
4532 : /// resetting the tenant.
4533 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
4534 0 : self: &Arc<Timeline>,
4535 0 : tenant: &crate::tenant::Tenant,
4536 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4537 0 : ctx: &RequestContext,
4538 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
4539 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
4540 0 : }
4541 :
4542 : /// Final step which unblocks the GC.
4543 : ///
4544 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
4545 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4546 0 : self: &Arc<Timeline>,
4547 0 : tenant: &crate::tenant::Tenant,
4548 0 : attempt: detach_ancestor::Attempt,
4549 0 : ctx: &RequestContext,
4550 0 : ) -> Result<(), detach_ancestor::Error> {
4551 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
4552 0 : }
4553 : }
4554 :
4555 : impl Drop for Timeline {
4556 10 : fn drop(&mut self) {
4557 10 : if let Some(ancestor) = &self.ancestor_timeline {
4558 : // This lock should never be poisoned, but in case it is we do a .map() instead of
4559 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
4560 4 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
4561 4 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
4562 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
4563 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
4564 4 : }
4565 0 : }
4566 6 : }
4567 10 : }
4568 : }
4569 :
4570 : /// Top-level failure to compact.
4571 0 : #[derive(Debug, thiserror::Error)]
4572 : pub(crate) enum CompactionError {
4573 : #[error("The timeline or pageserver is shutting down")]
4574 : ShuttingDown,
4575 : /// Compaction tried to offload a timeline and failed
4576 : #[error("Failed to offload timeline: {0}")]
4577 : Offload(OffloadError),
4578 : /// Compaction cannot be done right now; page reconstruction and so on.
4579 : #[error(transparent)]
4580 : Other(anyhow::Error),
4581 : }
4582 :
4583 : impl From<OffloadError> for CompactionError {
4584 0 : fn from(e: OffloadError) -> Self {
4585 0 : match e {
4586 0 : OffloadError::Cancelled => Self::ShuttingDown,
4587 0 : _ => Self::Offload(e),
4588 : }
4589 0 : }
4590 : }
4591 :
4592 : impl CompactionError {
4593 0 : pub fn is_cancelled(&self) -> bool {
4594 0 : matches!(self, CompactionError::ShuttingDown)
4595 0 : }
4596 : }
4597 :
4598 : impl From<CollectKeySpaceError> for CompactionError {
4599 0 : fn from(err: CollectKeySpaceError) -> Self {
4600 0 : match err {
4601 : CollectKeySpaceError::Cancelled
4602 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4603 0 : CompactionError::ShuttingDown
4604 : }
4605 0 : e => CompactionError::Other(e.into()),
4606 : }
4607 0 : }
4608 : }
4609 :
4610 : impl From<super::upload_queue::NotInitialized> for CompactionError {
4611 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
4612 0 : match value {
4613 : super::upload_queue::NotInitialized::Uninitialized => {
4614 0 : CompactionError::Other(anyhow::anyhow!(value))
4615 : }
4616 : super::upload_queue::NotInitialized::ShuttingDown
4617 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
4618 : }
4619 0 : }
4620 : }
4621 :
4622 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
4623 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
4624 0 : match e {
4625 : super::storage_layer::layer::DownloadError::TimelineShutdown
4626 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
4627 0 : CompactionError::ShuttingDown
4628 : }
4629 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
4630 : | super::storage_layer::layer::DownloadError::DownloadRequired
4631 : | super::storage_layer::layer::DownloadError::NotFile(_)
4632 : | super::storage_layer::layer::DownloadError::DownloadFailed
4633 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
4634 0 : CompactionError::Other(anyhow::anyhow!(e))
4635 : }
4636 : #[cfg(test)]
4637 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
4638 0 : CompactionError::Other(anyhow::anyhow!(e))
4639 : }
4640 : }
4641 0 : }
4642 : }
4643 :
4644 : impl From<layer_manager::Shutdown> for CompactionError {
4645 0 : fn from(_: layer_manager::Shutdown) -> Self {
4646 0 : CompactionError::ShuttingDown
4647 0 : }
4648 : }
4649 :
4650 : #[serde_as]
4651 196 : #[derive(serde::Serialize)]
4652 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4653 :
4654 : #[derive(Default)]
4655 : enum DurationRecorder {
4656 : #[default]
4657 : NotStarted,
4658 : Recorded(RecordedDuration, tokio::time::Instant),
4659 : }
4660 :
4661 : impl DurationRecorder {
4662 504 : fn till_now(&self) -> DurationRecorder {
4663 504 : match self {
4664 : DurationRecorder::NotStarted => {
4665 0 : panic!("must only call on recorded measurements")
4666 : }
4667 504 : DurationRecorder::Recorded(_, ended) => {
4668 504 : let now = tokio::time::Instant::now();
4669 504 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4670 504 : }
4671 504 : }
4672 504 : }
4673 196 : fn into_recorded(self) -> Option<RecordedDuration> {
4674 196 : match self {
4675 0 : DurationRecorder::NotStarted => None,
4676 196 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4677 : }
4678 196 : }
4679 : }
4680 :
4681 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
4682 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
4683 : /// the layer descriptor requires the user to provide the ranges, which should cover all
4684 : /// keys specified in the `data` field.
4685 : #[cfg(test)]
4686 : #[derive(Clone)]
4687 : pub struct DeltaLayerTestDesc {
4688 : pub lsn_range: Range<Lsn>,
4689 : pub key_range: Range<Key>,
4690 : pub data: Vec<(Key, Lsn, Value)>,
4691 : }
4692 :
4693 : #[cfg(test)]
4694 : impl DeltaLayerTestDesc {
4695 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
4696 2 : Self {
4697 2 : lsn_range,
4698 2 : key_range,
4699 2 : data,
4700 2 : }
4701 2 : }
4702 :
4703 72 : pub fn new_with_inferred_key_range(
4704 72 : lsn_range: Range<Lsn>,
4705 72 : data: Vec<(Key, Lsn, Value)>,
4706 72 : ) -> Self {
4707 196 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
4708 196 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
4709 72 : Self {
4710 72 : key_range: (*key_min)..(key_max.next()),
4711 72 : lsn_range,
4712 72 : data,
4713 72 : }
4714 72 : }
4715 :
4716 10 : pub(crate) fn layer_name(&self) -> LayerName {
4717 10 : LayerName::Delta(super::storage_layer::DeltaLayerName {
4718 10 : key_range: self.key_range.clone(),
4719 10 : lsn_range: self.lsn_range.clone(),
4720 10 : })
4721 10 : }
4722 : }
4723 :
4724 : impl Timeline {
4725 28 : async fn finish_compact_batch(
4726 28 : self: &Arc<Self>,
4727 28 : new_deltas: &[ResidentLayer],
4728 28 : new_images: &[ResidentLayer],
4729 28 : layers_to_remove: &[Layer],
4730 28 : ) -> Result<(), CompactionError> {
4731 28 : let mut guard = tokio::select! {
4732 28 : guard = self.layers.write() => guard,
4733 28 : _ = self.cancel.cancelled() => {
4734 0 : return Err(CompactionError::ShuttingDown);
4735 : }
4736 : };
4737 :
4738 28 : let mut duplicated_layers = HashSet::new();
4739 28 :
4740 28 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
4741 :
4742 336 : for l in new_deltas {
4743 308 : if guard.contains(l.as_ref()) {
4744 : // expected in tests
4745 0 : tracing::error!(layer=%l, "duplicated L1 layer");
4746 :
4747 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
4748 : // `guard` on self.layers. as of writing this, there are no error returns except
4749 : // for compact_level0_phase1 creating an L0, which does not happen in practice
4750 : // because we have not implemented L0 => L0 compaction.
