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 : pub(crate) mod import_pgdata;
8 : mod init;
9 : pub mod layer_manager;
10 : pub(crate) mod logical_size;
11 : pub mod offload;
12 : pub mod span;
13 : pub mod uninit;
14 : mod walreceiver;
15 :
16 : use anyhow::{anyhow, bail, ensure, Context, Result};
17 : use arc_swap::ArcSwap;
18 : use bytes::Bytes;
19 : use camino::Utf8Path;
20 : use chrono::{DateTime, Utc};
21 : use enumset::EnumSet;
22 : use fail::fail_point;
23 : use handle::ShardTimelineId;
24 : use offload::OffloadError;
25 : use once_cell::sync::Lazy;
26 : use pageserver_api::{
27 : config::tenant_conf_defaults::DEFAULT_COMPACTION_THRESHOLD,
28 : key::{
29 : KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
30 : NON_INHERITED_SPARSE_RANGE,
31 : },
32 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
33 : models::{
34 : CompactionAlgorithm, CompactionAlgorithmSettings, DownloadRemoteLayersTaskInfo,
35 : DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy, InMemoryLayerInfo, LayerMapInfo,
36 : LsnLease, TimelineState,
37 : },
38 : reltag::BlockNumber,
39 : shard::{ShardIdentity, ShardNumber, TenantShardId},
40 : };
41 : use rand::Rng;
42 : use remote_storage::DownloadError;
43 : use serde_with::serde_as;
44 : use storage_broker::BrokerClientChannel;
45 : use tokio::{
46 : runtime::Handle,
47 : sync::{oneshot, watch},
48 : };
49 : use tokio_util::sync::CancellationToken;
50 : use tracing::*;
51 : use utils::{
52 : fs_ext, pausable_failpoint,
53 : sync::gate::{Gate, GateGuard},
54 : };
55 : use wal_decoder::serialized_batch::SerializedValueBatch;
56 :
57 : use std::sync::atomic::Ordering as AtomicOrdering;
58 : use std::sync::{Arc, Mutex, RwLock, Weak};
59 : use std::time::{Duration, Instant, SystemTime};
60 : use std::{
61 : array,
62 : collections::{BTreeMap, HashMap, HashSet},
63 : sync::atomic::AtomicU64,
64 : };
65 : use std::{cmp::min, ops::ControlFlow};
66 : use std::{
67 : collections::btree_map::Entry,
68 : ops::{Deref, Range},
69 : };
70 : use std::{pin::pin, sync::OnceLock};
71 :
72 : use crate::{
73 : aux_file::AuxFileSizeEstimator,
74 : tenant::{
75 : config::AttachmentMode,
76 : layer_map::{LayerMap, SearchResult},
77 : metadata::TimelineMetadata,
78 : storage_layer::{inmemory_layer::IndexEntry, PersistentLayerDesc},
79 : },
80 : walingest::WalLagCooldown,
81 : walredo,
82 : };
83 : use crate::{
84 : context::{DownloadBehavior, RequestContext},
85 : disk_usage_eviction_task::DiskUsageEvictionInfo,
86 : pgdatadir_mapping::CollectKeySpaceError,
87 : };
88 : use crate::{
89 : disk_usage_eviction_task::finite_f32,
90 : tenant::storage_layer::{
91 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
92 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
93 : ValuesReconstructState,
94 : },
95 : };
96 : use crate::{
97 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
98 : };
99 : use crate::{
100 : l0_flush::{self, L0FlushGlobalState},
101 : metrics::GetKind,
102 : };
103 : use crate::{
104 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
105 : };
106 : use crate::{
107 : pgdatadir_mapping::DirectoryKind,
108 : virtual_file::{MaybeFatalIo, VirtualFile},
109 : };
110 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
111 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
112 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
113 :
114 : use crate::config::PageServerConf;
115 : use crate::keyspace::{KeyPartitioning, KeySpace};
116 : use crate::metrics::TimelineMetrics;
117 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
118 : use crate::tenant::config::TenantConfOpt;
119 : use pageserver_api::reltag::RelTag;
120 : use pageserver_api::shard::ShardIndex;
121 :
122 : use postgres_connection::PgConnectionConfig;
123 : use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
124 : use utils::{
125 : completion,
126 : generation::Generation,
127 : id::TimelineId,
128 : lsn::{AtomicLsn, Lsn, RecordLsn},
129 : seqwait::SeqWait,
130 : simple_rcu::{Rcu, RcuReadGuard},
131 : };
132 :
133 : use crate::task_mgr;
134 : use crate::task_mgr::TaskKind;
135 : use crate::tenant::gc_result::GcResult;
136 : use crate::ZERO_PAGE;
137 : use pageserver_api::key::Key;
138 :
139 : use self::delete::DeleteTimelineFlow;
140 : pub(super) use self::eviction_task::EvictionTaskTenantState;
141 : use self::eviction_task::EvictionTaskTimelineState;
142 : use self::layer_manager::LayerManager;
143 : use self::logical_size::LogicalSize;
144 : use self::walreceiver::{WalReceiver, WalReceiverConf};
145 :
146 : use super::{
147 : config::TenantConf, storage_layer::LayerVisibilityHint, upload_queue::NotInitialized,
148 : MaybeOffloaded,
149 : };
150 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
151 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
152 : use super::{
153 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
154 : storage_layer::ReadableLayer,
155 : };
156 : use super::{
157 : secondary::heatmap::{HeatMapLayer, HeatMapTimeline},
158 : GcError,
159 : };
160 :
161 : #[cfg(test)]
162 : use pageserver_api::value::Value;
163 :
164 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
165 : pub(crate) enum FlushLoopState {
166 : NotStarted,
167 : Running {
168 : #[cfg(test)]
169 : expect_initdb_optimization: bool,
170 : #[cfg(test)]
171 : initdb_optimization_count: usize,
172 : },
173 : Exited,
174 : }
175 :
176 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
177 : pub enum ImageLayerCreationMode {
178 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
179 : Try,
180 : /// Force creating the image layers if possible. For now, no image layers will be created
181 : /// for metadata keys. Used in compaction code path with force flag enabled.
182 : Force,
183 : /// Initial ingestion of the data, and no data should be dropped in this function. This
184 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
185 : /// code path.
186 : Initial,
187 : }
188 :
189 : impl std::fmt::Display for ImageLayerCreationMode {
190 716 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
191 716 : write!(f, "{:?}", self)
192 716 : }
193 : }
194 :
195 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
196 : /// Can be removed after all refactors are done.
197 28 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
198 28 : drop(rlock)
199 28 : }
200 :
201 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
202 : /// Can be removed after all refactors are done.
203 744 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
204 744 : drop(rlock)
205 744 : }
206 :
207 : /// The outward-facing resources required to build a Timeline
208 : pub struct TimelineResources {
209 : pub remote_client: RemoteTimelineClient,
210 : pub timeline_get_throttle:
211 : Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
212 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
213 : }
214 :
215 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
216 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
217 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
218 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
219 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
220 : pub(crate) struct RelSizeCache {
221 : pub(crate) complete_as_of: Lsn,
222 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
223 : }
224 :
225 : pub struct Timeline {
226 : pub(crate) conf: &'static PageServerConf,
227 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
228 :
229 : myself: Weak<Self>,
230 :
231 : pub(crate) tenant_shard_id: TenantShardId,
232 : pub timeline_id: TimelineId,
233 :
234 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
235 : /// Never changes for the lifetime of this [`Timeline`] object.
236 : ///
237 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
238 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
239 : pub(crate) generation: Generation,
240 :
241 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
242 : /// to shards, and is constant through the lifetime of this Timeline.
243 : shard_identity: ShardIdentity,
244 :
245 : pub pg_version: u32,
246 :
247 : /// The tuple has two elements.
248 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
249 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
250 : ///
251 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
252 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
253 : ///
254 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
255 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
256 : /// `PersistentLayerDesc`'s.
257 : ///
258 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
259 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
260 : /// runtime, e.g., during page reconstruction.
261 : ///
262 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
263 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
264 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
265 :
266 : last_freeze_at: AtomicLsn,
267 : // Atomic would be more appropriate here.
268 : last_freeze_ts: RwLock<Instant>,
269 :
270 : pub(crate) standby_horizon: AtomicLsn,
271 :
272 : // WAL redo manager. `None` only for broken tenants.
273 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
274 :
275 : /// Remote storage client.
276 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
277 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
278 :
279 : // What page versions do we hold in the repository? If we get a
280 : // request > last_record_lsn, we need to wait until we receive all
281 : // the WAL up to the request. The SeqWait provides functions for
282 : // that. TODO: If we get a request for an old LSN, such that the
283 : // versions have already been garbage collected away, we should
284 : // throw an error, but we don't track that currently.
285 : //
286 : // last_record_lsn.load().last points to the end of last processed WAL record.
287 : //
288 : // We also remember the starting point of the previous record in
289 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
290 : // first WAL record when the node is started up. But here, we just
291 : // keep track of it.
292 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
293 :
294 : // All WAL records have been processed and stored durably on files on
295 : // local disk, up to this LSN. On crash and restart, we need to re-process
296 : // the WAL starting from this point.
297 : //
298 : // Some later WAL records might have been processed and also flushed to disk
299 : // already, so don't be surprised to see some, but there's no guarantee on
300 : // them yet.
301 : disk_consistent_lsn: AtomicLsn,
302 :
303 : // Parent timeline that this timeline was branched from, and the LSN
304 : // of the branch point.
305 : ancestor_timeline: Option<Arc<Timeline>>,
306 : ancestor_lsn: Lsn,
307 :
308 : pub(super) metrics: TimelineMetrics,
309 :
310 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
311 : // in `crate::page_service` writes these metrics.
312 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
313 :
314 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
315 :
316 : /// Ensures layers aren't frozen by checkpointer between
317 : /// [`Timeline::get_layer_for_write`] and layer reads.
318 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
319 : /// Must always be acquired before the layer map/individual layer lock
320 : /// to avoid deadlock.
321 : ///
322 : /// The state is cleared upon freezing.
323 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
324 :
325 : /// Used to avoid multiple `flush_loop` tasks running
326 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
327 :
328 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
329 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
330 : /// The flush cycle counter is sent back on the layer_flush_done channel when
331 : /// the flush finishes. You can use that to wait for the flush to finish.
332 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
333 : /// read by whoever sends an update
334 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
335 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
336 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
337 :
338 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
339 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
340 :
341 : // List of child timelines and their branch points. This is needed to avoid
342 : // garbage collecting data that is still needed by the child timelines.
343 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
344 :
345 : // It may change across major versions so for simplicity
346 : // keep it after running initdb for a timeline.
347 : // It is needed in checks when we want to error on some operations
348 : // when they are requested for pre-initdb lsn.
349 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
350 : // though let's keep them both for better error visibility.
351 : pub initdb_lsn: Lsn,
352 :
353 : /// When did we last calculate the partitioning? Make it pub to test cases.
354 : pub(super) partitioning: tokio::sync::Mutex<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
355 :
356 : /// Configuration: how often should the partitioning be recalculated.
357 : repartition_threshold: u64,
358 :
359 : last_image_layer_creation_check_at: AtomicLsn,
360 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
361 :
362 : /// Current logical size of the "datadir", at the last LSN.
363 : current_logical_size: LogicalSize,
364 :
365 : /// Information about the last processed message by the WAL receiver,
366 : /// or None if WAL receiver has not received anything for this timeline
367 : /// yet.
368 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
369 : pub walreceiver: Mutex<Option<WalReceiver>>,
370 :
371 : /// Relation size cache
372 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
373 :
374 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
375 :
376 : state: watch::Sender<TimelineState>,
377 :
378 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
379 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
380 : pub delete_progress: TimelineDeleteProgress,
381 :
382 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
383 :
384 : /// Load or creation time information about the disk_consistent_lsn and when the loading
385 : /// happened. Used for consumption metrics.
386 : pub(crate) loaded_at: (Lsn, SystemTime),
387 :
388 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
389 : pub(crate) gate: Gate,
390 :
391 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
392 : /// to the timeline should drop out when this token fires.
393 : pub(crate) cancel: CancellationToken,
394 :
395 : /// Make sure we only have one running compaction at a time in tests.
396 : ///
397 : /// Must only be taken in two places:
398 : /// - [`Timeline::compact`] (this file)
399 : /// - [`delete::delete_local_timeline_directory`]
400 : ///
401 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
402 : compaction_lock: tokio::sync::Mutex<()>,
403 :
404 : /// Make sure we only have one running gc at a time.
405 : ///
406 : /// Must only be taken in two places:
407 : /// - [`Timeline::gc`] (this file)
408 : /// - [`delete::delete_local_timeline_directory`]
409 : ///
410 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
411 : gc_lock: tokio::sync::Mutex<()>,
412 :
413 : /// Cloned from [`super::Tenant::timeline_get_throttle`] on construction.
414 : timeline_get_throttle:
415 : Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
416 :
417 : /// Size estimator for aux file v2
418 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
419 :
420 : /// Some test cases directly place keys into the timeline without actually modifying the directory
421 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
422 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
423 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
424 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
425 : #[cfg(test)]
426 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
427 :
428 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
429 :
430 : pub(crate) handles: handle::PerTimelineState<crate::page_service::TenantManagerTypes>,
431 :
432 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
433 :
434 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
435 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
436 : }
437 :
438 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
439 :
440 : pub struct WalReceiverInfo {
441 : pub wal_source_connconf: PgConnectionConfig,
442 : pub last_received_msg_lsn: Lsn,
443 : pub last_received_msg_ts: u128,
444 : }
445 :
446 : /// Information about how much history needs to be retained, needed by
447 : /// Garbage Collection.
448 : #[derive(Default)]
449 : pub(crate) struct GcInfo {
450 : /// Specific LSNs that are needed.
451 : ///
452 : /// Currently, this includes all points where child branches have
453 : /// been forked off from. In the future, could also include
454 : /// explicit user-defined snapshot points.
455 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
456 :
457 : /// The cutoff coordinates, which are combined by selecting the minimum.
458 : pub(crate) cutoffs: GcCutoffs,
459 :
460 : /// Leases granted to particular LSNs.
461 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
462 :
463 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
464 : pub(crate) within_ancestor_pitr: bool,
465 : }
466 :
467 : impl GcInfo {
468 228 : pub(crate) fn min_cutoff(&self) -> Lsn {
469 228 : self.cutoffs.select_min()
470 228 : }
471 :
472 232 : pub(super) fn insert_child(
473 232 : &mut self,
474 232 : child_id: TimelineId,
475 232 : child_lsn: Lsn,
476 232 : is_offloaded: MaybeOffloaded,
477 232 : ) {
478 232 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
479 232 : self.retain_lsns.sort_by_key(|i| i.0);
480 232 : }
481 :
482 4 : pub(super) fn remove_child_maybe_offloaded(
483 4 : &mut self,
484 4 : child_id: TimelineId,
485 4 : maybe_offloaded: MaybeOffloaded,
486 4 : ) -> bool {
487 4 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
488 4 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
489 4 : let mut removed = false;
490 6 : self.retain_lsns.retain(|i| {
491 6 : if removed {
492 2 : return true;
493 4 : }
494 4 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
495 4 : removed |= remove;
496 4 : !remove
497 6 : });
498 4 : removed
499 4 : }
500 :
501 4 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
502 4 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
503 4 : }
504 :
505 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
506 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
507 0 : }
508 : }
509 :
510 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
511 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
512 : /// between time-based and space-based retention for observability and consumption metrics purposes.
513 : #[derive(Debug, Clone)]
514 : pub(crate) struct GcCutoffs {
515 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
516 : /// history we must keep to retain a specified number of bytes of WAL.
517 : pub(crate) space: Lsn,
518 :
519 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
520 : /// history we must keep to enable reading back at least the PITR interval duration.
521 : pub(crate) time: Lsn,
522 : }
523 :
524 : impl Default for GcCutoffs {
525 418 : fn default() -> Self {
526 418 : Self {
527 418 : space: Lsn::INVALID,
528 418 : time: Lsn::INVALID,
529 418 : }
530 418 : }
531 : }
532 :
533 : impl GcCutoffs {
534 268 : fn select_min(&self) -> Lsn {
535 268 : std::cmp::min(self.space, self.time)
536 268 : }
537 : }
538 :
539 : pub(crate) struct TimelineVisitOutcome {
540 : completed_keyspace: KeySpace,
541 : image_covered_keyspace: KeySpace,
542 : }
543 :
544 : /// An error happened in a get() operation.
545 2 : #[derive(thiserror::Error, Debug)]
546 : pub(crate) enum PageReconstructError {
547 : #[error(transparent)]
548 : Other(anyhow::Error),
549 :
550 : #[error("Ancestor LSN wait error: {0}")]
551 : AncestorLsnTimeout(WaitLsnError),
552 :
553 : #[error("timeline shutting down")]
554 : Cancelled,
555 :
556 : /// An error happened replaying WAL records
557 : #[error(transparent)]
558 : WalRedo(anyhow::Error),
559 :
560 : #[error("{0}")]
561 : MissingKey(MissingKeyError),
562 : }
563 :
564 : impl From<anyhow::Error> for PageReconstructError {
565 0 : fn from(value: anyhow::Error) -> Self {
566 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
567 0 : match value.downcast::<PageReconstructError>() {
568 0 : Ok(pre) => pre,
569 0 : Err(other) => PageReconstructError::Other(other),
570 : }
571 0 : }
572 : }
573 :
574 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
575 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
576 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
577 0 : }
578 : }
579 :
580 : impl From<layer_manager::Shutdown> for PageReconstructError {
581 0 : fn from(_: layer_manager::Shutdown) -> Self {
582 0 : PageReconstructError::Cancelled
583 0 : }
584 : }
585 :
586 : impl GetVectoredError {
587 : #[cfg(test)]
588 6 : pub(crate) fn is_missing_key_error(&self) -> bool {
589 6 : matches!(self, Self::MissingKey(_))
590 6 : }
591 : }
592 :
593 : impl From<layer_manager::Shutdown> for GetVectoredError {
594 0 : fn from(_: layer_manager::Shutdown) -> Self {
595 0 : GetVectoredError::Cancelled
596 0 : }
597 : }
598 :
599 : #[derive(thiserror::Error)]
600 : pub struct MissingKeyError {
601 : key: Key,
602 : shard: ShardNumber,
603 : cont_lsn: Lsn,
604 : request_lsn: Lsn,
605 : ancestor_lsn: Option<Lsn>,
606 : backtrace: Option<std::backtrace::Backtrace>,
607 : }
608 :
609 : impl std::fmt::Debug for MissingKeyError {
610 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
611 0 : write!(f, "{}", self)
612 0 : }
613 : }
614 :
615 : impl std::fmt::Display for MissingKeyError {
616 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
617 0 : write!(
618 0 : f,
619 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
620 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
621 0 : )?;
622 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
623 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
624 0 : }
625 :
626 0 : if let Some(ref backtrace) = self.backtrace {
627 0 : write!(f, "\n{}", backtrace)?;
628 0 : }
629 :
630 0 : Ok(())
631 0 : }
632 : }
633 :
634 : impl PageReconstructError {
635 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
636 0 : pub(crate) fn is_stopping(&self) -> bool {
637 : use PageReconstructError::*;
638 0 : match self {
639 0 : Cancelled => true,
640 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
641 : }
642 0 : }
643 : }
644 :
645 0 : #[derive(thiserror::Error, Debug)]
646 : pub(crate) enum CreateImageLayersError {
647 : #[error("timeline shutting down")]
648 : Cancelled,
649 :
650 : #[error("read failed")]
651 : GetVectoredError(#[source] GetVectoredError),
652 :
653 : #[error("reconstruction failed")]
654 : PageReconstructError(#[source] PageReconstructError),
655 :
656 : #[error(transparent)]
657 : Other(#[from] anyhow::Error),
658 : }
659 :
660 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
661 0 : fn from(_: layer_manager::Shutdown) -> Self {
662 0 : CreateImageLayersError::Cancelled
663 0 : }
664 : }
665 :
666 0 : #[derive(thiserror::Error, Debug, Clone)]
667 : pub(crate) enum FlushLayerError {
668 : /// Timeline cancellation token was cancelled
669 : #[error("timeline shutting down")]
670 : Cancelled,
671 :
672 : /// We tried to flush a layer while the Timeline is in an unexpected state
673 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
674 : NotRunning(FlushLoopState),
675 :
676 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
677 : // loop via a watch channel, where we can only borrow it.
678 : #[error("create image layers (shared)")]
679 : CreateImageLayersError(Arc<CreateImageLayersError>),
680 :
681 : #[error("other (shared)")]
682 : Other(#[from] Arc<anyhow::Error>),
683 : }
684 :
685 : impl FlushLayerError {
686 : // When crossing from generic anyhow errors to this error type, we explicitly check
687 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
688 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
689 0 : let cancelled = timeline.cancel.is_cancelled()
690 : // The upload queue might have been shut down before the official cancellation of the timeline.
691 0 : || err
692 0 : .downcast_ref::<NotInitialized>()
693 0 : .map(NotInitialized::is_stopping)
694 0 : .unwrap_or_default();
695 0 : if cancelled {
696 0 : Self::Cancelled
697 : } else {
698 0 : Self::Other(Arc::new(err))
699 : }
700 0 : }
701 : }
702 :
703 : impl From<layer_manager::Shutdown> for FlushLayerError {
704 0 : fn from(_: layer_manager::Shutdown) -> Self {
705 0 : FlushLayerError::Cancelled
706 0 : }
707 : }
708 :
709 0 : #[derive(thiserror::Error, Debug)]
710 : pub(crate) enum GetVectoredError {
711 : #[error("timeline shutting down")]
712 : Cancelled,
713 :
714 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
715 : Oversized(u64),
716 :
717 : #[error("requested at invalid LSN: {0}")]
718 : InvalidLsn(Lsn),
719 :
720 : #[error("requested key not found: {0}")]
721 : MissingKey(MissingKeyError),
722 :
723 : #[error("ancestry walk")]
724 : GetReadyAncestorError(#[source] GetReadyAncestorError),
725 :
726 : #[error(transparent)]
727 : Other(#[from] anyhow::Error),
728 : }
729 :
730 : impl From<GetReadyAncestorError> for GetVectoredError {
731 2 : fn from(value: GetReadyAncestorError) -> Self {
732 : use GetReadyAncestorError::*;
733 2 : match value {
734 0 : Cancelled => GetVectoredError::Cancelled,
735 : AncestorLsnTimeout(_) | BadState { .. } => {
736 2 : GetVectoredError::GetReadyAncestorError(value)
737 : }
738 : }
739 2 : }
740 : }
741 :
742 2 : #[derive(thiserror::Error, Debug)]
743 : pub(crate) enum GetReadyAncestorError {
744 : #[error("ancestor LSN wait error")]
745 : AncestorLsnTimeout(#[from] WaitLsnError),
746 :
747 : #[error("bad state on timeline {timeline_id}: {state:?}")]
748 : BadState {
749 : timeline_id: TimelineId,
750 : state: TimelineState,
751 : },
752 :
753 : #[error("cancelled")]
754 : Cancelled,
755 : }
756 :
757 : #[derive(Clone, Copy)]
758 : pub enum LogicalSizeCalculationCause {
759 : Initial,
760 : ConsumptionMetricsSyntheticSize,
761 : EvictionTaskImitation,
762 : TenantSizeHandler,
763 : }
764 :
765 : pub enum GetLogicalSizePriority {
766 : User,
767 : Background,
768 : }
769 :
770 0 : #[derive(enumset::EnumSetType)]
771 : pub(crate) enum CompactFlags {
772 : ForceRepartition,
773 : ForceImageLayerCreation,
774 : ForceL0Compaction,
775 : EnhancedGcBottomMostCompaction,
776 : DryRun,
777 : }
778 :
779 : #[serde_with::serde_as]
780 0 : #[derive(Debug, Clone, serde::Deserialize)]
781 : pub(crate) struct CompactRange {
782 : #[serde_as(as = "serde_with::DisplayFromStr")]
783 : pub start: Key,
784 : #[serde_as(as = "serde_with::DisplayFromStr")]
785 : pub end: Key,
786 : }
787 :
788 : #[derive(Clone, Default)]
789 : pub(crate) struct CompactOptions {
790 : pub flags: EnumSet<CompactFlags>,
791 : pub compact_range: Option<CompactRange>,
792 : }
793 :
794 : impl std::fmt::Debug for Timeline {
795 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
796 0 : write!(f, "Timeline<{}>", self.timeline_id)
797 0 : }
798 : }
799 :
800 0 : #[derive(thiserror::Error, Debug)]
801 : pub(crate) enum WaitLsnError {
802 : // Called on a timeline which is shutting down
803 : #[error("Shutdown")]
804 : Shutdown,
805 :
806 : // Called on an timeline not in active state or shutting down
807 : #[error("Bad timeline state: {0:?}")]
808 : BadState(TimelineState),
809 :
810 : // Timeout expired while waiting for LSN to catch up with goal.
811 : #[error("{0}")]
812 : Timeout(String),
813 : }
814 :
815 : // The impls below achieve cancellation mapping for errors.
816 : // Perhaps there's a way of achieving this with less cruft.
