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, ArcSwapOption};
18 : use bytes::Bytes;
19 : use camino::Utf8Path;
20 : use chrono::{DateTime, Utc};
21 : use enumset::EnumSet;
22 : use fail::fail_point;
23 : use futures::{stream::FuturesUnordered, StreamExt};
24 : use handle::ShardTimelineId;
25 : use layer_manager::Shutdown;
26 : use offload::OffloadError;
27 : use once_cell::sync::Lazy;
28 : use pageserver_api::models::PageTraceEvent;
29 : use pageserver_api::{
30 : key::{
31 : KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
32 : SPARSE_RANGE,
33 : },
34 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
35 : models::{
36 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
37 : DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy,
38 : InMemoryLayerInfo, LayerMapInfo, LsnLease, TimelineState,
39 : },
40 : reltag::BlockNumber,
41 : shard::{ShardIdentity, ShardNumber, TenantShardId},
42 : };
43 : use rand::Rng;
44 : use remote_storage::DownloadError;
45 : use serde_with::serde_as;
46 : use storage_broker::BrokerClientChannel;
47 : use tokio::sync::mpsc::Sender;
48 : use tokio::{
49 : runtime::Handle,
50 : sync::{oneshot, watch},
51 : };
52 : use tokio_util::sync::CancellationToken;
53 : use tracing::*;
54 : use utils::{
55 : fs_ext,
56 : guard_arc_swap::GuardArcSwap,
57 : pausable_failpoint,
58 : postgres_client::PostgresClientProtocol,
59 : sync::gate::{Gate, GateGuard},
60 : };
61 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
62 :
63 : use std::array;
64 : use std::cmp::{max, min};
65 : use std::collections::btree_map::Entry;
66 : use std::collections::{BTreeMap, HashMap, HashSet};
67 : use std::ops::{ControlFlow, Deref, Range};
68 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering as AtomicOrdering};
69 : use std::sync::{Arc, Mutex, OnceLock, RwLock, Weak};
70 : use std::time::{Duration, Instant, SystemTime};
71 :
72 : use crate::l0_flush::{self, L0FlushGlobalState};
73 : use crate::tenant::storage_layer::ImageLayerName;
74 : use crate::{
75 : aux_file::AuxFileSizeEstimator,
76 : page_service::TenantManagerTypes,
77 : tenant::{
78 : config::AttachmentMode,
79 : layer_map::{LayerMap, SearchResult},
80 : metadata::TimelineMetadata,
81 : storage_layer::{
82 : inmemory_layer::IndexEntry, BatchLayerWriter, IoConcurrency, PersistentLayerDesc,
83 : ValueReconstructSituation,
84 : },
85 : },
86 : walingest::WalLagCooldown,
87 : walredo,
88 : };
89 : use crate::{
90 : context::{DownloadBehavior, RequestContext},
91 : disk_usage_eviction_task::DiskUsageEvictionInfo,
92 : pgdatadir_mapping::CollectKeySpaceError,
93 : };
94 : use crate::{
95 : disk_usage_eviction_task::finite_f32,
96 : tenant::storage_layer::{
97 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
98 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
99 : ValuesReconstructState,
100 : },
101 : };
102 : use crate::{
103 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
104 : };
105 : use crate::{
106 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
107 : };
108 : use crate::{
109 : pgdatadir_mapping::DirectoryKind,
110 : virtual_file::{MaybeFatalIo, VirtualFile},
111 : };
112 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
113 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
114 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
115 :
116 : use crate::config::PageServerConf;
117 : use crate::keyspace::{KeyPartitioning, KeySpace};
118 : use crate::metrics::TimelineMetrics;
119 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
120 : use crate::tenant::config::TenantConfOpt;
121 : use pageserver_api::reltag::RelTag;
122 : use pageserver_api::shard::ShardIndex;
123 :
124 : use postgres_connection::PgConnectionConfig;
125 : use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
126 : use utils::{
127 : completion,
128 : generation::Generation,
129 : id::TimelineId,
130 : lsn::{AtomicLsn, Lsn, RecordLsn},
131 : seqwait::SeqWait,
132 : simple_rcu::{Rcu, RcuReadGuard},
133 : };
134 :
135 : use crate::task_mgr;
136 : use crate::task_mgr::TaskKind;
137 : use crate::tenant::gc_result::GcResult;
138 : use crate::ZERO_PAGE;
139 : use pageserver_api::key::Key;
140 :
141 : use self::delete::DeleteTimelineFlow;
142 : pub(super) use self::eviction_task::EvictionTaskTenantState;
143 : use self::eviction_task::EvictionTaskTimelineState;
144 : use self::layer_manager::LayerManager;
145 : use self::logical_size::LogicalSize;
146 : use self::walreceiver::{WalReceiver, WalReceiverConf};
147 :
148 : use super::{
149 : config::TenantConf, storage_layer::LayerVisibilityHint, upload_queue::NotInitialized,
150 : MaybeOffloaded,
151 : };
152 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
153 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
154 : use super::{
155 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
156 : storage_layer::ReadableLayer,
157 : };
158 : use super::{
159 : secondary::heatmap::{HeatMapLayer, HeatMapTimeline},
160 : GcError,
161 : };
162 :
163 : #[cfg(test)]
164 : use pageserver_api::value::Value;
165 :
166 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
167 : pub(crate) enum FlushLoopState {
168 : NotStarted,
169 : Running {
170 : #[cfg(test)]
171 : expect_initdb_optimization: bool,
172 : #[cfg(test)]
173 : initdb_optimization_count: usize,
174 : },
175 : Exited,
176 : }
177 :
178 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
179 : pub enum ImageLayerCreationMode {
180 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
181 : Try,
182 : /// Force creating the image layers if possible. For now, no image layers will be created
183 : /// for metadata keys. Used in compaction code path with force flag enabled.
184 : Force,
185 : /// Initial ingestion of the data, and no data should be dropped in this function. This
186 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
187 : /// code path.
188 : Initial,
189 : }
190 :
191 : impl std::fmt::Display for ImageLayerCreationMode {
192 1136 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
193 1136 : write!(f, "{:?}", self)
194 1136 : }
195 : }
196 :
197 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
198 : /// Can be removed after all refactors are done.
199 56 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
200 56 : drop(rlock)
201 56 : }
202 :
203 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
204 : /// Can be removed after all refactors are done.
205 1192 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
206 1192 : drop(rlock)
207 1192 : }
208 :
209 : /// The outward-facing resources required to build a Timeline
210 : pub struct TimelineResources {
211 : pub remote_client: RemoteTimelineClient,
212 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
213 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
214 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
215 : }
216 :
217 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
218 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
219 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
220 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
221 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
222 : pub(crate) struct RelSizeCache {
223 : pub(crate) complete_as_of: Lsn,
224 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
225 : }
226 :
227 : pub struct Timeline {
228 : pub(crate) conf: &'static PageServerConf,
229 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
230 :
231 : myself: Weak<Self>,
232 :
233 : pub(crate) tenant_shard_id: TenantShardId,
234 : pub timeline_id: TimelineId,
235 :
236 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
237 : /// Never changes for the lifetime of this [`Timeline`] object.
238 : ///
239 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
240 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
241 : pub(crate) generation: Generation,
242 :
243 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
244 : /// to shards, and is constant through the lifetime of this Timeline.
245 : shard_identity: ShardIdentity,
246 :
247 : pub pg_version: u32,
248 :
249 : /// The tuple has two elements.
250 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
251 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
252 : ///
253 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
254 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
255 : ///
256 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
257 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
258 : /// `PersistentLayerDesc`'s.
259 : ///
260 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
261 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
262 : /// runtime, e.g., during page reconstruction.
263 : ///
264 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
265 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
266 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
267 :
268 : last_freeze_at: AtomicLsn,
269 : // Atomic would be more appropriate here.
270 : last_freeze_ts: RwLock<Instant>,
271 :
272 : pub(crate) standby_horizon: AtomicLsn,
273 :
274 : // WAL redo manager. `None` only for broken tenants.
275 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
276 :
277 : /// Remote storage client.
278 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
279 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
280 :
281 : // What page versions do we hold in the repository? If we get a
282 : // request > last_record_lsn, we need to wait until we receive all
283 : // the WAL up to the request. The SeqWait provides functions for
284 : // that. TODO: If we get a request for an old LSN, such that the
285 : // versions have already been garbage collected away, we should
286 : // throw an error, but we don't track that currently.
287 : //
288 : // last_record_lsn.load().last points to the end of last processed WAL record.
289 : //
290 : // We also remember the starting point of the previous record in
291 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
292 : // first WAL record when the node is started up. But here, we just
293 : // keep track of it.
294 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
295 :
296 : // All WAL records have been processed and stored durably on files on
297 : // local disk, up to this LSN. On crash and restart, we need to re-process
298 : // the WAL starting from this point.
299 : //
300 : // Some later WAL records might have been processed and also flushed to disk
301 : // already, so don't be surprised to see some, but there's no guarantee on
302 : // them yet.
303 : disk_consistent_lsn: AtomicLsn,
304 :
305 : // Parent timeline that this timeline was branched from, and the LSN
306 : // of the branch point.
307 : ancestor_timeline: Option<Arc<Timeline>>,
308 : ancestor_lsn: Lsn,
309 :
310 : pub(super) metrics: TimelineMetrics,
311 :
312 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
313 : // in `crate::page_service` writes these metrics.
314 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
315 :
316 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
317 :
318 : /// Ensures layers aren't frozen by checkpointer between
319 : /// [`Timeline::get_layer_for_write`] and layer reads.
320 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
321 : /// Must always be acquired before the layer map/individual layer lock
322 : /// to avoid deadlock.
323 : ///
324 : /// The state is cleared upon freezing.
325 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
326 :
327 : /// Used to avoid multiple `flush_loop` tasks running
328 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
329 :
330 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
331 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
332 : /// The flush cycle counter is sent back on the layer_flush_done channel when
333 : /// the flush finishes. You can use that to wait for the flush to finish.
334 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
335 : /// read by whoever sends an update
336 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
337 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
338 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
339 :
340 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
341 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
342 :
343 : // List of child timelines and their branch points. This is needed to avoid
344 : // garbage collecting data that is still needed by the child timelines.
345 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
346 :
347 : // It may change across major versions so for simplicity
348 : // keep it after running initdb for a timeline.
349 : // It is needed in checks when we want to error on some operations
350 : // when they are requested for pre-initdb lsn.
351 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
352 : // though let's keep them both for better error visibility.
353 : pub initdb_lsn: Lsn,
354 :
355 : /// The repartitioning result. Allows a single writer and multiple readers.
356 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
357 :
358 : /// Configuration: how often should the partitioning be recalculated.
359 : repartition_threshold: u64,
360 :
361 : last_image_layer_creation_check_at: AtomicLsn,
362 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
363 :
364 : /// Current logical size of the "datadir", at the last LSN.
365 : current_logical_size: LogicalSize,
366 :
367 : /// Information about the last processed message by the WAL receiver,
368 : /// or None if WAL receiver has not received anything for this timeline
369 : /// yet.
370 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
371 : pub walreceiver: Mutex<Option<WalReceiver>>,
372 :
373 : /// Relation size cache
374 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
375 :
376 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
377 :
378 : state: watch::Sender<TimelineState>,
379 :
380 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
381 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
382 : pub delete_progress: TimelineDeleteProgress,
383 :
384 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
385 :
386 : /// Load or creation time information about the disk_consistent_lsn and when the loading
387 : /// happened. Used for consumption metrics.
388 : pub(crate) loaded_at: (Lsn, SystemTime),
389 :
390 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
391 : pub(crate) gate: Gate,
392 :
393 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
394 : /// to the timeline should drop out when this token fires.
395 : pub(crate) cancel: CancellationToken,
396 :
397 : /// Make sure we only have one running compaction at a time in tests.
398 : ///
399 : /// Must only be taken in two places:
400 : /// - [`Timeline::compact`] (this file)
401 : /// - [`delete::delete_local_timeline_directory`]
402 : ///
403 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
404 : compaction_lock: tokio::sync::Mutex<()>,
405 :
406 : /// If true, the last compaction failed.
407 : compaction_failed: AtomicBool,
408 :
409 : /// Make sure we only have one running gc at a time.
410 : ///
411 : /// Must only be taken in two places:
412 : /// - [`Timeline::gc`] (this file)
413 : /// - [`delete::delete_local_timeline_directory`]
414 : ///
415 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
416 : gc_lock: tokio::sync::Mutex<()>,
417 :
418 : /// Cloned from [`super::Tenant::pagestream_throttle`] on construction.
419 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
420 :
421 : /// Size estimator for aux file v2
422 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
423 :
424 : /// Some test cases directly place keys into the timeline without actually modifying the directory
425 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
426 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
427 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
428 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
429 : #[cfg(test)]
430 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
431 :
432 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
433 :
434 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
435 :
436 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
437 :
438 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
439 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
440 :
441 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
442 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
443 : }
444 :
445 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
446 :
447 : pub struct WalReceiverInfo {
448 : pub wal_source_connconf: PgConnectionConfig,
449 : pub last_received_msg_lsn: Lsn,
450 : pub last_received_msg_ts: u128,
451 : }
452 :
453 : /// Information about how much history needs to be retained, needed by
454 : /// Garbage Collection.
455 : #[derive(Default)]
456 : pub(crate) struct GcInfo {
457 : /// Specific LSNs that are needed.
458 : ///
459 : /// Currently, this includes all points where child branches have
460 : /// been forked off from. In the future, could also include
461 : /// explicit user-defined snapshot points.
462 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
463 :
464 : /// The cutoff coordinates, which are combined by selecting the minimum.
465 : pub(crate) cutoffs: GcCutoffs,
466 :
467 : /// Leases granted to particular LSNs.
468 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
469 :
470 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
471 : pub(crate) within_ancestor_pitr: bool,
472 : }
473 :
474 : impl GcInfo {
475 564 : pub(crate) fn min_cutoff(&self) -> Lsn {
476 564 : self.cutoffs.select_min()
477 564 : }
478 :
479 464 : pub(super) fn insert_child(
480 464 : &mut self,
481 464 : child_id: TimelineId,
482 464 : child_lsn: Lsn,
483 464 : is_offloaded: MaybeOffloaded,
484 464 : ) {
485 464 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
486 464 : self.retain_lsns.sort_by_key(|i| i.0);
487 464 : }
488 :
489 8 : pub(super) fn remove_child_maybe_offloaded(
490 8 : &mut self,
491 8 : child_id: TimelineId,
492 8 : maybe_offloaded: MaybeOffloaded,
493 8 : ) -> bool {
494 8 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
495 8 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
496 8 : let mut removed = false;
497 12 : self.retain_lsns.retain(|i| {
498 12 : if removed {
499 4 : return true;
500 8 : }
501 8 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
502 8 : removed |= remove;
503 8 : !remove
504 12 : });
505 8 : removed
506 8 : }
507 :
508 8 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
509 8 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
510 8 : }
511 :
512 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
513 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
514 0 : }
515 : }
516 :
517 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
518 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
519 : /// between time-based and space-based retention for observability and consumption metrics purposes.
520 : #[derive(Debug, Clone)]
521 : pub(crate) struct GcCutoffs {
522 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
523 : /// history we must keep to retain a specified number of bytes of WAL.
524 : pub(crate) space: Lsn,
525 :
526 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
527 : /// history we must keep to enable reading back at least the PITR interval duration.
528 : pub(crate) time: Lsn,
529 : }
530 :
531 : impl Default for GcCutoffs {
532 892 : fn default() -> Self {
533 892 : Self {
534 892 : space: Lsn::INVALID,
535 892 : time: Lsn::INVALID,
536 892 : }
537 892 : }
538 : }
539 :
540 : impl GcCutoffs {
541 564 : fn select_min(&self) -> Lsn {
542 564 : std::cmp::min(self.space, self.time)
543 564 : }
544 : }
545 :
546 : pub(crate) struct TimelineVisitOutcome {
547 : completed_keyspace: KeySpace,
548 : image_covered_keyspace: KeySpace,
549 : }
550 :
551 : /// An error happened in a get() operation.