4751 0 : duplicated_layers.insert(l.layer_desc().key());
4752 308 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
4753 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
4754 308 : } else {
4755 308 : insert_layers.push(l.clone());
4756 308 : }
4757 : }
4758 :
4759 : // only remove those inputs which were not outputs
4760 28 : let remove_layers: Vec<Layer> = layers_to_remove
4761 28 : .iter()
4762 402 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
4763 28 : .cloned()
4764 28 : .collect();
4765 28 :
4766 28 : if !new_images.is_empty() {
4767 0 : guard
4768 0 : .open_mut()?
4769 0 : .track_new_image_layers(new_images, &self.metrics);
4770 28 : }
4771 :
4772 28 : guard
4773 28 : .open_mut()?
4774 28 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
4775 28 :
4776 28 : self.remote_client
4777 28 : .schedule_compaction_update(&remove_layers, new_deltas)?;
4778 :
4779 28 : drop_wlock(guard);
4780 28 :
4781 28 : Ok(())
4782 28 : }
4783 :
4784 0 : async fn rewrite_layers(
4785 0 : self: &Arc<Self>,
4786 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
4787 0 : mut drop_layers: Vec<Layer>,
4788 0 : ) -> Result<(), CompactionError> {
4789 0 : let mut guard = self.layers.write().await;
4790 :
4791 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
4792 : // to avoid double-removing, and avoid rewriting something that was removed.
4793 0 : replace_layers.retain(|(l, _)| guard.contains(l));
4794 0 : drop_layers.retain(|l| guard.contains(l));
4795 0 :
4796 0 : guard
4797 0 : .open_mut()?
4798 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
4799 0 :
4800 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
4801 0 :
4802 0 : self.remote_client
4803 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
4804 :
4805 0 : Ok(())
4806 0 : }
4807 :
4808 : /// Schedules the uploads of the given image layers
4809 364 : fn upload_new_image_layers(
4810 364 : self: &Arc<Self>,
4811 364 : new_images: impl IntoIterator<Item = ResidentLayer>,
4812 364 : ) -> Result<(), super::upload_queue::NotInitialized> {
4813 390 : for layer in new_images {
4814 26 : self.remote_client.schedule_layer_file_upload(layer)?;
4815 : }
4816 : // should any new image layer been created, not uploading index_part will
4817 : // result in a mismatch between remote_physical_size and layermap calculated
4818 : // size, which will fail some tests, but should not be an issue otherwise.
4819 364 : self.remote_client
4820 364 : .schedule_index_upload_for_file_changes()?;
4821 364 : Ok(())
4822 364 : }
4823 :
4824 : /// Find the Lsns above which layer files need to be retained on
4825 : /// garbage collection.
4826 : ///
4827 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
4828 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
4829 : /// the space-based retention.
4830 : ///
4831 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
4832 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
4833 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
4834 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
4835 : /// in the response as the GC cutoff point for the timeline.
4836 4 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
4837 : pub(super) async fn find_gc_cutoffs(
4838 : &self,
4839 : space_cutoff: Lsn,
4840 : pitr: Duration,
4841 : cancel: &CancellationToken,
4842 : ctx: &RequestContext,
4843 : ) -> Result<GcCutoffs, PageReconstructError> {
4844 : let _timer = self
4845 : .metrics
4846 : .find_gc_cutoffs_histo
4847 : .start_timer()
4848 : .record_on_drop();
4849 :
4850 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
4851 :
4852 : if cfg!(test) {
4853 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
4854 : if pitr == Duration::ZERO {
4855 : return Ok(GcCutoffs {
4856 : time: self.get_last_record_lsn(),
4857 : space: space_cutoff,
4858 : });
4859 : }
4860 : }
4861 :
4862 : // Calculate a time-based limit on how much to retain:
4863 : // - if PITR interval is set, then this is our cutoff.
4864 : // - if PITR interval is not set, then we do a lookup
4865 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
4866 : let time_cutoff = {
4867 : let now = SystemTime::now();
4868 : let time_range = if pitr == Duration::ZERO {
4869 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
4870 : } else {
4871 : pitr
4872 : };
4873 :
4874 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
4875 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
4876 : let timestamp = to_pg_timestamp(time_cutoff);
4877 :
4878 : match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
4879 : LsnForTimestamp::Present(lsn) => Some(lsn),
4880 : LsnForTimestamp::Future(lsn) => {
4881 : // The timestamp is in the future. That sounds impossible,
4882 : // but what it really means is that there hasn't been
4883 : // any commits since the cutoff timestamp.
4884 : //
4885 : // In this case we should use the LSN of the most recent commit,
4886 : // which is implicitly the last LSN in the log.
4887 : debug!("future({})", lsn);
4888 : Some(self.get_last_record_lsn())
4889 : }
4890 : LsnForTimestamp::Past(lsn) => {
4891 : debug!("past({})", lsn);
4892 : None
4893 : }
4894 : LsnForTimestamp::NoData(lsn) => {
4895 : debug!("nodata({})", lsn);
4896 : None
4897 : }
4898 : }
4899 : };
4900 :
4901 : Ok(match (pitr, time_cutoff) {
4902 : (Duration::ZERO, Some(time_cutoff)) => {
4903 : // PITR is not set. Retain the size-based limit, or the default time retention,
4904 : // whichever requires less data.
4905 : GcCutoffs {
4906 : time: self.get_last_record_lsn(),
4907 : space: std::cmp::max(time_cutoff, space_cutoff),
4908 : }
4909 : }
4910 : (Duration::ZERO, None) => {
4911 : // PITR is not set, and time lookup failed
4912 : GcCutoffs {
4913 : time: self.get_last_record_lsn(),
4914 : space: space_cutoff,
4915 : }
4916 : }
4917 : (_, None) => {
4918 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
4919 : // cannot advance beyond what was already GC'd, and respect space-based retention
4920 : GcCutoffs {
4921 : time: *self.get_latest_gc_cutoff_lsn(),
4922 : space: space_cutoff,
4923 : }
4924 : }
4925 : (_, Some(time_cutoff)) => {
4926 : // PITR interval is set and we looked up timestamp successfully. Ignore
4927 : // size based retention and make time cutoff authoritative
4928 : GcCutoffs {
4929 : time: time_cutoff,
4930 : space: time_cutoff,
4931 : }
4932 : }
4933 : })
4934 : }
4935 :
4936 : /// Garbage collect layer files on a timeline that are no longer needed.