817 :
818 : impl From<CreateImageLayersError> for CompactionError {
819 0 : fn from(e: CreateImageLayersError) -> Self {
820 0 : match e {
821 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
822 0 : CreateImageLayersError::Other(e) => {
823 0 : CompactionError::Other(e.context("create image layers"))
824 : }
825 0 : _ => CompactionError::Other(e.into()),
826 : }
827 0 : }
828 : }
829 :
830 : impl From<CreateImageLayersError> for FlushLayerError {
831 0 : fn from(e: CreateImageLayersError) -> Self {
832 0 : match e {
833 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
834 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
835 : }
836 0 : }
837 : }
838 :
839 : impl From<PageReconstructError> for CreateImageLayersError {
840 0 : fn from(e: PageReconstructError) -> Self {
841 0 : match e {
842 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
843 0 : _ => CreateImageLayersError::PageReconstructError(e),
844 : }
845 0 : }
846 : }
847 :
848 : impl From<GetVectoredError> for CreateImageLayersError {
849 0 : fn from(e: GetVectoredError) -> Self {
850 0 : match e {
851 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
852 0 : _ => CreateImageLayersError::GetVectoredError(e),
853 : }
854 0 : }
855 : }
856 :
857 : impl From<GetVectoredError> for PageReconstructError {
858 6 : fn from(e: GetVectoredError) -> Self {
859 6 : match e {
860 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
861 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
862 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
863 4 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
864 2 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
865 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
866 : }
867 6 : }
868 : }
869 :
870 : impl From<GetReadyAncestorError> for PageReconstructError {
871 2 : fn from(e: GetReadyAncestorError) -> Self {
872 : use GetReadyAncestorError::*;
873 2 : match e {
874 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
875 2 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
876 0 : Cancelled => PageReconstructError::Cancelled,
877 : }
878 2 : }
879 : }
880 :
881 : pub(crate) enum WaitLsnWaiter<'a> {
882 : Timeline(&'a Timeline),
883 : Tenant,
884 : PageService,
885 : }
886 :
887 : /// Argument to [`Timeline::shutdown`].
888 : #[derive(Debug, Clone, Copy)]
889 : pub(crate) enum ShutdownMode {
890 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
891 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
892 : ///
893 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
894 : /// the call to [`Timeline::shutdown`].
895 : FreezeAndFlush,
896 : /// Only flush the layers to the remote storage without freezing any open layers. This is the
897 : /// mode used by ancestor detach and any other operations that reloads a tenant but not increasing
898 : /// the generation number.
899 : Flush,
900 : /// Shut down immediately, without waiting for any open layers to flush.
901 : Hard,
902 : }
903 :
904 : struct ImageLayerCreationOutcome {
905 : image: Option<ResidentLayer>,
906 : next_start_key: Key,
907 : }
908 :
909 : /// Public interface functions
910 : impl Timeline {
911 : /// Get the LSN where this branch was created
912 4 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
913 4 : self.ancestor_lsn
914 4 : }
915 :
916 : /// Get the ancestor's timeline id
917 12 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
918 12 : self.ancestor_timeline
919 12 : .as_ref()
920 12 : .map(|ancestor| ancestor.timeline_id)
921 12 : }
922 :
923 : /// Get the ancestor timeline
924 2 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
925 2 : self.ancestor_timeline.as_ref()
926 2 : }
927 :
928 : /// Get the bytes written since the PITR cutoff on this branch, and
929 : /// whether this branch's ancestor_lsn is within its parent's PITR.
930 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
931 0 : let gc_info = self.gc_info.read().unwrap();
932 0 : let history = self
933 0 : .get_last_record_lsn()
934 0 : .checked_sub(gc_info.cutoffs.time)
935 0 : .unwrap_or(Lsn(0))
936 0 : .0;
937 0 : (history, gc_info.within_ancestor_pitr)
938 0 : }
939 :
940 : /// Lock and get timeline's GC cutoff
941 248 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
942 248 : self.latest_gc_cutoff_lsn.read()
943 248 : }
944 :
945 : /// Look up given page version.
946 : ///
947 : /// If a remote layer file is needed, it is downloaded as part of this
948 : /// call.
949 : ///
950 : /// This method enforces [`Self::timeline_get_throttle`] internally.
951 : ///
952 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
953 : /// abstraction above this needs to store suitable metadata to track what
954 : /// data exists with what keys, in separate metadata entries. If a
955 : /// non-existent key is requested, we may incorrectly return a value from
956 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
957 : /// non-existing key.
958 : ///
959 : /// # Cancel-Safety
960 : ///
961 : /// This method is cancellation-safe.
962 : #[inline(always)]
963 606830 : pub(crate) async fn get(
964 606830 : &self,
965 606830 : key: Key,
966 606830 : lsn: Lsn,
967 606830 : ctx: &RequestContext,
968 606830 : ) -> Result<Bytes, PageReconstructError> {
969 606830 : if !lsn.is_valid() {
970 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
971 606830 : }
972 606830 :
973 606830 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
974 606830 : // already checked the key against the shard_identity when looking up the Timeline from
975 606830 : // page_service.
976 606830 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
977 :
978 606830 : self.timeline_get_throttle.throttle(ctx, 1).await;
979 :
980 606830 : let keyspace = KeySpace {
981 606830 : ranges: vec![key..key.next()],
982 606830 : };
983 606830 :
984 606830 : // Initialise the reconstruct state for the key with the cache
985 606830 : // entry returned above.
986 606830 : let mut reconstruct_state = ValuesReconstructState::new();
987 :
988 606830 : let vectored_res = self
989 606830 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
990 182956 : .await;
991 :
992 606830 : let key_value = vectored_res?.pop_first();
993 606824 : match key_value {
994 606812 : Some((got_key, value)) => {
995 606812 : if got_key != key {
996 0 : error!(
997 0 : "Expected {}, but singular vectored get returned {}",
998 : key, got_key
999 : );
1000 0 : Err(PageReconstructError::Other(anyhow!(
1001 0 : "Singular vectored get returned wrong key"
1002 0 : )))
1003 : } else {
1004 606812 : value
1005 : }
1006 : }
1007 12 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1008 12 : key,
1009 12 : shard: self.shard_identity.get_shard_number(&key),
1010 12 : cont_lsn: Lsn(0),
1011 12 : request_lsn: lsn,
1012 12 : ancestor_lsn: None,
1013 12 : backtrace: None,
1014 12 : })),
1015 : }
1016 606830 : }
1017 :
1018 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1019 : pub(crate) const VEC_GET_LAYERS_VISITED_WARN_THRESH: f64 = 512.0;
1020 :
1021 : /// Look up multiple page versions at a given LSN
1022 : ///
1023 : /// This naive implementation will be replaced with a more efficient one
1024 : /// which actually vectorizes the read path.
1025 19528 : pub(crate) async fn get_vectored(
1026 19528 : &self,
1027 19528 : keyspace: KeySpace,
1028 19528 : lsn: Lsn,
1029 19528 : ctx: &RequestContext,
1030 19528 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1031 19528 : if !lsn.is_valid() {
1032 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1033 19528 : }
1034 19528 :
1035 19528 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1036 19528 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1037 0 : return Err(GetVectoredError::Oversized(key_count));
1038 19528 : }
1039 :
1040 39056 : for range in &keyspace.ranges {
1041 19528 : let mut key = range.start;
1042 39262 : while key != range.end {
1043 19734 : assert!(!self.shard_identity.is_key_disposable(&key));
1044 19734 : key = key.next();
1045 : }
1046 : }
1047 :
1048 19528 : trace!(
1049 0 : "get vectored request for {:?}@{} from task kind {:?}",
1050 0 : keyspace,
1051 0 : lsn,
1052 0 : ctx.task_kind(),
1053 : );
1054 :
1055 19528 : let start = crate::metrics::GET_VECTORED_LATENCY
1056 19528 : .for_task_kind(ctx.task_kind())
1057 19528 : .map(|metric| (metric, Instant::now()));
1058 :
1059 : // start counting after throttle so that throttle time
1060 : // is always less than observation time
1061 19528 : let throttled = self
1062 19528 : .timeline_get_throttle
1063 19528 : .throttle(ctx, key_count as usize)
1064 0 : .await;
1065 :
1066 19528 : let res = self
1067 19528 : .get_vectored_impl(
1068 19528 : keyspace.clone(),
1069 19528 : lsn,
1070 19528 : &mut ValuesReconstructState::new(),
1071 19528 : ctx,
1072 19528 : )
1073 7037 : .await;
1074 :
1075 19528 : if let Some((metric, start)) = start {
1076 0 : let elapsed = start.elapsed();
1077 0 : let ex_throttled = if let Some(throttled) = throttled {
1078 0 : elapsed.checked_sub(throttled)
1079 : } else {
1080 0 : Some(elapsed)
1081 : };
1082 :
1083 0 : if let Some(ex_throttled) = ex_throttled {
1084 0 : metric.observe(ex_throttled.as_secs_f64());
1085 0 : } else {
1086 0 : use utils::rate_limit::RateLimit;
1087 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1088 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
1089 0 : let mut rate_limit = LOGGED.lock().unwrap();
1090 0 : rate_limit.call(|| {
1091 0 : warn!("error deducting time spent throttled; this message is logged at a global rate limit");
1092 0 : });
1093 0 : }
1094 19528 : }
1095 :
1096 19528 : res
1097 19528 : }
1098 :
1099 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1100 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1101 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1102 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1103 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1104 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1105 : /// the scan operation will not cause OOM in the future.
1106 12 : pub(crate) async fn scan(
1107 12 : &self,
1108 12 : keyspace: KeySpace,
1109 12 : lsn: Lsn,
1110 12 : ctx: &RequestContext,
1111 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1112 12 : if !lsn.is_valid() {
1113 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1114 12 : }
1115 12 :
1116 12 : trace!(
1117 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1118 0 : keyspace,
1119 0 : lsn,
1120 0 : ctx.task_kind()
1121 : );
1122 :
1123 : // We should generalize this into Keyspace::contains in the future.
1124 24 : for range in &keyspace.ranges {
1125 12 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1126 12 : || range.end.field1 > METADATA_KEY_END_PREFIX
1127 : {
1128 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1129 0 : "only metadata keyspace can be scanned"
1130 0 : )));
1131 12 : }
1132 : }
1133 :
1134 12 : let start = crate::metrics::SCAN_LATENCY
1135 12 : .for_task_kind(ctx.task_kind())
1136 12 : .map(ScanLatencyOngoingRecording::start_recording);
1137 :
1138 : // start counting after throttle so that throttle time
1139 : // is always less than observation time
1140 12 : let throttled = self
1141 12 : .timeline_get_throttle
1142 12 : // assume scan = 1 quota for now until we find a better way to process this
1143 12 : .throttle(ctx, 1)
1144 0 : .await;
1145 :
1146 12 : let vectored_res = self
1147 12 : .get_vectored_impl(
1148 12 : keyspace.clone(),
1149 12 : lsn,
1150 12 : &mut ValuesReconstructState::default(),
1151 12 : ctx,
1152 12 : )
1153 0 : .await;
1154 :
1155 12 : if let Some(recording) = start {
1156 0 : recording.observe(throttled);
1157 12 : }
1158 :
1159 12 : vectored_res
1160 12 : }
1161 :
1162 626672 : pub(super) async fn get_vectored_impl(
1163 626672 : &self,
1164 626672 : keyspace: KeySpace,
1165 626672 : lsn: Lsn,
1166 626672 : reconstruct_state: &mut ValuesReconstructState,
1167 626672 : ctx: &RequestContext,
1168 626672 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1169 626672 : let get_kind = if keyspace.total_raw_size() == 1 {
1170 626236 : GetKind::Singular
1171 : } else {
1172 436 : GetKind::Vectored
1173 : };
1174 :
1175 626672 : let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
1176 626672 : .for_get_kind(get_kind)
1177 626672 : .start_timer();
1178 626672 : self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1179 192172 : .await?;
1180 626656 : get_data_timer.stop_and_record();
1181 626656 :
1182 626656 : let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
1183 626656 : .for_get_kind(get_kind)
1184 626656 : .start_timer();
1185 626656 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
1186 626656 : let layers_visited = reconstruct_state.get_layers_visited();
1187 :
1188 666982 : for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
1189 666982 : match res {
1190 0 : Err(err) => {
1191 0 : results.insert(key, Err(err));
1192 0 : }
1193 666982 : Ok(state) => {
1194 666982 : let state = ValueReconstructState::from(state);
1195 :
1196 666982 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
1197 666982 : results.insert(key, reconstruct_res);
1198 : }
1199 : }
1200 : }
1201 626656 : reconstruct_timer.stop_and_record();
1202 626656 :
1203 626656 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1204 626656 : // when they're missing. Instead they are omitted from the resulting btree
1205 626656 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1206 626656 : // to avoid infinite results.
1207 626656 : if !results.is_empty() {
1208 626450 : let avg = layers_visited as f64 / results.len() as f64;
1209 626450 : if avg >= Self::VEC_GET_LAYERS_VISITED_WARN_THRESH {
1210 0 : use utils::rate_limit::RateLimit;
1211 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1212 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1213 0 : let mut rate_limit = LOGGED.lock().unwrap();
1214 0 : rate_limit.call(|| {
1215 0 : tracing::info!(
1216 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1217 0 : lsn = %lsn,
1218 0 : "Vectored read for {} visited {} layers on average per key and {} in total. {}/{} pages were returned",
1219 0 : keyspace, avg, layers_visited, results.len(), keyspace.total_raw_size());
1220 0 : });
1221 626450 : }
1222 :
1223 : // Note that this is an approximation. Tracking the exact number of layers visited
1224 : // per key requires virtually unbounded memory usage and is inefficient
1225 : // (i.e. segment tree tracking each range queried from a layer)
1226 626450 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED.observe(avg);
1227 206 : }
1228 :
1229 626656 : Ok(results)
1230 626672 : }
1231 :
1232 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1233 274290 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1234 274290 : self.last_record_lsn.load().last
1235 274290 : }
1236 :
1237 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1238 0 : self.last_record_lsn.load().prev
1239 0 : }
1240 :
1241 : /// Atomically get both last and prev.
1242 228 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1243 228 : self.last_record_lsn.load()
1244 228 : }
1245 :
1246 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1247 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1248 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1249 0 : self.last_record_lsn.status_receiver()
1250 0 : }
1251 :
1252 414 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1253 414 : self.disk_consistent_lsn.load()
1254 414 : }
1255 :
1256 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1257 : /// not validated with control plane yet.
1258 : /// See [`Self::get_remote_consistent_lsn_visible`].
1259 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1260 0 : self.remote_client.remote_consistent_lsn_projected()
1261 0 : }
1262 :
1263 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1264 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1265 : /// generation validation in the deletion queue.
1266 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1267 0 : self.remote_client.remote_consistent_lsn_visible()
1268 0 : }
1269 :
1270 : /// The sum of the file size of all historic layers in the layer map.
1271 : /// This method makes no distinction between local and remote layers.
1272 : /// Hence, the result **does not represent local filesystem usage**.
1273 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1274 0 : let guard = self.layers.read().await;
1275 0 : guard.layer_size_sum()
1276 0 : }
1277 :
1278 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1279 0 : self.metrics.resident_physical_size_get()
1280 0 : }
1281 :
1282 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1283 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1284 0 : }
1285 :
1286 : ///
1287 : /// Wait until WAL has been received and processed up to this LSN.
1288 : ///
1289 : /// You should call this before any of the other get_* or list_* functions. Calling
1290 : /// those functions with an LSN that has been processed yet is an error.
1291 : ///
1292 228125 : pub(crate) async fn wait_lsn(
1293 228125 : &self,
1294 228125 : lsn: Lsn,
1295 228125 : who_is_waiting: WaitLsnWaiter<'_>,
1296 228125 : ctx: &RequestContext, /* Prepare for use by cancellation */
1297 228125 : ) -> Result<(), WaitLsnError> {
1298 228125 : let state = self.current_state();
1299 228125 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1300 0 : return Err(WaitLsnError::Shutdown);
1301 228125 : } else if !matches!(state, TimelineState::Active) {
1302 0 : return Err(WaitLsnError::BadState(state));
1303 228125 : }
1304 228125 :
1305 228125 : if cfg!(debug_assertions) {
1306 228125 : match ctx.task_kind() {
1307 : TaskKind::WalReceiverManager
1308 : | TaskKind::WalReceiverConnectionHandler
1309 : | TaskKind::WalReceiverConnectionPoller => {
1310 0 : let is_myself = match who_is_waiting {
1311 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1312 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1313 : };
1314 0 : if is_myself {
1315 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1316 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1317 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1318 0 : }
1319 0 : } else {
1320 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1321 0 : // our walreceiver task can make progress independent of theirs
1322 0 : }
1323 : }
1324 228125 : _ => {}
1325 : }
1326 0 : }
1327 :
1328 228125 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1329 228125 :
1330 228125 : match self
1331 228125 : .last_record_lsn
1332 228125 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1333 0 : .await
1334 : {
1335 228125 : Ok(()) => Ok(()),
1336 0 : Err(e) => {
1337 : use utils::seqwait::SeqWaitError::*;
1338 0 : match e {
1339 0 : Shutdown => Err(WaitLsnError::Shutdown),
1340 : Timeout => {
1341 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1342 0 : drop(_timer);
1343 0 : let walreceiver_status = self.walreceiver_status();
1344 0 : Err(WaitLsnError::Timeout(format!(
1345 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1346 0 : lsn,
1347 0 : self.get_last_record_lsn(),
1348 0 : self.get_disk_consistent_lsn(),
1349 0 : walreceiver_status,
1350 0 : )))
1351 : }
1352 : }
1353 : }
1354 : }
1355 228125 : }
1356 :
1357 0 : pub(crate) fn walreceiver_status(&self) -> String {
1358 0 : match &*self.walreceiver.lock().unwrap() {
1359 0 : None => "stopping or stopped".to_string(),
1360 0 : Some(walreceiver) => match walreceiver.status() {
1361 0 : Some(status) => status.to_human_readable_string(),
1362 0 : None => "Not active".to_string(),
1363 : },
1364 : }
1365 0 : }
1366 :
1367 : /// Check that it is valid to request operations with that lsn.
1368 232 : pub(crate) fn check_lsn_is_in_scope(
1369 232 : &self,
1370 232 : lsn: Lsn,
1371 232 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1372 232 : ) -> anyhow::Result<()> {
1373 232 : ensure!(
1374 232 : lsn >= **latest_gc_cutoff_lsn,
1375 4 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1376 4 : lsn,
1377 4 : **latest_gc_cutoff_lsn,
1378 : );
1379 228 : Ok(())
1380 232 : }
1381 :
1382 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1383 10 : pub(crate) fn init_lsn_lease(
1384 10 : &self,
1385 10 : lsn: Lsn,
1386 10 : length: Duration,
1387 10 : ctx: &RequestContext,
1388 10 : ) -> anyhow::Result<LsnLease> {
1389 10 : self.make_lsn_lease(lsn, length, true, ctx)
1390 10 : }
1391 :
1392 : /// 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.
1393 4 : pub(crate) fn renew_lsn_lease(
1394 4 : &self,
1395 4 : lsn: Lsn,
1396 4 : length: Duration,
1397 4 : ctx: &RequestContext,
1398 4 : ) -> anyhow::Result<LsnLease> {
1399 4 : self.make_lsn_lease(lsn, length, false, ctx)
1400 4 : }
1401 :
1402 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1403 : ///
1404 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1405 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1406 : ///
1407 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1408 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1409 14 : fn make_lsn_lease(
1410 14 : &self,
1411 14 : lsn: Lsn,
1412 14 : length: Duration,
1413 14 : init: bool,
1414 14 : _ctx: &RequestContext,
1415 14 : ) -> anyhow::Result<LsnLease> {
1416 12 : let lease = {
1417 : // Normalize the requested LSN to be aligned, and move to the first record
1418 : // if it points to the beginning of the page (header).
1419 14 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1420 14 :
1421 14 : let mut gc_info = self.gc_info.write().unwrap();
1422 14 :
1423 14 : let valid_until = SystemTime::now() + length;
1424 14 :
1425 14 : let entry = gc_info.leases.entry(lsn);
1426 14 :
1427 14 : match entry {
1428 6 : Entry::Occupied(mut occupied) => {
1429 6 : let existing_lease = occupied.get_mut();
1430 6 : if valid_until > existing_lease.valid_until {
1431 2 : existing_lease.valid_until = valid_until;
1432 2 : let dt: DateTime<Utc> = valid_until.into();
1433 2 : info!("lease extended to {}", dt);
1434 : } else {
1435 4 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1436 4 : info!("existing lease covers greater length, valid until {}", dt);
1437 : }
1438 :
1439 6 : existing_lease.clone()
1440 : }
1441 8 : Entry::Vacant(vacant) => {
1442 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
1443 : // not blocked by the lsn lease deadline.
1444 8 : let validate = {
1445 8 : let conf = self.tenant_conf.load();
1446 8 : conf.location.attach_mode == AttachmentMode::Single
1447 8 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1448 : };
1449 :
1450 8 : if init || validate {
1451 8 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1452 8 : if lsn < *latest_gc_cutoff_lsn {
1453 2 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1454 6 : }
1455 0 : }
1456 :
1457 6 : let dt: DateTime<Utc> = valid_until.into();
1458 6 : info!("lease created, valid until {}", dt);
1459 6 : vacant.insert(LsnLease { valid_until }).clone()
1460 : }
1461 : }
1462 : };
1463 :
1464 12 : Ok(lease)
1465 14 : }
1466 :
1467 : /// Flush to disk all data that was written with the put_* functions
1468 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))]
1469 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1470 : self.freeze_and_flush0().await
1471 : }
1472 :
1473 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1474 : // polluting the span hierarchy.
1475 1092 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1476 1092 : let token = {
1477 : // Freeze the current open in-memory layer. It will be written to disk on next
1478 : // iteration.
1479 1092 : let mut g = self.write_lock.lock().await;
1480 :
1481 1092 : let to_lsn = self.get_last_record_lsn();
1482 1092 : self.freeze_inmem_layer_at(to_lsn, &mut g).await?
1483 : };
1484 1092 : self.wait_flush_completion(token).await
1485 1092 : }
1486 :
1487 : // Check if an open ephemeral layer should be closed: this provides
1488 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1489 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1490 : // ephemeral layer bytes has been breached.
1491 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1492 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1493 : // If the write lock is held, there is an active wal receiver: rolling open layers
1494 : // is their responsibility while they hold this lock.
1495 0 : return;
1496 : };
1497 :
1498 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1499 : // time, and this was missed.
1500 : // if write_guard.is_none() { return; }
1501 :
1502 0 : let Ok(layers_guard) = self.layers.try_read() else {
1503 : // Don't block if the layer lock is busy
1504 0 : return;
1505 : };
1506 :
1507 0 : let Ok(lm) = layers_guard.layer_map() else {
1508 0 : return;
1509 : };
1510 :
1511 0 : let Some(open_layer) = &lm.open_layer else {
1512 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1513 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1514 : // that didn't result in writes to this shard.
1515 :
1516 : // Must not hold the layers lock while waiting for a flush.
1517 0 : drop(layers_guard);
1518 0 :
1519 0 : let last_record_lsn = self.get_last_record_lsn();
1520 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1521 0 : if last_record_lsn > disk_consistent_lsn {
1522 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1523 : // we are a sharded tenant and have skipped some WAL
1524 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1525 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1526 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1527 : // without any data ingested (yet) doesn't write a remote index as soon as it
1528 : // sees its LSN advance: we only do this if we've been layer-less
1529 : // for some time.
1530 0 : tracing::debug!(
1531 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1532 : disk_consistent_lsn,
1533 : last_record_lsn
1534 : );
1535 :
1536 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1537 : // We know there is no open layer, so we can request freezing without actually
1538 : // freezing anything. This is true even if we have dropped the layers_guard, we
1539 : // still hold the write_guard.
1540 0 : let _ = async {
1541 0 : let token = self
1542 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1543 0 : .await?;
1544 0 : self.wait_flush_completion(token).await
1545 0 : }
1546 0 : .await;
1547 0 : }
1548 0 : }
1549 :
1550 0 : return;
1551 : };
1552 :
1553 0 : let Some(current_size) = open_layer.try_len() else {
1554 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1555 : // read lock to get size should always succeed.
1556 0 : tracing::warn!("Lock conflict while reading size of open layer");
1557 0 : return;
1558 : };
1559 :
1560 0 : let current_lsn = self.get_last_record_lsn();
1561 :
1562 0 : let checkpoint_distance_override = open_layer.tick().await;
1563 :
1564 0 : if let Some(size_override) = checkpoint_distance_override {
1565 0 : if current_size > size_override {
1566 : // This is not harmful, but it only happens in relatively rare cases where
1567 : // time-based checkpoints are not happening fast enough to keep the amount of
1568 : // ephemeral data within configured limits. It's a sign of stress on the system.
1569 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1570 0 : }
1571 0 : }
1572 :
1573 0 : let checkpoint_distance =
1574 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1575 0 :
1576 0 : if self.should_roll(
1577 0 : current_size,
1578 0 : current_size,
1579 0 : checkpoint_distance,
1580 0 : self.get_last_record_lsn(),
1581 0 : self.last_freeze_at.load(),
1582 0 : open_layer.get_opened_at(),
1583 0 : ) {
1584 0 : match open_layer.info() {
1585 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1586 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1587 0 : // happens asynchronously in the background.
1588 0 : tracing::debug!(
1589 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1590 : );
1591 : }
1592 : InMemoryLayerInfo::Open { .. } => {
1593 : // Upgrade to a write lock and freeze the layer
1594 0 : drop(layers_guard);
1595 0 : let res = self
1596 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1597 0 : .await;
1598 :
1599 0 : if let Err(e) = res {
1600 0 : tracing::info!(
1601 0 : "failed to flush frozen layer after background freeze: {e:#}"
1602 : );
1603 0 : }
1604 : }
1605 : }
1606 0 : }
1607 0 : }
1608 :
1609 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1610 : ///
1611 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1612 : /// child timelines that are not offloaded yet.
1613 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1614 0 : if self.remote_client.is_archived() != Some(true) {
1615 0 : return (false, "the timeline is not archived");
1616 0 : }
1617 0 : if !self.remote_client.no_pending_work() {
1618 : // if the remote client is still processing some work, we can't offload
1619 0 : return (false, "the upload queue is not drained yet");
1620 0 : }
1621 0 :
1622 0 : (true, "ok")
1623 0 : }
1624 :
1625 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1626 : /// compaction tasks.