552 : #[derive(thiserror::Error, Debug)]
553 : pub(crate) enum PageReconstructError {
554 : #[error(transparent)]
555 : Other(anyhow::Error),
556 :
557 : #[error("Ancestor LSN wait error: {0}")]
558 : AncestorLsnTimeout(WaitLsnError),
559 :
560 : #[error("timeline shutting down")]
561 : Cancelled,
562 :
563 : /// An error happened replaying WAL records
564 : #[error(transparent)]
565 : WalRedo(anyhow::Error),
566 :
567 : #[error("{0}")]
568 : MissingKey(MissingKeyError),
569 : }
570 :
571 : impl From<anyhow::Error> for PageReconstructError {
572 0 : fn from(value: anyhow::Error) -> Self {
573 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
574 0 : match value.downcast::<PageReconstructError>() {
575 0 : Ok(pre) => pre,
576 0 : Err(other) => PageReconstructError::Other(other),
577 : }
578 0 : }
579 : }
580 :
581 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
582 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
583 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
584 0 : }
585 : }
586 :
587 : impl From<layer_manager::Shutdown> for PageReconstructError {
588 0 : fn from(_: layer_manager::Shutdown) -> Self {
589 0 : PageReconstructError::Cancelled
590 0 : }
591 : }
592 :
593 : impl GetVectoredError {
594 : #[cfg(test)]
595 12 : pub(crate) fn is_missing_key_error(&self) -> bool {
596 12 : matches!(self, Self::MissingKey(_))
597 12 : }
598 : }
599 :
600 : impl From<layer_manager::Shutdown> for GetVectoredError {
601 0 : fn from(_: layer_manager::Shutdown) -> Self {
602 0 : GetVectoredError::Cancelled
603 0 : }
604 : }
605 :
606 : #[derive(thiserror::Error)]
607 : pub struct MissingKeyError {
608 : key: Key,
609 : shard: ShardNumber,
610 : cont_lsn: Lsn,
611 : request_lsn: Lsn,
612 : ancestor_lsn: Option<Lsn>,
613 : backtrace: Option<std::backtrace::Backtrace>,
614 : }
615 :
616 : impl std::fmt::Debug for MissingKeyError {
617 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
618 0 : write!(f, "{}", self)
619 0 : }
620 : }
621 :
622 : impl std::fmt::Display for MissingKeyError {
623 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
624 0 : write!(
625 0 : f,
626 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
627 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
628 0 : )?;
629 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
630 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
631 0 : }
632 :
633 0 : if let Some(ref backtrace) = self.backtrace {
634 0 : write!(f, "\n{}", backtrace)?;
635 0 : }
636 :
637 0 : Ok(())
638 0 : }
639 : }
640 :
641 : impl PageReconstructError {
642 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
643 0 : pub(crate) fn is_stopping(&self) -> bool {
644 : use PageReconstructError::*;
645 0 : match self {
646 0 : Cancelled => true,
647 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
648 : }
649 0 : }
650 : }
651 :
652 : #[derive(thiserror::Error, Debug)]
653 : pub(crate) enum CreateImageLayersError {
654 : #[error("timeline shutting down")]
655 : Cancelled,
656 :
657 : #[error("read failed")]
658 : GetVectoredError(#[source] GetVectoredError),
659 :
660 : #[error("reconstruction failed")]
661 : PageReconstructError(#[source] PageReconstructError),
662 :
663 : #[error(transparent)]
664 : Other(#[from] anyhow::Error),
665 : }
666 :
667 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
668 0 : fn from(_: layer_manager::Shutdown) -> Self {
669 0 : CreateImageLayersError::Cancelled
670 0 : }
671 : }
672 :
673 : #[derive(thiserror::Error, Debug, Clone)]
674 : pub(crate) enum FlushLayerError {
675 : /// Timeline cancellation token was cancelled
676 : #[error("timeline shutting down")]
677 : Cancelled,
678 :
679 : /// We tried to flush a layer while the Timeline is in an unexpected state
680 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
681 : NotRunning(FlushLoopState),
682 :
683 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
684 : // loop via a watch channel, where we can only borrow it.
685 : #[error("create image layers (shared)")]
686 : CreateImageLayersError(Arc<CreateImageLayersError>),
687 :
688 : #[error("other (shared)")]
689 : Other(#[from] Arc<anyhow::Error>),
690 : }
691 :
692 : impl FlushLayerError {
693 : // When crossing from generic anyhow errors to this error type, we explicitly check
694 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
695 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
696 0 : let cancelled = timeline.cancel.is_cancelled()
697 : // The upload queue might have been shut down before the official cancellation of the timeline.
698 0 : || err
699 0 : .downcast_ref::<NotInitialized>()
700 0 : .map(NotInitialized::is_stopping)
701 0 : .unwrap_or_default();
702 0 : if cancelled {
703 0 : Self::Cancelled
704 : } else {
705 0 : Self::Other(Arc::new(err))
706 : }
707 0 : }
708 : }
709 :
710 : impl From<layer_manager::Shutdown> for FlushLayerError {
711 0 : fn from(_: layer_manager::Shutdown) -> Self {
712 0 : FlushLayerError::Cancelled
713 0 : }
714 : }
715 :
716 : #[derive(thiserror::Error, Debug)]
717 : pub(crate) enum GetVectoredError {
718 : #[error("timeline shutting down")]
719 : Cancelled,
720 :
721 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
722 : Oversized(u64),
723 :
724 : #[error("requested at invalid LSN: {0}")]
725 : InvalidLsn(Lsn),
726 :
727 : #[error("requested key not found: {0}")]
728 : MissingKey(MissingKeyError),
729 :
730 : #[error("ancestry walk")]
731 : GetReadyAncestorError(#[source] GetReadyAncestorError),
732 :
733 : #[error(transparent)]
734 : Other(#[from] anyhow::Error),
735 : }
736 :
737 : impl From<GetReadyAncestorError> for GetVectoredError {
738 4 : fn from(value: GetReadyAncestorError) -> Self {
739 : use GetReadyAncestorError::*;
740 4 : match value {
741 0 : Cancelled => GetVectoredError::Cancelled,
742 : AncestorLsnTimeout(_) | BadState { .. } => {
743 4 : GetVectoredError::GetReadyAncestorError(value)
744 : }
745 : }
746 4 : }
747 : }
748 :
749 : #[derive(thiserror::Error, Debug)]
750 : pub(crate) enum GetReadyAncestorError {
751 : #[error("ancestor LSN wait error")]
752 : AncestorLsnTimeout(#[from] WaitLsnError),
753 :
754 : #[error("bad state on timeline {timeline_id}: {state:?}")]
755 : BadState {
756 : timeline_id: TimelineId,
757 : state: TimelineState,
758 : },
759 :
760 : #[error("cancelled")]
761 : Cancelled,
762 : }
763 :
764 : #[derive(Clone, Copy)]
765 : pub enum LogicalSizeCalculationCause {
766 : Initial,
767 : ConsumptionMetricsSyntheticSize,
768 : EvictionTaskImitation,
769 : TenantSizeHandler,
770 : }
771 :
772 : pub enum GetLogicalSizePriority {
773 : User,
774 : Background,
775 : }
776 :
777 0 : #[derive(Debug, enumset::EnumSetType)]
778 : pub(crate) enum CompactFlags {
779 : ForceRepartition,
780 : ForceImageLayerCreation,
781 : ForceL0Compaction,
782 : EnhancedGcBottomMostCompaction,
783 : DryRun,
784 : }
785 :
786 : #[serde_with::serde_as]
787 0 : #[derive(Debug, Clone, serde::Deserialize)]
788 : pub(crate) struct CompactRequest {
789 : pub compact_key_range: Option<CompactKeyRange>,
790 : pub compact_lsn_range: Option<CompactLsnRange>,
791 : /// Whether the compaction job should be scheduled.
792 : #[serde(default)]
793 : pub scheduled: bool,
794 : /// Whether the compaction job should be split across key ranges.
795 : #[serde(default)]
796 : pub sub_compaction: bool,
797 : /// Max job size for each subcompaction job.
798 : pub sub_compaction_max_job_size_mb: Option<u64>,
799 : }
800 :
801 : #[derive(Debug, Clone, Default)]
802 : pub(crate) struct CompactOptions {
803 : pub flags: EnumSet<CompactFlags>,
804 : /// If set, the compaction will only compact the key range specified by this option.
805 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
806 : pub compact_key_range: Option<CompactKeyRange>,
807 : /// If set, the compaction will only compact the LSN within this value.
808 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
809 : pub compact_lsn_range: Option<CompactLsnRange>,
810 : /// Enable sub-compaction (split compaction job across key ranges).
811 : /// This option is only used by GC compaction.
812 : pub sub_compaction: bool,
813 : /// Set job size for the GC compaction.
814 : /// This option is only used by GC compaction.
815 : pub sub_compaction_max_job_size_mb: Option<u64>,
816 : }
817 :
818 : impl std::fmt::Debug for Timeline {
819 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
820 0 : write!(f, "Timeline<{}>", self.timeline_id)
821 0 : }
822 : }
823 :
824 : #[derive(thiserror::Error, Debug)]
825 : pub(crate) enum WaitLsnError {
826 : // Called on a timeline which is shutting down
827 : #[error("Shutdown")]
828 : Shutdown,
829 :
830 : // Called on an timeline not in active state or shutting down
831 : #[error("Bad timeline state: {0:?}")]
832 : BadState(TimelineState),
833 :
834 : // Timeout expired while waiting for LSN to catch up with goal.
835 : #[error("{0}")]
836 : Timeout(String),
837 : }
838 :
839 : // The impls below achieve cancellation mapping for errors.
840 : // Perhaps there's a way of achieving this with less cruft.
841 :
842 : impl From<CreateImageLayersError> for CompactionError {
843 0 : fn from(e: CreateImageLayersError) -> Self {
844 0 : match e {
845 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
846 0 : CreateImageLayersError::Other(e) => {
847 0 : CompactionError::Other(e.context("create image layers"))
848 : }
849 0 : _ => CompactionError::Other(e.into()),
850 : }
851 0 : }
852 : }
853 :
854 : impl From<CreateImageLayersError> for FlushLayerError {
855 0 : fn from(e: CreateImageLayersError) -> Self {
856 0 : match e {
857 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
858 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
859 : }
860 0 : }
861 : }
862 :
863 : impl From<PageReconstructError> for CreateImageLayersError {
864 0 : fn from(e: PageReconstructError) -> Self {
865 0 : match e {
866 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
867 0 : _ => CreateImageLayersError::PageReconstructError(e),
868 : }
869 0 : }
870 : }
871 :
872 : impl From<GetVectoredError> for CreateImageLayersError {
873 0 : fn from(e: GetVectoredError) -> Self {
874 0 : match e {
875 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
876 0 : _ => CreateImageLayersError::GetVectoredError(e),
877 : }
878 0 : }
879 : }
880 :
881 : impl From<GetVectoredError> for PageReconstructError {
882 12 : fn from(e: GetVectoredError) -> Self {
883 12 : match e {
884 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
885 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
886 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
887 8 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
888 4 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
889 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
890 : }
891 12 : }
892 : }
893 :
894 : impl From<GetReadyAncestorError> for PageReconstructError {
895 4 : fn from(e: GetReadyAncestorError) -> Self {
896 : use GetReadyAncestorError::*;
897 4 : match e {
898 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
899 4 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
900 0 : Cancelled => PageReconstructError::Cancelled,
901 : }
902 4 : }
903 : }
904 :
905 : pub(crate) enum WaitLsnTimeout {
906 : Custom(Duration),
907 : // Use the [`PageServerConf::wait_lsn_timeout`] default
908 : Default,
909 : }
910 :
911 : pub(crate) enum WaitLsnWaiter<'a> {
912 : Timeline(&'a Timeline),
913 : Tenant,
914 : PageService,
915 : HttpEndpoint,
916 : }
917 :
918 : /// Argument to [`Timeline::shutdown`].
919 : #[derive(Debug, Clone, Copy)]
920 : pub(crate) enum ShutdownMode {
921 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
922 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
923 : ///
924 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
925 : /// the call to [`Timeline::shutdown`].
926 : FreezeAndFlush,
927 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
928 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
929 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
930 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
931 : Reload,
932 : /// Shut down immediately, without waiting for any open layers to flush.
933 : Hard,
934 : }
935 :
936 : struct ImageLayerCreationOutcome {
937 : unfinished_image_layer: Option<ImageLayerWriter>,
938 : next_start_key: Key,
939 : }
940 :
941 : /// Public interface functions
942 : impl Timeline {
943 : /// Get the LSN where this branch was created
944 8 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
945 8 : self.ancestor_lsn
946 8 : }
947 :
948 : /// Get the ancestor's timeline id
949 24 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
950 24 : self.ancestor_timeline
951 24 : .as_ref()
952 24 : .map(|ancestor| ancestor.timeline_id)
953 24 : }
954 :
955 : /// Get the ancestor timeline
956 4 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
957 4 : self.ancestor_timeline.as_ref()
958 4 : }
959 :
960 : /// Get the bytes written since the PITR cutoff on this branch, and
961 : /// whether this branch's ancestor_lsn is within its parent's PITR.
962 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
963 0 : let gc_info = self.gc_info.read().unwrap();
964 0 : let history = self
965 0 : .get_last_record_lsn()
966 0 : .checked_sub(gc_info.cutoffs.time)
967 0 : .unwrap_or(Lsn(0))
968 0 : .0;
969 0 : (history, gc_info.within_ancestor_pitr)
970 0 : }
971 :
972 : /// Lock and get timeline's GC cutoff
973 604 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
974 604 : self.latest_gc_cutoff_lsn.read()
975 604 : }
976 :
977 : /// Look up given page version.
978 : ///
979 : /// If a remote layer file is needed, it is downloaded as part of this
980 : /// call.
981 : ///
982 : /// This method enforces [`Self::pagestream_throttle`] internally.
983 : ///
984 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
985 : /// abstraction above this needs to store suitable metadata to track what
986 : /// data exists with what keys, in separate metadata entries. If a
987 : /// non-existent key is requested, we may incorrectly return a value from
988 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
989 : /// non-existing key.
990 : ///
991 : /// # Cancel-Safety
992 : ///
993 : /// This method is cancellation-safe.
994 : #[inline(always)]
995 1215202 : pub(crate) async fn get(
996 1215202 : &self,
997 1215202 : key: Key,
998 1215202 : lsn: Lsn,
999 1215202 : ctx: &RequestContext,
1000 1215202 : ) -> Result<Bytes, PageReconstructError> {
1001 1215202 : if !lsn.is_valid() {
1002 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1003 1215202 : }
1004 1215202 :
1005 1215202 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1006 1215202 : // already checked the key against the shard_identity when looking up the Timeline from
1007 1215202 : // page_service.
1008 1215202 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1009 :
1010 1215202 : let keyspace = KeySpace {
1011 1215202 : ranges: vec![key..key.next()],
1012 1215202 : };
1013 1215202 :
1014 1215202 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1015 :
1016 1215202 : let vectored_res = self
1017 1215202 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
1018 1215202 : .await;
1019 :
1020 1215202 : let key_value = vectored_res?.pop_first();
1021 1215190 : match key_value {
1022 1215166 : Some((got_key, value)) => {
1023 1215166 : if got_key != key {
1024 0 : error!(
1025 0 : "Expected {}, but singular vectored get returned {}",
1026 : key, got_key
1027 : );
1028 0 : Err(PageReconstructError::Other(anyhow!(
1029 0 : "Singular vectored get returned wrong key"
1030 0 : )))
1031 : } else {
1032 1215166 : value
1033 : }
1034 : }
1035 24 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1036 24 : key,
1037 24 : shard: self.shard_identity.get_shard_number(&key),
1038 24 : cont_lsn: Lsn(0),
1039 24 : request_lsn: lsn,
1040 24 : ancestor_lsn: None,
1041 24 : backtrace: None,
1042 24 : })),
1043 : }
1044 1215202 : }
1045 :
1046 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1047 : pub(crate) const VEC_GET_LAYERS_VISITED_WARN_THRESH: f64 = 512.0;
1048 :
1049 : /// Look up multiple page versions at a given LSN
1050 : ///
1051 : /// This naive implementation will be replaced with a more efficient one
1052 : /// which actually vectorizes the read path.