4937 : ///
4938 : /// Currently, we don't make any attempt at removing unneeded page versions
4939 : /// within a layer file. We can only remove the whole file if it's fully
4940 : /// obsolete.
4941 4 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
4942 : // this is most likely the background tasks, but it might be the spawned task from
4943 : // immediate_gc
4944 4 : let _g = tokio::select! {
4945 4 : guard = self.gc_lock.lock() => guard,
4946 4 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
4947 : };
4948 4 : let timer = self.metrics.garbage_collect_histo.start_timer();
4949 4 :
4950 4 : fail_point!("before-timeline-gc");
4951 4 :
4952 4 : // Is the timeline being deleted?
4953 4 : if self.is_stopping() {
4954 0 : return Err(GcError::TimelineCancelled);
4955 4 : }
4956 4 :
4957 4 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
4958 4 : let gc_info = self.gc_info.read().unwrap();
4959 4 :
4960 4 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
4961 4 : let time_cutoff = gc_info.cutoffs.time;
4962 4 : let retain_lsns = gc_info
4963 4 : .retain_lsns
4964 4 : .iter()
4965 4 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
4966 4 : .collect();
4967 4 :
4968 4 : // Gets the maximum LSN that holds the valid lease.
4969 4 : //
4970 4 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
4971 4 : // Here, we do not check for stale leases again.
4972 4 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
4973 4 :
4974 4 : (
4975 4 : space_cutoff,
4976 4 : time_cutoff,
4977 4 : retain_lsns,
4978 4 : max_lsn_with_valid_lease,
4979 4 : )
4980 4 : };
4981 4 :
4982 4 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
4983 4 : let standby_horizon = self.standby_horizon.load();
4984 4 : // Hold GC for the standby, but as a safety guard do it only within some
4985 4 : // reasonable lag.
4986 4 : if standby_horizon != Lsn::INVALID {
4987 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
4988 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
4989 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
4990 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
4991 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
4992 : } else {
4993 0 : warn!(
4994 0 : "standby is lagging for more than {}MB, not holding gc for it",
4995 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
4996 : )
4997 : }
4998 0 : }
4999 4 : }
5000 :
5001 : // Reset standby horizon to ignore it if it is not updated till next GC.
5002 : // It is an easy way to unset it when standby disappears without adding
5003 : // more conf options.
5004 4 : self.standby_horizon.store(Lsn::INVALID);
5005 4 : self.metrics
5006 4 : .standby_horizon_gauge
5007 4 : .set(Lsn::INVALID.0 as i64);
5008 :
5009 4 : let res = self
5010 4 : .gc_timeline(
5011 4 : space_cutoff,
5012 4 : time_cutoff,
5013 4 : retain_lsns,
5014 4 : max_lsn_with_valid_lease,
5015 4 : new_gc_cutoff,
5016 4 : )
5017 4 : .instrument(
5018 4 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5019 : )
5020 0 : .await?;
5021 :
5022 : // only record successes
5023 4 : timer.stop_and_record();
5024 4 :
5025 4 : Ok(res)
5026 4 : }
5027 :
5028 4 : async fn gc_timeline(
5029 4 : &self,
5030 4 : space_cutoff: Lsn,
5031 4 : time_cutoff: Lsn,
5032 4 : retain_lsns: Vec<Lsn>,
5033 4 : max_lsn_with_valid_lease: Option<Lsn>,
5034 4 : new_gc_cutoff: Lsn,
5035 4 : ) -> Result<GcResult, GcError> {
5036 4 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5037 4 :
5038 4 : let now = SystemTime::now();
5039 4 : let mut result: GcResult = GcResult::default();
5040 4 :
5041 4 : // Nothing to GC. Return early.
5042 4 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
5043 4 : if latest_gc_cutoff >= new_gc_cutoff {
5044 0 : info!(
5045 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5046 : );
5047 0 : return Ok(result);
5048 4 : }
5049 :
5050 : // We need to ensure that no one tries to read page versions or create
5051 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5052 : // for details. This will block until the old value is no longer in use.
5053 : //
5054 : // The GC cutoff should only ever move forwards.
5055 4 : let waitlist = {
5056 4 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
5057 4 : if *write_guard > new_gc_cutoff {
5058 0 : return Err(GcError::BadLsn {
5059 0 : why: format!(
5060 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5061 0 : *write_guard, new_gc_cutoff
5062 0 : ),
5063 0 : });
5064 4 : }
5065 4 :
5066 4 : write_guard.store_and_unlock(new_gc_cutoff)
5067 4 : };
5068 4 : waitlist.wait().await;
5069 :
5070 4 : info!("GC starting");
5071 :
5072 4 : debug!("retain_lsns: {:?}", retain_lsns);
5073 :
5074 4 : let mut layers_to_remove = Vec::new();
5075 :
5076 : // Scan all layers in the timeline (remote or on-disk).
5077 : //
5078 : // Garbage collect the layer if all conditions are satisfied:
5079 : // 1. it is older than cutoff LSN;
5080 : // 2. it is older than PITR interval;
5081 : // 3. it doesn't need to be retained for 'retain_lsns';
5082 : // 4. it does not need to be kept for LSNs holding valid leases.
5083 : // 5. newer on-disk image layers cover the layer's whole key range
5084 : //
5085 : // TODO holding a write lock is too agressive and avoidable
5086 4 : let mut guard = self.layers.write().await;
5087 4 : let layers = guard.layer_map()?;
5088 24 : 'outer: for l in layers.iter_historic_layers() {
5089 24 : result.layers_total += 1;
5090 24 :
5091 24 : // 1. Is it newer than GC horizon cutoff point?
5092 24 : if l.get_lsn_range().end > space_cutoff {
5093 2 : info!(
5094 0 : "keeping {} because it's newer than space_cutoff {}",
5095 0 : l.layer_name(),
5096 : space_cutoff,
5097 : );
5098 2 : result.layers_needed_by_cutoff += 1;
5099 2 : continue 'outer;
5100 22 : }
5101 22 :
5102 22 : // 2. It is newer than PiTR cutoff point?
5103 22 : if l.get_lsn_range().end > time_cutoff {
5104 0 : info!(
5105 0 : "keeping {} because it's newer than time_cutoff {}",
5106 0 : l.layer_name(),
5107 : time_cutoff,
5108 : );
5109 0 : result.layers_needed_by_pitr += 1;
5110 0 : continue 'outer;
5111 22 : }
5112 :
5113 : // 3. Is it needed by a child branch?