1627 364 : pub(crate) async fn compact(
1628 364 : self: &Arc<Self>,
1629 364 : cancel: &CancellationToken,
1630 364 : flags: EnumSet<CompactFlags>,
1631 364 : ctx: &RequestContext,
1632 364 : ) -> Result<bool, CompactionError> {
1633 364 : self.compact_with_options(
1634 364 : cancel,
1635 364 : CompactOptions {
1636 364 : flags,
1637 364 : compact_range: None,
1638 364 : },
1639 364 : ctx,
1640 364 : )
1641 37078 : .await
1642 364 : }
1643 :
1644 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1645 : /// compaction tasks.
1646 364 : pub(crate) async fn compact_with_options(
1647 364 : self: &Arc<Self>,
1648 364 : cancel: &CancellationToken,
1649 364 : options: CompactOptions,
1650 364 : ctx: &RequestContext,
1651 364 : ) -> Result<bool, CompactionError> {
1652 364 : // most likely the cancellation token is from background task, but in tests it could be the
1653 364 : // request task as well.
1654 364 :
1655 364 : let prepare = async move {
1656 364 : let guard = self.compaction_lock.lock().await;
1657 :
1658 364 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1659 364 : BackgroundLoopKind::Compaction,
1660 364 : ctx,
1661 364 : )
1662 0 : .await;
1663 :
1664 364 : (guard, permit)
1665 364 : };
1666 :
1667 : // this wait probably never needs any "long time spent" logging, because we already nag if
1668 : // compaction task goes over it's period (20s) which is quite often in production.
1669 364 : let (_guard, _permit) = tokio::select! {
1670 364 : tuple = prepare => { tuple },
1671 364 : _ = self.cancel.cancelled() => return Ok(false),
1672 364 : _ = cancel.cancelled() => return Ok(false),
1673 : };
1674 :
1675 364 : let last_record_lsn = self.get_last_record_lsn();
1676 364 :
1677 364 : // Last record Lsn could be zero in case the timeline was just created
1678 364 : if !last_record_lsn.is_valid() {
1679 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1680 0 : return Ok(false);
1681 364 : }
1682 364 :
1683 364 : match self.get_compaction_algorithm_settings().kind {
1684 : CompactionAlgorithm::Tiered => {
1685 0 : self.compact_tiered(cancel, ctx).await?;
1686 0 : Ok(false)
1687 : }
1688 37078 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1689 : }
1690 364 : }
1691 :
1692 : /// Mutate the timeline with a [`TimelineWriter`].
1693 5133162 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1694 5133162 : TimelineWriter {
1695 5133162 : tl: self,
1696 5133162 : write_guard: self.write_lock.lock().await,
1697 : }
1698 5133162 : }
1699 :
1700 0 : pub(crate) fn activate(
1701 0 : self: &Arc<Self>,
1702 0 : parent: Arc<crate::tenant::Tenant>,
1703 0 : broker_client: BrokerClientChannel,
1704 0 : background_jobs_can_start: Option<&completion::Barrier>,
1705 0 : ctx: &RequestContext,
1706 0 : ) {
1707 0 : if self.tenant_shard_id.is_shard_zero() {
1708 0 : // Logical size is only maintained accurately on shard zero.
1709 0 : self.spawn_initial_logical_size_computation_task(ctx);
1710 0 : }
1711 0 : self.launch_wal_receiver(ctx, broker_client);
1712 0 : self.set_state(TimelineState::Active);
1713 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1714 0 : }
1715 :
1716 : /// After this function returns, there are no timeline-scoped tasks are left running.
1717 : ///
1718 : /// The preferred pattern for is:
1719 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1720 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1721 : /// go the extra mile and keep track of JoinHandles
1722 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1723 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1724 : ///
1725 : /// For legacy reasons, we still have multiple tasks spawned using
1726 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1727 : /// We refer to these as "timeline-scoped task_mgr tasks".
1728 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1729 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1730 : /// or [`task_mgr::shutdown_watcher`].
1731 : /// We want to gradually convert the code base away from these.
1732 : ///
1733 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1734 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1735 : /// ones that aren't mentioned here):
1736 : /// - [`TaskKind::TimelineDeletionWorker`]
1737 : /// - NB: also used for tenant deletion
1738 : /// - [`TaskKind::RemoteUploadTask`]`
1739 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1740 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1741 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1742 : /// - [`TaskKind::Eviction`]
1743 : /// - [`TaskKind::LayerFlushTask`]
1744 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1745 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1746 10 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1747 10 : debug_assert_current_span_has_tenant_and_timeline_id();
1748 10 :
1749 10 : // Regardless of whether we're going to try_freeze_and_flush
1750 10 : // or not, stop ingesting any more data. Walreceiver only provides
1751 10 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1752 10 : // So, only after the self.gate.close() below will we know for sure that
1753 10 : // no walreceiver tasks are left.
1754 10 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1755 10 : // data during the call to `self.freeze_and_flush()` below.
1756 10 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1757 10 : // which is what we'd need to properly model early shutdown of the walreceiver
1758 10 : // task sub-tree before the other Timeline task sub-trees.
1759 10 : let walreceiver = self.walreceiver.lock().unwrap().take();
1760 10 : tracing::debug!(
1761 0 : is_some = walreceiver.is_some(),
1762 0 : "Waiting for WalReceiverManager..."
1763 : );
1764 10 : if let Some(walreceiver) = walreceiver {
1765 0 : walreceiver.cancel();
1766 10 : }
1767 : // ... and inform any waiters for newer LSNs that there won't be any.
1768 10 : self.last_record_lsn.shutdown();
1769 10 :
1770 10 : if let ShutdownMode::FreezeAndFlush = mode {
1771 6 : if let Some((open, frozen)) = self
1772 6 : .layers
1773 6 : .read()
1774 0 : .await
1775 6 : .layer_map()
1776 6 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1777 6 : .ok()
1778 6 : .filter(|(open, frozen)| *open || *frozen > 0)
1779 : {
1780 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1781 6 : } else {
1782 6 : // this is double-shutdown, ignore it
1783 6 : }
1784 :
1785 : // we shut down walreceiver above, so, we won't add anything more
1786 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1787 : // to reach the disk & upload queue, then shut the upload queue and
1788 : // wait for it to drain.
1789 6 : match self.freeze_and_flush().await {
1790 : Ok(_) => {
1791 : // drain the upload queue
1792 : // if we did not wait for completion here, it might be our shutdown process
1793 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1794 : // be spawned.
1795 : //
1796 : // what is problematic is the shutting down of RemoteTimelineClient, because
1797 : // obviously it does not make sense to stop while we wait for it, but what
1798 : // about corner cases like s3 suddenly hanging up?
1799 6 : self.remote_client.shutdown().await;
1800 : }
1801 : Err(FlushLayerError::Cancelled) => {
1802 : // this is likely the second shutdown, ignore silently.
1803 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1804 0 : debug_assert!(self.cancel.is_cancelled());
1805 : }
1806 0 : Err(e) => {
1807 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1808 0 : // we have some extra WAL replay to do next time the timeline starts.
1809 0 : warn!("failed to freeze and flush: {e:#}");
1810 : }
1811 : }
1812 :
1813 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
1814 : // we also do a final check here to ensure that the queue is empty.
1815 6 : if !self.remote_client.no_pending_work() {
1816 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1817 6 : }
1818 4 : }
1819 :
1820 10 : if let ShutdownMode::Flush = mode {
1821 : // drain the upload queue
1822 2 : self.remote_client.shutdown().await;
1823 2 : if !self.remote_client.no_pending_work() {
1824 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1825 2 : }
1826 8 : }
1827 :
1828 : // Signal any subscribers to our cancellation token to drop out
1829 10 : tracing::debug!("Cancelling CancellationToken");
1830 10 : self.cancel.cancel();
1831 10 :
1832 10 : // Ensure Prevent new page service requests from starting.
1833 10 : self.handles.shutdown();
1834 10 :
1835 10 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1836 10 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1837 10 : self.remote_client.stop();
1838 10 :
1839 10 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1840 10 : // to shut down the upload queue tasks.
1841 10 : // TODO: fix that, task management should be encapsulated inside remote_client.
1842 10 : task_mgr::shutdown_tasks(
1843 10 : Some(TaskKind::RemoteUploadTask),
1844 10 : Some(self.tenant_shard_id),
1845 10 : Some(self.timeline_id),
1846 10 : )
1847 0 : .await;
1848 :
1849 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
1850 10 : tracing::debug!("Waiting for tasks...");
1851 10 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1852 :
1853 : {
1854 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
1855 : // open.
1856 10 : let mut write_guard = self.write_lock.lock().await;
1857 10 : self.layers.write().await.shutdown(&mut write_guard);
1858 10 : }
1859 10 :
1860 10 : // Finally wait until any gate-holders are complete.
1861 10 : //
1862 10 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1863 10 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1864 10 : self.gate.close().await;
1865 :
1866 10 : self.metrics.shutdown();
1867 10 : }
1868 :
1869 420 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1870 420 : match (self.current_state(), new_state) {
1871 420 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1872 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1873 : }
1874 0 : (st, TimelineState::Loading) => {
1875 0 : error!("ignoring transition from {st:?} into Loading state");
1876 : }
1877 0 : (TimelineState::Broken { .. }, new_state) => {
1878 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1879 : }
1880 : (TimelineState::Stopping, TimelineState::Active) => {
1881 0 : error!("Not activating a Stopping timeline");
1882 : }
1883 418 : (_, new_state) => {
1884 418 : self.state.send_replace(new_state);
1885 418 : }
1886 : }
1887 420 : }
1888 :
1889 2 : pub(crate) fn set_broken(&self, reason: String) {
1890 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1891 2 : let broken_state = TimelineState::Broken {
1892 2 : reason,
1893 2 : backtrace: backtrace_str,
1894 2 : };
1895 2 : self.set_state(broken_state);
1896 2 :
1897 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1898 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1899 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1900 2 : self.cancel.cancel();
1901 2 : }
1902 :
1903 229135 : pub(crate) fn current_state(&self) -> TimelineState {
1904 229135 : self.state.borrow().clone()
1905 229135 : }
1906 :
1907 6 : pub(crate) fn is_broken(&self) -> bool {
1908 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1909 6 : }
1910 :
1911 222 : pub(crate) fn is_active(&self) -> bool {
1912 222 : self.current_state() == TimelineState::Active
1913 222 : }
1914 :
1915 2 : pub(crate) fn is_archived(&self) -> Option<bool> {
1916 2 : self.remote_client.is_archived()
1917 2 : }
1918 :
1919 368 : pub(crate) fn is_stopping(&self) -> bool {
1920 368 : self.current_state() == TimelineState::Stopping
1921 368 : }
1922 :
1923 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1924 0 : self.state.subscribe()
1925 0 : }
1926 :
1927 228127 : pub(crate) async fn wait_to_become_active(
1928 228127 : &self,
1929 228127 : _ctx: &RequestContext, // Prepare for use by cancellation
1930 228127 : ) -> Result<(), TimelineState> {
1931 228127 : let mut receiver = self.state.subscribe();
1932 : loop {
1933 228127 : let current_state = receiver.borrow().clone();
1934 228127 : match current_state {
1935 : TimelineState::Loading => {
1936 0 : receiver
1937 0 : .changed()
1938 0 : .await
1939 0 : .expect("holding a reference to self");
1940 : }
1941 : TimelineState::Active { .. } => {
1942 228125 : return Ok(());
1943 : }
1944 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1945 : // There's no chance the timeline can transition back into ::Active
1946 2 : return Err(current_state);
1947 : }
1948 : }
1949 : }
1950 228127 : }
1951 :
1952 0 : pub(crate) async fn layer_map_info(
1953 0 : &self,
1954 0 : reset: LayerAccessStatsReset,
1955 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
1956 0 : let guard = self.layers.read().await;
1957 0 : let layer_map = guard.layer_map()?;
1958 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
1959 0 : if let Some(open_layer) = &layer_map.open_layer {
1960 0 : in_memory_layers.push(open_layer.info());
1961 0 : }
1962 0 : for frozen_layer in &layer_map.frozen_layers {
1963 0 : in_memory_layers.push(frozen_layer.info());
1964 0 : }
1965 :
1966 0 : let historic_layers = layer_map
1967 0 : .iter_historic_layers()
1968 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
1969 0 : .collect();
1970 0 :
1971 0 : Ok(LayerMapInfo {
1972 0 : in_memory_layers,
1973 0 : historic_layers,
1974 0 : })
1975 0 : }
1976 :
1977 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))]
1978 : pub(crate) async fn download_layer(
1979 : &self,
1980 : layer_file_name: &LayerName,
1981 : ) -> anyhow::Result<Option<bool>> {
1982 : let Some(layer) = self.find_layer(layer_file_name).await? else {
1983 : return Ok(None);
1984 : };
1985 :
1986 : layer.download().await?;
1987 :
1988 : Ok(Some(true))
1989 : }
1990 :
1991 : /// Evict just one layer.
1992 : ///
1993 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
1994 0 : pub(crate) async fn evict_layer(
1995 0 : &self,
1996 0 : layer_file_name: &LayerName,
1997 0 : ) -> anyhow::Result<Option<bool>> {
1998 0 : let _gate = self
1999 0 : .gate
2000 0 : .enter()
2001 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2002 :
2003 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2004 0 : return Ok(None);
2005 : };
2006 :
2007 : // curl has this by default
2008 0 : let timeout = std::time::Duration::from_secs(120);
2009 0 :
2010 0 : match local_layer.evict_and_wait(timeout).await {
2011 0 : Ok(()) => Ok(Some(true)),
2012 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2013 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2014 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2015 : }
2016 0 : }
2017 :
2018 4803010 : fn should_roll(
2019 4803010 : &self,
2020 4803010 : layer_size: u64,
2021 4803010 : projected_layer_size: u64,
2022 4803010 : checkpoint_distance: u64,
2023 4803010 : projected_lsn: Lsn,
2024 4803010 : last_freeze_at: Lsn,
2025 4803010 : opened_at: Instant,
2026 4803010 : ) -> bool {
2027 4803010 : let distance = projected_lsn.widening_sub(last_freeze_at);
2028 4803010 :
2029 4803010 : // Rolling the open layer can be triggered by:
2030 4803010 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2031 4803010 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2032 4803010 : // account for how writes are distributed across shards: we expect each node to consume
2033 4803010 : // 1/count of the LSN on average.
2034 4803010 : // 2. The size of the currently open layer.
2035 4803010 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2036 4803010 : // up and suspend activity.
2037 4803010 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2038 0 : info!(
2039 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2040 : projected_lsn, layer_size, distance
2041 : );
2042 :
2043 0 : true
2044 4803010 : } else if projected_layer_size >= checkpoint_distance {
2045 : // NB: this check is relied upon by:
2046 80 : let _ = IndexEntry::validate_checkpoint_distance;
2047 80 : info!(
2048 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2049 : projected_lsn, layer_size, projected_layer_size
2050 : );
2051 :
2052 80 : true
2053 4802930 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2054 0 : info!(
2055 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2056 0 : projected_lsn,
2057 0 : layer_size,
2058 0 : opened_at.elapsed()
2059 : );
2060 :
2061 0 : true
2062 : } else {
2063 4802930 : false
2064 : }
2065 4803010 : }
2066 : }
2067 :
2068 : /// Number of times we will compute partition within a checkpoint distance.
2069 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2070 :
2071 : // Private functions
2072 : impl Timeline {
2073 12 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2074 12 : let tenant_conf = self.tenant_conf.load();
2075 12 : tenant_conf
2076 12 : .tenant_conf
2077 12 : .lsn_lease_length
2078 12 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2079 12 : }
2080 :
2081 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2082 0 : let tenant_conf = self.tenant_conf.load();
2083 0 : tenant_conf
2084 0 : .tenant_conf
2085 0 : .lsn_lease_length_for_ts
2086 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2087 0 : }
2088 :
2089 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2090 0 : let tenant_conf = self.tenant_conf.load();
2091 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2092 0 : }
2093 :
2094 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2095 0 : let tenant_conf = self.tenant_conf.load();
2096 0 : tenant_conf
2097 0 : .tenant_conf
2098 0 : .lazy_slru_download
2099 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2100 0 : }
2101 :
2102 4804616 : fn get_checkpoint_distance(&self) -> u64 {
2103 4804616 : let tenant_conf = self.tenant_conf.load();
2104 4804616 : tenant_conf
2105 4804616 : .tenant_conf
2106 4804616 : .checkpoint_distance
2107 4804616 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2108 4804616 : }
2109 :
2110 4802930 : fn get_checkpoint_timeout(&self) -> Duration {
2111 4802930 : let tenant_conf = self.tenant_conf.load();
2112 4802930 : tenant_conf
2113 4802930 : .tenant_conf
2114 4802930 : .checkpoint_timeout
2115 4802930 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2116 4802930 : }
2117 :
2118 618 : fn get_compaction_target_size(&self) -> u64 {
2119 618 : let tenant_conf = self.tenant_conf.load();
2120 618 : tenant_conf
2121 618 : .tenant_conf
2122 618 : .compaction_target_size
2123 618 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2124 618 : }
2125 :
2126 1536 : fn get_compaction_threshold(&self) -> usize {
2127 1536 : let tenant_conf = self.tenant_conf.load();
2128 1536 : tenant_conf
2129 1536 : .tenant_conf
2130 1536 : .compaction_threshold
2131 1536 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2132 1536 : }
2133 :
2134 14 : fn get_image_creation_threshold(&self) -> usize {
2135 14 : let tenant_conf = self.tenant_conf.load();
2136 14 : tenant_conf
2137 14 : .tenant_conf
2138 14 : .image_creation_threshold
2139 14 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2140 14 : }
2141 :
2142 364 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2143 364 : let tenant_conf = &self.tenant_conf.load();
2144 364 : tenant_conf
2145 364 : .tenant_conf
2146 364 : .compaction_algorithm
2147 364 : .as_ref()
2148 364 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2149 364 : .clone()
2150 364 : }
2151 :
2152 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2153 0 : let tenant_conf = self.tenant_conf.load();
2154 0 : tenant_conf
2155 0 : .tenant_conf
2156 0 : .eviction_policy
2157 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2158 0 : }
2159 :
2160 418 : fn get_evictions_low_residence_duration_metric_threshold(
2161 418 : tenant_conf: &TenantConfOpt,
2162 418 : default_tenant_conf: &TenantConf,
2163 418 : ) -> Duration {
2164 418 : tenant_conf
2165 418 : .evictions_low_residence_duration_metric_threshold
2166 418 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2167 418 : }
2168 :
2169 716 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2170 716 : let tenant_conf = self.tenant_conf.load();
2171 716 : tenant_conf
2172 716 : .tenant_conf
2173 716 : .image_layer_creation_check_threshold
2174 716 : .unwrap_or(
2175 716 : self.conf
2176 716 : .default_tenant_conf
2177 716 : .image_layer_creation_check_threshold,
2178 716 : )
2179 716 : }
2180 :
2181 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2182 0 : // NB: Most tenant conf options are read by background loops, so,
2183 0 : // changes will automatically be picked up.
2184 0 :
2185 0 : // The threshold is embedded in the metric. So, we need to update it.
2186 0 : {
2187 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2188 0 : &new_conf.tenant_conf,
2189 0 : &self.conf.default_tenant_conf,
2190 0 : );
2191 0 :
2192 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2193 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2194 0 :
2195 0 : let timeline_id_str = self.timeline_id.to_string();
2196 0 :
2197 0 : self.remote_client.update_config(&new_conf.location);
2198 0 :
2199 0 : self.metrics
2200 0 : .evictions_with_low_residence_duration
2201 0 : .write()
2202 0 : .unwrap()
2203 0 : .change_threshold(
2204 0 : &tenant_id_str,
2205 0 : &shard_id_str,
2206 0 : &timeline_id_str,
2207 0 : new_threshold,
2208 0 : );
2209 0 : }
2210 0 : }
2211 :
2212 : /// Open a Timeline handle.
2213 : ///
2214 : /// Loads the metadata for the timeline into memory, but not the layer map.
2215 : #[allow(clippy::too_many_arguments)]
2216 418 : pub(super) fn new(
2217 418 : conf: &'static PageServerConf,
2218 418 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2219 418 : metadata: &TimelineMetadata,
2220 418 : ancestor: Option<Arc<Timeline>>,
2221 418 : timeline_id: TimelineId,
2222 418 : tenant_shard_id: TenantShardId,
2223 418 : generation: Generation,
2224 418 : shard_identity: ShardIdentity,
2225 418 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2226 418 : resources: TimelineResources,
2227 418 : pg_version: u32,
2228 418 : state: TimelineState,
2229 418 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2230 418 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2231 418 : cancel: CancellationToken,
2232 418 : ) -> Arc<Self> {
2233 418 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2234 418 : let (state, _) = watch::channel(state);
2235 418 :
2236 418 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2237 418 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2238 418 :
2239 418 : let evictions_low_residence_duration_metric_threshold = {
2240 418 : let loaded_tenant_conf = tenant_conf.load();
2241 418 : Self::get_evictions_low_residence_duration_metric_threshold(
2242 418 : &loaded_tenant_conf.tenant_conf,
2243 418 : &conf.default_tenant_conf,
2244 418 : )
2245 : };
2246 :
2247 418 : if let Some(ancestor) = &ancestor {
2248 230 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2249 230 : // If we construct an explicit timeline object, it's obviously not offloaded
2250 230 : let is_offloaded = MaybeOffloaded::No;
2251 230 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2252 230 : }
2253 :
2254 418 : Arc::new_cyclic(|myself| {
2255 418 : let metrics = TimelineMetrics::new(
2256 418 : &tenant_shard_id,
2257 418 : &timeline_id,
2258 418 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2259 418 : "mtime",
2260 418 : evictions_low_residence_duration_metric_threshold,
2261 418 : ),
2262 418 : );
2263 418 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2264 :
2265 418 : let mut result = Timeline {
2266 418 : conf,
2267 418 : tenant_conf,
2268 418 : myself: myself.clone(),
2269 418 : timeline_id,
2270 418 : tenant_shard_id,
2271 418 : generation,
2272 418 : shard_identity,
2273 418 : pg_version,
2274 418 : layers: Default::default(),
2275 418 :
2276 418 : walredo_mgr,
2277 418 : walreceiver: Mutex::new(None),
2278 418 :
2279 418 : remote_client: Arc::new(resources.remote_client),
2280 418 :
2281 418 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2282 418 : last_record_lsn: SeqWait::new(RecordLsn {
2283 418 : last: disk_consistent_lsn,
2284 418 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2285 418 : }),
2286 418 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2287 418 :
2288 418 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2289 418 : last_freeze_ts: RwLock::new(Instant::now()),
2290 418 :
2291 418 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2292 418 :
2293 418 : ancestor_timeline: ancestor,
2294 418 : ancestor_lsn: metadata.ancestor_lsn(),
2295 418 :
2296 418 : metrics,
2297 418 :
2298 418 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2299 418 : &tenant_shard_id,
2300 418 : &timeline_id,
2301 418 : ),
2302 418 :
2303 2926 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2304 418 :
2305 418 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2306 418 :
2307 418 : layer_flush_start_tx,
2308 418 : layer_flush_done_tx,
2309 418 :
2310 418 : write_lock: tokio::sync::Mutex::new(None),
2311 418 :
2312 418 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2313 418 :
2314 418 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2315 418 : initdb_lsn: metadata.initdb_lsn(),
2316 418 :
2317 418 : current_logical_size: if disk_consistent_lsn.is_valid() {
2318 : // we're creating timeline data with some layer files existing locally,
2319 : // need to recalculate timeline's logical size based on data in the layers.
2320 234 : LogicalSize::deferred_initial(disk_consistent_lsn)
2321 : } else {
2322 : // we're creating timeline data without any layers existing locally,
2323 : // initial logical size is 0.
2324 184 : LogicalSize::empty_initial()
2325 : },
2326 418 : partitioning: tokio::sync::Mutex::new((
2327 418 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2328 418 : Lsn(0),
2329 418 : )),
2330 418 : repartition_threshold: 0,
2331 418 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2332 418 : last_image_layer_creation_check_instant: Mutex::new(None),
2333 418 :
2334 418 : last_received_wal: Mutex::new(None),
2335 418 : rel_size_cache: RwLock::new(RelSizeCache {
2336 418 : complete_as_of: disk_consistent_lsn,
2337 418 : map: HashMap::new(),
2338 418 : }),
2339 418 :
2340 418 : download_all_remote_layers_task_info: RwLock::new(None),
2341 418 :
2342 418 : state,
2343 418 :
2344 418 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2345 418 : EvictionTaskTimelineState::default(),
2346 418 : ),
2347 418 : delete_progress: TimelineDeleteProgress::default(),
2348 418 :
2349 418 : cancel,
2350 418 : gate: Gate::default(),
2351 418 :
2352 418 : compaction_lock: tokio::sync::Mutex::default(),
2353 418 : gc_lock: tokio::sync::Mutex::default(),
2354 418 :
2355 418 : standby_horizon: AtomicLsn::new(0),
2356 418 :
2357 418 : timeline_get_throttle: resources.timeline_get_throttle,
2358 418 :
2359 418 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2360 418 :
2361 418 : #[cfg(test)]
2362 418 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2363 418 :
2364 418 : l0_flush_global_state: resources.l0_flush_global_state,
2365 418 :
2366 418 : handles: Default::default(),
2367 418 :
2368 418 : attach_wal_lag_cooldown,
2369 418 :
2370 418 : create_idempotency,
2371 418 : };
2372 418 :
2373 418 : result.repartition_threshold =
2374 418 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2375 418 :
2376 418 : result
2377 418 : .metrics
2378 418 : .last_record_gauge
2379 418 : .set(disk_consistent_lsn.0 as i64);
2380 418 : result
2381 418 : })
2382 418 : }
2383 :
2384 588 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2385 588 : let Ok(guard) = self.gate.enter() else {
2386 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2387 0 : return;
2388 : };
2389 588 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2390 588 : match *flush_loop_state {
2391 412 : FlushLoopState::NotStarted => (),
2392 : FlushLoopState::Running { .. } => {
2393 176 : info!(
2394 0 : "skipping attempt to start flush_loop twice {}/{}",
2395 0 : self.tenant_shard_id, self.timeline_id
2396 : );
2397 176 : return;
2398 : }
2399 : FlushLoopState::Exited => {
2400 0 : warn!(
2401 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2402 0 : self.tenant_shard_id, self.timeline_id
2403 : );
2404 0 : return;
2405 : }
2406 : }
2407 :
2408 412 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2409 412 : let self_clone = Arc::clone(self);
2410 412 :
2411 412 : debug!("spawning flush loop");
2412 412 : *flush_loop_state = FlushLoopState::Running {
2413 412 : #[cfg(test)]
2414 412 : expect_initdb_optimization: false,
2415 412 : #[cfg(test)]
2416 412 : initdb_optimization_count: 0,
2417 412 : };
2418 412 : task_mgr::spawn(
2419 412 : task_mgr::BACKGROUND_RUNTIME.handle(),
2420 412 : task_mgr::TaskKind::LayerFlushTask,
2421 412 : self.tenant_shard_id,
2422 412 : Some(self.timeline_id),
2423 412 : "layer flush task",
2424 412 : async move {
2425 412 : let _guard = guard;
2426 412 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2427 18014 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2428 10 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2429 10 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2430 10 : *flush_loop_state = FlushLoopState::Exited;
2431 10 : Ok(())
2432 10 : }
2433 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))
2434 : );
2435 588 : }
2436 :
2437 : /// Creates and starts the wal receiver.