1053 39356 : pub(crate) async fn get_vectored(
1054 39356 : &self,
1055 39356 : keyspace: KeySpace,
1056 39356 : lsn: Lsn,
1057 39356 : io_concurrency: super::storage_layer::IoConcurrency,
1058 39356 : ctx: &RequestContext,
1059 39356 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1060 39356 : if !lsn.is_valid() {
1061 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1062 39356 : }
1063 39356 :
1064 39356 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1065 39356 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1066 0 : return Err(GetVectoredError::Oversized(key_count));
1067 39356 : }
1068 :
1069 78712 : for range in &keyspace.ranges {
1070 39356 : let mut key = range.start;
1071 79180 : while key != range.end {
1072 39824 : assert!(!self.shard_identity.is_key_disposable(&key));
1073 39824 : key = key.next();
1074 : }
1075 : }
1076 :
1077 39356 : trace!(
1078 0 : "get vectored request for {:?}@{} from task kind {:?}",
1079 0 : keyspace,
1080 0 : lsn,
1081 0 : ctx.task_kind(),
1082 : );
1083 :
1084 39356 : let start = crate::metrics::GET_VECTORED_LATENCY
1085 39356 : .for_task_kind(ctx.task_kind())
1086 39356 : .map(|metric| (metric, Instant::now()));
1087 :
1088 39356 : let res = self
1089 39356 : .get_vectored_impl(
1090 39356 : keyspace.clone(),
1091 39356 : lsn,
1092 39356 : &mut ValuesReconstructState::new(io_concurrency),
1093 39356 : ctx,
1094 39356 : )
1095 39356 : .await;
1096 :
1097 39356 : if let Some((metric, start)) = start {
1098 0 : let elapsed = start.elapsed();
1099 0 : metric.observe(elapsed.as_secs_f64());
1100 39356 : }
1101 :
1102 39356 : res
1103 39356 : }
1104 :
1105 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1106 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1107 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1108 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1109 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1110 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1111 : /// the scan operation will not cause OOM in the future.
1112 24 : pub(crate) async fn scan(
1113 24 : &self,
1114 24 : keyspace: KeySpace,
1115 24 : lsn: Lsn,
1116 24 : ctx: &RequestContext,
1117 24 : io_concurrency: super::storage_layer::IoConcurrency,
1118 24 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1119 24 : if !lsn.is_valid() {
1120 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1121 24 : }
1122 24 :
1123 24 : trace!(
1124 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1125 0 : keyspace,
1126 0 : lsn,
1127 0 : ctx.task_kind()
1128 : );
1129 :
1130 : // We should generalize this into Keyspace::contains in the future.
1131 48 : for range in &keyspace.ranges {
1132 24 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1133 24 : || range.end.field1 > METADATA_KEY_END_PREFIX
1134 : {
1135 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1136 0 : "only metadata keyspace can be scanned"
1137 0 : )));
1138 24 : }
1139 : }
1140 :
1141 24 : let start = crate::metrics::SCAN_LATENCY
1142 24 : .for_task_kind(ctx.task_kind())
1143 24 : .map(ScanLatencyOngoingRecording::start_recording);
1144 :
1145 24 : let vectored_res = self
1146 24 : .get_vectored_impl(
1147 24 : keyspace.clone(),
1148 24 : lsn,
1149 24 : &mut ValuesReconstructState::new(io_concurrency),
1150 24 : ctx,
1151 24 : )
1152 24 : .await;
1153 :
1154 24 : if let Some(recording) = start {
1155 0 : recording.observe();
1156 24 : }
1157 :
1158 24 : vectored_res
1159 24 : }
1160 :
1161 1255290 : pub(super) async fn get_vectored_impl(
1162 1255290 : &self,
1163 1255290 : keyspace: KeySpace,
1164 1255290 : lsn: Lsn,
1165 1255290 : reconstruct_state: &mut ValuesReconstructState,
1166 1255290 : ctx: &RequestContext,
1167 1255290 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1168 1255290 : let traversal_res: Result<(), _> = self
1169 1255290 : .get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1170 1255290 : .await;
1171 1255290 : if let Err(err) = traversal_res {
1172 : // Wait for all the spawned IOs to complete.
1173 : // See comments on `spawn_io` inside `storage_layer` for more details.
1174 32 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1175 32 : .into_values()
1176 32 : .map(|state| state.collect_pending_ios())
1177 32 : .collect::<FuturesUnordered<_>>();
1178 32 : while collect_futs.next().await.is_some() {}
1179 32 : return Err(err);
1180 1255258 : };
1181 1255258 :
1182 1255258 : let layers_visited = reconstruct_state.get_layers_visited();
1183 1255258 :
1184 1255258 : let futs = FuturesUnordered::new();
1185 1335926 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1186 1335926 : futs.push({
1187 1335926 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1188 1335926 : async move {
1189 1335926 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1190 :
1191 1335926 : let converted = match state.collect_pending_ios().await {
1192 1335926 : Ok(ok) => ok,
1193 0 : Err(err) => {
1194 0 : return (key, Err(err));
1195 : }
1196 : };
1197 :
1198 : // The walredo module expects the records to be descending in terms of Lsn.
1199 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1200 1335926 : debug_assert!(
1201 1335926 : converted
1202 1335926 : .records
1203 1335926 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1204 0 : "{converted:?}"
1205 : );
1206 :
1207 : (
1208 1335926 : key,
1209 1335926 : walredo_self.reconstruct_value(key, lsn, converted).await,
1210 : )
1211 1335926 : }
1212 1335926 : });
1213 1335926 : }
1214 :
1215 1255258 : let results = futs
1216 1255258 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1217 1255258 : .await;
1218 :
1219 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1220 : // when they're missing. Instead they are omitted from the resulting btree
1221 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1222 : // to avoid infinite results.
1223 1255258 : if !results.is_empty() {
1224 1254778 : let avg = layers_visited as f64 / results.len() as f64;
1225 1254778 : if avg >= Self::VEC_GET_LAYERS_VISITED_WARN_THRESH {
1226 0 : use utils::rate_limit::RateLimit;
1227 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1228 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1229 0 : let mut rate_limit = LOGGED.lock().unwrap();
1230 0 : rate_limit.call(|| {
1231 0 : tracing::info!(
1232 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1233 0 : lsn = %lsn,
1234 0 : "Vectored read for {} visited {} layers on average per key and {} in total. {}/{} pages were returned",
1235 0 : keyspace, avg, layers_visited, results.len(), keyspace.total_raw_size());
1236 0 : });
1237 1254778 : }
1238 :
1239 : // Note that this is an approximation. Tracking the exact number of layers visited
1240 : // per key requires virtually unbounded memory usage and is inefficient
1241 : // (i.e. segment tree tracking each range queried from a layer)
1242 1254778 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED.observe(avg);
1243 480 : }
1244 :
1245 1255258 : Ok(results)
1246 1255290 : }
1247 :
1248 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1249 548848 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1250 548848 : self.last_record_lsn.load().last
1251 548848 : }
1252 :
1253 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1254 0 : self.last_record_lsn.load().prev
1255 0 : }
1256 :
1257 : /// Atomically get both last and prev.
1258 456 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1259 456 : self.last_record_lsn.load()
1260 456 : }
1261 :
1262 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1263 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1264 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1265 0 : self.last_record_lsn.status_receiver()
1266 0 : }
1267 :
1268 884 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1269 884 : self.disk_consistent_lsn.load()
1270 884 : }
1271 :
1272 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1273 : /// not validated with control plane yet.
1274 : /// See [`Self::get_remote_consistent_lsn_visible`].
1275 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1276 0 : self.remote_client.remote_consistent_lsn_projected()
1277 0 : }
1278 :
1279 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1280 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1281 : /// generation validation in the deletion queue.
1282 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1283 0 : self.remote_client.remote_consistent_lsn_visible()
1284 0 : }
1285 :
1286 : /// The sum of the file size of all historic layers in the layer map.
1287 : /// This method makes no distinction between local and remote layers.
1288 : /// Hence, the result **does not represent local filesystem usage**.
1289 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1290 0 : let guard = self.layers.read().await;
1291 0 : guard.layer_size_sum()
1292 0 : }
1293 :
1294 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1295 0 : self.metrics.resident_physical_size_get()
1296 0 : }
1297 :
1298 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1299 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1300 0 : }
1301 :
1302 : ///
1303 : /// Wait until WAL has been received and processed up to this LSN.
1304 : ///
1305 : /// You should call this before any of the other get_* or list_* functions. Calling
1306 : /// those functions with an LSN that has been processed yet is an error.
1307 : ///
1308 452230 : pub(crate) async fn wait_lsn(
1309 452230 : &self,
1310 452230 : lsn: Lsn,
1311 452230 : who_is_waiting: WaitLsnWaiter<'_>,
1312 452230 : timeout: WaitLsnTimeout,
1313 452230 : ctx: &RequestContext, /* Prepare for use by cancellation */
1314 452230 : ) -> Result<(), WaitLsnError> {
1315 452230 : let state = self.current_state();
1316 452230 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1317 0 : return Err(WaitLsnError::Shutdown);
1318 452230 : } else if !matches!(state, TimelineState::Active) {
1319 0 : return Err(WaitLsnError::BadState(state));
1320 452230 : }
1321 452230 :
1322 452230 : if cfg!(debug_assertions) {
1323 452230 : match ctx.task_kind() {
1324 : TaskKind::WalReceiverManager
1325 : | TaskKind::WalReceiverConnectionHandler
1326 : | TaskKind::WalReceiverConnectionPoller => {
1327 0 : let is_myself = match who_is_waiting {
1328 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1329 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService | WaitLsnWaiter::HttpEndpoint => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1330 : };
1331 0 : if is_myself {
1332 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1333 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1334 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1335 0 : }
1336 0 : } else {
1337 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1338 0 : // our walreceiver task can make progress independent of theirs
1339 0 : }
1340 : }
1341 452230 : _ => {}
1342 : }
1343 0 : }
1344 :
1345 452230 : let timeout = match timeout {
1346 0 : WaitLsnTimeout::Custom(t) => t,
1347 452230 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1348 : };
1349 :
1350 452230 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1351 452230 :
1352 452230 : match self.last_record_lsn.wait_for_timeout(lsn, timeout).await {
1353 452230 : Ok(()) => Ok(()),
1354 0 : Err(e) => {
1355 : use utils::seqwait::SeqWaitError::*;
1356 0 : match e {
1357 0 : Shutdown => Err(WaitLsnError::Shutdown),
1358 : Timeout => {
1359 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1360 0 : drop(_timer);
1361 0 : let walreceiver_status = self.walreceiver_status();
1362 0 : Err(WaitLsnError::Timeout(format!(
1363 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1364 0 : lsn,
1365 0 : self.get_last_record_lsn(),
1366 0 : self.get_disk_consistent_lsn(),
1367 0 : walreceiver_status,
1368 0 : )))
1369 : }
1370 : }
1371 : }
1372 : }
1373 452230 : }
1374 :
1375 0 : pub(crate) fn walreceiver_status(&self) -> String {
1376 0 : match &*self.walreceiver.lock().unwrap() {
1377 0 : None => "stopping or stopped".to_string(),
1378 0 : Some(walreceiver) => match walreceiver.status() {
1379 0 : Some(status) => status.to_human_readable_string(),
1380 0 : None => "Not active".to_string(),
1381 : },
1382 : }
1383 0 : }
1384 :
1385 : /// Check that it is valid to request operations with that lsn.
1386 464 : pub(crate) fn check_lsn_is_in_scope(
1387 464 : &self,
1388 464 : lsn: Lsn,
1389 464 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1390 464 : ) -> anyhow::Result<()> {
1391 464 : ensure!(
1392 464 : lsn >= **latest_gc_cutoff_lsn,
1393 8 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1394 8 : lsn,
1395 8 : **latest_gc_cutoff_lsn,
1396 : );
1397 456 : Ok(())
1398 464 : }
1399 :
1400 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1401 20 : pub(crate) fn init_lsn_lease(
1402 20 : &self,
1403 20 : lsn: Lsn,
1404 20 : length: Duration,
1405 20 : ctx: &RequestContext,
1406 20 : ) -> anyhow::Result<LsnLease> {
1407 20 : self.make_lsn_lease(lsn, length, true, ctx)
1408 20 : }
1409 :
1410 : /// 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.
1411 8 : pub(crate) fn renew_lsn_lease(
1412 8 : &self,
1413 8 : lsn: Lsn,
1414 8 : length: Duration,
1415 8 : ctx: &RequestContext,
1416 8 : ) -> anyhow::Result<LsnLease> {
1417 8 : self.make_lsn_lease(lsn, length, false, ctx)
1418 8 : }
1419 :
1420 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1421 : ///
1422 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1423 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1424 : ///
1425 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1426 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1427 28 : fn make_lsn_lease(
1428 28 : &self,
1429 28 : lsn: Lsn,
1430 28 : length: Duration,
1431 28 : init: bool,
1432 28 : _ctx: &RequestContext,
1433 28 : ) -> anyhow::Result<LsnLease> {
1434 24 : let lease = {
1435 : // Normalize the requested LSN to be aligned, and move to the first record
1436 : // if it points to the beginning of the page (header).
1437 28 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1438 28 :
1439 28 : let mut gc_info = self.gc_info.write().unwrap();
1440 28 :
1441 28 : let valid_until = SystemTime::now() + length;
1442 28 :
1443 28 : let entry = gc_info.leases.entry(lsn);
1444 28 :
1445 28 : match entry {
1446 12 : Entry::Occupied(mut occupied) => {
1447 12 : let existing_lease = occupied.get_mut();
1448 12 : if valid_until > existing_lease.valid_until {
1449 4 : existing_lease.valid_until = valid_until;
1450 4 : let dt: DateTime<Utc> = valid_until.into();
1451 4 : info!("lease extended to {}", dt);
1452 : } else {
1453 8 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1454 8 : info!("existing lease covers greater length, valid until {}", dt);
1455 : }
1456 :
1457 12 : existing_lease.clone()
1458 : }
1459 16 : Entry::Vacant(vacant) => {
1460 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
1461 : // not blocked by the lsn lease deadline.
1462 16 : let validate = {
1463 16 : let conf = self.tenant_conf.load();
1464 16 : conf.location.attach_mode == AttachmentMode::Single
1465 16 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1466 : };
1467 :
1468 16 : if init || validate {
1469 16 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1470 16 : if lsn < *latest_gc_cutoff_lsn {
1471 4 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1472 12 : }
1473 0 : }
1474 :
1475 12 : let dt: DateTime<Utc> = valid_until.into();
1476 12 : info!("lease created, valid until {}", dt);
1477 12 : vacant.insert(LsnLease { valid_until }).clone()
1478 : }
1479 : }
1480 : };
1481 :
1482 24 : Ok(lease)
1483 28 : }
1484 :
1485 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1486 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1487 : #[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))]
1488 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1489 : self.freeze0().await
1490 : }
1491 :
1492 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1493 : #[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))]
1494 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1495 : self.freeze_and_flush0().await
1496 : }
1497 :
1498 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1499 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1500 2240 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1501 2240 : let mut g = self.write_lock.lock().await;
1502 2240 : let to_lsn = self.get_last_record_lsn();
1503 2240 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1504 2240 : }
1505 :
1506 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1507 : // polluting the span hierarchy.
1508 2240 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1509 2240 : let token = self.freeze0().await?;
1510 2240 : self.wait_flush_completion(token).await
1511 2240 : }
1512 :
1513 : // Check if an open ephemeral layer should be closed: this provides
1514 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1515 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1516 : // ephemeral layer bytes has been breached.
1517 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1518 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1519 : // If the write lock is held, there is an active wal receiver: rolling open layers
1520 : // is their responsibility while they hold this lock.
1521 0 : return;
1522 : };
1523 :
1524 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1525 : // time, and this was missed.
1526 : // if write_guard.is_none() { return; }
1527 :
1528 0 : let Ok(layers_guard) = self.layers.try_read() else {
1529 : // Don't block if the layer lock is busy
1530 0 : return;
1531 : };
1532 :
1533 0 : let Ok(lm) = layers_guard.layer_map() else {
1534 0 : return;
1535 : };
1536 :
1537 0 : let Some(open_layer) = &lm.open_layer else {
1538 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1539 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1540 : // that didn't result in writes to this shard.