5114 : // NOTE With that we would keep data that
5115 : // might be referenced by child branches forever.
5116 : // We can track this in child timeline GC and delete parent layers when
5117 : // they are no longer needed. This might be complicated with long inheritance chains.
5118 : //
5119 : // TODO Vec is not a great choice for `retain_lsns`
5120 22 : for retain_lsn in &retain_lsns {
5121 : // start_lsn is inclusive
5122 0 : if &l.get_lsn_range().start <= retain_lsn {
5123 0 : info!(
5124 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5125 0 : l.layer_name(),
5126 0 : retain_lsn,
5127 0 : l.is_incremental(),
5128 : );
5129 0 : result.layers_needed_by_branches += 1;
5130 0 : continue 'outer;
5131 0 : }
5132 : }
5133 :
5134 : // 4. Is there a valid lease that requires us to keep this layer?
5135 22 : if let Some(lsn) = &max_lsn_with_valid_lease {
5136 : // keep if layer start <= any of the lease
5137 18 : if &l.get_lsn_range().start <= lsn {
5138 14 : info!(
5139 0 : "keeping {} because there is a valid lease preventing GC at {}",
5140 0 : l.layer_name(),
5141 : lsn,
5142 : );
5143 14 : result.layers_needed_by_leases += 1;
5144 14 : continue 'outer;
5145 4 : }
5146 4 : }
5147 :
5148 : // 5. Is there a later on-disk layer for this relation?
5149 : //
5150 : // The end-LSN is exclusive, while disk_consistent_lsn is
5151 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5152 : // OK for a delta layer to have end LSN 101, but if the end LSN
5153 : // is 102, then it might not have been fully flushed to disk
5154 : // before crash.
5155 : //
5156 : // For example, imagine that the following layers exist:
5157 : //
5158 : // 1000 - image (A)
5159 : // 1000-2000 - delta (B)
5160 : // 2000 - image (C)
5161 : // 2000-3000 - delta (D)
5162 : // 3000 - image (E)
5163 : //
5164 : // If GC horizon is at 2500, we can remove layers A and B, but
5165 : // we cannot remove C, even though it's older than 2500, because
5166 : // the delta layer 2000-3000 depends on it.
5167 8 : if !layers
5168 8 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5169 : {
5170 6 : info!("keeping {} because it is the latest layer", l.layer_name());
5171 6 : result.layers_not_updated += 1;
5172 6 : continue 'outer;
5173 2 : }
5174 2 :
5175 2 : // We didn't find any reason to keep this file, so remove it.
5176 2 : info!(
5177 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5178 0 : l.layer_name(),
5179 0 : l.is_incremental(),
5180 : );
5181 2 : layers_to_remove.push(l);
5182 : }
5183 :
5184 4 : if !layers_to_remove.is_empty() {
5185 : // Persist the new GC cutoff value before we actually remove anything.
5186 : // This unconditionally schedules also an index_part.json update, even though, we will
5187 : // be doing one a bit later with the unlinked gc'd layers.
5188 2 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5189 2 : self.schedule_uploads(disk_consistent_lsn, None)
5190 2 : .map_err(|e| {
5191 0 : if self.cancel.is_cancelled() {
5192 0 : GcError::TimelineCancelled
5193 : } else {
5194 0 : GcError::Remote(e)
5195 : }
5196 2 : })?;
5197 :
5198 2 : let gc_layers = layers_to_remove
5199 2 : .iter()
5200 2 : .map(|x| guard.get_from_desc(x))
5201 2 : .collect::<Vec<Layer>>();
5202 2 :
5203 2 : result.layers_removed = gc_layers.len() as u64;
5204 2 :
5205 2 : self.remote_client.schedule_gc_update(&gc_layers)?;
5206 :
5207 2 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5208 2 :
5209 2 : #[cfg(feature = "testing")]
5210 2 : {
5211 2 : result.doomed_layers = gc_layers;
5212 2 : }
5213 2 : }
5214 :
5215 4 : info!(
5216 0 : "GC completed removing {} layers, cutoff {}",
5217 : result.layers_removed, new_gc_cutoff
5218 : );
5219 :
5220 4 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5221 4 : Ok(result)
5222 4 : }
5223 :
5224 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5225 667436 : async fn reconstruct_value(
5226 667436 : &self,
5227 667436 : key: Key,
5228 667436 : request_lsn: Lsn,
5229 667436 : mut data: ValueReconstructState,
5230 667436 : ) -> Result<Bytes, PageReconstructError> {
5231 667436 : // Perform WAL redo if needed
5232 667436 : data.records.reverse();
5233 667436 :
5234 667436 : // If we have a page image, and no WAL, we're all set
5235 667436 : if data.records.is_empty() {
5236 667026 : if let Some((img_lsn, img)) = &data.img {
5237 667026 : trace!(
5238 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5239 : key,
5240 : img_lsn,
5241 : request_lsn,
5242 : );
5243 667026 : Ok(img.clone())
5244 : } else {
5245 0 : Err(PageReconstructError::from(anyhow!(
5246 0 : "base image for {key} at {request_lsn} not found"
5247 0 : )))
5248 : }
5249 : } else {
5250 : // We need to do WAL redo.
5251 : //
5252 : // If we don't have a base image, then the oldest WAL record better initialize
5253 : // the page
5254 410 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5255 0 : Err(PageReconstructError::from(anyhow!(
5256 0 : "Base image for {} at {} not found, but got {} WAL records",
5257 0 : key,
5258 0 : request_lsn,
5259 0 : data.records.len()
5260 0 : )))
5261 : } else {
5262 410 : if data.img.is_some() {
5263 344 : trace!(
5264 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5265 0 : data.records.len(),
5266 : key,
5267 : request_lsn
5268 : );
5269 : } else {
5270 66 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5271 : };
5272 410 : let res = self
5273 410 : .walredo_mgr
5274 410 : .as_ref()
5275 410 : .context("timeline has no walredo manager")
5276 410 : .map_err(PageReconstructError::WalRedo)?
5277 410 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5278 0 : .await;
5279 410 : let img = match res {
5280 410 : Ok(img) => img,
5281 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5282 0 : Err(walredo::Error::Other(e)) => {
5283 0 : return Err(PageReconstructError::WalRedo(
5284 0 : e.context("reconstruct a page image"),
5285 0 : ))
5286 : }
5287 : };
5288 410 : Ok(img)
5289 : }
5290 : }
5291 667436 : }
5292 :
5293 0 : pub(crate) async fn spawn_download_all_remote_layers(
5294 0 : self: Arc<Self>,
5295 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5296 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5297 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5298 :
5299 : // this is not really needed anymore; it has tests which really check the return value from
5300 : // http api. it would be better not to maintain this anymore.