2438 : ///
2439 : /// This function is expected to be called at most once per Timeline's lifecycle
2440 : /// when the timeline is activated.
2441 0 : fn launch_wal_receiver(
2442 0 : self: &Arc<Self>,
2443 0 : ctx: &RequestContext,
2444 0 : broker_client: BrokerClientChannel,
2445 0 : ) {
2446 0 : info!(
2447 0 : "launching WAL receiver for timeline {} of tenant {}",
2448 0 : self.timeline_id, self.tenant_shard_id
2449 : );
2450 :
2451 0 : let tenant_conf = self.tenant_conf.load();
2452 0 : let wal_connect_timeout = tenant_conf
2453 0 : .tenant_conf
2454 0 : .walreceiver_connect_timeout
2455 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2456 0 : let lagging_wal_timeout = tenant_conf
2457 0 : .tenant_conf
2458 0 : .lagging_wal_timeout
2459 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2460 0 : let max_lsn_wal_lag = tenant_conf
2461 0 : .tenant_conf
2462 0 : .max_lsn_wal_lag
2463 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2464 0 :
2465 0 : let mut guard = self.walreceiver.lock().unwrap();
2466 0 : assert!(
2467 0 : guard.is_none(),
2468 0 : "multiple launches / re-launches of WAL receiver are not supported"
2469 : );
2470 0 : *guard = Some(WalReceiver::start(
2471 0 : Arc::clone(self),
2472 0 : WalReceiverConf {
2473 0 : protocol: self.conf.wal_receiver_protocol,
2474 0 : wal_connect_timeout,
2475 0 : lagging_wal_timeout,
2476 0 : max_lsn_wal_lag,
2477 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2478 0 : availability_zone: self.conf.availability_zone.clone(),
2479 0 : ingest_batch_size: self.conf.ingest_batch_size,
2480 0 : },
2481 0 : broker_client,
2482 0 : ctx,
2483 0 : ));
2484 0 : }
2485 :
2486 : /// Initialize with an empty layer map. Used when creating a new timeline.
2487 412 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2488 412 : let mut layers = self.layers.try_write().expect(
2489 412 : "in the context where we call this function, no other task has access to the object",
2490 412 : );
2491 412 : layers
2492 412 : .open_mut()
2493 412 : .expect("in this context the LayerManager must still be open")
2494 412 : .initialize_empty(Lsn(start_lsn.0));
2495 412 : }
2496 :
2497 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2498 : /// files.
2499 6 : pub(super) async fn load_layer_map(
2500 6 : &self,
2501 6 : disk_consistent_lsn: Lsn,
2502 6 : index_part: IndexPart,
2503 6 : ) -> anyhow::Result<()> {
2504 : use init::{Decision::*, Discovered, DismissedLayer};
2505 : use LayerName::*;
2506 :
2507 6 : let mut guard = self.layers.write().await;
2508 :
2509 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2510 6 :
2511 6 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2512 6 : // structs representing all files on disk
2513 6 : let timeline_path = self
2514 6 : .conf
2515 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2516 6 : let conf = self.conf;
2517 6 : let span = tracing::Span::current();
2518 6 :
2519 6 : // Copy to move into the task we're about to spawn
2520 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2521 :
2522 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2523 6 : move || {
2524 6 : let _g = span.entered();
2525 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2526 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2527 6 : let mut unrecognized_files = Vec::new();
2528 6 :
2529 6 : let mut path = timeline_path;
2530 :
2531 22 : for discovered in discovered {
2532 16 : let (name, kind) = match discovered {
2533 16 : Discovered::Layer(layer_file_name, local_metadata) => {
2534 16 : discovered_layers.push((layer_file_name, local_metadata));
2535 16 : continue;
2536 : }
2537 0 : Discovered::IgnoredBackup(path) => {
2538 0 : std::fs::remove_file(path)
2539 0 : .or_else(fs_ext::ignore_not_found)
2540 0 : .fatal_err("Removing .old file");
2541 0 : continue;
2542 : }
2543 0 : Discovered::Unknown(file_name) => {
2544 0 : // we will later error if there are any
2545 0 : unrecognized_files.push(file_name);
2546 0 : continue;
2547 : }
2548 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2549 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2550 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2551 : };
2552 0 : path.push(Utf8Path::new(&name));
2553 0 : init::cleanup(&path, kind)?;
2554 0 : path.pop();
2555 : }
2556 :
2557 6 : if !unrecognized_files.is_empty() {
2558 : // assume that if there are any there are many many.
2559 0 : let n = unrecognized_files.len();
2560 0 : let first = &unrecognized_files[..n.min(10)];
2561 0 : anyhow::bail!(
2562 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2563 0 : );
2564 6 : }
2565 6 :
2566 6 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2567 6 :
2568 6 : let mut loaded_layers = Vec::new();
2569 6 : let mut needs_cleanup = Vec::new();
2570 6 : let mut total_physical_size = 0;
2571 :
2572 22 : for (name, decision) in decided {
2573 16 : let decision = match decision {
2574 16 : Ok(decision) => decision,
2575 0 : Err(DismissedLayer::Future { local }) => {
2576 0 : if let Some(local) = local {
2577 0 : init::cleanup_future_layer(
2578 0 : &local.local_path,
2579 0 : &name,
2580 0 : disk_consistent_lsn,
2581 0 : )?;
2582 0 : }
2583 0 : needs_cleanup.push(name);
2584 0 : continue;
2585 : }
2586 0 : Err(DismissedLayer::LocalOnly(local)) => {
2587 0 : init::cleanup_local_only_file(&name, &local)?;
2588 : // this file never existed remotely, we will have to do rework
2589 0 : continue;
2590 : }
2591 0 : Err(DismissedLayer::BadMetadata(local)) => {
2592 0 : init::cleanup_local_file_for_remote(&local)?;
2593 : // this file never existed remotely, we will have to do rework
2594 0 : continue;
2595 : }
2596 : };
2597 :
2598 16 : match &name {
2599 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2600 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2601 : }
2602 :
2603 16 : tracing::debug!(layer=%name, ?decision, "applied");
2604 :
2605 16 : let layer = match decision {
2606 16 : Resident { local, remote } => {
2607 16 : total_physical_size += local.file_size;
2608 16 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2609 16 : .drop_eviction_guard()
2610 : }
2611 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2612 : };
2613 :
2614 16 : loaded_layers.push(layer);
2615 : }
2616 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2617 6 : }
2618 6 : })
2619 5 : .await
2620 6 : .map_err(anyhow::Error::new)
2621 6 : .and_then(|x| x)?;
2622 :
2623 6 : let num_layers = loaded_layers.len();
2624 6 :
2625 6 : guard
2626 6 : .open_mut()
2627 6 : .expect("layermanager must be open during init")
2628 6 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2629 6 :
2630 6 : self.remote_client
2631 6 : .schedule_layer_file_deletion(&needs_cleanup)?;
2632 6 : self.remote_client
2633 6 : .schedule_index_upload_for_file_changes()?;
2634 : // This barrier orders above DELETEs before any later operations.
2635 : // This is critical because code executing after the barrier might
2636 : // create again objects with the same key that we just scheduled for deletion.
2637 : // For example, if we just scheduled deletion of an image layer "from the future",
2638 : // later compaction might run again and re-create the same image layer.
2639 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2640 : // "same" here means same key range and LSN.
2641 : //
2642 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2643 : // the upload queue may execute the PUT first, then the DELETE.
2644 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2645 : //
2646 : // 1. a future image layer is created and uploaded
2647 : // 2. ps restart
2648 : // 3. the future layer from (1) is deleted during load layer map
2649 : // 4. image layer is re-created and uploaded
2650 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2651 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2652 : //
2653 : // See https://github.com/neondatabase/neon/issues/5878
2654 : //
2655 : // NB: generation numbers naturally protect against this because they disambiguate
2656 : // (1) and (4)
2657 : // TODO: this is basically a no-op now, should we remove it?
2658 6 : self.remote_client.schedule_barrier()?;
2659 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2660 : // on retry.
2661 :
2662 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2663 6 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2664 6 : self.update_layer_visibility().await?;
2665 :
2666 6 : info!(
2667 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2668 : num_layers, disk_consistent_lsn, total_physical_size
2669 : );
2670 :
2671 6 : timer.stop_and_record();
2672 6 : Ok(())
2673 6 : }
2674 :
2675 : /// Retrieve current logical size of the timeline.
2676 : ///
2677 : /// The size could be lagging behind the actual number, in case
2678 : /// the initial size calculation has not been run (gets triggered on the first size access).
2679 : ///
2680 : /// return size and boolean flag that shows if the size is exact
2681 0 : pub(crate) fn get_current_logical_size(
2682 0 : self: &Arc<Self>,
2683 0 : priority: GetLogicalSizePriority,
2684 0 : ctx: &RequestContext,
2685 0 : ) -> logical_size::CurrentLogicalSize {
2686 0 : if !self.tenant_shard_id.is_shard_zero() {
2687 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2688 : // when HTTP API is serving a GET for timeline zero, return zero
2689 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2690 0 : }
2691 0 :
2692 0 : let current_size = self.current_logical_size.current_size();
2693 0 : debug!("Current size: {current_size:?}");
2694 :
2695 0 : match (current_size.accuracy(), priority) {
2696 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2697 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2698 0 : // background task will eventually deliver an exact value, we're in no rush
2699 0 : }
2700 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2701 : // background task is not ready, but user is asking for it now;
2702 : // => make the background task skip the line
2703 : // (The alternative would be to calculate the size here, but,
2704 : // it can actually take a long time if the user has a lot of rels.
2705 : // And we'll inevitable need it again; So, let the background task do the work.)
2706 0 : match self
2707 0 : .current_logical_size
2708 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2709 0 : .get()
2710 : {
2711 0 : Some(cancel) => cancel.cancel(),
2712 : None => {
2713 0 : match self.current_state() {
2714 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2715 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2716 0 : // Don't make noise.
2717 0 : }
2718 : TimelineState::Loading => {
2719 : // Import does not return an activated timeline.
2720 0 : info!("discarding priority boost for logical size calculation because timeline is not yet active");
2721 : }
2722 : TimelineState::Active => {
2723 : // activation should be setting the once cell
2724 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2725 0 : debug_assert!(false);
2726 : }
2727 : }
2728 : }
2729 : }
2730 : }
2731 : }
2732 :
2733 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2734 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2735 0 : let first = self
2736 0 : .current_logical_size
2737 0 : .did_return_approximate_to_walreceiver
2738 0 : .compare_exchange(
2739 0 : false,
2740 0 : true,
2741 0 : AtomicOrdering::Relaxed,
2742 0 : AtomicOrdering::Relaxed,
2743 0 : )
2744 0 : .is_ok();
2745 0 : if first {
2746 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2747 0 : }
2748 0 : }
2749 0 : }
2750 :
2751 0 : current_size
2752 0 : }
2753 :
2754 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2755 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2756 : // nothing to do for freshly created timelines;
2757 0 : assert_eq!(
2758 0 : self.current_logical_size.current_size().accuracy(),
2759 0 : logical_size::Accuracy::Exact,
2760 0 : );
2761 0 : self.current_logical_size.initialized.add_permits(1);
2762 0 : return;
2763 : };
2764 :
2765 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2766 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2767 0 : self.current_logical_size
2768 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2769 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2770 0 :
2771 0 : let self_clone = Arc::clone(self);
2772 0 : let background_ctx = ctx.detached_child(
2773 0 : TaskKind::InitialLogicalSizeCalculation,
2774 0 : DownloadBehavior::Download,
2775 0 : );
2776 0 : task_mgr::spawn(
2777 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2778 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2779 0 : self.tenant_shard_id,
2780 0 : Some(self.timeline_id),
2781 0 : "initial size calculation",
2782 : // NB: don't log errors here, task_mgr will do that.
2783 0 : async move {
2784 0 : let cancel = task_mgr::shutdown_token();
2785 0 : self_clone
2786 0 : .initial_logical_size_calculation_task(
2787 0 : initial_part_end,
2788 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2789 0 : cancel,
2790 0 : background_ctx,
2791 0 : )
2792 0 : .await;
2793 0 : Ok(())
2794 0 : }
2795 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)),
2796 : );
2797 0 : }
2798 :
2799 0 : async fn initial_logical_size_calculation_task(
2800 0 : self: Arc<Self>,
2801 0 : initial_part_end: Lsn,
2802 0 : skip_concurrency_limiter: CancellationToken,
2803 0 : cancel: CancellationToken,
2804 0 : background_ctx: RequestContext,
2805 0 : ) {
2806 0 : scopeguard::defer! {
2807 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2808 0 : self.current_logical_size.initialized.add_permits(1);
2809 0 : }
2810 0 :
2811 0 : let try_once = |attempt: usize| {
2812 0 : let background_ctx = &background_ctx;
2813 0 : let self_ref = &self;
2814 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2815 0 : async move {
2816 0 : let cancel = task_mgr::shutdown_token();
2817 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2818 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2819 0 : background_ctx,
2820 0 : );
2821 :
2822 : use crate::metrics::initial_logical_size::StartCircumstances;
2823 0 : let (_maybe_permit, circumstances) = tokio::select! {
2824 0 : permit = wait_for_permit => {
2825 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2826 : }
2827 0 : _ = self_ref.cancel.cancelled() => {
2828 0 : return Err(CalculateLogicalSizeError::Cancelled);
2829 : }
2830 0 : _ = cancel.cancelled() => {
2831 0 : return Err(CalculateLogicalSizeError::Cancelled);
2832 : },
2833 0 : () = skip_concurrency_limiter.cancelled() => {
2834 : // Some action that is part of a end user interaction requested logical size
2835 : // => break out of the rate limit
2836 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2837 : // but then again what happens if they cancel; also, we should just be using
2838 : // one runtime across the entire process, so, let's leave this for now.
2839 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2840 : }
2841 : };
2842 :
2843 0 : let metrics_guard = if attempt == 1 {
2844 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2845 : } else {
2846 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2847 : };
2848 :
2849 0 : let calculated_size = self_ref
2850 0 : .logical_size_calculation_task(
2851 0 : initial_part_end,
2852 0 : LogicalSizeCalculationCause::Initial,
2853 0 : background_ctx,
2854 0 : )
2855 0 : .await?;
2856 :
2857 0 : self_ref
2858 0 : .trigger_aux_file_size_computation(initial_part_end, background_ctx)
2859 0 : .await?;
2860 :
2861 : // TODO: add aux file size to logical size
2862 :
2863 0 : Ok((calculated_size, metrics_guard))
2864 0 : }
2865 0 : };
2866 :
2867 0 : let retrying = async {
2868 0 : let mut attempt = 0;
2869 : loop {
2870 0 : attempt += 1;
2871 0 :
2872 0 : match try_once(attempt).await {
2873 0 : Ok(res) => return ControlFlow::Continue(res),
2874 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
2875 : Err(
2876 0 : e @ (CalculateLogicalSizeError::Decode(_)
2877 0 : | CalculateLogicalSizeError::PageRead(_)),
2878 0 : ) => {
2879 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2880 : // exponential back-off doesn't make sense at these long intervals;
2881 : // use fixed retry interval with generous jitter instead
2882 0 : let sleep_duration = Duration::from_secs(
2883 0 : u64::try_from(
2884 0 : // 1hour base
2885 0 : (60_i64 * 60_i64)
2886 0 : // 10min jitter
2887 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2888 0 : )
2889 0 : .expect("10min < 1hour"),
2890 0 : );
2891 0 : tokio::time::sleep(sleep_duration).await;
2892 : }
2893 : }
2894 : }
2895 0 : };
2896 :
2897 0 : let (calculated_size, metrics_guard) = tokio::select! {
2898 0 : res = retrying => {
2899 0 : match res {
2900 0 : ControlFlow::Continue(calculated_size) => calculated_size,
2901 0 : ControlFlow::Break(()) => return,
2902 : }
2903 : }
2904 0 : _ = cancel.cancelled() => {
2905 0 : return;
2906 : }
2907 : };
2908 :
2909 : // we cannot query current_logical_size.current_size() to know the current
2910 : // *negative* value, only truncated to u64.
2911 0 : let added = self
2912 0 : .current_logical_size
2913 0 : .size_added_after_initial
2914 0 : .load(AtomicOrdering::Relaxed);
2915 0 :
2916 0 : let sum = calculated_size.saturating_add_signed(added);
2917 0 :
2918 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2919 0 : self.metrics.current_logical_size_gauge.set(sum);
2920 0 :
2921 0 : self.current_logical_size
2922 0 : .initial_logical_size
2923 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
2924 0 : .ok()
2925 0 : .expect("only this task sets it");
2926 0 : }
2927 :
2928 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
2929 0 : self: &Arc<Self>,
2930 0 : lsn: Lsn,
2931 0 : cause: LogicalSizeCalculationCause,
2932 0 : ctx: RequestContext,
2933 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
2934 0 : let (sender, receiver) = oneshot::channel();
2935 0 : let self_clone = Arc::clone(self);
2936 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
2937 0 : // we should stop the size calculation work and return an error.
2938 0 : // That would require restructuring this function's API to
2939 0 : // return the result directly, instead of a Receiver for the result.
2940 0 : let ctx = ctx.detached_child(
2941 0 : TaskKind::OndemandLogicalSizeCalculation,
2942 0 : DownloadBehavior::Download,
2943 0 : );
2944 0 : task_mgr::spawn(
2945 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2946 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
2947 0 : self.tenant_shard_id,
2948 0 : Some(self.timeline_id),
2949 0 : "ondemand logical size calculation",
2950 0 : async move {
2951 0 : let res = self_clone
2952 0 : .logical_size_calculation_task(lsn, cause, &ctx)
2953 0 : .await;
2954 0 : let _ = sender.send(res).ok();
2955 0 : Ok(()) // Receiver is responsible for handling errors
2956 0 : }
2957 0 : .in_current_span(),
2958 0 : );
2959 0 : receiver
2960 0 : }
2961 :
2962 : /// # Cancel-Safety
2963 : ///
2964 : /// This method is cancellation-safe.
2965 0 : #[instrument(skip_all)]
2966 : async fn logical_size_calculation_task(
2967 : self: &Arc<Self>,
2968 : lsn: Lsn,
2969 : cause: LogicalSizeCalculationCause,
2970 : ctx: &RequestContext,
2971 : ) -> Result<u64, CalculateLogicalSizeError> {
2972 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
2973 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
2974 : // accurate relation sizes, and they do not emit consumption metrics.
2975 : debug_assert!(self.tenant_shard_id.is_shard_zero());
2976 :
2977 : let guard = self
2978 : .gate
2979 : .enter()
2980 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
2981 :
2982 : let self_calculation = Arc::clone(self);
2983 :
2984 0 : let mut calculation = pin!(async {
2985 0 : let ctx = ctx.attached_child();
2986 0 : self_calculation
2987 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
2988 0 : .await
2989 0 : });
2990 :
2991 : tokio::select! {
2992 : res = &mut calculation => { res }
2993 : _ = self.cancel.cancelled() => {
2994 : debug!("cancelling logical size calculation for timeline shutdown");
2995 : calculation.await
2996 : }
2997 : }
2998 : }
2999 :
3000 : /// Calculate the logical size of the database at the latest LSN.
3001 : ///
3002 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3003 : /// especially if we need to download remote layers.
3004 : ///
3005 : /// # Cancel-Safety
3006 : ///
3007 : /// This method is cancellation-safe.
3008 0 : async fn calculate_logical_size(
3009 0 : &self,
3010 0 : up_to_lsn: Lsn,
3011 0 : cause: LogicalSizeCalculationCause,
3012 0 : _guard: &GateGuard,
3013 0 : ctx: &RequestContext,
3014 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3015 0 : info!(
3016 0 : "Calculating logical size for timeline {} at {}",
3017 : self.timeline_id, up_to_lsn
3018 : );
3019 :
3020 0 : pausable_failpoint!("timeline-calculate-logical-size-pause");
3021 :
3022 : // See if we've already done the work for initial size calculation.
3023 : // This is a short-cut for timelines that are mostly unused.
3024 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3025 0 : return Ok(size);
3026 0 : }
3027 0 : let storage_time_metrics = match cause {
3028 : LogicalSizeCalculationCause::Initial
3029 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3030 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3031 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3032 0 : &self.metrics.imitate_logical_size_histo
3033 : }
3034 : };
3035 0 : let timer = storage_time_metrics.start_timer();
3036 0 : let logical_size = self
3037 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3038 0 : .await?;
3039 0 : debug!("calculated logical size: {logical_size}");
3040 0 : timer.stop_and_record();
3041 0 : Ok(logical_size)
3042 0 : }
3043 :
3044 : /// Update current logical size, adding `delta' to the old value.
3045 270570 : fn update_current_logical_size(&self, delta: i64) {
3046 270570 : let logical_size = &self.current_logical_size;
3047 270570 : logical_size.increment_size(delta);
3048 270570 :
3049 270570 : // Also set the value in the prometheus gauge. Note that
3050 270570 : // there is a race condition here: if this is is called by two
3051 270570 : // threads concurrently, the prometheus gauge might be set to
3052 270570 : // one value while current_logical_size is set to the
3053 270570 : // other.
3054 270570 : match logical_size.current_size() {
3055 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
3056 270570 : .metrics
3057 270570 : .current_logical_size_gauge
3058 270570 : .set(new_current_size.into()),
3059 0 : CurrentLogicalSize::Approximate(_) => {
3060 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3061 0 : // forth between the initial size calculation task.
3062 0 : }
3063 : }
3064 270570 : }
3065 :
3066 2834 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3067 2834 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3068 2834 : let aux_metric =
3069 2834 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3070 2834 :
3071 2834 : let sum_of_entries = self
3072 2834 : .directory_metrics
3073 2834 : .iter()
3074 19838 : .map(|v| v.load(AtomicOrdering::Relaxed))
3075 2834 : .sum();
3076 : // Set a high general threshold and a lower threshold for the auxiliary files,
3077 : // as we can have large numbers of relations in the db directory.
3078 : const SUM_THRESHOLD: u64 = 5000;
3079 : const AUX_THRESHOLD: u64 = 1000;
3080 2834 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3081 0 : self.metrics
3082 0 : .directory_entries_count_gauge
3083 0 : .set(sum_of_entries);
3084 2834 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3085 0 : metric.set(sum_of_entries);
3086 2834 : }
3087 2834 : }
3088 :
3089 0 : async fn find_layer(
3090 0 : &self,
3091 0 : layer_name: &LayerName,
3092 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3093 0 : let guard = self.layers.read().await;
3094 0 : let layer = guard
3095 0 : .layer_map()?
3096 0 : .iter_historic_layers()
3097 0 : .find(|l| &l.layer_name() == layer_name)
3098 0 : .map(|found| guard.get_from_desc(&found));
3099 0 : Ok(layer)
3100 0 : }
3101 :
3102 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3103 : /// indicating which layers are currently on-disk on the primary.
3104 : ///
3105 : /// None is returned if the Timeline is in a state where uploading a heatmap
3106 : /// doesn't make sense, such as shutting down or initializing. The caller
3107 : /// should treat this as a cue to simply skip doing any heatmap uploading
3108 : /// for this timeline.
3109 2 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3110 2 : if !self.is_active() {
3111 0 : return None;
3112 2 : }
3113 :
3114 2 : let guard = self.layers.read().await;
3115 :
3116 10 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3117 10 : match layer.visibility() {
3118 : LayerVisibilityHint::Visible => {
3119 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3120 8 : let last_activity_ts = layer.latest_activity();
3121 8 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
3122 : }
3123 : LayerVisibilityHint::Covered => {
3124 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3125 2 : None
3126 : }
3127 : }
3128 10 : });
3129 2 :
3130 2 : let mut layers = resident.collect::<Vec<_>>();
3131 2 :
3132 2 : // Sort layers in order of which to download first. For a large set of layers to download, we
3133 2 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3134 2 : // or hours later:
3135 2 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3136 2 : // only exist for a few minutes before being compacted into L1s.
3137 2 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3138 2 : // the layer is likely to be covered by an image layer during compaction.