1541 :
1542 : // Must not hold the layers lock while waiting for a flush.
1543 0 : drop(layers_guard);
1544 0 :
1545 0 : let last_record_lsn = self.get_last_record_lsn();
1546 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1547 0 : if last_record_lsn > disk_consistent_lsn {
1548 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1549 : // we are a sharded tenant and have skipped some WAL
1550 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1551 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1552 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1553 : // without any data ingested (yet) doesn't write a remote index as soon as it
1554 : // sees its LSN advance: we only do this if we've been layer-less
1555 : // for some time.
1556 0 : tracing::debug!(
1557 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1558 : disk_consistent_lsn,
1559 : last_record_lsn
1560 : );
1561 :
1562 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1563 : // We know there is no open layer, so we can request freezing without actually
1564 : // freezing anything. This is true even if we have dropped the layers_guard, we
1565 : // still hold the write_guard.
1566 0 : let _ = async {
1567 0 : let token = self
1568 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1569 0 : .await?;
1570 0 : self.wait_flush_completion(token).await
1571 0 : }
1572 0 : .await;
1573 0 : }
1574 0 : }
1575 :
1576 0 : return;
1577 : };
1578 :
1579 0 : let Some(current_size) = open_layer.try_len() else {
1580 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1581 : // read lock to get size should always succeed.
1582 0 : tracing::warn!("Lock conflict while reading size of open layer");
1583 0 : return;
1584 : };
1585 :
1586 0 : let current_lsn = self.get_last_record_lsn();
1587 :
1588 0 : let checkpoint_distance_override = open_layer.tick().await;
1589 :
1590 0 : if let Some(size_override) = checkpoint_distance_override {
1591 0 : if current_size > size_override {
1592 : // This is not harmful, but it only happens in relatively rare cases where
1593 : // time-based checkpoints are not happening fast enough to keep the amount of
1594 : // ephemeral data within configured limits. It's a sign of stress on the system.
1595 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1596 0 : }
1597 0 : }
1598 :
1599 0 : let checkpoint_distance =
1600 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1601 0 :
1602 0 : if self.should_roll(
1603 0 : current_size,
1604 0 : current_size,
1605 0 : checkpoint_distance,
1606 0 : self.get_last_record_lsn(),
1607 0 : self.last_freeze_at.load(),
1608 0 : open_layer.get_opened_at(),
1609 0 : ) {
1610 0 : match open_layer.info() {
1611 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1612 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1613 0 : // happens asynchronously in the background.
1614 0 : tracing::debug!(
1615 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1616 : );
1617 : }
1618 : InMemoryLayerInfo::Open { .. } => {
1619 : // Upgrade to a write lock and freeze the layer
1620 0 : drop(layers_guard);
1621 0 : let res = self
1622 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1623 0 : .await;
1624 :
1625 0 : if let Err(e) = res {
1626 0 : tracing::info!(
1627 0 : "failed to flush frozen layer after background freeze: {e:#}"
1628 : );
1629 0 : }
1630 : }
1631 : }
1632 0 : }
1633 0 : }
1634 :
1635 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1636 : ///
1637 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1638 : /// child timelines that are not offloaded yet.
1639 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1640 0 : if self.remote_client.is_archived() != Some(true) {
1641 0 : return (false, "the timeline is not archived");
1642 0 : }
1643 0 : if !self.remote_client.no_pending_work() {
1644 : // if the remote client is still processing some work, we can't offload
1645 0 : return (false, "the upload queue is not drained yet");
1646 0 : }
1647 0 :
1648 0 : (true, "ok")
1649 0 : }
1650 :
1651 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1652 : /// compaction tasks.
1653 728 : pub(crate) async fn compact(
1654 728 : self: &Arc<Self>,
1655 728 : cancel: &CancellationToken,
1656 728 : flags: EnumSet<CompactFlags>,
1657 728 : ctx: &RequestContext,
1658 728 : ) -> Result<bool, CompactionError> {
1659 728 : self.compact_with_options(
1660 728 : cancel,
1661 728 : CompactOptions {
1662 728 : flags,
1663 728 : compact_key_range: None,
1664 728 : compact_lsn_range: None,
1665 728 : sub_compaction: false,
1666 728 : sub_compaction_max_job_size_mb: None,
1667 728 : },
1668 728 : ctx,
1669 728 : )
1670 728 : .await
1671 728 : }
1672 :
1673 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1674 : /// compaction tasks.
1675 728 : pub(crate) async fn compact_with_options(
1676 728 : self: &Arc<Self>,
1677 728 : cancel: &CancellationToken,
1678 728 : options: CompactOptions,
1679 728 : ctx: &RequestContext,
1680 728 : ) -> Result<bool, CompactionError> {
1681 728 : // most likely the cancellation token is from background task, but in tests it could be the
1682 728 : // request task as well.
1683 728 :
1684 728 : let prepare = async move {
1685 728 : let guard = self.compaction_lock.lock().await;
1686 :
1687 728 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1688 728 : BackgroundLoopKind::Compaction,
1689 728 : ctx,
1690 728 : )
1691 728 : .await;
1692 :
1693 728 : (guard, permit)
1694 728 : };
1695 :
1696 : // this wait probably never needs any "long time spent" logging, because we already nag if
1697 : // compaction task goes over it's period (20s) which is quite often in production.
1698 728 : let (_guard, _permit) = tokio::select! {
1699 728 : tuple = prepare => { tuple },
1700 728 : _ = self.cancel.cancelled() => return Ok(false),
1701 728 : _ = cancel.cancelled() => return Ok(false),
1702 : };
1703 :
1704 728 : let last_record_lsn = self.get_last_record_lsn();
1705 728 :
1706 728 : // Last record Lsn could be zero in case the timeline was just created
1707 728 : if !last_record_lsn.is_valid() {
1708 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1709 0 : return Ok(false);
1710 728 : }
1711 :
1712 728 : let result = match self.get_compaction_algorithm_settings().kind {
1713 : CompactionAlgorithm::Tiered => {
1714 0 : self.compact_tiered(cancel, ctx).await?;
1715 0 : Ok(false)
1716 : }
1717 728 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1718 : };
1719 :
1720 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
1721 0 : match result {
1722 728 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
1723 : Err(CompactionError::Other(_)) => {
1724 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
1725 : }
1726 : // Don't change the current value on offload failure or shutdown. We don't want to
1727 : // abruptly stall nor resume L0 flushes in these cases.
1728 0 : Err(CompactionError::Offload(_)) => {}
1729 0 : Err(CompactionError::ShuttingDown) => {}
1730 : };
1731 :
1732 728 : result
1733 728 : }
1734 :
1735 : /// Mutate the timeline with a [`TimelineWriter`].
1736 10266380 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1737 10266380 : TimelineWriter {
1738 10266380 : tl: self,
1739 10266380 : write_guard: self.write_lock.lock().await,
1740 : }
1741 10266380 : }
1742 :
1743 0 : pub(crate) fn activate(
1744 0 : self: &Arc<Self>,
1745 0 : parent: Arc<crate::tenant::Tenant>,
1746 0 : broker_client: BrokerClientChannel,
1747 0 : background_jobs_can_start: Option<&completion::Barrier>,
1748 0 : ctx: &RequestContext,
1749 0 : ) {
1750 0 : if self.tenant_shard_id.is_shard_zero() {
1751 0 : // Logical size is only maintained accurately on shard zero.
1752 0 : self.spawn_initial_logical_size_computation_task(ctx);
1753 0 : }
1754 0 : self.launch_wal_receiver(ctx, broker_client);
1755 0 : self.set_state(TimelineState::Active);
1756 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1757 0 : }
1758 :
1759 : /// After this function returns, there are no timeline-scoped tasks are left running.
1760 : ///
1761 : /// The preferred pattern for is:
1762 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1763 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1764 : /// go the extra mile and keep track of JoinHandles
1765 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1766 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1767 : ///
1768 : /// For legacy reasons, we still have multiple tasks spawned using
1769 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1770 : /// We refer to these as "timeline-scoped task_mgr tasks".
1771 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1772 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1773 : /// or [`task_mgr::shutdown_watcher`].
1774 : /// We want to gradually convert the code base away from these.
1775 : ///
1776 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1777 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1778 : /// ones that aren't mentioned here):
1779 : /// - [`TaskKind::TimelineDeletionWorker`]
1780 : /// - NB: also used for tenant deletion
1781 : /// - [`TaskKind::RemoteUploadTask`]`
1782 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1783 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1784 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1785 : /// - [`TaskKind::Eviction`]
1786 : /// - [`TaskKind::LayerFlushTask`]
1787 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1788 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1789 20 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1790 20 : debug_assert_current_span_has_tenant_and_timeline_id();
1791 20 :
1792 20 : // Regardless of whether we're going to try_freeze_and_flush
1793 20 : // or not, stop ingesting any more data. Walreceiver only provides
1794 20 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1795 20 : // So, only after the self.gate.close() below will we know for sure that
1796 20 : // no walreceiver tasks are left.
1797 20 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1798 20 : // data during the call to `self.freeze_and_flush()` below.
1799 20 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1800 20 : // which is what we'd need to properly model early shutdown of the walreceiver
1801 20 : // task sub-tree before the other Timeline task sub-trees.
1802 20 : let walreceiver = self.walreceiver.lock().unwrap().take();
1803 20 : tracing::debug!(
1804 0 : is_some = walreceiver.is_some(),
1805 0 : "Waiting for WalReceiverManager..."
1806 : );
1807 20 : if let Some(walreceiver) = walreceiver {
1808 0 : walreceiver.cancel();
1809 20 : }
1810 : // ... and inform any waiters for newer LSNs that there won't be any.
1811 20 : self.last_record_lsn.shutdown();
1812 20 :
1813 20 : if let ShutdownMode::FreezeAndFlush = mode {
1814 12 : if let Some((open, frozen)) = self
1815 12 : .layers
1816 12 : .read()
1817 12 : .await
1818 12 : .layer_map()
1819 12 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1820 12 : .ok()
1821 12 : .filter(|(open, frozen)| *open || *frozen > 0)
1822 : {
1823 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1824 12 : } else {
1825 12 : // this is double-shutdown, ignore it
1826 12 : }
1827 :
1828 : // we shut down walreceiver above, so, we won't add anything more
1829 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1830 : // to reach the disk & upload queue, then shut the upload queue and
1831 : // wait for it to drain.
1832 12 : match self.freeze_and_flush().await {
1833 : Ok(_) => {
1834 : // drain the upload queue
1835 : // if we did not wait for completion here, it might be our shutdown process
1836 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1837 : // be spawned.
1838 : //
1839 : // what is problematic is the shutting down of RemoteTimelineClient, because
1840 : // obviously it does not make sense to stop while we wait for it, but what
1841 : // about corner cases like s3 suddenly hanging up?
1842 12 : self.remote_client.shutdown().await;
1843 : }
1844 : Err(FlushLayerError::Cancelled) => {
1845 : // this is likely the second shutdown, ignore silently.
1846 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1847 0 : debug_assert!(self.cancel.is_cancelled());
1848 : }
1849 0 : Err(e) => {
1850 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1851 0 : // we have some extra WAL replay to do next time the timeline starts.
1852 0 : warn!("failed to freeze and flush: {e:#}");
1853 : }
1854 : }
1855 :
1856 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
1857 : // we also do a final check here to ensure that the queue is empty.
1858 12 : if !self.remote_client.no_pending_work() {
1859 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1860 12 : }
1861 8 : }
1862 :
1863 20 : if let ShutdownMode::Reload = mode {
1864 : // drain the upload queue
1865 4 : self.remote_client.shutdown().await;
1866 4 : if !self.remote_client.no_pending_work() {
1867 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1868 4 : }
1869 16 : }
1870 :
1871 : // Signal any subscribers to our cancellation token to drop out
1872 20 : tracing::debug!("Cancelling CancellationToken");
1873 20 : self.cancel.cancel();
1874 20 :
1875 20 : // Ensure Prevent new page service requests from starting.
1876 20 : self.handles.shutdown();
1877 20 :
1878 20 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1879 20 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1880 20 : self.remote_client.stop();
1881 20 :
1882 20 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1883 20 : // to shut down the upload queue tasks.
1884 20 : // TODO: fix that, task management should be encapsulated inside remote_client.
1885 20 : task_mgr::shutdown_tasks(
1886 20 : Some(TaskKind::RemoteUploadTask),
1887 20 : Some(self.tenant_shard_id),
1888 20 : Some(self.timeline_id),
1889 20 : )
1890 20 : .await;
1891 :
1892 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
1893 20 : tracing::debug!("Waiting for tasks...");
1894 20 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1895 :
1896 : {
1897 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
1898 : // open.
1899 20 : let mut write_guard = self.write_lock.lock().await;
1900 20 : self.layers.write().await.shutdown(&mut write_guard);
1901 20 : }
1902 20 :
1903 20 : // Finally wait until any gate-holders are complete.
1904 20 : //
1905 20 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1906 20 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1907 20 : self.gate.close().await;
1908 :
1909 20 : self.metrics.shutdown();
1910 20 : }
1911 :
1912 896 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1913 896 : match (self.current_state(), new_state) {
1914 896 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1915 4 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1916 : }
1917 0 : (st, TimelineState::Loading) => {
1918 0 : error!("ignoring transition from {st:?} into Loading state");
1919 : }
1920 0 : (TimelineState::Broken { .. }, new_state) => {
1921 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1922 : }
1923 : (TimelineState::Stopping, TimelineState::Active) => {
1924 0 : error!("Not activating a Stopping timeline");
1925 : }
1926 892 : (_, new_state) => {
1927 892 : self.state.send_replace(new_state);
1928 892 : }
1929 : }
1930 896 : }
1931 :
1932 4 : pub(crate) fn set_broken(&self, reason: String) {
1933 4 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1934 4 : let broken_state = TimelineState::Broken {
1935 4 : reason,
1936 4 : backtrace: backtrace_str,
1937 4 : };
1938 4 : self.set_state(broken_state);
1939 4 :
1940 4 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1941 4 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1942 4 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1943 4 : self.cancel.cancel();
1944 4 : }
1945 :
1946 454306 : pub(crate) fn current_state(&self) -> TimelineState {
1947 454306 : self.state.borrow().clone()
1948 454306 : }
1949 :
1950 12 : pub(crate) fn is_broken(&self) -> bool {
1951 12 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1952 12 : }
1953 :
1954 444 : pub(crate) fn is_active(&self) -> bool {
1955 444 : self.current_state() == TimelineState::Active
1956 444 : }
1957 :
1958 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
1959 0 : self.remote_client.is_archived()
1960 0 : }
1961 :
1962 736 : pub(crate) fn is_stopping(&self) -> bool {
1963 736 : self.current_state() == TimelineState::Stopping
1964 736 : }
1965 :
1966 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1967 0 : self.state.subscribe()
1968 0 : }
1969 :
1970 452234 : pub(crate) async fn wait_to_become_active(
1971 452234 : &self,
1972 452234 : _ctx: &RequestContext, // Prepare for use by cancellation
1973 452234 : ) -> Result<(), TimelineState> {
1974 452234 : let mut receiver = self.state.subscribe();
1975 : loop {
1976 452234 : let current_state = receiver.borrow().clone();
1977 452234 : match current_state {
1978 : TimelineState::Loading => {
1979 0 : receiver
1980 0 : .changed()
1981 0 : .await
1982 0 : .expect("holding a reference to self");
1983 : }
1984 : TimelineState::Active { .. } => {
1985 452230 : return Ok(());
1986 : }
1987 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1988 : // There's no chance the timeline can transition back into ::Active
1989 4 : return Err(current_state);
1990 : }
1991 : }
1992 : }
1993 452234 : }
1994 :
1995 0 : pub(crate) async fn layer_map_info(
1996 0 : &self,
1997 0 : reset: LayerAccessStatsReset,
1998 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
1999 0 : let guard = self.layers.read().await;
2000 0 : let layer_map = guard.layer_map()?;
2001 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2002 0 : if let Some(open_layer) = &layer_map.open_layer {
2003 0 : in_memory_layers.push(open_layer.info());
2004 0 : }
2005 0 : for frozen_layer in &layer_map.frozen_layers {
2006 0 : in_memory_layers.push(frozen_layer.info());
2007 0 : }
2008 :
2009 0 : let historic_layers = layer_map
2010 0 : .iter_historic_layers()
2011 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2012 0 : .collect();
2013 0 :
2014 0 : Ok(LayerMapInfo {
2015 0 : in_memory_layers,
2016 0 : historic_layers,
2017 0 : })
2018 0 : }
2019 :
2020 : #[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))]
2021 : pub(crate) async fn download_layer(
2022 : &self,
2023 : layer_file_name: &LayerName,
2024 : ) -> anyhow::Result<Option<bool>> {
2025 : let Some(layer) = self.find_layer(layer_file_name).await? else {
2026 : return Ok(None);
2027 : };
2028 :
2029 : layer.download().await?;
2030 :
2031 : Ok(Some(true))
2032 : }
2033 :
2034 : /// Evict just one layer.