5301 :
5302 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5303 0 : if let Some(st) = &*status_guard {
5304 0 : match &st.state {
5305 : DownloadRemoteLayersTaskState::Running => {
5306 0 : return Err(st.clone());
5307 : }
5308 : DownloadRemoteLayersTaskState::ShutDown
5309 0 : | DownloadRemoteLayersTaskState::Completed => {
5310 0 : *status_guard = None;
5311 0 : }
5312 : }
5313 0 : }
5314 :
5315 0 : let self_clone = Arc::clone(&self);
5316 0 : let task_id = task_mgr::spawn(
5317 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5318 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5319 0 : self.tenant_shard_id,
5320 0 : Some(self.timeline_id),
5321 0 : "download all remote layers task",
5322 0 : async move {
5323 0 : self_clone.download_all_remote_layers(request).await;
5324 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5325 0 : match &mut *status_guard {
5326 : None => {
5327 0 : warn!("tasks status is supposed to be Some(), since we are running");
5328 : }
5329 0 : Some(st) => {
5330 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5331 0 : if st.task_id != exp_task_id {
5332 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5333 0 : } else {
5334 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5335 0 : }
5336 : }
5337 : };
5338 0 : Ok(())
5339 0 : }
5340 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))
5341 : );
5342 :
5343 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5344 0 : task_id: format!("{task_id}"),
5345 0 : state: DownloadRemoteLayersTaskState::Running,
5346 0 : total_layer_count: 0,
5347 0 : successful_download_count: 0,
5348 0 : failed_download_count: 0,
5349 0 : };
5350 0 : *status_guard = Some(initial_info.clone());
5351 0 :
5352 0 : Ok(initial_info)
5353 0 : }
5354 :
5355 0 : async fn download_all_remote_layers(
5356 0 : self: &Arc<Self>,
5357 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5358 0 : ) {
5359 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5360 :
5361 0 : let remaining = {
5362 0 : let guard = self.layers.read().await;
5363 0 : let Ok(lm) = guard.layer_map() else {
5364 : // technically here we could look into iterating accessible layers, but downloading
5365 : // all layers of a shutdown timeline makes no sense regardless.
5366 0 : tracing::info!("attempted to download all layers of shutdown timeline");
5367 0 : return;
5368 : };
5369 0 : lm.iter_historic_layers()
5370 0 : .map(|desc| guard.get_from_desc(&desc))
5371 0 : .collect::<Vec<_>>()
5372 0 : };
5373 0 : let total_layer_count = remaining.len();
5374 :
5375 : macro_rules! lock_status {
5376 : ($st:ident) => {
5377 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5378 : let st = st
5379 : .as_mut()
5380 : .expect("this function is only called after the task has been spawned");
5381 : assert_eq!(
5382 : st.task_id,
5383 : format!(
5384 : "{}",
5385 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5386 : )
5387 : );
5388 : let $st = st;
5389 : };
5390 : }
5391 :
5392 : {
5393 0 : lock_status!(st);
5394 0 : st.total_layer_count = total_layer_count as u64;
5395 0 : }
5396 0 :
5397 0 : let mut remaining = remaining.into_iter();
5398 0 : let mut have_remaining = true;
5399 0 : let mut js = tokio::task::JoinSet::new();
5400 0 :
5401 0 : let cancel = task_mgr::shutdown_token();
5402 0 :
5403 0 : let limit = request.max_concurrent_downloads;
5404 :
5405 : loop {
5406 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5407 0 : let Some(next) = remaining.next() else {
5408 0 : have_remaining = false;
5409 0 : break;
5410 : };
5411 :
5412 0 : let span = tracing::info_span!("download", layer = %next);
5413 :
5414 0 : js.spawn(
5415 0 : async move {
5416 0 : let res = next.download().await;
5417 0 : (next, res)
5418 0 : }
5419 0 : .instrument(span),
5420 0 : );
5421 0 : }
5422 :
5423 0 : while let Some(res) = js.join_next().await {
5424 0 : match res {
5425 : Ok((_, Ok(_))) => {
5426 0 : lock_status!(st);
5427 0 : st.successful_download_count += 1;
5428 : }
5429 0 : Ok((layer, Err(e))) => {
5430 0 : tracing::error!(%layer, "download failed: {e:#}");
5431 0 : lock_status!(st);
5432 0 : st.failed_download_count += 1;
5433 : }
5434 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5435 0 : Err(je) if je.is_panic() => {
5436 0 : lock_status!(st);
5437 0 : st.failed_download_count += 1;
5438 : }
5439 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5440 : }
5441 : }
5442 :
5443 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5444 0 : break;
5445 0 : }
5446 : }
5447 :
5448 : {
5449 0 : lock_status!(st);
5450 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5451 : }
5452 0 : }
5453 :
5454 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5455 0 : &self,
5456 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5457 0 : self.download_all_remote_layers_task_info
5458 0 : .read()
5459 0 : .unwrap()
5460 0 : .clone()
5461 0 : }
5462 : }
5463 :
5464 : impl Timeline {
5465 : /// Returns non-remote layers for eviction.
5466 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5467 0 : let guard = self.layers.read().await;
5468 0 : let mut max_layer_size: Option<u64> = None;
5469 0 :
5470 0 : let resident_layers = guard
5471 0 : .likely_resident_layers()
5472 0 : .map(|layer| {
5473 0 : let file_size = layer.layer_desc().file_size;
5474 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5475 0 :
5476 0 : let last_activity_ts = layer.latest_activity();
5477 0 :
5478 0 : EvictionCandidate {
5479 0 : layer: layer.to_owned().into(),
5480 0 : last_activity_ts,
5481 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5482 0 : visibility: layer.visibility(),
5483 0 : }
5484 0 : })
5485 0 : .collect();
5486 0 :
5487 0 : DiskUsageEvictionInfo {
5488 0 : max_layer_size,
5489 0 : resident_layers,
5490 0 : }
5491 0 : }
5492 :
5493 1730 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5494 1730 : ShardIndex {
5495 1730 : shard_number: self.tenant_shard_id.shard_number,
5496 1730 : shard_count: self.tenant_shard_id.shard_count,
5497 1730 : }
5498 1730 : }
5499 :
5500 : /// Persistently blocks gc for `Manual` reason.
5501 : ///
5502 : /// Returns true if no such block existed before, false otherwise.
5503 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
5504 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5505 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5506 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
5507 0 : }
5508 :
5509 : /// Persistently unblocks gc for `Manual` reason.
5510 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
5511 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5512 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5513 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
5514 0 : }
5515 :
5516 : #[cfg(test)]
5517 40 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5518 40 : self.last_record_lsn.advance(new_lsn);
5519 40 : }
5520 :
5521 : #[cfg(test)]
5522 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
5523 2 : self.disk_consistent_lsn.store(new_value);
5524 2 : }
5525 :
5526 : /// Force create an image layer and place it into the layer map.
5527 : ///
5528 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5529 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5530 : /// placed into the layer map in one run AND be validated.