3139 20 : layers.sort_by_key(|(desc, _meta, _atime)| {
3140 20 : std::cmp::Reverse((
3141 20 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3142 20 : desc.lsn_range.end,
3143 20 : ))
3144 20 : });
3145 2 :
3146 2 : let layers = layers
3147 2 : .into_iter()
3148 8 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3149 2 : .collect();
3150 2 :
3151 2 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3152 2 : }
3153 :
3154 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3155 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3156 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3157 0 : // branchpoint in the value in IndexPart::lineage
3158 0 : self.ancestor_lsn == lsn
3159 0 : || (self.ancestor_lsn == Lsn::INVALID
3160 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3161 0 : }
3162 : }
3163 :
3164 : impl Timeline {
3165 : #[allow(clippy::doc_lazy_continuation)]
3166 : /// Get the data needed to reconstruct all keys in the provided keyspace
3167 : ///
3168 : /// The algorithm is as follows:
3169 : /// 1. While some keys are still not done and there's a timeline to visit:
3170 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3171 : /// 2.1: Build the fringe for the current keyspace
3172 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3173 : /// intersects
3174 : /// 2.3. Pop the timeline from the fringe
3175 : /// 2.4. If the fringe is empty, go back to 1
3176 626672 : async fn get_vectored_reconstruct_data(
3177 626672 : &self,
3178 626672 : mut keyspace: KeySpace,
3179 626672 : request_lsn: Lsn,
3180 626672 : reconstruct_state: &mut ValuesReconstructState,
3181 626672 : ctx: &RequestContext,
3182 626672 : ) -> Result<(), GetVectoredError> {
3183 626672 : let mut timeline_owned: Arc<Timeline>;
3184 626672 : let mut timeline = self;
3185 626672 :
3186 626672 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3187 :
3188 626670 : let missing_keyspace = loop {
3189 854797 : if self.cancel.is_cancelled() {
3190 0 : return Err(GetVectoredError::Cancelled);
3191 854797 : }
3192 :
3193 : let TimelineVisitOutcome {
3194 854797 : completed_keyspace: completed,
3195 854797 : image_covered_keyspace,
3196 854797 : } = Self::get_vectored_reconstruct_data_timeline(
3197 854797 : timeline,
3198 854797 : keyspace.clone(),
3199 854797 : cont_lsn,
3200 854797 : reconstruct_state,
3201 854797 : &self.cancel,
3202 854797 : ctx,
3203 854797 : )
3204 192172 : .await?;
3205 :
3206 854797 : keyspace.remove_overlapping_with(&completed);
3207 854797 :
3208 854797 : // Do not descend into the ancestor timeline for aux files.
3209 854797 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3210 854797 : // stalling compaction.
3211 854797 : keyspace.remove_overlapping_with(&KeySpace {
3212 854797 : ranges: vec![NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE],
3213 854797 : });
3214 854797 :
3215 854797 : // Keyspace is fully retrieved
3216 854797 : if keyspace.is_empty() {
3217 626656 : break None;
3218 228141 : }
3219 :
3220 228141 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3221 : // Not fully retrieved but no ancestor timeline.
3222 14 : break Some(keyspace);
3223 : };
3224 :
3225 : // Now we see if there are keys covered by the image layer but does not exist in the
3226 : // image layer, which means that the key does not exist.
3227 :
3228 : // The block below will stop the vectored search if any of the keys encountered an image layer
3229 : // which did not contain a snapshot for said key. Since we have already removed all completed
3230 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3231 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3232 : // and stop the search as a result of that.
3233 228127 : let removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3234 228127 : if !removed.is_empty() {
3235 0 : break Some(removed);
3236 228127 : }
3237 228127 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3238 228127 : // keyspace.
3239 228127 :
3240 228127 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3241 228127 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3242 228127 : timeline_owned = timeline
3243 228127 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3244 2 : .await?;
3245 228125 : timeline = &*timeline_owned;
3246 : };
3247 :
3248 626670 : if let Some(missing_keyspace) = missing_keyspace {
3249 14 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3250 14 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3251 14 : shard: self
3252 14 : .shard_identity
3253 14 : .get_shard_number(&missing_keyspace.start().unwrap()),
3254 14 : cont_lsn,
3255 14 : request_lsn,
3256 14 : ancestor_lsn: Some(timeline.ancestor_lsn),
3257 14 : backtrace: None,
3258 14 : }));
3259 626656 : }
3260 626656 :
3261 626656 : Ok(())
3262 626672 : }
3263 :
3264 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3265 : ///
3266 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3267 : /// the current keyspace. The current keyspace of the search at any given timeline
3268 : /// is the original keyspace minus all the keys that have been completed minus
3269 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3270 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3271 : ///
3272 : /// This is basically a depth-first search visitor implementation where a vertex
3273 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3274 : ///
3275 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3276 : /// and get all the required reconstruct data from the layer in one go.
3277 : ///
3278 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3279 : /// decides how to deal with these two keyspaces.
3280 854797 : async fn get_vectored_reconstruct_data_timeline(
3281 854797 : timeline: &Timeline,
3282 854797 : keyspace: KeySpace,
3283 854797 : mut cont_lsn: Lsn,
3284 854797 : reconstruct_state: &mut ValuesReconstructState,
3285 854797 : cancel: &CancellationToken,
3286 854797 : ctx: &RequestContext,
3287 854797 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3288 854797 : let mut unmapped_keyspace = keyspace.clone();
3289 854797 : let mut fringe = LayerFringe::new();
3290 854797 :
3291 854797 : let mut completed_keyspace = KeySpace::default();
3292 854797 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3293 :
3294 : loop {
3295 1699945 : if cancel.is_cancelled() {
3296 0 : return Err(GetVectoredError::Cancelled);
3297 1699945 : }
3298 1699945 :
3299 1699945 : let (keys_done_last_step, keys_with_image_coverage) =
3300 1699945 : reconstruct_state.consume_done_keys();
3301 1699945 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3302 1699945 : completed_keyspace.merge(&keys_done_last_step);
3303 1699945 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3304 21808 : unmapped_keyspace
3305 21808 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3306 21808 : image_covered_keyspace.add_range(keys_with_image_coverage);
3307 1678137 : }
3308 :
3309 : // Do not descent any further if the last layer we visited
3310 : // completed all keys in the keyspace it inspected. This is not
3311 : // required for correctness, but avoids visiting extra layers
3312 : // which turns out to be a perf bottleneck in some cases.
3313 1699945 : if !unmapped_keyspace.is_empty() {
3314 1075533 : let guard = timeline.layers.read().await;
3315 1075533 : let layers = guard.layer_map()?;
3316 :
3317 1075533 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3318 915673 : let start_lsn = l.get_lsn_range().start;
3319 915673 : cont_lsn > start_lsn
3320 1075533 : });
3321 1075533 :
3322 1075533 : match in_memory_layer {
3323 606344 : Some(l) => {
3324 606344 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3325 606344 : fringe.update(
3326 606344 : ReadableLayer::InMemoryLayer(l),
3327 606344 : unmapped_keyspace.clone(),
3328 606344 : lsn_range,
3329 606344 : );
3330 606344 : }
3331 : None => {
3332 469211 : for range in unmapped_keyspace.ranges.iter() {
3333 469211 : let results = layers.range_search(range.clone(), cont_lsn);
3334 469211 :
3335 469211 : results
3336 469211 : .found
3337 469211 : .into_iter()
3338 469211 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3339 238818 : (
3340 238818 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3341 238818 : keyspace_accum.to_keyspace(),
3342 238818 : lsn_floor..cont_lsn,
3343 238818 : )
3344 469211 : })
3345 469211 : .for_each(|(layer, keyspace, lsn_range)| {
3346 238818 : fringe.update(layer, keyspace, lsn_range)
3347 469211 : });
3348 469211 : }
3349 : }
3350 : }
3351 :
3352 : // It's safe to drop the layer map lock after planning the next round of reads.
3353 : // The fringe keeps readable handles for the layers which are safe to read even
3354 : // if layers were compacted or flushed.
3355 : //
3356 : // The more interesting consideration is: "Why is the read algorithm still correct
3357 : // if the layer map changes while it is operating?". Doing a vectored read on a
3358 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3359 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3360 : // range at *a particular point in time*. It is fine for the answer to be different
3361 : // at two different time points.
3362 1075533 : drop(guard);
3363 624412 : }
3364 :
3365 1699945 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3366 845148 : let next_cont_lsn = lsn_range.start;
3367 845148 : layer_to_read
3368 845148 : .get_values_reconstruct_data(
3369 845148 : keyspace_to_read.clone(),
3370 845148 : lsn_range,
3371 845148 : reconstruct_state,
3372 845148 : ctx,
3373 845148 : )
3374 182882 : .await?;
3375 :
3376 845148 : unmapped_keyspace = keyspace_to_read;
3377 845148 : cont_lsn = next_cont_lsn;
3378 845148 :
3379 845148 : reconstruct_state.on_layer_visited(&layer_to_read);
3380 : } else {
3381 854797 : break;
3382 854797 : }
3383 854797 : }
3384 854797 :
3385 854797 : Ok(TimelineVisitOutcome {
3386 854797 : completed_keyspace,
3387 854797 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3388 854797 : })
3389 854797 : }
3390 :
3391 228127 : async fn get_ready_ancestor_timeline(
3392 228127 : &self,
3393 228127 : ancestor: &Arc<Timeline>,
3394 228127 : ctx: &RequestContext,
3395 228127 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3396 228127 : // It's possible that the ancestor timeline isn't active yet, or
3397 228127 : // is active but hasn't yet caught up to the branch point. Wait
3398 228127 : // for it.
3399 228127 : //
3400 228127 : // This cannot happen while the pageserver is running normally,
3401 228127 : // because you cannot create a branch from a point that isn't
3402 228127 : // present in the pageserver yet. However, we don't wait for the
3403 228127 : // branch point to be uploaded to cloud storage before creating
3404 228127 : // a branch. I.e., the branch LSN need not be remote consistent
3405 228127 : // for the branching operation to succeed.
3406 228127 : //
3407 228127 : // Hence, if we try to load a tenant in such a state where
3408 228127 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3409 228127 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3410 228127 : // then we will need to wait for the ancestor timeline to
3411 228127 : // re-stream WAL up to branch_lsn before we access it.
3412 228127 : //
3413 228127 : // How can a tenant get in such a state?
3414 228127 : // - ungraceful pageserver process exit
3415 228127 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3416 228127 : //
3417 228127 : // NB: this could be avoided by requiring
3418 228127 : // branch_lsn >= remote_consistent_lsn
3419 228127 : // during branch creation.
3420 228127 : match ancestor.wait_to_become_active(ctx).await {
3421 228125 : Ok(()) => {}
3422 : Err(TimelineState::Stopping) => {
3423 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3424 0 : return Err(GetReadyAncestorError::Cancelled);
3425 : }
3426 2 : Err(state) => {
3427 2 : return Err(GetReadyAncestorError::BadState {
3428 2 : timeline_id: ancestor.timeline_id,
3429 2 : state,
3430 2 : });
3431 : }
3432 : }
3433 228125 : ancestor
3434 228125 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3435 0 : .await
3436 228125 : .map_err(|e| match e {
3437 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3438 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3439 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3440 0 : timeline_id: ancestor.timeline_id,
3441 0 : state,
3442 0 : },
3443 228125 : })?;
3444 :
3445 228125 : Ok(ancestor.clone())
3446 228127 : }
3447 :
3448 151304 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3449 151304 : &self.shard_identity
3450 151304 : }
3451 :
3452 : #[inline(always)]
3453 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3454 0 : ShardTimelineId {
3455 0 : shard_index: ShardIndex {
3456 0 : shard_number: self.shard_identity.number,
3457 0 : shard_count: self.shard_identity.count,
3458 0 : },
3459 0 : timeline_id: self.timeline_id,
3460 0 : }
3461 0 : }
3462 :
3463 : /// Returns a non-frozen open in-memory layer for ingestion.
3464 : ///
3465 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3466 : /// this function without holding the mutex.
3467 1268 : async fn get_layer_for_write(
3468 1268 : &self,
3469 1268 : lsn: Lsn,
3470 1268 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3471 1268 : ctx: &RequestContext,
3472 1268 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3473 1268 : let mut guard = self.layers.write().await;
3474 1268 : let gate_guard = self.gate.enter().context("enter gate for inmem layer")?;
3475 :
3476 1268 : let last_record_lsn = self.get_last_record_lsn();
3477 1268 : ensure!(
3478 1268 : lsn > last_record_lsn,
3479 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3480 : lsn,
3481 : last_record_lsn,
3482 : );
3483 :
3484 1268 : let layer = guard
3485 1268 : .open_mut()?
3486 1268 : .get_layer_for_write(
3487 1268 : lsn,
3488 1268 : self.conf,
3489 1268 : self.timeline_id,
3490 1268 : self.tenant_shard_id,
3491 1268 : gate_guard,
3492 1268 : ctx,
3493 1268 : )
3494 719 : .await?;
3495 1268 : Ok(layer)
3496 1268 : }
3497 :
3498 5279064 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3499 5279064 : assert!(new_lsn.is_aligned());
3500 :
3501 5279064 : self.metrics.last_record_gauge.set(new_lsn.0 as i64);
3502 5279064 : self.last_record_lsn.advance(new_lsn);
3503 5279064 : }
3504 :
3505 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3506 : ///
3507 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3508 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3509 1172 : async fn freeze_inmem_layer_at(
3510 1172 : &self,
3511 1172 : at: Lsn,
3512 1172 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3513 1172 : ) -> Result<u64, FlushLayerError> {
3514 1172 : let frozen = {
3515 1172 : let mut guard = self.layers.write().await;
3516 1172 : guard
3517 1172 : .open_mut()?
3518 1172 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock)
3519 2 : .await
3520 : };
3521 :
3522 1172 : if frozen {
3523 1144 : let now = Instant::now();
3524 1144 : *(self.last_freeze_ts.write().unwrap()) = now;
3525 1144 : }
3526 :
3527 : // Increment the flush cycle counter and wake up the flush task.
3528 : // Remember the new value, so that when we listen for the flush
3529 : // to finish, we know when the flush that we initiated has
3530 : // finished, instead of some other flush that was started earlier.
3531 1172 : let mut my_flush_request = 0;
3532 1172 :
3533 1172 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3534 1172 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3535 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3536 1172 : }
3537 1172 :
3538 1172 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3539 1172 : my_flush_request = *counter + 1;
3540 1172 : *counter = my_flush_request;
3541 1172 : *lsn = std::cmp::max(at, *lsn);
3542 1172 : });
3543 1172 :
3544 1172 : assert_ne!(my_flush_request, 0);
3545 :
3546 1172 : Ok(my_flush_request)
3547 1172 : }
3548 :
3549 : /// Layer flusher task's main loop.
3550 412 : async fn flush_loop(
3551 412 : self: &Arc<Self>,
3552 412 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3553 412 : ctx: &RequestContext,
3554 412 : ) {
3555 412 : info!("started flush loop");
3556 : loop {
3557 1543 : tokio::select! {
3558 1543 : _ = self.cancel.cancelled() => {
3559 10 : info!("shutting down layer flush task due to Timeline::cancel");
3560 10 : break;
3561 : },
3562 1543 : _ = layer_flush_start_rx.changed() => {}
3563 1131 : }
3564 1131 : trace!("waking up");
3565 1131 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3566 1131 :
3567 1131 : // The highest LSN to which we flushed in the loop over frozen layers
3568 1131 : let mut flushed_to_lsn = Lsn(0);
3569 :
3570 1131 : let result = loop {
3571 2275 : if self.cancel.is_cancelled() {
3572 0 : info!("dropping out of flush loop for timeline shutdown");
3573 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3574 : // anyone waiting on that will respect self.cancel as well: they will stop
3575 : // waiting at the same time we as drop out of this loop.
3576 0 : return;
3577 2275 : }
3578 2275 :
3579 2275 : let timer = self.metrics.flush_time_histo.start_timer();
3580 :
3581 : let num_frozen_layers;
3582 : let frozen_layer_total_size;
3583 2275 : let layer_to_flush = {
3584 2275 : let guard = self.layers.read().await;
3585 2275 : let Ok(lm) = guard.layer_map() else {
3586 0 : info!("dropping out of flush loop for timeline shutdown");
3587 0 : return;
3588 : };
3589 2275 : num_frozen_layers = lm.frozen_layers.len();
3590 2275 : frozen_layer_total_size = lm
3591 2275 : .frozen_layers
3592 2275 : .iter()
3593 2275 : .map(|l| l.estimated_in_mem_size())
3594 2275 : .sum::<u64>();
3595 2275 : lm.frozen_layers.front().cloned()
3596 : // drop 'layers' lock to allow concurrent reads and writes
3597 : };
3598 2275 : let Some(layer_to_flush) = layer_to_flush else {
3599 1131 : break Ok(());
3600 : };
3601 1144 : if num_frozen_layers
3602 1144 : > std::cmp::max(
3603 1144 : self.get_compaction_threshold(),
3604 1144 : DEFAULT_COMPACTION_THRESHOLD,
3605 1144 : )
3606 0 : && frozen_layer_total_size >= /* 128 MB */ 128000000
3607 : {
3608 0 : tracing::warn!(
3609 0 : "too many frozen layers: {num_frozen_layers} layers with estimated in-mem size of {frozen_layer_total_size} bytes",
3610 : );
3611 1144 : }
3612 17053 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3613 1144 : Ok(this_layer_to_lsn) => {
3614 1144 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3615 1144 : }
3616 : Err(FlushLayerError::Cancelled) => {
3617 0 : info!("dropping out of flush loop for timeline shutdown");
3618 0 : return;
3619 : }
3620 0 : err @ Err(
3621 0 : FlushLayerError::NotRunning(_)
3622 0 : | FlushLayerError::Other(_)
3623 0 : | FlushLayerError::CreateImageLayersError(_),
3624 0 : ) => {
3625 0 : error!("could not flush frozen layer: {err:?}");
3626 0 : break err.map(|_| ());
3627 : }
3628 : }
3629 1144 : timer.stop_and_record();
3630 : };
3631 :
3632 : // Unsharded tenants should never advance their LSN beyond the end of the
3633 : // highest layer they write: such gaps between layer data and the frozen LSN
3634 : // are only legal on sharded tenants.
3635 1131 : debug_assert!(
3636 1131 : self.shard_identity.count.count() > 1
3637 1131 : || flushed_to_lsn >= frozen_to_lsn
3638 65 : || !flushed_to_lsn.is_valid()
3639 : );
3640 :
3641 1131 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3642 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3643 : // to us via layer_flush_start_rx, then advance it here.
3644 : //
3645 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3646 : // never encounter a gap in the wal.
3647 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3648 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3649 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3650 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3651 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3652 0 : }
3653 0 : }
3654 1131 : }
3655 :
3656 : // Notify any listeners that we're done
3657 1131 : let _ = self
3658 1131 : .layer_flush_done_tx
3659 1131 : .send_replace((flush_counter, result));
3660 : }
3661 10 : }
3662 :
3663 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
3664 1092 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3665 1092 : let mut rx = self.layer_flush_done_tx.subscribe();
3666 : loop {
3667 : {
3668 2180 : let (last_result_counter, last_result) = &*rx.borrow();
3669 2180 : if *last_result_counter >= request {
3670 1092 : if let Err(err) = last_result {
3671 : // We already logged the original error in
3672 : // flush_loop. We cannot propagate it to the caller
3673 : // here, because it might not be Cloneable
3674 0 : return Err(err.clone());
3675 : } else {
3676 1092 : return Ok(());
3677 : }
3678 1088 : }
3679 1088 : }
3680 1088 : trace!("waiting for flush to complete");
3681 1088 : tokio::select! {
3682 1088 : rx_e = rx.changed() => {
3683 1088 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3684 : },
3685 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3686 : // the notification from [`flush_loop`] that it completed.
3687 1088 : _ = self.cancel.cancelled() => {
3688 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3689 0 : return Ok(())
3690 : }
3691 : };
3692 1088 : trace!("done")
3693 : }
3694 1092 : }
3695 :
3696 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3697 : ///
3698 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3699 1144 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3700 : async fn flush_frozen_layer(
3701 : self: &Arc<Self>,
3702 : frozen_layer: Arc<InMemoryLayer>,
3703 : ctx: &RequestContext,
3704 : ) -> Result<Lsn, FlushLayerError> {
3705 : debug_assert_current_span_has_tenant_and_timeline_id();
3706 :
3707 : // As a special case, when we have just imported an image into the repository,
3708 : // instead of writing out a L0 delta layer, we directly write out image layer
3709 : // files instead. This is possible as long as *all* the data imported into the
3710 : // repository have the same LSN.
3711 : let lsn_range = frozen_layer.get_lsn_range();
3712 :
3713 : // Whether to directly create image layers for this flush, or flush them as delta layers
3714 : let create_image_layer =
3715 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3716 :
3717 : #[cfg(test)]
3718 : {
3719 : match &mut *self.flush_loop_state.lock().unwrap() {
3720 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3721 : panic!("flush loop not running")
3722 : }
3723 : FlushLoopState::Running {
3724 : expect_initdb_optimization,
3725 : initdb_optimization_count,
3726 : ..
3727 : } => {
3728 : if create_image_layer {
3729 : *initdb_optimization_count += 1;
3730 : } else {
3731 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3732 : }
3733 : }
3734 : }
3735 : }
3736 :
3737 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3738 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3739 : // require downloading anything during initial import.
3740 : let ((rel_partition, metadata_partition), _lsn) = self
3741 : .repartition(
3742 : self.initdb_lsn,
3743 : self.get_compaction_target_size(),
3744 : EnumSet::empty(),
3745 : ctx,
3746 : )
3747 : .await
3748 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
3749 :
3750 : if self.cancel.is_cancelled() {
3751 : return Err(FlushLayerError::Cancelled);
3752 : }
3753 :
3754 : let mut layers_to_upload = Vec::new();
3755 : layers_to_upload.extend(
3756 : self.create_image_layers(
3757 : &rel_partition,
3758 : self.initdb_lsn,
3759 : ImageLayerCreationMode::Initial,
3760 : ctx,
3761 : )
3762 : .await?,
3763 : );
3764 : if !metadata_partition.parts.is_empty() {
3765 : assert_eq!(
3766 : metadata_partition.parts.len(),
3767 : 1,
3768 : "currently sparse keyspace should only contain a single metadata keyspace"
3769 : );
3770 : layers_to_upload.extend(
3771 : self.create_image_layers(
3772 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
3773 : // every single key within the keyspace, and therefore, it's safe to force converting it
3774 : // into a dense keyspace before calling this function.
3775 : &metadata_partition.into_dense(),
3776 : self.initdb_lsn,
3777 : ImageLayerCreationMode::Initial,
3778 : ctx,
3779 : )
3780 : .await?,
3781 : );
3782 : }
3783 :
3784 : (layers_to_upload, None)
3785 : } else {
3786 : // Normal case, write out a L0 delta layer file.
3787 : // `create_delta_layer` will not modify the layer map.
3788 : // We will remove frozen layer and add delta layer in one atomic operation later.
3789 : let Some(layer) = self
3790 : .create_delta_layer(&frozen_layer, None, ctx)
3791 : .await
3792 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
3793 : else {
3794 : panic!("delta layer cannot be empty if no filter is applied");
3795 : };
3796 : (
3797 : // FIXME: even though we have a single image and single delta layer assumption
3798 : // we push them to vec
3799 : vec![layer.clone()],
3800 : Some(layer),
3801 : )
3802 : };
3803 :
3804 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
3805 :
3806 : if self.cancel.is_cancelled() {
3807 : return Err(FlushLayerError::Cancelled);
3808 : }
3809 :
3810 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
3811 :
3812 : // The new on-disk layers are now in the layer map. We can remove the
3813 : // in-memory layer from the map now. The flushed layer is stored in
3814 : // the mapping in `create_delta_layer`.
3815 : {
3816 : let mut guard = self.layers.write().await;
3817 :
3818 : guard.open_mut()?.finish_flush_l0_layer(
3819 : delta_layer_to_add.as_ref(),
3820 : &frozen_layer,
3821 : &self.metrics,
3822 : );
3823 :
3824 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
3825 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
3826 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
3827 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
3828 : }
3829 : // release lock on 'layers'
3830 : };
3831 :
3832 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
3833 : // This makes us refuse ingest until the new layers have been persisted to the remote.
3834 : self.remote_client
3835 : .wait_completion()
3836 : .await
3837 0 : .map_err(|e| match e {
3838 : WaitCompletionError::UploadQueueShutDownOrStopped
3839 : | WaitCompletionError::NotInitialized(
3840 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
3841 0 : ) => FlushLayerError::Cancelled,
3842 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
3843 0 : FlushLayerError::Other(anyhow!(e).into())
3844 : }
3845 0 : })?;
3846 :
3847 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
3848 : // a compaction can delete the file and then it won't be available for uploads any more.
3849 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
3850 : // race situation.
3851 : // See https://github.com/neondatabase/neon/issues/4526
3852 : pausable_failpoint!("flush-frozen-pausable");
3853 :
3854 : // This failpoint is used by another test case `test_pageserver_recovery`.
3855 : fail_point!("flush-frozen-exit");
3856 :
3857 : Ok(Lsn(lsn_range.end.0 - 1))
3858 : }
3859 :
3860 : /// Return true if the value changed
3861 : ///
3862 : /// This function must only be used from the layer flush task.
3863 1144 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
3864 1144 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
3865 1144 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
3866 1144 : new_value != old_value
3867 1144 : }
3868 :
3869 : /// Update metadata file
3870 1146 : fn schedule_uploads(
3871 1146 : &self,
3872 1146 : disk_consistent_lsn: Lsn,
3873 1146 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
3874 1146 : ) -> anyhow::Result<()> {
3875 1146 : // We can only save a valid 'prev_record_lsn' value on disk if we
3876 1146 : // flushed *all* in-memory changes to disk. We only track
3877 1146 : // 'prev_record_lsn' in memory for the latest processed record, so we
3878 1146 : // don't remember what the correct value that corresponds to some old
3879 1146 : // LSN is. But if we flush everything, then the value corresponding
3880 1146 : // current 'last_record_lsn' is correct and we can store it on disk.