2035 : ///
2036 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2037 0 : pub(crate) async fn evict_layer(
2038 0 : &self,
2039 0 : layer_file_name: &LayerName,
2040 0 : ) -> anyhow::Result<Option<bool>> {
2041 0 : let _gate = self
2042 0 : .gate
2043 0 : .enter()
2044 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2045 :
2046 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2047 0 : return Ok(None);
2048 : };
2049 :
2050 : // curl has this by default
2051 0 : let timeout = std::time::Duration::from_secs(120);
2052 0 :
2053 0 : match local_layer.evict_and_wait(timeout).await {
2054 0 : Ok(()) => Ok(Some(true)),
2055 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2056 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2057 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2058 : }
2059 0 : }
2060 :
2061 9606020 : fn should_roll(
2062 9606020 : &self,
2063 9606020 : layer_size: u64,
2064 9606020 : projected_layer_size: u64,
2065 9606020 : checkpoint_distance: u64,
2066 9606020 : projected_lsn: Lsn,
2067 9606020 : last_freeze_at: Lsn,
2068 9606020 : opened_at: Instant,
2069 9606020 : ) -> bool {
2070 9606020 : let distance = projected_lsn.widening_sub(last_freeze_at);
2071 9606020 :
2072 9606020 : // Rolling the open layer can be triggered by:
2073 9606020 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2074 9606020 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2075 9606020 : // account for how writes are distributed across shards: we expect each node to consume
2076 9606020 : // 1/count of the LSN on average.
2077 9606020 : // 2. The size of the currently open layer.
2078 9606020 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2079 9606020 : // up and suspend activity.
2080 9606020 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2081 0 : info!(
2082 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2083 : projected_lsn, layer_size, distance
2084 : );
2085 :
2086 0 : true
2087 9606020 : } else if projected_layer_size >= checkpoint_distance {
2088 : // NB: this check is relied upon by:
2089 160 : let _ = IndexEntry::validate_checkpoint_distance;
2090 160 : info!(
2091 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2092 : projected_lsn, layer_size, projected_layer_size
2093 : );
2094 :
2095 160 : true
2096 9605860 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2097 0 : info!(
2098 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2099 0 : projected_lsn,
2100 0 : layer_size,
2101 0 : opened_at.elapsed()
2102 : );
2103 :
2104 0 : true
2105 : } else {
2106 9605860 : false
2107 : }
2108 9606020 : }
2109 : }
2110 :
2111 : /// Number of times we will compute partition within a checkpoint distance.
2112 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2113 :
2114 : // Private functions
2115 : impl Timeline {
2116 24 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2117 24 : let tenant_conf = self.tenant_conf.load();
2118 24 : tenant_conf
2119 24 : .tenant_conf
2120 24 : .lsn_lease_length
2121 24 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2122 24 : }
2123 :
2124 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2125 0 : let tenant_conf = self.tenant_conf.load();
2126 0 : tenant_conf
2127 0 : .tenant_conf
2128 0 : .lsn_lease_length_for_ts
2129 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2130 0 : }
2131 :
2132 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2133 0 : let tenant_conf = self.tenant_conf.load();
2134 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2135 0 : }
2136 :
2137 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2138 0 : let tenant_conf = self.tenant_conf.load();
2139 0 : tenant_conf
2140 0 : .tenant_conf
2141 0 : .lazy_slru_download
2142 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2143 0 : }
2144 :
2145 9608992 : fn get_checkpoint_distance(&self) -> u64 {
2146 9608992 : let tenant_conf = self.tenant_conf.load();
2147 9608992 : tenant_conf
2148 9608992 : .tenant_conf
2149 9608992 : .checkpoint_distance
2150 9608992 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2151 9608992 : }
2152 :
2153 9605860 : fn get_checkpoint_timeout(&self) -> Duration {
2154 9605860 : let tenant_conf = self.tenant_conf.load();
2155 9605860 : tenant_conf
2156 9605860 : .tenant_conf
2157 9605860 : .checkpoint_timeout
2158 9605860 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2159 9605860 : }
2160 :
2161 5008 : fn get_compaction_period(&self) -> Duration {
2162 5008 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2163 5008 : tenant_conf
2164 5008 : .compaction_period
2165 5008 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2166 5008 : }
2167 :
2168 1324 : fn get_compaction_target_size(&self) -> u64 {
2169 1324 : let tenant_conf = self.tenant_conf.load();
2170 1324 : tenant_conf
2171 1324 : .tenant_conf
2172 1324 : .compaction_target_size
2173 1324 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2174 1324 : }
2175 :
2176 728 : fn get_compaction_threshold(&self) -> usize {
2177 728 : let tenant_conf = self.tenant_conf.load();
2178 728 : tenant_conf
2179 728 : .tenant_conf
2180 728 : .compaction_threshold
2181 728 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2182 728 : }
2183 :
2184 56 : fn get_compaction_upper_limit(&self) -> usize {
2185 56 : let tenant_conf = self.tenant_conf.load();
2186 56 : tenant_conf
2187 56 : .tenant_conf
2188 56 : .compaction_upper_limit
2189 56 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2190 56 : }
2191 :
2192 2504 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2193 : // Disable L0 flushes by default. This and compaction needs further tuning.
2194 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 0; // TODO: default to e.g. 3
2195 :
2196 : // If compaction is disabled, don't delay.
2197 2504 : if self.get_compaction_period() == Duration::ZERO {
2198 2504 : return None;
2199 0 : }
2200 0 :
2201 0 : let compaction_threshold = self.get_compaction_threshold();
2202 0 : let tenant_conf = self.tenant_conf.load();
2203 0 : let l0_flush_delay_threshold = tenant_conf
2204 0 : .tenant_conf
2205 0 : .l0_flush_delay_threshold
2206 0 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2207 0 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2208 0 :
2209 0 : // 0 disables backpressure.
2210 0 : if l0_flush_delay_threshold == 0 {
2211 0 : return None;
2212 0 : }
2213 0 :
2214 0 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2215 0 : // backpressure flushes below this.
2216 0 : // TODO: the tenant config should have validation to prevent this instead.
2217 0 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2218 0 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2219 2504 : }
2220 :
2221 2504 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2222 : // Disable L0 stalls by default. In ingest benchmarks, we see image compaction take >10
2223 : // minutes, blocking L0 compaction, and we can't stall L0 flushes for that long.
2224 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2225 :
2226 : // If compaction is disabled, don't stall.
2227 2504 : if self.get_compaction_period() == Duration::ZERO {
2228 2504 : return None;
2229 0 : }
2230 0 :
2231 0 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2232 0 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2233 0 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2234 0 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2235 0 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2236 0 : return None;
2237 0 : }
2238 0 :
2239 0 : let compaction_threshold = self.get_compaction_threshold();
2240 0 : let tenant_conf = self.tenant_conf.load();
2241 0 : let l0_flush_stall_threshold = tenant_conf
2242 0 : .tenant_conf
2243 0 : .l0_flush_stall_threshold
2244 0 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2245 0 :
2246 0 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2247 0 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2248 0 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2249 0 : if cfg!(feature = "testing")
2250 0 : && compaction_threshold == 1
2251 0 : && l0_flush_stall_threshold.is_none()
2252 : {
2253 0 : return None;
2254 0 : }
2255 0 :
2256 0 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2257 0 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2258 0 :
2259 0 : // 0 disables backpressure.
2260 0 : if l0_flush_stall_threshold == 0 {
2261 0 : return None;
2262 0 : }
2263 0 :
2264 0 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2265 0 : // backpressure flushes below this.
2266 0 : // TODO: the tenant config should have validation to prevent this instead.
2267 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2268 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2269 2504 : }
2270 :
2271 2344 : fn get_l0_flush_wait_upload(&self) -> bool {
2272 2344 : let tenant_conf = self.tenant_conf.load();
2273 2344 : tenant_conf
2274 2344 : .tenant_conf
2275 2344 : .l0_flush_wait_upload
2276 2344 : .unwrap_or(self.conf.default_tenant_conf.l0_flush_wait_upload)
2277 2344 : }
2278 :
2279 28 : fn get_image_creation_threshold(&self) -> usize {
2280 28 : let tenant_conf = self.tenant_conf.load();
2281 28 : tenant_conf
2282 28 : .tenant_conf
2283 28 : .image_creation_threshold
2284 28 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2285 28 : }
2286 :
2287 728 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2288 728 : let tenant_conf = &self.tenant_conf.load();
2289 728 : tenant_conf
2290 728 : .tenant_conf
2291 728 : .compaction_algorithm
2292 728 : .as_ref()
2293 728 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2294 728 : .clone()
2295 728 : }
2296 :
2297 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2298 0 : let tenant_conf = self.tenant_conf.load();
2299 0 : tenant_conf
2300 0 : .tenant_conf
2301 0 : .eviction_policy
2302 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2303 0 : }
2304 :
2305 892 : fn get_evictions_low_residence_duration_metric_threshold(
2306 892 : tenant_conf: &TenantConfOpt,
2307 892 : default_tenant_conf: &TenantConf,
2308 892 : ) -> Duration {
2309 892 : tenant_conf
2310 892 : .evictions_low_residence_duration_metric_threshold
2311 892 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2312 892 : }
2313 :
2314 1136 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2315 1136 : let tenant_conf = self.tenant_conf.load();
2316 1136 : tenant_conf
2317 1136 : .tenant_conf
2318 1136 : .image_layer_creation_check_threshold
2319 1136 : .unwrap_or(
2320 1136 : self.conf
2321 1136 : .default_tenant_conf
2322 1136 : .image_layer_creation_check_threshold,
2323 1136 : )
2324 1136 : }
2325 :
2326 : /// Resolve the effective WAL receiver protocol to use for this tenant.
2327 : ///
2328 : /// Priority order is:
2329 : /// 1. Tenant config override
2330 : /// 2. Default value for tenant config override
2331 : /// 3. Pageserver config override
2332 : /// 4. Pageserver config default
2333 0 : pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
2334 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2335 0 : tenant_conf
2336 0 : .wal_receiver_protocol_override
2337 0 : .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
2338 0 : .unwrap_or(self.conf.wal_receiver_protocol)
2339 0 : }
2340 :
2341 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2342 0 : // NB: Most tenant conf options are read by background loops, so,
2343 0 : // changes will automatically be picked up.
2344 0 :
2345 0 : // The threshold is embedded in the metric. So, we need to update it.
2346 0 : {
2347 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2348 0 : &new_conf.tenant_conf,
2349 0 : &self.conf.default_tenant_conf,
2350 0 : );
2351 0 :
2352 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2353 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2354 0 :
2355 0 : let timeline_id_str = self.timeline_id.to_string();
2356 0 :
2357 0 : self.remote_client.update_config(&new_conf.location);
2358 0 :
2359 0 : self.metrics
2360 0 : .evictions_with_low_residence_duration
2361 0 : .write()
2362 0 : .unwrap()
2363 0 : .change_threshold(
2364 0 : &tenant_id_str,
2365 0 : &shard_id_str,
2366 0 : &timeline_id_str,
2367 0 : new_threshold,
2368 0 : );
2369 0 : }
2370 0 : }
2371 :
2372 : /// Open a Timeline handle.
2373 : ///
2374 : /// Loads the metadata for the timeline into memory, but not the layer map.
2375 : #[allow(clippy::too_many_arguments)]
2376 892 : pub(super) fn new(
2377 892 : conf: &'static PageServerConf,
2378 892 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2379 892 : metadata: &TimelineMetadata,
2380 892 : ancestor: Option<Arc<Timeline>>,
2381 892 : timeline_id: TimelineId,
2382 892 : tenant_shard_id: TenantShardId,
2383 892 : generation: Generation,
2384 892 : shard_identity: ShardIdentity,
2385 892 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2386 892 : resources: TimelineResources,
2387 892 : pg_version: u32,
2388 892 : state: TimelineState,
2389 892 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2390 892 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2391 892 : cancel: CancellationToken,
2392 892 : ) -> Arc<Self> {
2393 892 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2394 892 : let (state, _) = watch::channel(state);
2395 892 :
2396 892 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2397 892 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2398 892 :
2399 892 : let evictions_low_residence_duration_metric_threshold = {
2400 892 : let loaded_tenant_conf = tenant_conf.load();
2401 892 : Self::get_evictions_low_residence_duration_metric_threshold(
2402 892 : &loaded_tenant_conf.tenant_conf,
2403 892 : &conf.default_tenant_conf,
2404 892 : )
2405 : };
2406 :
2407 892 : if let Some(ancestor) = &ancestor {
2408 460 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2409 460 : // If we construct an explicit timeline object, it's obviously not offloaded
2410 460 : let is_offloaded = MaybeOffloaded::No;
2411 460 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2412 460 : }
2413 :
2414 892 : Arc::new_cyclic(|myself| {
2415 892 : let metrics = TimelineMetrics::new(
2416 892 : &tenant_shard_id,
2417 892 : &timeline_id,
2418 892 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2419 892 : "mtime",
2420 892 : evictions_low_residence_duration_metric_threshold,
2421 892 : ),
2422 892 : );
2423 892 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2424 :
2425 892 : let mut result = Timeline {
2426 892 : conf,
2427 892 : tenant_conf,
2428 892 : myself: myself.clone(),
2429 892 : timeline_id,
2430 892 : tenant_shard_id,
2431 892 : generation,
2432 892 : shard_identity,
2433 892 : pg_version,
2434 892 : layers: Default::default(),
2435 892 :
2436 892 : walredo_mgr,
2437 892 : walreceiver: Mutex::new(None),
2438 892 :
2439 892 : remote_client: Arc::new(resources.remote_client),
2440 892 :
2441 892 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2442 892 : last_record_lsn: SeqWait::new(RecordLsn {
2443 892 : last: disk_consistent_lsn,
2444 892 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2445 892 : }),
2446 892 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2447 892 :
2448 892 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2449 892 : last_freeze_ts: RwLock::new(Instant::now()),
2450 892 :
2451 892 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2452 892 :
2453 892 : ancestor_timeline: ancestor,
2454 892 : ancestor_lsn: metadata.ancestor_lsn(),
2455 892 :
2456 892 : metrics,
2457 892 :
2458 892 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2459 892 : &tenant_shard_id,
2460 892 : &timeline_id,
2461 892 : resources.pagestream_throttle_metrics,
2462 892 : ),
2463 892 :
2464 6244 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2465 892 :
2466 892 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2467 892 :
2468 892 : layer_flush_start_tx,
2469 892 : layer_flush_done_tx,
2470 892 :
2471 892 : write_lock: tokio::sync::Mutex::new(None),
2472 892 :
2473 892 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2474 892 :
2475 892 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2476 892 : initdb_lsn: metadata.initdb_lsn(),
2477 892 :
2478 892 : current_logical_size: if disk_consistent_lsn.is_valid() {
2479 : // we're creating timeline data with some layer files existing locally,
2480 : // need to recalculate timeline's logical size based on data in the layers.
2481 468 : LogicalSize::deferred_initial(disk_consistent_lsn)
2482 : } else {
2483 : // we're creating timeline data without any layers existing locally,
2484 : // initial logical size is 0.