5531 : #[cfg(test)]
5532 52 : pub(super) async fn force_create_image_layer(
5533 52 : self: &Arc<Timeline>,
5534 52 : lsn: Lsn,
5535 52 : mut images: Vec<(Key, Bytes)>,
5536 52 : check_start_lsn: Option<Lsn>,
5537 52 : ctx: &RequestContext,
5538 52 : ) -> anyhow::Result<()> {
5539 52 : let last_record_lsn = self.get_last_record_lsn();
5540 52 : assert!(
5541 52 : lsn <= last_record_lsn,
5542 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5543 : );
5544 52 : if let Some(check_start_lsn) = check_start_lsn {
5545 52 : assert!(lsn >= check_start_lsn);
5546 0 : }
5547 126 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5548 52 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5549 52 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5550 52 : let mut image_layer_writer = ImageLayerWriter::new(
5551 52 : self.conf,
5552 52 : self.timeline_id,
5553 52 : self.tenant_shard_id,
5554 52 : &(min_key..end_key),
5555 52 : lsn,
5556 52 : ctx,
5557 52 : )
5558 26 : .await?;
5559 230 : for (key, img) in images {
5560 178 : image_layer_writer.put_image(key, img, ctx).await?;
5561 : }
5562 104 : let (desc, path) = image_layer_writer.finish(ctx).await?;
5563 52 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5564 52 : info!("force created image layer {}", image_layer.local_path());
5565 : {
5566 52 : let mut guard = self.layers.write().await;
5567 52 : guard.open_mut().unwrap().force_insert_layer(image_layer);
5568 52 : }
5569 52 :
5570 52 : Ok(())
5571 52 : }
5572 :
5573 : /// Force create a delta layer and place it into the layer map.
5574 : ///
5575 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5576 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5577 : /// placed into the layer map in one run AND be validated.
5578 : #[cfg(test)]
5579 74 : pub(super) async fn force_create_delta_layer(
5580 74 : self: &Arc<Timeline>,
5581 74 : mut deltas: DeltaLayerTestDesc,
5582 74 : check_start_lsn: Option<Lsn>,
5583 74 : ctx: &RequestContext,
5584 74 : ) -> anyhow::Result<()> {
5585 74 : let last_record_lsn = self.get_last_record_lsn();
5586 74 : deltas
5587 74 : .data
5588 124 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5589 74 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5590 74 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5591 272 : for (_, lsn, _) in &deltas.data {
5592 198 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5593 : }
5594 74 : assert!(
5595 74 : deltas.lsn_range.end <= last_record_lsn,
5596 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
5597 : deltas.lsn_range.end,
5598 : last_record_lsn
5599 : );
5600 74 : if let Some(check_start_lsn) = check_start_lsn {
5601 74 : assert!(deltas.lsn_range.start >= check_start_lsn);
5602 0 : }
5603 74 : let mut delta_layer_writer = DeltaLayerWriter::new(
5604 74 : self.conf,
5605 74 : self.timeline_id,
5606 74 : self.tenant_shard_id,
5607 74 : deltas.key_range.start,
5608 74 : deltas.lsn_range,
5609 74 : ctx,
5610 74 : )
5611 37 : .await?;
5612 272 : for (key, lsn, val) in deltas.data {
5613 198 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5614 : }
5615 185 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
5616 74 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5617 74 : info!("force created delta layer {}", delta_layer.local_path());
5618 : {
5619 74 : let mut guard = self.layers.write().await;
5620 74 : guard.open_mut().unwrap().force_insert_layer(delta_layer);
5621 74 : }
5622 74 :
5623 74 : Ok(())
5624 74 : }
5625 :
5626 : /// Return all keys at the LSN in the image layers
5627 : #[cfg(test)]
5628 6 : pub(crate) async fn inspect_image_layers(
5629 6 : self: &Arc<Timeline>,
5630 6 : lsn: Lsn,
5631 6 : ctx: &RequestContext,
5632 6 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
5633 6 : let mut all_data = Vec::new();
5634 6 : let guard = self.layers.read().await;
5635 34 : for layer in guard.layer_map()?.iter_historic_layers() {
5636 34 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
5637 8 : let layer = guard.get_from_desc(&layer);
5638 8 : let mut reconstruct_data = ValuesReconstructState::default();
5639 8 : layer
5640 8 : .get_values_reconstruct_data(
5641 8 : KeySpace::single(Key::MIN..Key::MAX),
5642 8 : lsn..Lsn(lsn.0 + 1),
5643 8 : &mut reconstruct_data,
5644 8 : ctx,
5645 8 : )
5646 13 : .await?;
5647 74 : for (k, v) in reconstruct_data.keys {
5648 66 : all_data.push((k, v?.img.unwrap().1));
5649 : }
5650 26 : }
5651 : }
5652 6 : all_data.sort();
5653 6 : Ok(all_data)
5654 6 : }
5655 :
5656 : /// Get all historic layer descriptors in the layer map
5657 : #[cfg(test)]
5658 12 : pub(crate) async fn inspect_historic_layers(
5659 12 : self: &Arc<Timeline>,
5660 12 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
5661 12 : let mut layers = Vec::new();
5662 12 : let guard = self.layers.read().await;
5663 70 : for layer in guard.layer_map()?.iter_historic_layers() {
5664 70 : layers.push(layer.key());
5665 70 : }
5666 12 : Ok(layers)
5667 12 : }
5668 :
5669 : #[cfg(test)]
5670 10 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
5671 10 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
5672 10 : keyspace.merge(&ks);
5673 10 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
5674 10 : }
5675 : }
5676 :
5677 : /// Tracking writes ingestion does to a particular in-memory layer.
5678 : ///
5679 : /// Cleared upon freezing a layer.
5680 : pub(crate) struct TimelineWriterState {
5681 : open_layer: Arc<InMemoryLayer>,
5682 : current_size: u64,
5683 : // Previous Lsn which passed through
5684 : prev_lsn: Option<Lsn>,
5685 : // Largest Lsn which passed through the current writer
5686 : max_lsn: Option<Lsn>,
5687 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
5688 : cached_last_freeze_at: Lsn,
5689 : }
5690 :
5691 : impl TimelineWriterState {
5692 1268 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
5693 1268 : Self {
5694 1268 : open_layer,
5695 1268 : current_size,
5696 1268 : prev_lsn: None,
5697 1268 : max_lsn: None,
5698 1268 : cached_last_freeze_at: last_freeze_at,
5699 1268 : }
5700 1268 : }
5701 : }
5702 :
5703 : /// Various functions to mutate the timeline.
5704 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
5705 : // This is probably considered a bad practice in Rust and should be fixed eventually,
5706 : // but will cause large code changes.