3881 1146 : let RecordLsn {
3882 1146 : last: last_record_lsn,
3883 1146 : prev: prev_record_lsn,
3884 1146 : } = self.last_record_lsn.load();
3885 1146 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
3886 1067 : Some(prev_record_lsn)
3887 : } else {
3888 79 : None
3889 : };
3890 :
3891 1146 : let update = crate::tenant::metadata::MetadataUpdate::new(
3892 1146 : disk_consistent_lsn,
3893 1146 : ondisk_prev_record_lsn,
3894 1146 : *self.latest_gc_cutoff_lsn.read(),
3895 1146 : );
3896 1146 :
3897 1146 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
3898 0 : "{}",
3899 0 : x.unwrap()
3900 1146 : ));
3901 :
3902 2302 : for layer in layers_to_upload {
3903 1156 : self.remote_client.schedule_layer_file_upload(layer)?;
3904 : }
3905 1146 : self.remote_client
3906 1146 : .schedule_index_upload_for_metadata_update(&update)?;
3907 :
3908 1146 : Ok(())
3909 1146 : }
3910 :
3911 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
3912 0 : self.remote_client
3913 0 : .preserve_initdb_archive(
3914 0 : &self.tenant_shard_id.tenant_id,
3915 0 : &self.timeline_id,
3916 0 : &self.cancel,
3917 0 : )
3918 0 : .await
3919 0 : }
3920 :
3921 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
3922 : // in layer map immediately. The caller is responsible to put it into the layer map.
3923 968 : async fn create_delta_layer(
3924 968 : self: &Arc<Self>,
3925 968 : frozen_layer: &Arc<InMemoryLayer>,
3926 968 : key_range: Option<Range<Key>>,
3927 968 : ctx: &RequestContext,
3928 968 : ) -> anyhow::Result<Option<ResidentLayer>> {
3929 968 : let self_clone = Arc::clone(self);
3930 968 : let frozen_layer = Arc::clone(frozen_layer);
3931 968 : let ctx = ctx.attached_child();
3932 968 : let work = async move {
3933 968 : let Some((desc, path)) = frozen_layer
3934 968 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
3935 10241 : .await?
3936 : else {
3937 0 : return Ok(None);
3938 : };
3939 968 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
3940 :
3941 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
3942 : // We just need to fsync the directory in which these inodes are linked,
3943 : // which we know to be the timeline directory.
3944 : //
3945 : // We use fatal_err() below because the after write_to_disk returns with success,
3946 : // the in-memory state of the filesystem already has the layer file in its final place,
3947 : // and subsequent pageserver code could think it's durable while it really isn't.
3948 968 : let timeline_dir = VirtualFile::open(
3949 968 : &self_clone
3950 968 : .conf
3951 968 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
3952 968 : &ctx,
3953 968 : )
3954 489 : .await
3955 968 : .fatal_err("VirtualFile::open for timeline dir fsync");
3956 968 : timeline_dir
3957 968 : .sync_all()
3958 484 : .await
3959 968 : .fatal_err("VirtualFile::sync_all timeline dir");
3960 968 : anyhow::Ok(Some(new_delta))
3961 968 : };
3962 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
3963 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
3964 : use crate::virtual_file::io_engine::IoEngine;
3965 968 : match crate::virtual_file::io_engine::get() {
3966 0 : IoEngine::NotSet => panic!("io engine not set"),
3967 : IoEngine::StdFs => {
3968 484 : let span = tracing::info_span!("blocking");
3969 484 : tokio::task::spawn_blocking({
3970 484 : move || Handle::current().block_on(work.instrument(span))
3971 484 : })
3972 484 : .await
3973 484 : .context("spawn_blocking")
3974 484 : .and_then(|x| x)
3975 : }
3976 : #[cfg(target_os = "linux")]
3977 11209 : IoEngine::TokioEpollUring => work.await,
3978 : }
3979 968 : }
3980 :
3981 540 : async fn repartition(
3982 540 : &self,
3983 540 : lsn: Lsn,
3984 540 : partition_size: u64,
3985 540 : flags: EnumSet<CompactFlags>,
3986 540 : ctx: &RequestContext,
3987 540 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
3988 540 : let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
3989 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
3990 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
3991 : // and hence before the compaction task starts.
3992 0 : return Err(CompactionError::Other(anyhow!(
3993 0 : "repartition() called concurrently, this should not happen"
3994 0 : )));
3995 : };
3996 540 : let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
3997 540 : if lsn < *partition_lsn {
3998 0 : return Err(CompactionError::Other(anyhow!(
3999 0 : "repartition() called with LSN going backwards, this should not happen"
4000 0 : )));
4001 540 : }
4002 540 :
4003 540 : let distance = lsn.0 - partition_lsn.0;
4004 540 : if *partition_lsn != Lsn(0)
4005 262 : && distance <= self.repartition_threshold
4006 262 : && !flags.contains(CompactFlags::ForceRepartition)
4007 : {
4008 248 : debug!(
4009 : distance,
4010 : threshold = self.repartition_threshold,
4011 0 : "no repartitioning needed"
4012 : );
4013 248 : return Ok((
4014 248 : (dense_partition.clone(), sparse_partition.clone()),
4015 248 : *partition_lsn,
4016 248 : ));
4017 292 : }
4018 :
4019 15726 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4020 292 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4021 292 : let sparse_partitioning = SparseKeyPartitioning {
4022 292 : parts: vec![sparse_ks],
4023 292 : }; // no partitioning for metadata keys for now
4024 292 : *partitioning_guard = ((dense_partitioning, sparse_partitioning), lsn);
4025 292 :
4026 292 : Ok((partitioning_guard.0.clone(), partitioning_guard.1))
4027 540 : }
4028 :
4029 : // Is it time to create a new image layer for the given partition?
4030 14 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4031 14 : let threshold = self.get_image_creation_threshold();
4032 :
4033 14 : let guard = self.layers.read().await;
4034 14 : let Ok(layers) = guard.layer_map() else {
4035 0 : return false;
4036 : };
4037 :
4038 14 : let mut max_deltas = 0;
4039 28 : for part_range in &partition.ranges {
4040 14 : let image_coverage = layers.image_coverage(part_range, lsn);
4041 28 : for (img_range, last_img) in image_coverage {
4042 14 : let img_lsn = if let Some(last_img) = last_img {
4043 0 : last_img.get_lsn_range().end
4044 : } else {
4045 14 : Lsn(0)
4046 : };
4047 : // Let's consider an example:
4048 : //
4049 : // delta layer with LSN range 71-81
4050 : // delta layer with LSN range 81-91
4051 : // delta layer with LSN range 91-101
4052 : // image layer at LSN 100
4053 : //
4054 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4055 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4056 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4057 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4058 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4059 14 : if img_lsn < lsn {
4060 14 : let num_deltas =
4061 14 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4062 14 :
4063 14 : max_deltas = max_deltas.max(num_deltas);
4064 14 : if num_deltas >= threshold {
4065 0 : debug!(
4066 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4067 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4068 : );
4069 0 : return true;
4070 14 : }
4071 0 : }
4072 : }
4073 : }
4074 :
4075 14 : debug!(
4076 : max_deltas,
4077 0 : "none of the partitioned ranges had >= {threshold} deltas"
4078 : );
4079 14 : false
4080 14 : }
4081 :
4082 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4083 : /// so that at most one image layer will be produced from this function.
4084 202 : async fn create_image_layer_for_rel_blocks(
4085 202 : self: &Arc<Self>,
4086 202 : partition: &KeySpace,
4087 202 : mut image_layer_writer: ImageLayerWriter,
4088 202 : lsn: Lsn,
4089 202 : ctx: &RequestContext,
4090 202 : img_range: Range<Key>,
4091 202 : start: Key,
4092 202 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4093 202 : let mut wrote_keys = false;
4094 202 :
4095 202 : let mut key_request_accum = KeySpaceAccum::new();
4096 1344 : for range in &partition.ranges {
4097 1142 : let mut key = range.start;
4098 2472 : while key < range.end {
4099 : // Decide whether to retain this key: usually we do, but sharded tenants may
4100 : // need to drop keys that don't belong to them. If we retain the key, add it
4101 : // to `key_request_accum` for later issuing a vectored get
4102 1330 : if self.shard_identity.is_key_disposable(&key) {
4103 0 : debug!(
4104 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4105 0 : key,
4106 0 : self.shard_identity.get_shard_number(&key)
4107 : );
4108 1330 : } else {
4109 1330 : key_request_accum.add_key(key);
4110 1330 : }
4111 :
4112 1330 : let last_key_in_range = key.next() == range.end;
4113 1330 : key = key.next();
4114 1330 :
4115 1330 : // Maybe flush `key_rest_accum`
4116 1330 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4117 1330 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4118 : {
4119 1142 : let results = self
4120 1142 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
4121 49 : .await?;
4122 :
4123 1142 : if self.cancel.is_cancelled() {
4124 0 : return Err(CreateImageLayersError::Cancelled);
4125 1142 : }
4126 :
4127 2472 : for (img_key, img) in results {
4128 1330 : let img = match img {
4129 1330 : Ok(img) => img,
4130 0 : Err(err) => {
4131 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4132 0 : // page without losing any actual user data. That seems better
4133 0 : // than failing repeatedly and getting stuck.
4134 0 : //
4135 0 : // We had a bug at one point, where we truncated the FSM and VM
4136 0 : // in the pageserver, but the Postgres didn't know about that
4137 0 : // and continued to generate incremental WAL records for pages
4138 0 : // that didn't exist in the pageserver. Trying to replay those
4139 0 : // WAL records failed to find the previous image of the page.
4140 0 : // This special case allows us to recover from that situation.
4141 0 : // See https://github.com/neondatabase/neon/issues/2601.
4142 0 : //
4143 0 : // Unfortunately we cannot do this for the main fork, or for
4144 0 : // any metadata keys, keys, as that would lead to actual data
4145 0 : // loss.
4146 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4147 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4148 0 : ZERO_PAGE.clone()
4149 : } else {
4150 0 : return Err(CreateImageLayersError::from(err));
4151 : }
4152 : }
4153 : };
4154 :
4155 : // Write all the keys we just read into our new image layer.
4156 1460 : image_layer_writer.put_image(img_key, img, ctx).await?;
4157 1330 : wrote_keys = true;
4158 : }
4159 188 : }
4160 : }
4161 : }
4162 :
4163 202 : if wrote_keys {
4164 : // Normal path: we have written some data into the new image layer for this
4165 : // partition, so flush it to disk.
4166 407 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4167 202 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4168 202 : info!("created image layer for rel {}", image_layer.local_path());
4169 202 : Ok(ImageLayerCreationOutcome {
4170 202 : image: Some(image_layer),
4171 202 : next_start_key: img_range.end,
4172 202 : })
4173 : } else {
4174 : // Special case: the image layer may be empty if this is a sharded tenant and the
4175 : // partition does not cover any keys owned by this shard. In this case, to ensure
4176 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4177 : // layer we write will cover the key range that we just scanned.
4178 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4179 0 : Ok(ImageLayerCreationOutcome {
4180 0 : image: None,
4181 0 : next_start_key: start,
4182 0 : })
4183 : }
4184 202 : }
4185 :
4186 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4187 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4188 : /// would not be too large to fit in a single image layer.
4189 : #[allow(clippy::too_many_arguments)]
4190 192 : async fn create_image_layer_for_metadata_keys(
4191 192 : self: &Arc<Self>,
4192 192 : partition: &KeySpace,
4193 192 : mut image_layer_writer: ImageLayerWriter,
4194 192 : lsn: Lsn,
4195 192 : ctx: &RequestContext,
4196 192 : img_range: Range<Key>,
4197 192 : mode: ImageLayerCreationMode,
4198 192 : start: Key,
4199 192 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4200 192 : // Metadata keys image layer creation.
4201 192 : let mut reconstruct_state = ValuesReconstructState::default();
4202 192 : let begin = Instant::now();
4203 192 : let data = self
4204 192 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4205 1075 : .await?;
4206 192 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4207 192 : let mut new_data = BTreeMap::new();
4208 192 : let mut total_kb_retrieved = 0;
4209 192 : let mut total_keys_retrieved = 0;
4210 10204 : for (k, v) in data {
4211 10012 : let v = v?;
4212 10012 : total_kb_retrieved += KEY_SIZE + v.len();
4213 10012 : total_keys_retrieved += 1;
4214 10012 : new_data.insert(k, v);
4215 : }
4216 192 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4217 192 : };
4218 192 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4219 192 : let elapsed = begin.elapsed();
4220 192 :
4221 192 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4222 192 : info!(
4223 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()
4224 : );
4225 :
4226 192 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4227 2 : return Ok(ImageLayerCreationOutcome {
4228 2 : image: None,
4229 2 : next_start_key: img_range.end,
4230 2 : });
4231 190 : }
4232 190 : if self.cancel.is_cancelled() {
4233 0 : return Err(CreateImageLayersError::Cancelled);
4234 190 : }
4235 190 : let mut wrote_any_image = false;
4236 10202 : for (k, v) in data {
4237 10012 : if v.is_empty() {
4238 : // the key has been deleted, it does not need an image
4239 : // in metadata keyspace, an empty image == tombstone
4240 8 : continue;
4241 10004 : }
4242 10004 : wrote_any_image = true;
4243 10004 :
4244 10004 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4245 10004 :
4246 10004 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4247 10004 : // on the normal data path either.
4248 10162 : image_layer_writer.put_image(k, v, ctx).await?;
4249 : }
4250 :
4251 190 : if wrote_any_image {
4252 : // Normal path: we have written some data into the new image layer for this
4253 : // partition, so flush it to disk.
4254 24 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4255 12 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4256 12 : info!(
4257 0 : "created image layer for metadata {}",
4258 0 : image_layer.local_path()
4259 : );
4260 12 : Ok(ImageLayerCreationOutcome {
4261 12 : image: Some(image_layer),
4262 12 : next_start_key: img_range.end,
4263 12 : })
4264 : } else {
4265 : // Special case: the image layer may be empty if this is a sharded tenant and the
4266 : // partition does not cover any keys owned by this shard. In this case, to ensure
4267 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4268 : // layer we write will cover the key range that we just scanned.
4269 178 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4270 178 : Ok(ImageLayerCreationOutcome {
4271 178 : image: None,
4272 178 : next_start_key: start,
4273 178 : })
4274 : }
4275 192 : }
4276 :
4277 : /// Predicate function which indicates whether we should check if new image layers
4278 : /// are required. Since checking if new image layers are required is expensive in
4279 : /// terms of CPU, we only do it in the following cases:
4280 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4281 : /// 2. If enough time has passed since the last check:
4282 : /// 1. For large tenants, we wish to perform the check more often since they
4283 : /// suffer from the lack of image layers
4284 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4285 716 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4286 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4287 :
4288 716 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4289 716 : let distance = lsn
4290 716 : .checked_sub(last_checks_at)
4291 716 : .expect("Attempt to compact with LSN going backwards");
4292 716 : let min_distance =
4293 716 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4294 716 :
4295 716 : let distance_based_decision = distance.0 >= min_distance;
4296 716 :
4297 716 : let mut time_based_decision = false;
4298 716 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4299 716 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4300 614 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4301 : {
4302 0 : self.get_checkpoint_timeout()
4303 : } else {
4304 614 : Duration::from_secs(3600 * 48)
4305 : };
4306 :
4307 614 : time_based_decision = match *last_check_instant {
4308 438 : Some(last_check) => {
4309 438 : let elapsed = last_check.elapsed();
4310 438 : elapsed >= check_required_after
4311 : }
4312 176 : None => true,
4313 : };
4314 102 : }
4315 :
4316 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4317 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4318 : // check.
4319 716 : let decision = distance_based_decision || time_based_decision;
4320 :
4321 716 : if decision {
4322 178 : self.last_image_layer_creation_check_at.store(lsn);
4323 178 : *last_check_instant = Some(Instant::now());
4324 538 : }
4325 :
4326 716 : decision
4327 716 : }
4328 :
4329 716 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4330 : async fn create_image_layers(
4331 : self: &Arc<Timeline>,
4332 : partitioning: &KeyPartitioning,
4333 : lsn: Lsn,
4334 : mode: ImageLayerCreationMode,
4335 : ctx: &RequestContext,
4336 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4337 : let timer = self.metrics.create_images_time_histo.start_timer();
4338 : let mut image_layers = Vec::new();
4339 :
4340 : // We need to avoid holes between generated image layers.
4341 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4342 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4343 : //
4344 : // How such hole between partitions can appear?
4345 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4346 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4347 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4348 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4349 : let mut start = Key::MIN;
4350 :
4351 : let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
4352 :
4353 : for partition in partitioning.parts.iter() {
4354 : if self.cancel.is_cancelled() {
4355 : return Err(CreateImageLayersError::Cancelled);
4356 : }
4357 :
4358 : let img_range = start..partition.ranges.last().unwrap().end;
4359 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4360 : if compact_metadata {
4361 : for range in &partition.ranges {
4362 : assert!(
4363 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4364 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4365 : "metadata keys must be partitioned separately"
4366 : );
4367 : }
4368 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4369 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4370 : // might mess up with evictions.
4371 : start = img_range.end;
4372 : continue;
4373 : }
4374 : // For initial and force modes, we always generate image layers for metadata keys.
4375 : } else if let ImageLayerCreationMode::Try = mode {
4376 : // check_for_image_layers = false -> skip
4377 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4378 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4379 : start = img_range.end;
4380 : continue;
4381 : }
4382 : }
4383 : if let ImageLayerCreationMode::Force = mode {
4384 : // When forced to create image layers, we might try and create them where they already
4385 : // exist. This mode is only used in tests/debug.
4386 : let layers = self.layers.read().await;
4387 : if layers.contains_key(&PersistentLayerKey {
4388 : key_range: img_range.clone(),
4389 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4390 : is_delta: false,
4391 : }) {
4392 : tracing::info!(
4393 : "Skipping image layer at {lsn} {}..{}, already exists",
4394 : img_range.start,
4395 : img_range.end
4396 : );
4397 : start = img_range.end;
4398 : continue;
4399 : }
4400 : }
4401 :
4402 : let image_layer_writer = ImageLayerWriter::new(
4403 : self.conf,
4404 : self.timeline_id,
4405 : self.tenant_shard_id,
4406 : &img_range,
4407 : lsn,
4408 : ctx,
4409 : )
4410 : .await?;
4411 :
4412 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4413 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4414 0 : "failpoint image-layer-writer-fail-before-finish"
4415 0 : )))
4416 0 : });
4417 :
4418 : if !compact_metadata {
4419 : let ImageLayerCreationOutcome {
4420 : image,
4421 : next_start_key,
4422 : } = self
4423 : .create_image_layer_for_rel_blocks(
4424 : partition,
4425 : image_layer_writer,
4426 : lsn,
4427 : ctx,
4428 : img_range,
4429 : start,
4430 : )
4431 : .await?;
4432 :
4433 : start = next_start_key;
4434 : image_layers.extend(image);
4435 : } else {
4436 : let ImageLayerCreationOutcome {
4437 : image,
4438 : next_start_key,
4439 : } = self
4440 : .create_image_layer_for_metadata_keys(
4441 : partition,
4442 : image_layer_writer,
4443 : lsn,
4444 : ctx,
4445 : img_range,
4446 : mode,
4447 : start,
4448 : )
4449 : .await?;
4450 : start = next_start_key;
4451 : image_layers.extend(image);
4452 : }
4453 : }
4454 :
4455 : let mut guard = self.layers.write().await;
4456 :
4457 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4458 : // now they are being scheduled outside of write lock; current way is inconsistent with
4459 : // compaction lock order.
4460 : guard
4461 : .open_mut()?
4462 : .track_new_image_layers(&image_layers, &self.metrics);
4463 : drop_wlock(guard);
4464 : timer.stop_and_record();
4465 :
4466 : // Creating image layers may have caused some previously visible layers to be covered
4467 : if !image_layers.is_empty() {
4468 : self.update_layer_visibility().await?;
4469 : }
4470 :
4471 : Ok(image_layers)
4472 : }
4473 :
4474 : /// Wait until the background initial logical size calculation is complete, or
4475 : /// this Timeline is shut down. Calling this function will cause the initial
4476 : /// logical size calculation to skip waiting for the background jobs barrier.
4477 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4478 0 : if !self.shard_identity.is_shard_zero() {
4479 : // We don't populate logical size on shard >0: skip waiting for it.
4480 0 : return;
4481 0 : }
4482 0 :
4483 0 : if self.remote_client.is_deleting() {
4484 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4485 0 : return;
4486 0 : }
4487 0 :
4488 0 : if self.current_logical_size.current_size().is_exact() {
4489 : // root timelines are initialized with exact count, but never start the background
4490 : // calculation
4491 0 : return;
4492 0 : }
4493 :
4494 0 : if let Some(await_bg_cancel) = self
4495 0 : .current_logical_size
4496 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4497 0 : .get()
4498 0 : {
4499 0 : await_bg_cancel.cancel();
4500 0 : } else {
4501 : // We should not wait if we were not able to explicitly instruct
4502 : // the logical size cancellation to skip the concurrency limit semaphore.
4503 : // TODO: this is an unexpected case. We should restructure so that it
4504 : // can't happen.
4505 0 : tracing::warn!(
4506 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4507 : );
4508 0 : debug_assert!(false);
4509 : }
4510 :
4511 0 : tokio::select!(
4512 0 : _ = self.current_logical_size.initialized.acquire() => {},
4513 0 : _ = self.cancel.cancelled() => {}
4514 : )
4515 0 : }
4516 :
4517 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4518 : /// Timelines layers up to the ancestor_lsn.
4519 : ///
4520 : /// Requires a timeline that:
4521 : /// - has an ancestor to detach from
4522 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4523 : /// a technical requirement
4524 : ///
4525 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4526 : /// polled again until completion.
4527 : ///
4528 : /// During the operation all timelines sharing the data with this timeline will be reparented
4529 : /// from our ancestor to be branches of this timeline.
4530 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4531 0 : self: &Arc<Timeline>,
4532 0 : tenant: &crate::tenant::Tenant,
4533 0 : options: detach_ancestor::Options,
4534 0 : ctx: &RequestContext,
4535 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
4536 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4537 0 : }
4538 :
4539 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
4540 : /// reparents any reparentable children of previous ancestor.
4541 : ///
4542 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
4543 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
4544 : /// successfully, tenant must be reloaded.
4545 : ///
4546 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
4547 : /// resetting the tenant.
4548 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
4549 0 : self: &Arc<Timeline>,
4550 0 : tenant: &crate::tenant::Tenant,
4551 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4552 0 : ctx: &RequestContext,
4553 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
4554 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
4555 0 : }
4556 :
4557 : /// Final step which unblocks the GC.
4558 : ///
4559 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
4560 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4561 0 : self: &Arc<Timeline>,
4562 0 : tenant: &crate::tenant::Tenant,
4563 0 : attempt: detach_ancestor::Attempt,
4564 0 : ctx: &RequestContext,
4565 0 : ) -> Result<(), detach_ancestor::Error> {
4566 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
4567 0 : }
4568 : }
4569 :
4570 : impl Drop for Timeline {
4571 10 : fn drop(&mut self) {
4572 10 : if let Some(ancestor) = &self.ancestor_timeline {
4573 : // This lock should never be poisoned, but in case it is we do a .map() instead of
4574 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
4575 4 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
4576 4 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
4577 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
4578 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
4579 4 : }
4580 0 : }
4581 6 : }
4582 10 : }
4583 : }
4584 :
4585 : /// Top-level failure to compact.
4586 0 : #[derive(Debug, thiserror::Error)]
4587 : pub(crate) enum CompactionError {
4588 : #[error("The timeline or pageserver is shutting down")]
4589 : ShuttingDown,
4590 : /// Compaction tried to offload a timeline and failed
4591 : #[error("Failed to offload timeline: {0}")]
4592 : Offload(OffloadError),
4593 : /// Compaction cannot be done right now; page reconstruction and so on.
4594 : #[error(transparent)]
4595 : Other(anyhow::Error),
4596 : }
4597 :
4598 : impl From<OffloadError> for CompactionError {
4599 0 : fn from(e: OffloadError) -> Self {
4600 0 : match e {
4601 0 : OffloadError::Cancelled => Self::ShuttingDown,
4602 0 : _ => Self::Offload(e),
4603 : }
4604 0 : }
4605 : }
4606 :
4607 : impl CompactionError {
4608 0 : pub fn is_cancelled(&self) -> bool {
4609 0 : matches!(self, CompactionError::ShuttingDown)
4610 0 : }
4611 : }
4612 :
4613 : impl From<CollectKeySpaceError> for CompactionError {
4614 0 : fn from(err: CollectKeySpaceError) -> Self {
4615 0 : match err {
4616 : CollectKeySpaceError::Cancelled
4617 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4618 0 : CompactionError::ShuttingDown
4619 : }
4620 0 : e => CompactionError::Other(e.into()),
4621 : }
4622 0 : }
4623 : }
4624 :
4625 : impl From<super::upload_queue::NotInitialized> for CompactionError {
4626 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
4627 0 : match value {
4628 : super::upload_queue::NotInitialized::Uninitialized => {
4629 0 : CompactionError::Other(anyhow::anyhow!(value))
4630 : }
4631 : super::upload_queue::NotInitialized::ShuttingDown
4632 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
4633 : }
4634 0 : }
4635 : }
4636 :
4637 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
4638 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
4639 0 : match e {
4640 : super::storage_layer::layer::DownloadError::TimelineShutdown
4641 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
4642 0 : CompactionError::ShuttingDown
4643 : }
4644 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
4645 : | super::storage_layer::layer::DownloadError::DownloadRequired
4646 : | super::storage_layer::layer::DownloadError::NotFile(_)
4647 : | super::storage_layer::layer::DownloadError::DownloadFailed
4648 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
4649 0 : CompactionError::Other(anyhow::anyhow!(e))
4650 : }
4651 : #[cfg(test)]
4652 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
4653 0 : CompactionError::Other(anyhow::anyhow!(e))
4654 : }
4655 : }
4656 0 : }
4657 : }
4658 :
4659 : impl From<layer_manager::Shutdown> for CompactionError {
4660 0 : fn from(_: layer_manager::Shutdown) -> Self {
4661 0 : CompactionError::ShuttingDown
4662 0 : }
4663 : }
4664 :
4665 : #[serde_as]
4666 196 : #[derive(serde::Serialize)]
4667 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4668 :
4669 : #[derive(Default)]
4670 : enum DurationRecorder {
4671 : #[default]
4672 : NotStarted,
4673 : Recorded(RecordedDuration, tokio::time::Instant),
4674 : }
4675 :
4676 : impl DurationRecorder {
4677 504 : fn till_now(&self) -> DurationRecorder {
4678 504 : match self {
4679 : DurationRecorder::NotStarted => {
4680 0 : panic!("must only call on recorded measurements")
4681 : }
4682 504 : DurationRecorder::Recorded(_, ended) => {
4683 504 : let now = tokio::time::Instant::now();
4684 504 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4685 504 : }
4686 504 : }
4687 504 : }
4688 196 : fn into_recorded(self) -> Option<RecordedDuration> {
4689 196 : match self {
4690 0 : DurationRecorder::NotStarted => None,
4691 196 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4692 : }
4693 196 : }
4694 : }
4695 :
4696 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
4697 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
4698 : /// the layer descriptor requires the user to provide the ranges, which should cover all
4699 : /// keys specified in the `data` field.