2485 424 : LogicalSize::empty_initial()
2486 : },
2487 :
2488 892 : partitioning: GuardArcSwap::new((
2489 892 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2490 892 : Lsn(0),
2491 892 : )),
2492 892 : repartition_threshold: 0,
2493 892 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2494 892 : last_image_layer_creation_check_instant: Mutex::new(None),
2495 892 :
2496 892 : last_received_wal: Mutex::new(None),
2497 892 : rel_size_cache: RwLock::new(RelSizeCache {
2498 892 : complete_as_of: disk_consistent_lsn,
2499 892 : map: HashMap::new(),
2500 892 : }),
2501 892 :
2502 892 : download_all_remote_layers_task_info: RwLock::new(None),
2503 892 :
2504 892 : state,
2505 892 :
2506 892 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2507 892 : EvictionTaskTimelineState::default(),
2508 892 : ),
2509 892 : delete_progress: TimelineDeleteProgress::default(),
2510 892 :
2511 892 : cancel,
2512 892 : gate: Gate::default(),
2513 892 :
2514 892 : compaction_lock: tokio::sync::Mutex::default(),
2515 892 : compaction_failed: AtomicBool::default(),
2516 892 : gc_lock: tokio::sync::Mutex::default(),
2517 892 :
2518 892 : standby_horizon: AtomicLsn::new(0),
2519 892 :
2520 892 : pagestream_throttle: resources.pagestream_throttle,
2521 892 :
2522 892 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2523 892 :
2524 892 : #[cfg(test)]
2525 892 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2526 892 :
2527 892 : l0_flush_global_state: resources.l0_flush_global_state,
2528 892 :
2529 892 : handles: Default::default(),
2530 892 :
2531 892 : attach_wal_lag_cooldown,
2532 892 :
2533 892 : create_idempotency,
2534 892 :
2535 892 : page_trace: Default::default(),
2536 892 : };
2537 892 :
2538 892 : result.repartition_threshold =
2539 892 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2540 892 :
2541 892 : result
2542 892 : .metrics
2543 892 : .last_record_lsn_gauge
2544 892 : .set(disk_consistent_lsn.0 as i64);
2545 892 : result
2546 892 : })
2547 892 : }
2548 :
2549 1288 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2550 1288 : let Ok(guard) = self.gate.enter() else {
2551 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2552 0 : return;
2553 : };
2554 1288 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2555 1288 : match *flush_loop_state {
2556 880 : FlushLoopState::NotStarted => (),
2557 : FlushLoopState::Running { .. } => {
2558 408 : info!(
2559 0 : "skipping attempt to start flush_loop twice {}/{}",
2560 0 : self.tenant_shard_id, self.timeline_id
2561 : );
2562 408 : return;
2563 : }
2564 : FlushLoopState::Exited => {
2565 0 : warn!(
2566 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2567 0 : self.tenant_shard_id, self.timeline_id
2568 : );
2569 0 : return;
2570 : }
2571 : }
2572 :
2573 880 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2574 880 : let self_clone = Arc::clone(self);
2575 880 :
2576 880 : debug!("spawning flush loop");
2577 880 : *flush_loop_state = FlushLoopState::Running {
2578 880 : #[cfg(test)]
2579 880 : expect_initdb_optimization: false,
2580 880 : #[cfg(test)]
2581 880 : initdb_optimization_count: 0,
2582 880 : };
2583 880 : task_mgr::spawn(
2584 880 : task_mgr::BACKGROUND_RUNTIME.handle(),
2585 880 : task_mgr::TaskKind::LayerFlushTask,
2586 880 : self.tenant_shard_id,
2587 880 : Some(self.timeline_id),
2588 880 : "layer flush task",
2589 880 : async move {
2590 880 : let _guard = guard;
2591 880 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2592 880 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2593 20 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2594 20 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2595 20 : *flush_loop_state = FlushLoopState::Exited;
2596 20 : Ok(())
2597 20 : }
2598 880 : .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))
2599 : );
2600 1288 : }
2601 :
2602 : /// Creates and starts the wal receiver.
2603 : ///
2604 : /// This function is expected to be called at most once per Timeline's lifecycle
2605 : /// when the timeline is activated.
2606 0 : fn launch_wal_receiver(
2607 0 : self: &Arc<Self>,
2608 0 : ctx: &RequestContext,
2609 0 : broker_client: BrokerClientChannel,
2610 0 : ) {
2611 0 : info!(
2612 0 : "launching WAL receiver for timeline {} of tenant {}",
2613 0 : self.timeline_id, self.tenant_shard_id
2614 : );
2615 :
2616 0 : let tenant_conf = self.tenant_conf.load();
2617 0 : let wal_connect_timeout = tenant_conf
2618 0 : .tenant_conf
2619 0 : .walreceiver_connect_timeout
2620 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2621 0 : let lagging_wal_timeout = tenant_conf
2622 0 : .tenant_conf
2623 0 : .lagging_wal_timeout
2624 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2625 0 : let max_lsn_wal_lag = tenant_conf
2626 0 : .tenant_conf
2627 0 : .max_lsn_wal_lag
2628 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2629 0 :
2630 0 : let mut guard = self.walreceiver.lock().unwrap();
2631 0 : assert!(
2632 0 : guard.is_none(),
2633 0 : "multiple launches / re-launches of WAL receiver are not supported"
2634 : );
2635 0 : *guard = Some(WalReceiver::start(
2636 0 : Arc::clone(self),
2637 0 : WalReceiverConf {
2638 0 : protocol: self.resolve_wal_receiver_protocol(),
2639 0 : wal_connect_timeout,
2640 0 : lagging_wal_timeout,
2641 0 : max_lsn_wal_lag,
2642 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2643 0 : availability_zone: self.conf.availability_zone.clone(),
2644 0 : ingest_batch_size: self.conf.ingest_batch_size,
2645 0 : },
2646 0 : broker_client,
2647 0 : ctx,
2648 0 : ));
2649 0 : }
2650 :
2651 : /// Initialize with an empty layer map. Used when creating a new timeline.
2652 880 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2653 880 : let mut layers = self.layers.try_write().expect(
2654 880 : "in the context where we call this function, no other task has access to the object",
2655 880 : );
2656 880 : layers
2657 880 : .open_mut()
2658 880 : .expect("in this context the LayerManager must still be open")
2659 880 : .initialize_empty(Lsn(start_lsn.0));
2660 880 : }
2661 :
2662 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2663 : /// files.
2664 12 : pub(super) async fn load_layer_map(
2665 12 : &self,
2666 12 : disk_consistent_lsn: Lsn,
2667 12 : index_part: IndexPart,
2668 12 : ) -> anyhow::Result<()> {
2669 : use init::{Decision::*, Discovered, DismissedLayer};
2670 : use LayerName::*;
2671 :
2672 12 : let mut guard = self.layers.write().await;
2673 :
2674 12 : let timer = self.metrics.load_layer_map_histo.start_timer();
2675 12 :
2676 12 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2677 12 : // structs representing all files on disk
2678 12 : let timeline_path = self
2679 12 : .conf
2680 12 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2681 12 : let conf = self.conf;
2682 12 : let span = tracing::Span::current();
2683 12 :
2684 12 : // Copy to move into the task we're about to spawn
2685 12 : let this = self.myself.upgrade().expect("&self method holds the arc");
2686 :
2687 12 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2688 12 : move || {
2689 12 : let _g = span.entered();
2690 12 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2691 12 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2692 12 : let mut unrecognized_files = Vec::new();
2693 12 :
2694 12 : let mut path = timeline_path;
2695 :
2696 44 : for discovered in discovered {
2697 32 : let (name, kind) = match discovered {
2698 32 : Discovered::Layer(layer_file_name, local_metadata) => {
2699 32 : discovered_layers.push((layer_file_name, local_metadata));
2700 32 : continue;
2701 : }
2702 0 : Discovered::IgnoredBackup(path) => {
2703 0 : std::fs::remove_file(path)
2704 0 : .or_else(fs_ext::ignore_not_found)
2705 0 : .fatal_err("Removing .old file");
2706 0 : continue;
2707 : }
2708 0 : Discovered::Unknown(file_name) => {
2709 0 : // we will later error if there are any
2710 0 : unrecognized_files.push(file_name);
2711 0 : continue;
2712 : }
2713 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2714 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2715 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2716 : };
2717 0 : path.push(Utf8Path::new(&name));
2718 0 : init::cleanup(&path, kind)?;
2719 0 : path.pop();
2720 : }
2721 :
2722 12 : if !unrecognized_files.is_empty() {
2723 : // assume that if there are any there are many many.
2724 0 : let n = unrecognized_files.len();
2725 0 : let first = &unrecognized_files[..n.min(10)];
2726 0 : anyhow::bail!(
2727 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2728 0 : );
2729 12 : }
2730 12 :
2731 12 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2732 12 :
2733 12 : let mut loaded_layers = Vec::new();
2734 12 : let mut needs_cleanup = Vec::new();
2735 12 : let mut total_physical_size = 0;
2736 :
2737 44 : for (name, decision) in decided {
2738 32 : let decision = match decision {
2739 32 : Ok(decision) => decision,
2740 0 : Err(DismissedLayer::Future { local }) => {
2741 0 : if let Some(local) = local {
2742 0 : init::cleanup_future_layer(
2743 0 : &local.local_path,
2744 0 : &name,
2745 0 : disk_consistent_lsn,
2746 0 : )?;
2747 0 : }
2748 0 : needs_cleanup.push(name);
2749 0 : continue;
2750 : }
2751 0 : Err(DismissedLayer::LocalOnly(local)) => {
2752 0 : init::cleanup_local_only_file(&name, &local)?;
2753 : // this file never existed remotely, we will have to do rework
2754 0 : continue;
2755 : }
2756 0 : Err(DismissedLayer::BadMetadata(local)) => {
2757 0 : init::cleanup_local_file_for_remote(&local)?;
2758 : // this file never existed remotely, we will have to do rework
2759 0 : continue;
2760 : }
2761 : };
2762 :
2763 32 : match &name {
2764 24 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2765 8 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2766 : }
2767 :
2768 32 : tracing::debug!(layer=%name, ?decision, "applied");
2769 :
2770 32 : let layer = match decision {
2771 32 : Resident { local, remote } => {
2772 32 : total_physical_size += local.file_size;
2773 32 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2774 32 : .drop_eviction_guard()
2775 : }
2776 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2777 : };
2778 :
2779 32 : loaded_layers.push(layer);
2780 : }
2781 12 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2782 12 : }
2783 12 : })
2784 12 : .await
2785 12 : .map_err(anyhow::Error::new)
2786 12 : .and_then(|x| x)?;
2787 :
2788 12 : let num_layers = loaded_layers.len();
2789 12 :
2790 12 : guard
2791 12 : .open_mut()
2792 12 : .expect("layermanager must be open during init")
2793 12 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2794 12 :
2795 12 : self.remote_client
2796 12 : .schedule_layer_file_deletion(&needs_cleanup)?;
2797 12 : self.remote_client
2798 12 : .schedule_index_upload_for_file_changes()?;
2799 : // This barrier orders above DELETEs before any later operations.
2800 : // This is critical because code executing after the barrier might
2801 : // create again objects with the same key that we just scheduled for deletion.
2802 : // For example, if we just scheduled deletion of an image layer "from the future",
2803 : // later compaction might run again and re-create the same image layer.
2804 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2805 : // "same" here means same key range and LSN.
2806 : //
2807 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2808 : // the upload queue may execute the PUT first, then the DELETE.
2809 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2810 : //
2811 : // 1. a future image layer is created and uploaded
2812 : // 2. ps restart
2813 : // 3. the future layer from (1) is deleted during load layer map
2814 : // 4. image layer is re-created and uploaded
2815 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2816 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2817 : //
2818 : // See https://github.com/neondatabase/neon/issues/5878
2819 : //
2820 : // NB: generation numbers naturally protect against this because they disambiguate
2821 : // (1) and (4)
2822 : // TODO: this is basically a no-op now, should we remove it?
2823 12 : self.remote_client.schedule_barrier()?;
2824 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2825 : // on retry.
2826 :
2827 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2828 12 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2829 12 : self.update_layer_visibility().await?;
2830 :
2831 12 : info!(
2832 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2833 : num_layers, disk_consistent_lsn, total_physical_size
2834 : );
2835 :
2836 12 : timer.stop_and_record();
2837 12 : Ok(())
2838 12 : }
2839 :
2840 : /// Retrieve current logical size of the timeline.
2841 : ///
2842 : /// The size could be lagging behind the actual number, in case
2843 : /// the initial size calculation has not been run (gets triggered on the first size access).
2844 : ///
2845 : /// return size and boolean flag that shows if the size is exact
2846 0 : pub(crate) fn get_current_logical_size(
2847 0 : self: &Arc<Self>,
2848 0 : priority: GetLogicalSizePriority,
2849 0 : ctx: &RequestContext,
2850 0 : ) -> logical_size::CurrentLogicalSize {
2851 0 : if !self.tenant_shard_id.is_shard_zero() {
2852 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2853 : // when HTTP API is serving a GET for timeline zero, return zero
2854 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2855 0 : }
2856 0 :
2857 0 : let current_size = self.current_logical_size.current_size();
2858 0 : debug!("Current size: {current_size:?}");
2859 :
2860 0 : match (current_size.accuracy(), priority) {
2861 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2862 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2863 0 : // background task will eventually deliver an exact value, we're in no rush
2864 0 : }
2865 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2866 : // background task is not ready, but user is asking for it now;
2867 : // => make the background task skip the line
2868 : // (The alternative would be to calculate the size here, but,
2869 : // it can actually take a long time if the user has a lot of rels.
2870 : // And we'll inevitable need it again; So, let the background task do the work.)
2871 0 : match self
2872 0 : .current_logical_size
2873 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2874 0 : .get()
2875 : {
2876 0 : Some(cancel) => cancel.cancel(),
2877 : None => {
2878 0 : match self.current_state() {
2879 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2880 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2881 0 : // Don't make noise.
2882 0 : }
2883 : TimelineState::Loading => {
2884 : // Import does not return an activated timeline.
2885 0 : info!("discarding priority boost for logical size calculation because timeline is not yet active");
2886 : }
2887 : TimelineState::Active => {
2888 : // activation should be setting the once cell
2889 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2890 0 : debug_assert!(false);
2891 : }
2892 : }
2893 : }
2894 : }
2895 : }
2896 : }
2897 :
2898 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2899 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2900 0 : let first = self
2901 0 : .current_logical_size
2902 0 : .did_return_approximate_to_walreceiver
2903 0 : .compare_exchange(
2904 0 : false,
2905 0 : true,
2906 0 : AtomicOrdering::Relaxed,
2907 0 : AtomicOrdering::Relaxed,
2908 0 : )
2909 0 : .is_ok();
2910 0 : if first {
2911 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2912 0 : }
2913 0 : }
2914 0 : }
2915 :
2916 0 : current_size
2917 0 : }
2918 :
2919 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2920 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2921 : // nothing to do for freshly created timelines;
2922 0 : assert_eq!(
2923 0 : self.current_logical_size.current_size().accuracy(),
2924 0 : logical_size::Accuracy::Exact,
2925 0 : );
2926 0 : self.current_logical_size.initialized.add_permits(1);
2927 0 : return;
2928 : };
2929 :
2930 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2931 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2932 0 : self.current_logical_size
2933 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2934 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2935 0 :
2936 0 : let self_clone = Arc::clone(self);
2937 0 : let background_ctx = ctx.detached_child(
2938 0 : TaskKind::InitialLogicalSizeCalculation,
2939 0 : DownloadBehavior::Download,
2940 0 : );
2941 0 : task_mgr::spawn(
2942 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2943 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2944 0 : self.tenant_shard_id,
2945 0 : Some(self.timeline_id),
2946 0 : "initial size calculation",
2947 : // NB: don't log errors here, task_mgr will do that.
2948 0 : async move {
2949 0 : self_clone
2950 0 : .initial_logical_size_calculation_task(
2951 0 : initial_part_end,
2952 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2953 0 : background_ctx,
2954 0 : )
2955 0 : .await;
2956 0 : Ok(())
2957 0 : }
2958 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)),
2959 : );
2960 0 : }
2961 :
2962 : /// # Cancellation
2963 : ///
2964 : /// This method is sensitive to `Timeline::cancel`.