5707 : pub(crate) struct TimelineWriter<'a> {
5708 : tl: &'a Timeline,
5709 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
5710 : }
5711 :
5712 : impl Deref for TimelineWriter<'_> {
5713 : type Target = Timeline;
5714 :
5715 4807192 : fn deref(&self) -> &Self::Target {
5716 4807192 : self.tl
5717 4807192 : }
5718 : }
5719 :
5720 : #[derive(PartialEq)]
5721 : enum OpenLayerAction {
5722 : Roll,
5723 : Open,
5724 : None,
5725 : }
5726 :
5727 : impl<'a> TimelineWriter<'a> {
5728 4804204 : async fn handle_open_layer_action(
5729 4804204 : &mut self,
5730 4804204 : at: Lsn,
5731 4804204 : action: OpenLayerAction,
5732 4804204 : ctx: &RequestContext,
5733 4804204 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
5734 4804204 : match action {
5735 : OpenLayerAction::Roll => {
5736 80 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
5737 80 : self.roll_layer(freeze_at).await?;
5738 80 : self.open_layer(at, ctx).await?;
5739 : }
5740 1188 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
5741 : OpenLayerAction::None => {
5742 4802936 : assert!(self.write_guard.is_some());
5743 : }
5744 : }
5745 :
5746 4804204 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
5747 4804204 : }
5748 :
5749 1268 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
5750 1268 : let layer = self
5751 1268 : .tl
5752 1268 : .get_layer_for_write(at, &self.write_guard, ctx)
5753 719 : .await?;
5754 1268 : let initial_size = layer.size().await?;
5755 :
5756 1268 : let last_freeze_at = self.last_freeze_at.load();
5757 1268 : self.write_guard.replace(TimelineWriterState::new(
5758 1268 : layer,
5759 1268 : initial_size,
5760 1268 : last_freeze_at,
5761 1268 : ));
5762 1268 :
5763 1268 : Ok(())
5764 1268 : }
5765 :
5766 80 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
5767 80 : let current_size = self.write_guard.as_ref().unwrap().current_size;
5768 80 :
5769 80 : // self.write_guard will be taken by the freezing
5770 80 : self.tl
5771 80 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
5772 2 : .await?;
5773 :
5774 80 : assert!(self.write_guard.is_none());
5775 :
5776 80 : if current_size >= self.get_checkpoint_distance() * 2 {
5777 0 : warn!("Flushed oversized open layer with size {}", current_size)
5778 80 : }
5779 :
5780 80 : Ok(())
5781 80 : }
5782 :
5783 4804204 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
5784 4804204 : let state = &*self.write_guard;
5785 4804204 : let Some(state) = &state else {
5786 1188 : return OpenLayerAction::Open;
5787 : };
5788 :
5789 : #[cfg(feature = "testing")]
5790 4803016 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
5791 : // this check and assertion are not really needed because
5792 : // LayerManager::try_freeze_in_memory_layer will always clear out the
5793 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
5794 : // is no TimelineWriterState.
5795 0 : assert!(
5796 0 : state.open_layer.end_lsn.get().is_some(),
5797 0 : "our open_layer must be outdated"
5798 : );
5799 :
5800 : // this would be a memory leak waiting to happen because the in-memory layer always has
5801 : // an index
5802 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
5803 4803016 : }
5804 4803016 :
5805 4803016 : if state.prev_lsn == Some(lsn) {
5806 : // Rolling mid LSN is not supported by [downstream code].
5807 : // Hence, only roll at LSN boundaries.
5808 : //
5809 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
5810 6 : return OpenLayerAction::None;
5811 4803010 : }
5812 4803010 :
5813 4803010 : if state.current_size == 0 {
5814 : // Don't roll empty layers
5815 0 : return OpenLayerAction::None;
5816 4803010 : }
5817 4803010 :
5818 4803010 : if self.tl.should_roll(
5819 4803010 : state.current_size,
5820 4803010 : state.current_size + new_value_size,
5821 4803010 : self.get_checkpoint_distance(),
5822 4803010 : lsn,
5823 4803010 : state.cached_last_freeze_at,
5824 4803010 : state.open_layer.get_opened_at(),
5825 4803010 : ) {
5826 80 : OpenLayerAction::Roll
5827 : } else {
5828 4802930 : OpenLayerAction::None
5829 : }
5830 4804204 : }
5831 :
5832 : /// Put a batch of keys at the specified Lsns.
5833 4804202 : pub(crate) async fn put_batch(
5834 4804202 : &mut self,
5835 4804202 : batch: SerializedValueBatch,
5836 4804202 : ctx: &RequestContext,
5837 4804202 : ) -> anyhow::Result<()> {
5838 4804202 : if batch.is_empty() {
5839 0 : return Ok(());
5840 4804202 : }
5841 4804202 :
5842 4804202 : let batch_max_lsn = batch.max_lsn;
5843 4804202 : let buf_size: u64 = batch.buffer_size() as u64;
5844 4804202 :
5845 4804202 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
5846 4804202 : let layer = self
5847 4804202 : .handle_open_layer_action(batch_max_lsn, action, ctx)
5848 721 : .await?;
5849 :
5850 4804202 : let res = layer.put_batch(batch, ctx).await;
5851 :
5852 4804202 : if res.is_ok() {
5853 4804202 : // Update the current size only when the entire write was ok.
5854 4804202 : // In case of failures, we may have had partial writes which
5855 4804202 : // render the size tracking out of sync. That's ok because
5856 4804202 : // the checkpoint distance should be significantly smaller
5857 4804202 : // than the S3 single shot upload limit of 5GiB.
5858 4804202 : let state = self.write_guard.as_mut().unwrap();
5859 4804202 :
5860 4804202 : state.current_size += buf_size;
5861 4804202 : state.prev_lsn = Some(batch_max_lsn);
5862 4804202 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
5863 4804202 : }
5864 :
5865 4804202 : res
5866 4804202 : }
5867 :
5868 : #[cfg(test)]
5869 : /// Test helper, for tests that would like to poke individual values without composing a batch
5870 4390154 : pub(crate) async fn put(
5871 4390154 : &mut self,
5872 4390154 : key: Key,
5873 4390154 : lsn: Lsn,
5874 4390154 : value: &Value,
5875 4390154 : ctx: &RequestContext,
5876 4390154 : ) -> anyhow::Result<()> {
5877 : use utils::bin_ser::BeSer;
5878 4390154 : if !key.is_valid_key_on_write_path() {
5879 0 : bail!(
5880 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
5881 0 : key
5882 0 : );
5883 4390154 : }
5884 4390154 : let val_ser_size = value.serialized_size().unwrap() as usize;
5885 4390154 : let batch = SerializedValueBatch::from_values(vec![(
5886 4390154 : key.to_compact(),
5887 4390154 : lsn,
5888 4390154 : val_ser_size,
5889 4390154 : value.clone(),
5890 4390154 : )]);
5891 4390154 :
5892 4390154 : self.put_batch(batch, ctx).await
5893 4390154 : }
5894 :
5895 2 : pub(crate) async fn delete_batch(
5896 2 : &mut self,
5897 2 : batch: &[(Range<Key>, Lsn)],
5898 2 : ctx: &RequestContext,
5899 2 : ) -> anyhow::Result<()> {
5900 2 : if let Some((_, lsn)) = batch.first() {
5901 2 : let action = self.get_open_layer_action(*lsn, 0);
5902 2 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
5903 2 : layer.put_tombstones(batch).await?;
5904 0 : }
5905 :
5906 2 : Ok(())
5907 2 : }
5908 :
5909 : /// Track the end of the latest digested WAL record.