4700 : #[cfg(test)]
4701 : #[derive(Clone)]
4702 : pub struct DeltaLayerTestDesc {
4703 : pub lsn_range: Range<Lsn>,
4704 : pub key_range: Range<Key>,
4705 : pub data: Vec<(Key, Lsn, Value)>,
4706 : }
4707 :
4708 : #[cfg(test)]
4709 : impl DeltaLayerTestDesc {
4710 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
4711 2 : Self {
4712 2 : lsn_range,
4713 2 : key_range,
4714 2 : data,
4715 2 : }
4716 2 : }
4717 :
4718 72 : pub fn new_with_inferred_key_range(
4719 72 : lsn_range: Range<Lsn>,
4720 72 : data: Vec<(Key, Lsn, Value)>,
4721 72 : ) -> Self {
4722 196 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
4723 196 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
4724 72 : Self {
4725 72 : key_range: (*key_min)..(key_max.next()),
4726 72 : lsn_range,
4727 72 : data,
4728 72 : }
4729 72 : }
4730 :
4731 10 : pub(crate) fn layer_name(&self) -> LayerName {
4732 10 : LayerName::Delta(super::storage_layer::DeltaLayerName {
4733 10 : key_range: self.key_range.clone(),
4734 10 : lsn_range: self.lsn_range.clone(),
4735 10 : })
4736 10 : }
4737 : }
4738 :
4739 : impl Timeline {
4740 28 : async fn finish_compact_batch(
4741 28 : self: &Arc<Self>,
4742 28 : new_deltas: &[ResidentLayer],
4743 28 : new_images: &[ResidentLayer],
4744 28 : layers_to_remove: &[Layer],
4745 28 : ) -> Result<(), CompactionError> {
4746 28 : let mut guard = tokio::select! {
4747 28 : guard = self.layers.write() => guard,
4748 28 : _ = self.cancel.cancelled() => {
4749 0 : return Err(CompactionError::ShuttingDown);
4750 : }
4751 : };
4752 :
4753 28 : let mut duplicated_layers = HashSet::new();
4754 28 :
4755 28 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
4756 :
4757 336 : for l in new_deltas {
4758 308 : if guard.contains(l.as_ref()) {
4759 : // expected in tests
4760 0 : tracing::error!(layer=%l, "duplicated L1 layer");
4761 :
4762 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
4763 : // `guard` on self.layers. as of writing this, there are no error returns except
4764 : // for compact_level0_phase1 creating an L0, which does not happen in practice
4765 : // because we have not implemented L0 => L0 compaction.
4766 0 : duplicated_layers.insert(l.layer_desc().key());
4767 308 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
4768 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
4769 308 : } else {
4770 308 : insert_layers.push(l.clone());
4771 308 : }
4772 : }
4773 :
4774 : // only remove those inputs which were not outputs
4775 28 : let remove_layers: Vec<Layer> = layers_to_remove
4776 28 : .iter()
4777 402 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
4778 28 : .cloned()
4779 28 : .collect();
4780 28 :
4781 28 : if !new_images.is_empty() {
4782 0 : guard
4783 0 : .open_mut()?
4784 0 : .track_new_image_layers(new_images, &self.metrics);
4785 28 : }
4786 :
4787 28 : guard
4788 28 : .open_mut()?
4789 28 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
4790 28 :
4791 28 : self.remote_client
4792 28 : .schedule_compaction_update(&remove_layers, new_deltas)?;
4793 :
4794 28 : drop_wlock(guard);
4795 28 :
4796 28 : Ok(())
4797 28 : }
4798 :
4799 0 : async fn rewrite_layers(
4800 0 : self: &Arc<Self>,
4801 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
4802 0 : mut drop_layers: Vec<Layer>,
4803 0 : ) -> Result<(), CompactionError> {
4804 0 : let mut guard = self.layers.write().await;
4805 :
4806 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
4807 : // to avoid double-removing, and avoid rewriting something that was removed.
4808 0 : replace_layers.retain(|(l, _)| guard.contains(l));
4809 0 : drop_layers.retain(|l| guard.contains(l));
4810 0 :
4811 0 : guard
4812 0 : .open_mut()?
4813 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
4814 0 :
4815 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
4816 0 :
4817 0 : self.remote_client
4818 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
4819 :
4820 0 : Ok(())
4821 0 : }
4822 :
4823 : /// Schedules the uploads of the given image layers
4824 364 : fn upload_new_image_layers(
4825 364 : self: &Arc<Self>,
4826 364 : new_images: impl IntoIterator<Item = ResidentLayer>,
4827 364 : ) -> Result<(), super::upload_queue::NotInitialized> {
4828 390 : for layer in new_images {
4829 26 : self.remote_client.schedule_layer_file_upload(layer)?;
4830 : }
4831 : // should any new image layer been created, not uploading index_part will
4832 : // result in a mismatch between remote_physical_size and layermap calculated
4833 : // size, which will fail some tests, but should not be an issue otherwise.
4834 364 : self.remote_client
4835 364 : .schedule_index_upload_for_file_changes()?;
4836 364 : Ok(())
4837 364 : }
4838 :
4839 0 : async fn find_gc_time_cutoff(
4840 0 : &self,
4841 0 : pitr: Duration,
4842 0 : cancel: &CancellationToken,
4843 0 : ctx: &RequestContext,
4844 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
4845 0 : debug_assert_current_span_has_tenant_and_timeline_id();
4846 0 : if self.shard_identity.is_shard_zero() {
4847 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
4848 0 : let now = SystemTime::now();
4849 0 : let time_range = if pitr == Duration::ZERO {
4850 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
4851 : } else {
4852 0 : pitr
4853 : };
4854 :
4855 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
4856 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
4857 0 : let timestamp = to_pg_timestamp(time_cutoff);
4858 :
4859 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
4860 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
4861 0 : LsnForTimestamp::Future(lsn) => {
4862 0 : // The timestamp is in the future. That sounds impossible,
4863 0 : // but what it really means is that there hasn't been
4864 0 : // any commits since the cutoff timestamp.
4865 0 : //
4866 0 : // In this case we should use the LSN of the most recent commit,
4867 0 : // which is implicitly the last LSN in the log.
4868 0 : debug!("future({})", lsn);
4869 0 : Some(self.get_last_record_lsn())
4870 : }
4871 0 : LsnForTimestamp::Past(lsn) => {
4872 0 : debug!("past({})", lsn);
4873 0 : None
4874 : }
4875 0 : LsnForTimestamp::NoData(lsn) => {
4876 0 : debug!("nodata({})", lsn);
4877 0 : None
4878 : }
4879 : };
4880 0 : Ok(time_cutoff)
4881 : } else {
4882 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
4883 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
4884 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
4885 : // space cutoff that we would also have respected ourselves.
4886 0 : match self
4887 0 : .remote_client
4888 0 : .download_foreign_index(ShardNumber(0), cancel)
4889 0 : .await
4890 : {
4891 0 : Ok((index_part, index_generation, _index_mtime)) => {
4892 0 : tracing::info!("GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
4893 0 : index_part.metadata.latest_gc_cutoff_lsn());
4894 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
4895 : }
4896 : Err(DownloadError::NotFound) => {
4897 : // This is unexpected, because during timeline creations shard zero persists to remote
4898 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
4899 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
4900 : // state, then shard zero should _eventually_ write an index when it GCs.
4901 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
4902 0 : Ok(None)
4903 : }
4904 0 : Err(e) => {
4905 0 : // TODO: this function should return a different error type than page reconstruct error
4906 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
4907 : }
4908 : }
4909 :
4910 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
4911 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
4912 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
4913 : // new location. This is legal in principle, but problematic in practice because it might result
4914 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
4915 : // because they have GC'd past the branch point.
4916 : }
4917 0 : }
4918 :
4919 : /// Find the Lsns above which layer files need to be retained on
4920 : /// garbage collection.
4921 : ///
4922 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
4923 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
4924 : /// the space-based retention.
4925 : ///
4926 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
4927 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
4928 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
4929 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
4930 : /// in the response as the GC cutoff point for the timeline.
4931 4 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
4932 : pub(super) async fn find_gc_cutoffs(
4933 : &self,
4934 : space_cutoff: Lsn,
4935 : pitr: Duration,
4936 : cancel: &CancellationToken,
4937 : ctx: &RequestContext,
4938 : ) -> Result<GcCutoffs, PageReconstructError> {
4939 : let _timer = self
4940 : .metrics
4941 : .find_gc_cutoffs_histo
4942 : .start_timer()
4943 : .record_on_drop();
4944 :
4945 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
4946 :
4947 : if cfg!(test) {
4948 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
4949 : if pitr == Duration::ZERO {
4950 : return Ok(GcCutoffs {
4951 : time: self.get_last_record_lsn(),
4952 : space: space_cutoff,
4953 : });
4954 : }
4955 : }
4956 :
4957 : // Calculate a time-based limit on how much to retain:
4958 : // - if PITR interval is set, then this is our cutoff.
4959 : // - if PITR interval is not set, then we do a lookup
4960 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
4961 : let time_cutoff = self.find_gc_time_cutoff(pitr, cancel, ctx).await?;
4962 :
4963 : Ok(match (pitr, time_cutoff) {
4964 : (Duration::ZERO, Some(time_cutoff)) => {
4965 : // PITR is not set. Retain the size-based limit, or the default time retention,
4966 : // whichever requires less data.
4967 : GcCutoffs {
4968 : time: self.get_last_record_lsn(),
4969 : space: std::cmp::max(time_cutoff, space_cutoff),
4970 : }
4971 : }
4972 : (Duration::ZERO, None) => {
4973 : // PITR is not set, and time lookup failed
4974 : GcCutoffs {
4975 : time: self.get_last_record_lsn(),
4976 : space: space_cutoff,
4977 : }
4978 : }
4979 : (_, None) => {
4980 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
4981 : // cannot advance beyond what was already GC'd, and respect space-based retention
4982 : GcCutoffs {
4983 : time: *self.get_latest_gc_cutoff_lsn(),
4984 : space: space_cutoff,
4985 : }
4986 : }
4987 : (_, Some(time_cutoff)) => {
4988 : // PITR interval is set and we looked up timestamp successfully. Ignore
4989 : // size based retention and make time cutoff authoritative
4990 : GcCutoffs {
4991 : time: time_cutoff,
4992 : space: time_cutoff,
4993 : }
4994 : }
4995 : })
4996 : }
4997 :
4998 : /// Garbage collect layer files on a timeline that are no longer needed.
4999 : ///
5000 : /// Currently, we don't make any attempt at removing unneeded page versions
5001 : /// within a layer file. We can only remove the whole file if it's fully
5002 : /// obsolete.
5003 4 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
5004 : // this is most likely the background tasks, but it might be the spawned task from
5005 : // immediate_gc
5006 4 : let _g = tokio::select! {
5007 4 : guard = self.gc_lock.lock() => guard,
5008 4 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
5009 : };
5010 4 : let timer = self.metrics.garbage_collect_histo.start_timer();
5011 4 :
5012 4 : fail_point!("before-timeline-gc");
5013 4 :
5014 4 : // Is the timeline being deleted?
5015 4 : if self.is_stopping() {
5016 0 : return Err(GcError::TimelineCancelled);
5017 4 : }
5018 4 :
5019 4 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
5020 4 : let gc_info = self.gc_info.read().unwrap();
5021 4 :
5022 4 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
5023 4 : let time_cutoff = gc_info.cutoffs.time;
5024 4 : let retain_lsns = gc_info
5025 4 : .retain_lsns
5026 4 : .iter()
5027 4 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
5028 4 : .collect();
5029 4 :
5030 4 : // Gets the maximum LSN that holds the valid lease.
5031 4 : //
5032 4 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
5033 4 : // Here, we do not check for stale leases again.
5034 4 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
5035 4 :
5036 4 : (
5037 4 : space_cutoff,
5038 4 : time_cutoff,
5039 4 : retain_lsns,
5040 4 : max_lsn_with_valid_lease,
5041 4 : )
5042 4 : };
5043 4 :
5044 4 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
5045 4 : let standby_horizon = self.standby_horizon.load();
5046 4 : // Hold GC for the standby, but as a safety guard do it only within some
5047 4 : // reasonable lag.
5048 4 : if standby_horizon != Lsn::INVALID {
5049 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
5050 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
5051 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
5052 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
5053 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
5054 : } else {
5055 0 : warn!(
5056 0 : "standby is lagging for more than {}MB, not holding gc for it",
5057 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
5058 : )
5059 : }
5060 0 : }
5061 4 : }
5062 :
5063 : // Reset standby horizon to ignore it if it is not updated till next GC.
5064 : // It is an easy way to unset it when standby disappears without adding
5065 : // more conf options.
5066 4 : self.standby_horizon.store(Lsn::INVALID);
5067 4 : self.metrics
5068 4 : .standby_horizon_gauge
5069 4 : .set(Lsn::INVALID.0 as i64);
5070 :
5071 4 : let res = self
5072 4 : .gc_timeline(
5073 4 : space_cutoff,
5074 4 : time_cutoff,
5075 4 : retain_lsns,
5076 4 : max_lsn_with_valid_lease,
5077 4 : new_gc_cutoff,
5078 4 : )
5079 4 : .instrument(
5080 4 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5081 : )
5082 0 : .await?;
5083 :
5084 : // only record successes
5085 4 : timer.stop_and_record();
5086 4 :
5087 4 : Ok(res)
5088 4 : }
5089 :
5090 4 : async fn gc_timeline(
5091 4 : &self,
5092 4 : space_cutoff: Lsn,
5093 4 : time_cutoff: Lsn,
5094 4 : retain_lsns: Vec<Lsn>,
5095 4 : max_lsn_with_valid_lease: Option<Lsn>,
5096 4 : new_gc_cutoff: Lsn,
5097 4 : ) -> Result<GcResult, GcError> {
5098 4 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5099 4 :
5100 4 : let now = SystemTime::now();
5101 4 : let mut result: GcResult = GcResult::default();
5102 4 :
5103 4 : // Nothing to GC. Return early.
5104 4 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
5105 4 : if latest_gc_cutoff >= new_gc_cutoff {
5106 0 : info!(
5107 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5108 : );
5109 0 : return Ok(result);
5110 4 : }
5111 :
5112 : // We need to ensure that no one tries to read page versions or create
5113 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5114 : // for details. This will block until the old value is no longer in use.
5115 : //
5116 : // The GC cutoff should only ever move forwards.
5117 4 : let waitlist = {
5118 4 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
5119 4 : if *write_guard > new_gc_cutoff {
5120 0 : return Err(GcError::BadLsn {
5121 0 : why: format!(
5122 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5123 0 : *write_guard, new_gc_cutoff
5124 0 : ),
5125 0 : });
5126 4 : }
5127 4 :
5128 4 : write_guard.store_and_unlock(new_gc_cutoff)
5129 4 : };
5130 4 : waitlist.wait().await;
5131 :
5132 4 : info!("GC starting");
5133 :
5134 4 : debug!("retain_lsns: {:?}", retain_lsns);
5135 :
5136 4 : let mut layers_to_remove = Vec::new();
5137 :
5138 : // Scan all layers in the timeline (remote or on-disk).
5139 : //
5140 : // Garbage collect the layer if all conditions are satisfied:
5141 : // 1. it is older than cutoff LSN;
5142 : // 2. it is older than PITR interval;
5143 : // 3. it doesn't need to be retained for 'retain_lsns';
5144 : // 4. it does not need to be kept for LSNs holding valid leases.
5145 : // 5. newer on-disk image layers cover the layer's whole key range
5146 : //
5147 : // TODO holding a write lock is too agressive and avoidable
5148 4 : let mut guard = self.layers.write().await;
5149 4 : let layers = guard.layer_map()?;
5150 24 : 'outer: for l in layers.iter_historic_layers() {
5151 24 : result.layers_total += 1;
5152 24 :
5153 24 : // 1. Is it newer than GC horizon cutoff point?
5154 24 : if l.get_lsn_range().end > space_cutoff {
5155 2 : info!(
5156 0 : "keeping {} because it's newer than space_cutoff {}",
5157 0 : l.layer_name(),
5158 : space_cutoff,
5159 : );
5160 2 : result.layers_needed_by_cutoff += 1;
5161 2 : continue 'outer;
5162 22 : }
5163 22 :
5164 22 : // 2. It is newer than PiTR cutoff point?
5165 22 : if l.get_lsn_range().end > time_cutoff {
5166 0 : info!(
5167 0 : "keeping {} because it's newer than time_cutoff {}",
5168 0 : l.layer_name(),
5169 : time_cutoff,
5170 : );
5171 0 : result.layers_needed_by_pitr += 1;
5172 0 : continue 'outer;
5173 22 : }
5174 :
5175 : // 3. Is it needed by a child branch?
5176 : // NOTE With that we would keep data that
5177 : // might be referenced by child branches forever.
5178 : // We can track this in child timeline GC and delete parent layers when
5179 : // they are no longer needed. This might be complicated with long inheritance chains.
5180 : //
5181 : // TODO Vec is not a great choice for `retain_lsns`
5182 22 : for retain_lsn in &retain_lsns {
5183 : // start_lsn is inclusive
5184 0 : if &l.get_lsn_range().start <= retain_lsn {
5185 0 : info!(
5186 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5187 0 : l.layer_name(),
5188 0 : retain_lsn,
5189 0 : l.is_incremental(),
5190 : );
5191 0 : result.layers_needed_by_branches += 1;
5192 0 : continue 'outer;
5193 0 : }
5194 : }
5195 :
5196 : // 4. Is there a valid lease that requires us to keep this layer?
5197 22 : if let Some(lsn) = &max_lsn_with_valid_lease {
5198 : // keep if layer start <= any of the lease
5199 18 : if &l.get_lsn_range().start <= lsn {
5200 14 : info!(
5201 0 : "keeping {} because there is a valid lease preventing GC at {}",
5202 0 : l.layer_name(),
5203 : lsn,
5204 : );
5205 14 : result.layers_needed_by_leases += 1;
5206 14 : continue 'outer;
5207 4 : }
5208 4 : }
5209 :
5210 : // 5. Is there a later on-disk layer for this relation?
5211 : //
5212 : // The end-LSN is exclusive, while disk_consistent_lsn is
5213 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5214 : // OK for a delta layer to have end LSN 101, but if the end LSN
5215 : // is 102, then it might not have been fully flushed to disk
5216 : // before crash.
5217 : //
5218 : // For example, imagine that the following layers exist:
5219 : //
5220 : // 1000 - image (A)
5221 : // 1000-2000 - delta (B)
5222 : // 2000 - image (C)
5223 : // 2000-3000 - delta (D)
5224 : // 3000 - image (E)
5225 : //
5226 : // If GC horizon is at 2500, we can remove layers A and B, but
5227 : // we cannot remove C, even though it's older than 2500, because
5228 : // the delta layer 2000-3000 depends on it.
5229 8 : if !layers
5230 8 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5231 : {
5232 6 : info!("keeping {} because it is the latest layer", l.layer_name());
5233 6 : result.layers_not_updated += 1;
5234 6 : continue 'outer;
5235 2 : }
5236 2 :
5237 2 : // We didn't find any reason to keep this file, so remove it.
5238 2 : info!(
5239 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5240 0 : l.layer_name(),
5241 0 : l.is_incremental(),
5242 : );
5243 2 : layers_to_remove.push(l);
5244 : }
5245 :
5246 4 : if !layers_to_remove.is_empty() {
5247 : // Persist the new GC cutoff value before we actually remove anything.
5248 : // This unconditionally schedules also an index_part.json update, even though, we will
5249 : // be doing one a bit later with the unlinked gc'd layers.
5250 2 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5251 2 : self.schedule_uploads(disk_consistent_lsn, None)
5252 2 : .map_err(|e| {
5253 0 : if self.cancel.is_cancelled() {
5254 0 : GcError::TimelineCancelled
5255 : } else {
5256 0 : GcError::Remote(e)
5257 : }
5258 2 : })?;
5259 :
5260 2 : let gc_layers = layers_to_remove
5261 2 : .iter()
5262 2 : .map(|x| guard.get_from_desc(x))
5263 2 : .collect::<Vec<Layer>>();
5264 2 :
5265 2 : result.layers_removed = gc_layers.len() as u64;
5266 2 :
5267 2 : self.remote_client.schedule_gc_update(&gc_layers)?;
5268 :
5269 2 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5270 2 :
5271 2 : #[cfg(feature = "testing")]
5272 2 : {
5273 2 : result.doomed_layers = gc_layers;
5274 2 : }
5275 2 : }
5276 :
5277 4 : info!(
5278 0 : "GC completed removing {} layers, cutoff {}",
5279 : result.layers_removed, new_gc_cutoff
5280 : );
5281 :
5282 4 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5283 4 : Ok(result)
5284 4 : }
5285 :
5286 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5287 667508 : async fn reconstruct_value(
5288 667508 : &self,
5289 667508 : key: Key,
5290 667508 : request_lsn: Lsn,
5291 667508 : mut data: ValueReconstructState,
5292 667508 : ) -> Result<Bytes, PageReconstructError> {
5293 667508 : // Perform WAL redo if needed
5294 667508 : data.records.reverse();
5295 667508 :
5296 667508 : // If we have a page image, and no WAL, we're all set
5297 667508 : if data.records.is_empty() {
5298 667098 : if let Some((img_lsn, img)) = &data.img {
5299 667098 : trace!(
5300 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5301 : key,
5302 : img_lsn,
5303 : request_lsn,
5304 : );
5305 667098 : Ok(img.clone())
5306 : } else {
5307 0 : Err(PageReconstructError::from(anyhow!(
5308 0 : "base image for {key} at {request_lsn} not found"
5309 0 : )))
5310 : }
5311 : } else {
5312 : // We need to do WAL redo.
5313 : //
5314 : // If we don't have a base image, then the oldest WAL record better initialize
5315 : // the page
5316 410 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5317 0 : Err(PageReconstructError::from(anyhow!(
5318 0 : "Base image for {} at {} not found, but got {} WAL records",
5319 0 : key,
5320 0 : request_lsn,
5321 0 : data.records.len()
5322 0 : )))
5323 : } else {
5324 410 : if data.img.is_some() {
5325 344 : trace!(
5326 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5327 0 : data.records.len(),
5328 : key,
5329 : request_lsn
5330 : );
5331 : } else {
5332 66 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5333 : };
5334 410 : let res = self
5335 410 : .walredo_mgr
5336 410 : .as_ref()
5337 410 : .context("timeline has no walredo manager")
5338 410 : .map_err(PageReconstructError::WalRedo)?
5339 410 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5340 0 : .await;
5341 410 : let img = match res {
5342 410 : Ok(img) => img,
5343 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5344 0 : Err(walredo::Error::Other(e)) => {
5345 0 : return Err(PageReconstructError::WalRedo(
5346 0 : e.context("reconstruct a page image"),
5347 0 : ))
5348 : }
5349 : };
5350 410 : Ok(img)
5351 : }
5352 : }
5353 667508 : }
5354 :
5355 0 : pub(crate) async fn spawn_download_all_remote_layers(
5356 0 : self: Arc<Self>,
5357 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5358 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5359 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5360 :
5361 : // this is not really needed anymore; it has tests which really check the return value from
5362 : // http api. it would be better not to maintain this anymore.
5363 :
5364 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5365 0 : if let Some(st) = &*status_guard {
5366 0 : match &st.state {
5367 : DownloadRemoteLayersTaskState::Running => {
5368 0 : return Err(st.clone());
5369 : }
5370 : DownloadRemoteLayersTaskState::ShutDown
5371 0 : | DownloadRemoteLayersTaskState::Completed => {
5372 0 : *status_guard = None;
5373 0 : }
5374 : }
5375 0 : }
5376 :
5377 0 : let self_clone = Arc::clone(&self);
5378 0 : let task_id = task_mgr::spawn(
5379 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5380 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5381 0 : self.tenant_shard_id,
5382 0 : Some(self.timeline_id),
5383 0 : "download all remote layers task",
5384 0 : async move {
5385 0 : self_clone.download_all_remote_layers(request).await;
5386 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5387 0 : match &mut *status_guard {
5388 : None => {
5389 0 : warn!("tasks status is supposed to be Some(), since we are running");
5390 : }
5391 0 : Some(st) => {
5392 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5393 0 : if st.task_id != exp_task_id {
5394 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5395 0 : } else {
5396 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5397 0 : }
5398 : }
5399 : };
5400 0 : Ok(())
5401 0 : }
5402 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))
5403 : );
5404 :
5405 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5406 0 : task_id: format!("{task_id}"),
5407 0 : state: DownloadRemoteLayersTaskState::Running,
5408 0 : total_layer_count: 0,
5409 0 : successful_download_count: 0,
5410 0 : failed_download_count: 0,
5411 0 : };
5412 0 : *status_guard = Some(initial_info.clone());
5413 0 :
5414 0 : Ok(initial_info)
5415 0 : }
5416 :
5417 0 : async fn download_all_remote_layers(
5418 0 : self: &Arc<Self>,
5419 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5420 0 : ) {
5421 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5422 :
5423 0 : let remaining = {
5424 0 : let guard = self.layers.read().await;
5425 0 : let Ok(lm) = guard.layer_map() else {
5426 : // technically here we could look into iterating accessible layers, but downloading
5427 : // all layers of a shutdown timeline makes no sense regardless.