2965 : ///
2966 : /// It is _not_ sensitive to task_mgr::shutdown_token().
2967 : ///
2968 : /// # Cancel-Safety
2969 : ///
2970 : /// It does Timeline IO, hence this should be polled to completion because
2971 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
2972 : /// to reason about.
2973 0 : async fn initial_logical_size_calculation_task(
2974 0 : self: Arc<Self>,
2975 0 : initial_part_end: Lsn,
2976 0 : skip_concurrency_limiter: CancellationToken,
2977 0 : background_ctx: RequestContext,
2978 0 : ) {
2979 0 : scopeguard::defer! {
2980 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2981 0 : self.current_logical_size.initialized.add_permits(1);
2982 0 : }
2983 0 :
2984 0 : let try_once = |attempt: usize| {
2985 0 : let background_ctx = &background_ctx;
2986 0 : let self_ref = &self;
2987 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2988 0 : async move {
2989 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2990 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2991 0 : background_ctx,
2992 0 : );
2993 :
2994 : use crate::metrics::initial_logical_size::StartCircumstances;
2995 0 : let (_maybe_permit, circumstances) = tokio::select! {
2996 0 : permit = wait_for_permit => {
2997 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2998 : }
2999 0 : _ = self_ref.cancel.cancelled() => {
3000 0 : return Err(CalculateLogicalSizeError::Cancelled);
3001 : }
3002 0 : () = skip_concurrency_limiter.cancelled() => {
3003 : // Some action that is part of a end user interaction requested logical size
3004 : // => break out of the rate limit
3005 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3006 : // but then again what happens if they cancel; also, we should just be using
3007 : // one runtime across the entire process, so, let's leave this for now.
3008 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3009 : }
3010 : };
3011 :
3012 0 : let metrics_guard = if attempt == 1 {
3013 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3014 : } else {
3015 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3016 : };
3017 :
3018 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3019 0 : self_ref.conf,
3020 0 : self_ref
3021 0 : .gate
3022 0 : .enter()
3023 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3024 : );
3025 :
3026 0 : let calculated_size = self_ref
3027 0 : .logical_size_calculation_task(
3028 0 : initial_part_end,
3029 0 : LogicalSizeCalculationCause::Initial,
3030 0 : background_ctx,
3031 0 : )
3032 0 : .await?;
3033 :
3034 0 : self_ref
3035 0 : .trigger_aux_file_size_computation(
3036 0 : initial_part_end,
3037 0 : background_ctx,
3038 0 : io_concurrency,
3039 0 : )
3040 0 : .await?;
3041 :
3042 : // TODO: add aux file size to logical size
3043 :
3044 0 : Ok((calculated_size, metrics_guard))
3045 0 : }
3046 0 : };
3047 :
3048 0 : let retrying = async {
3049 0 : let mut attempt = 0;
3050 : loop {
3051 0 : attempt += 1;
3052 0 :
3053 0 : match try_once(attempt).await {
3054 0 : Ok(res) => return ControlFlow::Continue(res),
3055 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3056 : Err(
3057 0 : e @ (CalculateLogicalSizeError::Decode(_)
3058 0 : | CalculateLogicalSizeError::PageRead(_)),
3059 0 : ) => {
3060 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3061 : // exponential back-off doesn't make sense at these long intervals;
3062 : // use fixed retry interval with generous jitter instead
3063 0 : let sleep_duration = Duration::from_secs(
3064 0 : u64::try_from(
3065 0 : // 1hour base
3066 0 : (60_i64 * 60_i64)
3067 0 : // 10min jitter
3068 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3069 0 : )
3070 0 : .expect("10min < 1hour"),
3071 0 : );
3072 0 : tokio::select! {
3073 0 : _ = tokio::time::sleep(sleep_duration) => {}
3074 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3075 : }
3076 : }
3077 : }
3078 : }
3079 0 : };
3080 :
3081 0 : let (calculated_size, metrics_guard) = match retrying.await {
3082 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3083 0 : ControlFlow::Break(()) => return,
3084 : };
3085 :
3086 : // we cannot query current_logical_size.current_size() to know the current
3087 : // *negative* value, only truncated to u64.
3088 0 : let added = self
3089 0 : .current_logical_size
3090 0 : .size_added_after_initial
3091 0 : .load(AtomicOrdering::Relaxed);
3092 0 :
3093 0 : let sum = calculated_size.saturating_add_signed(added);
3094 0 :
3095 0 : // set the gauge value before it can be set in `update_current_logical_size`.
3096 0 : self.metrics.current_logical_size_gauge.set(sum);
3097 0 :
3098 0 : self.current_logical_size
3099 0 : .initial_logical_size
3100 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3101 0 : .ok()
3102 0 : .expect("only this task sets it");
3103 0 : }
3104 :
3105 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3106 0 : self: &Arc<Self>,
3107 0 : lsn: Lsn,
3108 0 : cause: LogicalSizeCalculationCause,
3109 0 : ctx: RequestContext,
3110 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3111 0 : let (sender, receiver) = oneshot::channel();
3112 0 : let self_clone = Arc::clone(self);
3113 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3114 0 : // we should stop the size calculation work and return an error.
3115 0 : // That would require restructuring this function's API to
3116 0 : // return the result directly, instead of a Receiver for the result.
3117 0 : let ctx = ctx.detached_child(
3118 0 : TaskKind::OndemandLogicalSizeCalculation,
3119 0 : DownloadBehavior::Download,
3120 0 : );
3121 0 : task_mgr::spawn(
3122 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3123 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3124 0 : self.tenant_shard_id,
3125 0 : Some(self.timeline_id),
3126 0 : "ondemand logical size calculation",
3127 0 : async move {
3128 0 : let res = self_clone
3129 0 : .logical_size_calculation_task(lsn, cause, &ctx)
3130 0 : .await;
3131 0 : let _ = sender.send(res).ok();
3132 0 : Ok(()) // Receiver is responsible for handling errors
3133 0 : }
3134 0 : .in_current_span(),
3135 0 : );
3136 0 : receiver
3137 0 : }
3138 :
3139 : #[instrument(skip_all)]
3140 : async fn logical_size_calculation_task(
3141 : self: &Arc<Self>,
3142 : lsn: Lsn,
3143 : cause: LogicalSizeCalculationCause,
3144 : ctx: &RequestContext,
3145 : ) -> Result<u64, CalculateLogicalSizeError> {
3146 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3147 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3148 : // accurate relation sizes, and they do not emit consumption metrics.
3149 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3150 :
3151 : let guard = self
3152 : .gate
3153 : .enter()
3154 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3155 :
3156 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3157 : }
3158 :
3159 : /// Calculate the logical size of the database at the latest LSN.
3160 : ///
3161 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3162 : /// especially if we need to download remote layers.
3163 0 : async fn calculate_logical_size(
3164 0 : &self,
3165 0 : up_to_lsn: Lsn,
3166 0 : cause: LogicalSizeCalculationCause,
3167 0 : _guard: &GateGuard,
3168 0 : ctx: &RequestContext,
3169 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3170 0 : info!(
3171 0 : "Calculating logical size for timeline {} at {}",
3172 : self.timeline_id, up_to_lsn
3173 : );
3174 :
3175 0 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
3176 : {
3177 0 : return Err(CalculateLogicalSizeError::Cancelled);
3178 0 : }
3179 :
3180 : // See if we've already done the work for initial size calculation.
3181 : // This is a short-cut for timelines that are mostly unused.
3182 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3183 0 : return Ok(size);
3184 0 : }
3185 0 : let storage_time_metrics = match cause {
3186 : LogicalSizeCalculationCause::Initial
3187 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3188 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3189 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3190 0 : &self.metrics.imitate_logical_size_histo
3191 : }
3192 : };
3193 0 : let timer = storage_time_metrics.start_timer();
3194 0 : let logical_size = self
3195 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3196 0 : .await?;
3197 0 : debug!("calculated logical size: {logical_size}");
3198 0 : timer.stop_and_record();
3199 0 : Ok(logical_size)
3200 0 : }
3201 :
3202 : /// Update current logical size, adding `delta' to the old value.
3203 541140 : fn update_current_logical_size(&self, delta: i64) {
3204 541140 : let logical_size = &self.current_logical_size;
3205 541140 : logical_size.increment_size(delta);
3206 541140 :
3207 541140 : // Also set the value in the prometheus gauge. Note that
3208 541140 : // there is a race condition here: if this is is called by two
3209 541140 : // threads concurrently, the prometheus gauge might be set to
3210 541140 : // one value while current_logical_size is set to the
3211 541140 : // other.
3212 541140 : match logical_size.current_size() {
3213 541140 : CurrentLogicalSize::Exact(ref new_current_size) => self
3214 541140 : .metrics
3215 541140 : .current_logical_size_gauge
3216 541140 : .set(new_current_size.into()),
3217 0 : CurrentLogicalSize::Approximate(_) => {
3218 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3219 0 : // forth between the initial size calculation task.
3220 0 : }
3221 : }
3222 541140 : }
3223 :
3224 5912 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3225 5912 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3226 5912 : let aux_metric =
3227 5912 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3228 5912 :
3229 5912 : let sum_of_entries = self
3230 5912 : .directory_metrics
3231 5912 : .iter()
3232 41384 : .map(|v| v.load(AtomicOrdering::Relaxed))
3233 5912 : .sum();
3234 : // Set a high general threshold and a lower threshold for the auxiliary files,
3235 : // as we can have large numbers of relations in the db directory.
3236 : const SUM_THRESHOLD: u64 = 5000;
3237 : const AUX_THRESHOLD: u64 = 1000;
3238 5912 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3239 0 : self.metrics
3240 0 : .directory_entries_count_gauge
3241 0 : .set(sum_of_entries);
3242 5912 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3243 0 : metric.set(sum_of_entries);
3244 5912 : }
3245 5912 : }
3246 :
3247 0 : async fn find_layer(
3248 0 : &self,
3249 0 : layer_name: &LayerName,
3250 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3251 0 : let guard = self.layers.read().await;
3252 0 : let layer = guard
3253 0 : .layer_map()?
3254 0 : .iter_historic_layers()
3255 0 : .find(|l| &l.layer_name() == layer_name)
3256 0 : .map(|found| guard.get_from_desc(&found));
3257 0 : Ok(layer)
3258 0 : }
3259 :
3260 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3261 : /// indicating which layers are currently on-disk on the primary.
3262 : ///
3263 : /// None is returned if the Timeline is in a state where uploading a heatmap
3264 : /// doesn't make sense, such as shutting down or initializing. The caller
3265 : /// should treat this as a cue to simply skip doing any heatmap uploading
3266 : /// for this timeline.
3267 4 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3268 4 : if !self.is_active() {
3269 0 : return None;
3270 4 : }
3271 :
3272 4 : let guard = self.layers.read().await;
3273 :
3274 20 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3275 20 : match layer.visibility() {
3276 : LayerVisibilityHint::Visible => {
3277 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3278 16 : let last_activity_ts = layer.latest_activity();
3279 16 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
3280 : }
3281 : LayerVisibilityHint::Covered => {
3282 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3283 4 : None
3284 : }
3285 : }
3286 20 : });
3287 4 :
3288 4 : let mut layers = resident.collect::<Vec<_>>();
3289 4 :
3290 4 : // Sort layers in order of which to download first. For a large set of layers to download, we
3291 4 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3292 4 : // or hours later:
3293 4 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3294 4 : // only exist for a few minutes before being compacted into L1s.
3295 4 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3296 4 : // the layer is likely to be covered by an image layer during compaction.
3297 40 : layers.sort_by_key(|(desc, _meta, _atime)| {
3298 40 : std::cmp::Reverse((
3299 40 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3300 40 : desc.lsn_range.end,
3301 40 : ))
3302 40 : });
3303 4 :
3304 4 : let layers = layers
3305 4 : .into_iter()
3306 16 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3307 4 : .collect();
3308 4 :
3309 4 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3310 4 : }
3311 :
3312 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3313 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3314 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3315 0 : // branchpoint in the value in IndexPart::lineage
3316 0 : self.ancestor_lsn == lsn
3317 0 : || (self.ancestor_lsn == Lsn::INVALID
3318 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3319 0 : }
3320 : }
3321 :
3322 : impl Timeline {
3323 : #[allow(clippy::doc_lazy_continuation)]
3324 : /// Get the data needed to reconstruct all keys in the provided keyspace
3325 : ///
3326 : /// The algorithm is as follows:
3327 : /// 1. While some keys are still not done and there's a timeline to visit:
3328 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3329 : /// 2.1: Build the fringe for the current keyspace
3330 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3331 : /// intersects
3332 : /// 2.3. Pop the timeline from the fringe
3333 : /// 2.4. If the fringe is empty, go back to 1
3334 1255290 : async fn get_vectored_reconstruct_data(
3335 1255290 : &self,
3336 1255290 : mut keyspace: KeySpace,
3337 1255290 : request_lsn: Lsn,
3338 1255290 : reconstruct_state: &mut ValuesReconstructState,
3339 1255290 : ctx: &RequestContext,
3340 1255290 : ) -> Result<(), GetVectoredError> {
3341 1255290 : let mut timeline_owned: Arc<Timeline>;
3342 1255290 : let mut timeline = self;
3343 1255290 :
3344 1255290 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3345 :
3346 1255286 : let missing_keyspace = loop {
3347 1707520 : if self.cancel.is_cancelled() {
3348 0 : return Err(GetVectoredError::Cancelled);
3349 1707520 : }
3350 :
3351 : let TimelineVisitOutcome {
3352 1707520 : completed_keyspace: completed,
3353 1707520 : image_covered_keyspace,
3354 1707520 : } = Self::get_vectored_reconstruct_data_timeline(
3355 1707520 : timeline,
3356 1707520 : keyspace.clone(),
3357 1707520 : cont_lsn,
3358 1707520 : reconstruct_state,
3359 1707520 : &self.cancel,
3360 1707520 : ctx,
3361 1707520 : )
3362 1707520 : .await?;
3363 :
3364 1707520 : keyspace.remove_overlapping_with(&completed);
3365 1707520 :
3366 1707520 : // Do not descend into the ancestor timeline for aux files.
3367 1707520 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3368 1707520 : // stalling compaction.
3369 1707520 : keyspace.remove_overlapping_with(&KeySpace {
3370 1707520 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
3371 1707520 : });
3372 1707520 :
3373 1707520 : // Keyspace is fully retrieved
3374 1707520 : if keyspace.is_empty() {
3375 1255230 : break None;
3376 452290 : }
3377 :
3378 452290 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3379 : // Not fully retrieved but no ancestor timeline.
3380 56 : break Some(keyspace);
3381 : };
3382 :
3383 : // Now we see if there are keys covered by the image layer but does not exist in the
3384 : // image layer, which means that the key does not exist.
3385 :
3386 : // The block below will stop the vectored search if any of the keys encountered an image layer
3387 : // which did not contain a snapshot for said key. Since we have already removed all completed
3388 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3389 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3390 : // and stop the search as a result of that.
3391 452234 : let mut removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3392 452234 : // Do not fire missing key error for sparse keys.
3393 452234 : removed.remove_overlapping_with(&KeySpace {
3394 452234 : ranges: vec![SPARSE_RANGE],
3395 452234 : });
3396 452234 : if !removed.is_empty() {
3397 0 : break Some(removed);
3398 452234 : }
3399 452234 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3400 452234 : // keyspace.
3401 452234 :
3402 452234 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3403 452234 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3404 452234 : timeline_owned = timeline
3405 452234 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3406 452234 : .await?;
3407 452230 : timeline = &*timeline_owned;
3408 : };
3409 :
3410 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
3411 1255286 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
3412 56 : let mut missing_keyspace = missing_keyspace;
3413 56 : missing_keyspace.remove_overlapping_with(&KeySpace {
3414 56 : ranges: vec![SPARSE_RANGE],
3415 56 : });
3416 56 : if missing_keyspace.is_empty() {
3417 28 : None
3418 : } else {
3419 28 : Some(missing_keyspace)
3420 : }
3421 : } else {
3422 1255230 : None
3423 : };
3424 :
3425 1255286 : if let Some(missing_keyspace) = missing_keyspace {
3426 28 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3427 28 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3428 28 : shard: self
3429 28 : .shard_identity
3430 28 : .get_shard_number(&missing_keyspace.start().unwrap()),
3431 28 : cont_lsn,
3432 28 : request_lsn,
3433 28 : ancestor_lsn: Some(timeline.ancestor_lsn),
3434 28 : backtrace: None,
3435 28 : }));
3436 1255258 : }
3437 1255258 :
3438 1255258 : Ok(())
3439 1255290 : }
3440 :
3441 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3442 : ///
3443 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3444 : /// the current keyspace. The current keyspace of the search at any given timeline
3445 : /// is the original keyspace minus all the keys that have been completed minus
3446 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3447 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3448 : ///
3449 : /// This is basically a depth-first search visitor implementation where a vertex
3450 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3451 : ///
3452 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3453 : /// and get all the required reconstruct data from the layer in one go.