5910 : /// Remember the (end of) last valid WAL record remembered in the timeline.
5911 : ///
5912 : /// Call this after you have finished writing all the WAL up to 'lsn'.
5913 : ///
5914 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
5915 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
5916 : /// the latest and can be read.
5917 5279064 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
5918 5279064 : self.tl.finish_write(new_lsn);
5919 5279064 : }
5920 :
5921 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
5922 270570 : self.tl.update_current_logical_size(delta)
5923 270570 : }
5924 : }
5925 :
5926 : // We need TimelineWriter to be send in upcoming conversion of
5927 : // Timeline::layers to tokio::sync::RwLock.
5928 : #[test]
5929 2 : fn is_send() {
5930 2 : fn _assert_send<T: Send>() {}
5931 2 : _assert_send::<TimelineWriter<'_>>();
5932 2 : }
5933 :
5934 : #[cfg(test)]
5935 : mod tests {
5936 : use pageserver_api::key::Key;
5937 : use pageserver_api::value::Value;
5938 : use utils::{id::TimelineId, lsn::Lsn};
5939 :
5940 : use crate::tenant::{
5941 : harness::{test_img, TenantHarness},
5942 : layer_map::LayerMap,
5943 : storage_layer::{Layer, LayerName},
5944 : timeline::{DeltaLayerTestDesc, EvictionError},
5945 : Timeline,
5946 : };
5947 :
5948 : #[tokio::test]
5949 2 : async fn test_heatmap_generation() {
5950 2 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
5951 2 :
5952 2 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
5953 2 : Lsn(0x10)..Lsn(0x20),
5954 2 : vec![(
5955 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
5956 2 : Lsn(0x11),
5957 2 : Value::Image(test_img("foo")),
5958 2 : )],
5959 2 : );
5960 2 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
5961 2 : Lsn(0x10)..Lsn(0x20),
5962 2 : vec![(
5963 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
5964 2 : Lsn(0x11),
5965 2 : Value::Image(test_img("foo")),
5966 2 : )],
5967 2 : );
5968 2 : let l0_delta = DeltaLayerTestDesc::new(
5969 2 : Lsn(0x20)..Lsn(0x30),
5970 2 : Key::from_hex("000000000000000000000000000000000000").unwrap()
5971 2 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
5972 2 : vec![(
5973 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
5974 2 : Lsn(0x25),
5975 2 : Value::Image(test_img("foo")),
5976 2 : )],
5977 2 : );
5978 2 : let delta_layers = vec![
5979 2 : covered_delta.clone(),
5980 2 : visible_delta.clone(),
5981 2 : l0_delta.clone(),
5982 2 : ];
5983 2 :
5984 2 : let image_layer = (
5985 2 : Lsn(0x40),
5986 2 : vec![(
5987 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
5988 2 : test_img("bar"),
5989 2 : )],
5990 2 : );
5991 2 : let image_layers = vec![image_layer];
5992 2 :
5993 20 : let (tenant, ctx) = harness.load().await;
5994 2 : let timeline = tenant
5995 2 : .create_test_timeline_with_layers(
5996 2 : TimelineId::generate(),
5997 2 : Lsn(0x10),
5998 2 : 14,
5999 2 : &ctx,
6000 2 : delta_layers,
6001 2 : image_layers,
6002 2 : Lsn(0x100),
6003 2 : )
6004 31 : .await
6005 2 : .unwrap();
6006 2 :
6007 2 : // Layer visibility is an input to heatmap generation, so refresh it first
6008 2 : timeline.update_layer_visibility().await.unwrap();
6009 2 :
6010 2 : let heatmap = timeline
6011 2 : .generate_heatmap()
6012 2 : .await
6013 2 : .expect("Infallible while timeline is not shut down");
6014 2 :
6015 2 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
6016 2 :
6017 2 : // L0 should come last
6018 2 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
6019 2 :
6020 2 : let mut last_lsn = Lsn::MAX;
6021 10 : for layer in heatmap.layers {
6022 2 : // Covered layer should be omitted
6023 8 : assert!(layer.name != covered_delta.layer_name());
6024 2 :
6025 8 : let layer_lsn = match &layer.name {
6026 4 : LayerName::Delta(d) => d.lsn_range.end,
6027 4 : LayerName::Image(i) => i.lsn,
6028 2 : };
6029 2 :
6030 2 : // Apart from L0s, newest Layers should come first
6031 8 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
6032 6 : assert!(layer_lsn <= last_lsn);
6033 6 : last_lsn = layer_lsn;
6034 2 : }
6035 2 : }
6036 2 : }
6037 :
6038 : #[tokio::test]
6039 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
6040 2 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6041 2 : .await
6042 2 : .unwrap();
6043 2 :
6044 20 : let (tenant, ctx) = harness.load().await;
6045 2 : let timeline = tenant
6046 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6047 6 : .await
6048 2 : .unwrap();
6049 2 :
6050 2 : let layer = find_some_layer(&timeline).await;
6051 2 : let layer = layer
6052 2 : .keep_resident()
6053 2 : .await
6054 2 : .expect("no download => no downloading errors")
6055 2 : .drop_eviction_guard();
6056 2 :
6057 2 : let forever = std::time::Duration::from_secs(120);
6058 2 :
6059 2 : let first = layer.evict_and_wait(forever);
6060 2 : let second = layer.evict_and_wait(forever);
6061 2 :
6062 2 : let (first, second) = tokio::join!(first, second);
6063 2 :
6064 2 : let res = layer.keep_resident().await;
6065 2 : assert!(res.is_none(), "{res:?}");
6066 2 :
6067 2 : match (first, second) {
6068 2 : (Ok(()), Ok(())) => {
6069 2 : // because there are no more timeline locks being taken on eviction path, we can
6070 2 : // witness all three outcomes here.
6071 2 : }
6072 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
6073 0 : // if one completes before the other, this is fine just as well.
6074 0 : }
6075 2 : other => unreachable!("unexpected {:?}", other),
6076 2 : }
6077 2 : }
6078 :
6079 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
6080 2 : let layers = timeline.layers.read().await;
6081 2 : let desc = layers
6082 2 : .layer_map()
6083 2 : .unwrap()
6084 2 : .iter_historic_layers()
6085 2 : .next()
6086 2 : .expect("must find one layer to evict");
6087 2 :
6088 2 : layers.get_from_desc(&desc)
6089 2 : }
6090 : }
|