5428 0 : tracing::info!("attempted to download all layers of shutdown timeline");
5429 0 : return;
5430 : };
5431 0 : lm.iter_historic_layers()
5432 0 : .map(|desc| guard.get_from_desc(&desc))
5433 0 : .collect::<Vec<_>>()
5434 0 : };
5435 0 : let total_layer_count = remaining.len();
5436 :
5437 : macro_rules! lock_status {
5438 : ($st:ident) => {
5439 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5440 : let st = st
5441 : .as_mut()
5442 : .expect("this function is only called after the task has been spawned");
5443 : assert_eq!(
5444 : st.task_id,
5445 : format!(
5446 : "{}",
5447 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5448 : )
5449 : );
5450 : let $st = st;
5451 : };
5452 : }
5453 :
5454 : {
5455 0 : lock_status!(st);
5456 0 : st.total_layer_count = total_layer_count as u64;
5457 0 : }
5458 0 :
5459 0 : let mut remaining = remaining.into_iter();
5460 0 : let mut have_remaining = true;
5461 0 : let mut js = tokio::task::JoinSet::new();
5462 0 :
5463 0 : let cancel = task_mgr::shutdown_token();
5464 0 :
5465 0 : let limit = request.max_concurrent_downloads;
5466 :
5467 : loop {
5468 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5469 0 : let Some(next) = remaining.next() else {
5470 0 : have_remaining = false;
5471 0 : break;
5472 : };
5473 :
5474 0 : let span = tracing::info_span!("download", layer = %next);
5475 :
5476 0 : js.spawn(
5477 0 : async move {
5478 0 : let res = next.download().await;
5479 0 : (next, res)
5480 0 : }
5481 0 : .instrument(span),
5482 0 : );
5483 0 : }
5484 :
5485 0 : while let Some(res) = js.join_next().await {
5486 0 : match res {
5487 : Ok((_, Ok(_))) => {
5488 0 : lock_status!(st);
5489 0 : st.successful_download_count += 1;
5490 : }
5491 0 : Ok((layer, Err(e))) => {
5492 0 : tracing::error!(%layer, "download failed: {e:#}");
5493 0 : lock_status!(st);
5494 0 : st.failed_download_count += 1;
5495 : }
5496 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5497 0 : Err(je) if je.is_panic() => {
5498 0 : lock_status!(st);
5499 0 : st.failed_download_count += 1;
5500 : }
5501 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5502 : }
5503 : }
5504 :
5505 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5506 0 : break;
5507 0 : }
5508 : }
5509 :
5510 : {
5511 0 : lock_status!(st);
5512 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5513 : }
5514 0 : }
5515 :
5516 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5517 0 : &self,
5518 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5519 0 : self.download_all_remote_layers_task_info
5520 0 : .read()
5521 0 : .unwrap()
5522 0 : .clone()
5523 0 : }
5524 : }
5525 :
5526 : impl Timeline {
5527 : /// Returns non-remote layers for eviction.
5528 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5529 0 : let guard = self.layers.read().await;
5530 0 : let mut max_layer_size: Option<u64> = None;
5531 0 :
5532 0 : let resident_layers = guard
5533 0 : .likely_resident_layers()
5534 0 : .map(|layer| {
5535 0 : let file_size = layer.layer_desc().file_size;
5536 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5537 0 :
5538 0 : let last_activity_ts = layer.latest_activity();
5539 0 :
5540 0 : EvictionCandidate {
5541 0 : layer: layer.to_owned().into(),
5542 0 : last_activity_ts,
5543 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5544 0 : visibility: layer.visibility(),
5545 0 : }
5546 0 : })
5547 0 : .collect();
5548 0 :
5549 0 : DiskUsageEvictionInfo {
5550 0 : max_layer_size,
5551 0 : resident_layers,
5552 0 : }
5553 0 : }
5554 :
5555 1730 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5556 1730 : ShardIndex {
5557 1730 : shard_number: self.tenant_shard_id.shard_number,
5558 1730 : shard_count: self.tenant_shard_id.shard_count,
5559 1730 : }
5560 1730 : }
5561 :
5562 : /// Persistently blocks gc for `Manual` reason.
5563 : ///
5564 : /// Returns true if no such block existed before, false otherwise.
5565 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
5566 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5567 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5568 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
5569 0 : }
5570 :
5571 : /// Persistently unblocks gc for `Manual` reason.
5572 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
5573 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5574 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5575 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
5576 0 : }
5577 :
5578 : #[cfg(test)]
5579 40 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5580 40 : self.last_record_lsn.advance(new_lsn);
5581 40 : }
5582 :
5583 : #[cfg(test)]
5584 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
5585 2 : self.disk_consistent_lsn.store(new_value);
5586 2 : }
5587 :
5588 : /// Force create an image layer and place it into the layer map.
5589 : ///
5590 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5591 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5592 : /// placed into the layer map in one run AND be validated.
5593 : #[cfg(test)]
5594 52 : pub(super) async fn force_create_image_layer(
5595 52 : self: &Arc<Timeline>,
5596 52 : lsn: Lsn,
5597 52 : mut images: Vec<(Key, Bytes)>,
5598 52 : check_start_lsn: Option<Lsn>,
5599 52 : ctx: &RequestContext,
5600 52 : ) -> anyhow::Result<()> {
5601 52 : let last_record_lsn = self.get_last_record_lsn();
5602 52 : assert!(
5603 52 : lsn <= last_record_lsn,
5604 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5605 : );
5606 52 : if let Some(check_start_lsn) = check_start_lsn {
5607 52 : assert!(lsn >= check_start_lsn);
5608 0 : }
5609 126 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5610 52 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5611 52 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5612 52 : let mut image_layer_writer = ImageLayerWriter::new(
5613 52 : self.conf,
5614 52 : self.timeline_id,
5615 52 : self.tenant_shard_id,
5616 52 : &(min_key..end_key),
5617 52 : lsn,
5618 52 : ctx,
5619 52 : )
5620 26 : .await?;
5621 230 : for (key, img) in images {
5622 178 : image_layer_writer.put_image(key, img, ctx).await?;
5623 : }
5624 104 : let (desc, path) = image_layer_writer.finish(ctx).await?;
5625 52 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5626 52 : info!("force created image layer {}", image_layer.local_path());
5627 : {
5628 52 : let mut guard = self.layers.write().await;
5629 52 : guard.open_mut().unwrap().force_insert_layer(image_layer);
5630 52 : }
5631 52 :
5632 52 : Ok(())
5633 52 : }
5634 :
5635 : /// Force create a delta layer and place it into the layer map.
5636 : ///
5637 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5638 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5639 : /// placed into the layer map in one run AND be validated.
5640 : #[cfg(test)]
5641 74 : pub(super) async fn force_create_delta_layer(
5642 74 : self: &Arc<Timeline>,
5643 74 : mut deltas: DeltaLayerTestDesc,
5644 74 : check_start_lsn: Option<Lsn>,
5645 74 : ctx: &RequestContext,
5646 74 : ) -> anyhow::Result<()> {
5647 74 : let last_record_lsn = self.get_last_record_lsn();
5648 74 : deltas
5649 74 : .data
5650 124 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5651 74 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5652 74 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5653 272 : for (_, lsn, _) in &deltas.data {
5654 198 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5655 : }
5656 74 : assert!(
5657 74 : deltas.lsn_range.end <= last_record_lsn,
5658 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
5659 : deltas.lsn_range.end,
5660 : last_record_lsn
5661 : );
5662 74 : if let Some(check_start_lsn) = check_start_lsn {
5663 74 : assert!(deltas.lsn_range.start >= check_start_lsn);
5664 0 : }
5665 74 : let mut delta_layer_writer = DeltaLayerWriter::new(
5666 74 : self.conf,
5667 74 : self.timeline_id,
5668 74 : self.tenant_shard_id,
5669 74 : deltas.key_range.start,
5670 74 : deltas.lsn_range,
5671 74 : ctx,
5672 74 : )
5673 37 : .await?;
5674 272 : for (key, lsn, val) in deltas.data {
5675 198 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5676 : }
5677 185 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
5678 74 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5679 74 : info!("force created delta layer {}", delta_layer.local_path());
5680 : {
5681 74 : let mut guard = self.layers.write().await;
5682 74 : guard.open_mut().unwrap().force_insert_layer(delta_layer);
5683 74 : }
5684 74 :
5685 74 : Ok(())
5686 74 : }
5687 :
5688 : /// Return all keys at the LSN in the image layers
5689 : #[cfg(test)]
5690 6 : pub(crate) async fn inspect_image_layers(
5691 6 : self: &Arc<Timeline>,
5692 6 : lsn: Lsn,
5693 6 : ctx: &RequestContext,
5694 6 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
5695 6 : let mut all_data = Vec::new();
5696 6 : let guard = self.layers.read().await;
5697 34 : for layer in guard.layer_map()?.iter_historic_layers() {
5698 34 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
5699 8 : let layer = guard.get_from_desc(&layer);
5700 8 : let mut reconstruct_data = ValuesReconstructState::default();
5701 8 : layer
5702 8 : .get_values_reconstruct_data(
5703 8 : KeySpace::single(Key::MIN..Key::MAX),
5704 8 : lsn..Lsn(lsn.0 + 1),
5705 8 : &mut reconstruct_data,
5706 8 : ctx,
5707 8 : )
5708 13 : .await?;
5709 74 : for (k, v) in reconstruct_data.keys {
5710 66 : all_data.push((k, v?.img.unwrap().1));
5711 : }
5712 26 : }
5713 : }
5714 6 : all_data.sort();
5715 6 : Ok(all_data)
5716 6 : }
5717 :
5718 : /// Get all historic layer descriptors in the layer map
5719 : #[cfg(test)]
5720 12 : pub(crate) async fn inspect_historic_layers(
5721 12 : self: &Arc<Timeline>,
5722 12 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
5723 12 : let mut layers = Vec::new();
5724 12 : let guard = self.layers.read().await;
5725 70 : for layer in guard.layer_map()?.iter_historic_layers() {
5726 70 : layers.push(layer.key());
5727 70 : }
5728 12 : Ok(layers)
5729 12 : }
5730 :
5731 : #[cfg(test)]
5732 10 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
5733 10 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
5734 10 : keyspace.merge(&ks);
5735 10 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
5736 10 : }
5737 : }
5738 :
5739 : /// Tracking writes ingestion does to a particular in-memory layer.
5740 : ///
5741 : /// Cleared upon freezing a layer.
5742 : pub(crate) struct TimelineWriterState {
5743 : open_layer: Arc<InMemoryLayer>,
5744 : current_size: u64,
5745 : // Previous Lsn which passed through
5746 : prev_lsn: Option<Lsn>,
5747 : // Largest Lsn which passed through the current writer
5748 : max_lsn: Option<Lsn>,
5749 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
5750 : cached_last_freeze_at: Lsn,
5751 : }
5752 :
5753 : impl TimelineWriterState {
5754 1268 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
5755 1268 : Self {
5756 1268 : open_layer,
5757 1268 : current_size,
5758 1268 : prev_lsn: None,
5759 1268 : max_lsn: None,
5760 1268 : cached_last_freeze_at: last_freeze_at,
5761 1268 : }
5762 1268 : }
5763 : }
5764 :
5765 : /// Various functions to mutate the timeline.
5766 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
5767 : // This is probably considered a bad practice in Rust and should be fixed eventually,
5768 : // but will cause large code changes.
5769 : pub(crate) struct TimelineWriter<'a> {
5770 : tl: &'a Timeline,
5771 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
5772 : }
5773 :
5774 : impl Deref for TimelineWriter<'_> {
5775 : type Target = Timeline;
5776 :
5777 4807192 : fn deref(&self) -> &Self::Target {
5778 4807192 : self.tl
5779 4807192 : }
5780 : }
5781 :
5782 : #[derive(PartialEq)]
5783 : enum OpenLayerAction {
5784 : Roll,
5785 : Open,
5786 : None,
5787 : }
5788 :
5789 : impl<'a> TimelineWriter<'a> {
5790 4804204 : async fn handle_open_layer_action(
5791 4804204 : &mut self,
5792 4804204 : at: Lsn,
5793 4804204 : action: OpenLayerAction,
5794 4804204 : ctx: &RequestContext,
5795 4804204 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
5796 4804204 : match action {
5797 : OpenLayerAction::Roll => {
5798 80 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
5799 80 : self.roll_layer(freeze_at).await?;
5800 80 : self.open_layer(at, ctx).await?;
5801 : }
5802 1188 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
5803 : OpenLayerAction::None => {
5804 4802936 : assert!(self.write_guard.is_some());
5805 : }
5806 : }
5807 :
5808 4804204 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
5809 4804204 : }
5810 :
5811 1268 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
5812 1268 : let layer = self
5813 1268 : .tl
5814 1268 : .get_layer_for_write(at, &self.write_guard, ctx)
5815 720 : .await?;
5816 1268 : let initial_size = layer.size().await?;
5817 :
5818 1268 : let last_freeze_at = self.last_freeze_at.load();
5819 1268 : self.write_guard.replace(TimelineWriterState::new(
5820 1268 : layer,
5821 1268 : initial_size,
5822 1268 : last_freeze_at,
5823 1268 : ));
5824 1268 :
5825 1268 : Ok(())
5826 1268 : }
5827 :
5828 80 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
5829 80 : let current_size = self.write_guard.as_ref().unwrap().current_size;
5830 80 :
5831 80 : // self.write_guard will be taken by the freezing
5832 80 : self.tl
5833 80 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
5834 2 : .await?;
5835 :
5836 80 : assert!(self.write_guard.is_none());
5837 :
5838 80 : if current_size >= self.get_checkpoint_distance() * 2 {
5839 0 : warn!("Flushed oversized open layer with size {}", current_size)
5840 80 : }
5841 :
5842 80 : Ok(())
5843 80 : }
5844 :
5845 4804204 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
5846 4804204 : let state = &*self.write_guard;
5847 4804204 : let Some(state) = &state else {
5848 1188 : return OpenLayerAction::Open;
5849 : };
5850 :
5851 : #[cfg(feature = "testing")]
5852 4803016 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
5853 : // this check and assertion are not really needed because
5854 : // LayerManager::try_freeze_in_memory_layer will always clear out the
5855 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
5856 : // is no TimelineWriterState.
5857 0 : assert!(
5858 0 : state.open_layer.end_lsn.get().is_some(),
5859 0 : "our open_layer must be outdated"
5860 : );
5861 :
5862 : // this would be a memory leak waiting to happen because the in-memory layer always has
5863 : // an index
5864 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
5865 4803016 : }
5866 4803016 :
5867 4803016 : if state.prev_lsn == Some(lsn) {
5868 : // Rolling mid LSN is not supported by [downstream code].
5869 : // Hence, only roll at LSN boundaries.
5870 : //
5871 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
5872 6 : return OpenLayerAction::None;
5873 4803010 : }
5874 4803010 :
5875 4803010 : if state.current_size == 0 {
5876 : // Don't roll empty layers
5877 0 : return OpenLayerAction::None;
5878 4803010 : }
5879 4803010 :
5880 4803010 : if self.tl.should_roll(
5881 4803010 : state.current_size,
5882 4803010 : state.current_size + new_value_size,
5883 4803010 : self.get_checkpoint_distance(),
5884 4803010 : lsn,
5885 4803010 : state.cached_last_freeze_at,
5886 4803010 : state.open_layer.get_opened_at(),
5887 4803010 : ) {
5888 80 : OpenLayerAction::Roll
5889 : } else {
5890 4802930 : OpenLayerAction::None
5891 : }
5892 4804204 : }
5893 :
5894 : /// Put a batch of keys at the specified Lsns.
5895 4804202 : pub(crate) async fn put_batch(
5896 4804202 : &mut self,
5897 4804202 : batch: SerializedValueBatch,
5898 4804202 : ctx: &RequestContext,
5899 4804202 : ) -> anyhow::Result<()> {
5900 4804202 : if !batch.has_data() {
5901 0 : return Ok(());
5902 4804202 : }
5903 4804202 :
5904 4804202 : let batch_max_lsn = batch.max_lsn;
5905 4804202 : let buf_size: u64 = batch.buffer_size() as u64;
5906 4804202 :
5907 4804202 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
5908 4804202 : let layer = self
5909 4804202 : .handle_open_layer_action(batch_max_lsn, action, ctx)
5910 722 : .await?;
5911 :
5912 4804202 : let res = layer.put_batch(batch, ctx).await;
5913 :
5914 4804202 : if res.is_ok() {
5915 4804202 : // Update the current size only when the entire write was ok.
5916 4804202 : // In case of failures, we may have had partial writes which
5917 4804202 : // render the size tracking out of sync. That's ok because
5918 4804202 : // the checkpoint distance should be significantly smaller
5919 4804202 : // than the S3 single shot upload limit of 5GiB.
5920 4804202 : let state = self.write_guard.as_mut().unwrap();
5921 4804202 :
5922 4804202 : state.current_size += buf_size;
5923 4804202 : state.prev_lsn = Some(batch_max_lsn);
5924 4804202 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
5925 4804202 : }
5926 :
5927 4804202 : res
5928 4804202 : }
5929 :
5930 : #[cfg(test)]
5931 : /// Test helper, for tests that would like to poke individual values without composing a batch
5932 4390154 : pub(crate) async fn put(
5933 4390154 : &mut self,
5934 4390154 : key: Key,
5935 4390154 : lsn: Lsn,
5936 4390154 : value: &Value,
5937 4390154 : ctx: &RequestContext,
5938 4390154 : ) -> anyhow::Result<()> {
5939 : use utils::bin_ser::BeSer;
5940 4390154 : if !key.is_valid_key_on_write_path() {
5941 0 : bail!(
5942 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
5943 0 : key
5944 0 : );
5945 4390154 : }
5946 4390154 : let val_ser_size = value.serialized_size().unwrap() as usize;
5947 4390154 : let batch = SerializedValueBatch::from_values(vec![(
5948 4390154 : key.to_compact(),
5949 4390154 : lsn,
5950 4390154 : val_ser_size,
5951 4390154 : value.clone(),
5952 4390154 : )]);
5953 4390154 :
5954 4390154 : self.put_batch(batch, ctx).await
5955 4390154 : }
5956 :
5957 2 : pub(crate) async fn delete_batch(
5958 2 : &mut self,
5959 2 : batch: &[(Range<Key>, Lsn)],
5960 2 : ctx: &RequestContext,
5961 2 : ) -> anyhow::Result<()> {
5962 2 : if let Some((_, lsn)) = batch.first() {
5963 2 : let action = self.get_open_layer_action(*lsn, 0);
5964 2 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
5965 2 : layer.put_tombstones(batch).await?;
5966 0 : }
5967 :
5968 2 : Ok(())
5969 2 : }
5970 :
5971 : /// Track the end of the latest digested WAL record.
5972 : /// Remember the (end of) last valid WAL record remembered in the timeline.
5973 : ///
5974 : /// Call this after you have finished writing all the WAL up to 'lsn'.
5975 : ///
5976 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
5977 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
5978 : /// the latest and can be read.
5979 5279064 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
5980 5279064 : self.tl.finish_write(new_lsn);
5981 5279064 : }
5982 :
5983 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
5984 270570 : self.tl.update_current_logical_size(delta)
5985 270570 : }
5986 : }
5987 :
5988 : // We need TimelineWriter to be send in upcoming conversion of
5989 : // Timeline::layers to tokio::sync::RwLock.
5990 : #[test]
5991 2 : fn is_send() {
5992 2 : fn _assert_send<T: Send>() {}
5993 2 : _assert_send::<TimelineWriter<'_>>();
5994 2 : }
5995 :
5996 : #[cfg(test)]
5997 : mod tests {
5998 : use pageserver_api::key::Key;
5999 : use pageserver_api::value::Value;
6000 : use utils::{id::TimelineId, lsn::Lsn};
6001 :
6002 : use crate::tenant::{
6003 : harness::{test_img, TenantHarness},
6004 : layer_map::LayerMap,
6005 : storage_layer::{Layer, LayerName},
6006 : timeline::{DeltaLayerTestDesc, EvictionError},
6007 : Timeline,
6008 : };
6009 :
6010 : #[tokio::test]
6011 2 : async fn test_heatmap_generation() {
6012 2 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
6013 2 :
6014 2 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6015 2 : Lsn(0x10)..Lsn(0x20),
6016 2 : vec![(
6017 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6018 2 : Lsn(0x11),
6019 2 : Value::Image(test_img("foo")),
6020 2 : )],
6021 2 : );
6022 2 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6023 2 : Lsn(0x10)..Lsn(0x20),
6024 2 : vec![(
6025 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6026 2 : Lsn(0x11),
6027 2 : Value::Image(test_img("foo")),
6028 2 : )],
6029 2 : );
6030 2 : let l0_delta = DeltaLayerTestDesc::new(
6031 2 : Lsn(0x20)..Lsn(0x30),
6032 2 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6033 2 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6034 2 : vec![(
6035 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6036 2 : Lsn(0x25),
6037 2 : Value::Image(test_img("foo")),
6038 2 : )],
6039 2 : );
6040 2 : let delta_layers = vec![
6041 2 : covered_delta.clone(),
6042 2 : visible_delta.clone(),
6043 2 : l0_delta.clone(),
6044 2 : ];
6045 2 :
6046 2 : let image_layer = (
6047 2 : Lsn(0x40),
6048 2 : vec![(
6049 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6050 2 : test_img("bar"),
6051 2 : )],
6052 2 : );
6053 2 : let image_layers = vec![image_layer];
6054 2 :
6055 20 : let (tenant, ctx) = harness.load().await;
6056 2 : let timeline = tenant
6057 2 : .create_test_timeline_with_layers(
6058 2 : TimelineId::generate(),
6059 2 : Lsn(0x10),
6060 2 : 14,
6061 2 : &ctx,
6062 2 : delta_layers,
6063 2 : image_layers,
6064 2 : Lsn(0x100),
6065 2 : )
6066 31 : .await
6067 2 : .unwrap();
6068 2 :
6069 2 : // Layer visibility is an input to heatmap generation, so refresh it first
6070 2 : timeline.update_layer_visibility().await.unwrap();
6071 2 :
6072 2 : let heatmap = timeline
6073 2 : .generate_heatmap()
6074 2 : .await
6075 2 : .expect("Infallible while timeline is not shut down");
6076 2 :
6077 2 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
6078 2 :
6079 2 : // L0 should come last
6080 2 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
6081 2 :
6082 2 : let mut last_lsn = Lsn::MAX;
6083 10 : for layer in heatmap.layers {
6084 2 : // Covered layer should be omitted
6085 8 : assert!(layer.name != covered_delta.layer_name());
6086 2 :
6087 8 : let layer_lsn = match &layer.name {
6088 4 : LayerName::Delta(d) => d.lsn_range.end,
6089 4 : LayerName::Image(i) => i.lsn,
6090 2 : };
6091 2 :
6092 2 : // Apart from L0s, newest Layers should come first
6093 8 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
6094 6 : assert!(layer_lsn <= last_lsn);
6095 6 : last_lsn = layer_lsn;
6096 2 : }
6097 2 : }
6098 2 : }
6099 :
6100 : #[tokio::test]
6101 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
6102 2 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6103 2 : .await
6104 2 : .unwrap();
6105 2 :
6106 20 : let (tenant, ctx) = harness.load().await;
6107 2 : let timeline = tenant
6108 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6109 6 : .await
6110 2 : .unwrap();
6111 2 :
6112 2 : let layer = find_some_layer(&timeline).await;
6113 2 : let layer = layer
6114 2 : .keep_resident()
6115 2 : .await
6116 2 : .expect("no download => no downloading errors")
6117 2 : .drop_eviction_guard();
6118 2 :
6119 2 : let forever = std::time::Duration::from_secs(120);
6120 2 :
6121 2 : let first = layer.evict_and_wait(forever);
6122 2 : let second = layer.evict_and_wait(forever);
6123 2 :
6124 2 : let (first, second) = tokio::join!(first, second);
6125 2 :
6126 2 : let res = layer.keep_resident().await;
6127 2 : assert!(res.is_none(), "{res:?}");
6128 2 :
6129 2 : match (first, second) {
6130 2 : (Ok(()), Ok(())) => {
6131 2 : // because there are no more timeline locks being taken on eviction path, we can
6132 2 : // witness all three outcomes here.
6133 2 : }
6134 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
6135 0 : // if one completes before the other, this is fine just as well.
6136 0 : }
6137 2 : other => unreachable!("unexpected {:?}", other),
6138 2 : }
6139 2 : }
6140 :
6141 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
6142 2 : let layers = timeline.layers.read().await;
6143 2 : let desc = layers
6144 2 : .layer_map()
6145 2 : .unwrap()
6146 2 : .iter_historic_layers()
6147 2 : .next()
6148 2 : .expect("must find one layer to evict");
6149 2 :
6150 2 : layers.get_from_desc(&desc)
6151 2 : }
6152 : }
|