3454 : ///
3455 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3456 : /// decides how to deal with these two keyspaces.
3457 1707520 : async fn get_vectored_reconstruct_data_timeline(
3458 1707520 : timeline: &Timeline,
3459 1707520 : keyspace: KeySpace,
3460 1707520 : mut cont_lsn: Lsn,
3461 1707520 : reconstruct_state: &mut ValuesReconstructState,
3462 1707520 : cancel: &CancellationToken,
3463 1707520 : ctx: &RequestContext,
3464 1707520 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3465 1707520 : let mut unmapped_keyspace = keyspace.clone();
3466 1707520 : let mut fringe = LayerFringe::new();
3467 1707520 :
3468 1707520 : let mut completed_keyspace = KeySpace::default();
3469 1707520 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3470 :
3471 : loop {
3472 3403623 : if cancel.is_cancelled() {
3473 0 : return Err(GetVectoredError::Cancelled);
3474 3403623 : }
3475 3403623 :
3476 3403623 : let (keys_done_last_step, keys_with_image_coverage) =
3477 3403623 : reconstruct_state.consume_done_keys();
3478 3403623 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3479 3403623 : completed_keyspace.merge(&keys_done_last_step);
3480 3403623 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3481 45092 : unmapped_keyspace
3482 45092 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3483 45092 : image_covered_keyspace.add_range(keys_with_image_coverage);
3484 3358531 : }
3485 :
3486 : // Do not descent any further if the last layer we visited
3487 : // completed all keys in the keyspace it inspected. This is not
3488 : // required for correctness, but avoids visiting extra layers
3489 : // which turns out to be a perf bottleneck in some cases.
3490 3403623 : if !unmapped_keyspace.is_empty() {
3491 2152877 : let guard = timeline.layers.read().await;
3492 2152877 : let layers = guard.layer_map()?;
3493 :
3494 2152877 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3495 1830155 : let start_lsn = l.get_lsn_range().start;
3496 1830155 : cont_lsn > start_lsn
3497 2152877 : });
3498 2152877 :
3499 2152877 : match in_memory_layer {
3500 1213252 : Some(l) => {
3501 1213252 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3502 1213252 : fringe.update(
3503 1213252 : ReadableLayer::InMemoryLayer(l),
3504 1213252 : unmapped_keyspace.clone(),
3505 1213252 : lsn_range,
3506 1213252 : );
3507 1213252 : }
3508 : None => {
3509 939673 : for range in unmapped_keyspace.ranges.iter() {
3510 939673 : let results = layers.range_search(range.clone(), cont_lsn);
3511 939673 :
3512 939673 : results
3513 939673 : .found
3514 939673 : .into_iter()
3515 939673 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3516 482879 : (
3517 482879 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3518 482879 : keyspace_accum.to_keyspace(),
3519 482879 : lsn_floor..cont_lsn,
3520 482879 : )
3521 939673 : })
3522 939673 : .for_each(|(layer, keyspace, lsn_range)| {
3523 482879 : fringe.update(layer, keyspace, lsn_range)
3524 939673 : });
3525 939673 : }
3526 : }
3527 : }
3528 :
3529 : // It's safe to drop the layer map lock after planning the next round of reads.
3530 : // The fringe keeps readable handles for the layers which are safe to read even
3531 : // if layers were compacted or flushed.
3532 : //
3533 : // The more interesting consideration is: "Why is the read algorithm still correct
3534 : // if the layer map changes while it is operating?". Doing a vectored read on a
3535 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3536 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3537 : // range at *a particular point in time*. It is fine for the answer to be different
3538 : // at two different time points.
3539 2152877 : drop(guard);
3540 1250746 : }
3541 :
3542 3403623 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3543 1696103 : let next_cont_lsn = lsn_range.start;
3544 1696103 : layer_to_read
3545 1696103 : .get_values_reconstruct_data(
3546 1696103 : keyspace_to_read.clone(),
3547 1696103 : lsn_range,
3548 1696103 : reconstruct_state,
3549 1696103 : ctx,
3550 1696103 : )
3551 1696103 : .await?;
3552 :
3553 1696103 : unmapped_keyspace = keyspace_to_read;
3554 1696103 : cont_lsn = next_cont_lsn;
3555 1696103 :
3556 1696103 : reconstruct_state.on_layer_visited(&layer_to_read);
3557 : } else {
3558 1707520 : break;
3559 1707520 : }
3560 1707520 : }
3561 1707520 :
3562 1707520 : Ok(TimelineVisitOutcome {
3563 1707520 : completed_keyspace,
3564 1707520 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3565 1707520 : })
3566 1707520 : }
3567 :
3568 452234 : async fn get_ready_ancestor_timeline(
3569 452234 : &self,
3570 452234 : ancestor: &Arc<Timeline>,
3571 452234 : ctx: &RequestContext,
3572 452234 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3573 452234 : // It's possible that the ancestor timeline isn't active yet, or
3574 452234 : // is active but hasn't yet caught up to the branch point. Wait
3575 452234 : // for it.
3576 452234 : //
3577 452234 : // This cannot happen while the pageserver is running normally,
3578 452234 : // because you cannot create a branch from a point that isn't
3579 452234 : // present in the pageserver yet. However, we don't wait for the
3580 452234 : // branch point to be uploaded to cloud storage before creating
3581 452234 : // a branch. I.e., the branch LSN need not be remote consistent
3582 452234 : // for the branching operation to succeed.
3583 452234 : //
3584 452234 : // Hence, if we try to load a tenant in such a state where
3585 452234 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3586 452234 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3587 452234 : // then we will need to wait for the ancestor timeline to
3588 452234 : // re-stream WAL up to branch_lsn before we access it.
3589 452234 : //
3590 452234 : // How can a tenant get in such a state?
3591 452234 : // - ungraceful pageserver process exit
3592 452234 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3593 452234 : //
3594 452234 : // NB: this could be avoided by requiring
3595 452234 : // branch_lsn >= remote_consistent_lsn
3596 452234 : // during branch creation.
3597 452234 : match ancestor.wait_to_become_active(ctx).await {
3598 452230 : Ok(()) => {}
3599 : Err(TimelineState::Stopping) => {
3600 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3601 0 : return Err(GetReadyAncestorError::Cancelled);
3602 : }
3603 4 : Err(state) => {
3604 4 : return Err(GetReadyAncestorError::BadState {
3605 4 : timeline_id: ancestor.timeline_id,
3606 4 : state,
3607 4 : });
3608 : }
3609 : }
3610 452230 : ancestor
3611 452230 : .wait_lsn(
3612 452230 : self.ancestor_lsn,
3613 452230 : WaitLsnWaiter::Timeline(self),
3614 452230 : WaitLsnTimeout::Default,
3615 452230 : ctx,
3616 452230 : )
3617 452230 : .await
3618 452230 : .map_err(|e| match e {
3619 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3620 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3621 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3622 0 : timeline_id: ancestor.timeline_id,
3623 0 : state,
3624 0 : },
3625 452230 : })?;
3626 :
3627 452230 : Ok(ancestor.clone())
3628 452234 : }
3629 :
3630 594312 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3631 594312 : &self.shard_identity
3632 594312 : }
3633 :
3634 : #[inline(always)]
3635 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3636 0 : ShardTimelineId {
3637 0 : shard_index: ShardIndex {
3638 0 : shard_number: self.shard_identity.number,
3639 0 : shard_count: self.shard_identity.count,
3640 0 : },
3641 0 : timeline_id: self.timeline_id,
3642 0 : }
3643 0 : }
3644 :
3645 : /// Returns a non-frozen open in-memory layer for ingestion.
3646 : ///
3647 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3648 : /// this function without holding the mutex.
3649 2592 : async fn get_layer_for_write(
3650 2592 : &self,
3651 2592 : lsn: Lsn,
3652 2592 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3653 2592 : ctx: &RequestContext,
3654 2592 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3655 2592 : let mut guard = self.layers.write().await;
3656 :
3657 2592 : let last_record_lsn = self.get_last_record_lsn();
3658 2592 : ensure!(
3659 2592 : lsn > last_record_lsn,
3660 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3661 : lsn,
3662 : last_record_lsn,
3663 : );
3664 :
3665 2592 : let layer = guard
3666 2592 : .open_mut()?
3667 2592 : .get_layer_for_write(
3668 2592 : lsn,
3669 2592 : self.conf,
3670 2592 : self.timeline_id,
3671 2592 : self.tenant_shard_id,
3672 2592 : &self.gate,
3673 2592 : ctx,
3674 2592 : )
3675 2592 : .await?;
3676 2592 : Ok(layer)
3677 2592 : }
3678 :
3679 10558184 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3680 10558184 : assert!(new_lsn.is_aligned());
3681 :
3682 10558184 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
3683 10558184 : self.last_record_lsn.advance(new_lsn);
3684 10558184 : }
3685 :
3686 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3687 : ///
3688 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3689 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3690 2400 : async fn freeze_inmem_layer_at(
3691 2400 : &self,
3692 2400 : at: Lsn,
3693 2400 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3694 2400 : ) -> Result<u64, FlushLayerError> {
3695 2400 : let frozen = {
3696 2400 : let mut guard = self.layers.write().await;
3697 2400 : guard
3698 2400 : .open_mut()?
3699 2400 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock)
3700 2400 : .await
3701 : };
3702 :
3703 2400 : if frozen {
3704 2344 : let now = Instant::now();
3705 2344 : *(self.last_freeze_ts.write().unwrap()) = now;
3706 2344 : }
3707 :
3708 : // Increment the flush cycle counter and wake up the flush task.
3709 : // Remember the new value, so that when we listen for the flush
3710 : // to finish, we know when the flush that we initiated has
3711 : // finished, instead of some other flush that was started earlier.
3712 2400 : let mut my_flush_request = 0;
3713 2400 :
3714 2400 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3715 2400 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3716 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3717 2400 : }
3718 2400 :
3719 2400 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3720 2400 : my_flush_request = *counter + 1;
3721 2400 : *counter = my_flush_request;
3722 2400 : *lsn = std::cmp::max(at, *lsn);
3723 2400 : });
3724 2400 :
3725 2400 : assert_ne!(my_flush_request, 0);
3726 :
3727 2400 : Ok(my_flush_request)
3728 2400 : }
3729 :
3730 : /// Layer flusher task's main loop.
3731 880 : async fn flush_loop(
3732 880 : self: &Arc<Self>,
3733 880 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3734 880 : ctx: &RequestContext,
3735 880 : ) {
3736 : // Subscribe to L0 delta layer updates, for compaction backpressure.
3737 880 : let mut watch_l0 = match self.layers.read().await.layer_map() {
3738 880 : Ok(lm) => lm.watch_level0_deltas(),
3739 0 : Err(Shutdown) => return,
3740 : };
3741 :
3742 880 : info!("started flush loop");
3743 : loop {
3744 3207 : tokio::select! {
3745 3207 : _ = self.cancel.cancelled() => {
3746 20 : info!("shutting down layer flush task due to Timeline::cancel");
3747 20 : break;
3748 : },
3749 3207 : _ = layer_flush_start_rx.changed() => {}
3750 2327 : }
3751 2327 : trace!("waking up");
3752 2327 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3753 2327 :
3754 2327 : // The highest LSN to which we flushed in the loop over frozen layers
3755 2327 : let mut flushed_to_lsn = Lsn(0);
3756 :
3757 2327 : let result = loop {
3758 4671 : if self.cancel.is_cancelled() {
3759 0 : info!("dropping out of flush loop for timeline shutdown");
3760 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3761 : // anyone waiting on that will respect self.cancel as well: they will stop
3762 : // waiting at the same time we as drop out of this loop.
3763 0 : return;
3764 4671 : }
3765 4671 :
3766 4671 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
3767 4671 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
3768 4671 : if flushed_to_lsn >= frozen_to_lsn {
3769 2271 : break Ok(());
3770 2400 : }
3771 :
3772 : // Fetch the next layer to flush, if any.
3773 2400 : let (layer, l0_count, frozen_count, frozen_size) = {
3774 2400 : let layers = self.layers.read().await;
3775 2400 : let Ok(lm) = layers.layer_map() else {
3776 0 : info!("dropping out of flush loop for timeline shutdown");
3777 0 : return;
3778 : };
3779 2400 : let l0_count = lm.level0_deltas().len();
3780 2400 : let frozen_count = lm.frozen_layers.len();
3781 2400 : let frozen_size: u64 = lm
3782 2400 : .frozen_layers
3783 2400 : .iter()
3784 2418 : .map(|l| l.estimated_in_mem_size())
3785 2400 : .sum();
3786 2400 : let layer = lm.frozen_layers.front().cloned();
3787 2400 : (layer, l0_count, frozen_count, frozen_size)
3788 2400 : // drop 'layers' lock
3789 2400 : };
3790 2400 : let Some(layer) = layer else {
3791 56 : break Ok(());
3792 : };
3793 :
3794 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
3795 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
3796 : //
3797 : // NB: the compaction loop only checks `compaction_threshold` every 20 seconds, so
3798 : // we can end up stalling before compaction even starts. Consider making it more
3799 : // responsive (e.g. via `watch_level0_deltas`).
3800 2344 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
3801 0 : if l0_count >= stall_threshold {
3802 0 : warn!(
3803 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
3804 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
3805 : );
3806 0 : let stall_timer = self
3807 0 : .metrics
3808 0 : .flush_delay_histo
3809 0 : .start_timer()
3810 0 : .record_on_drop();
3811 0 : tokio::select! {
3812 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
3813 0 : if let Ok(l0) = result.as_deref() {
3814 0 : let delay = stall_timer.elapsed().as_secs_f64();
3815 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
3816 0 : }
3817 : },
3818 0 : _ = self.cancel.cancelled() => {},
3819 : }
3820 0 : continue; // check again
3821 0 : }
3822 2344 : }
3823 :
3824 : // Flush the layer.
3825 2344 : let flush_timer = self.metrics.flush_time_histo.start_timer();
3826 2344 : match self.flush_frozen_layer(layer, ctx).await {
3827 2344 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
3828 : Err(FlushLayerError::Cancelled) => {
3829 0 : info!("dropping out of flush loop for timeline shutdown");
3830 0 : return;
3831 : }
3832 0 : err @ Err(
3833 0 : FlushLayerError::NotRunning(_)
3834 0 : | FlushLayerError::Other(_)
3835 0 : | FlushLayerError::CreateImageLayersError(_),
3836 0 : ) => {
3837 0 : error!("could not flush frozen layer: {err:?}");
3838 0 : break err.map(|_| ());
3839 : }
3840 : }
3841 2344 : let flush_duration = flush_timer.stop_and_record();
3842 :
3843 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
3844 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
3845 : // ingestion by having ephemeral layer rolls wait for flushes.
3846 2344 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
3847 0 : if l0_count >= delay_threshold {
3848 0 : let delay = flush_duration.as_secs_f64();
3849 0 : info!(
3850 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
3851 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
3852 : );
3853 0 : let _delay_timer = self
3854 0 : .metrics
3855 0 : .flush_delay_histo
3856 0 : .start_timer()
3857 0 : .record_on_drop();
3858 0 : tokio::select! {
3859 0 : _ = tokio::time::sleep(flush_duration) => {},
3860 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
3861 0 : _ = self.cancel.cancelled() => {},
3862 : }
3863 0 : }
3864 2344 : }
3865 : };
3866 :
3867 : // Unsharded tenants should never advance their LSN beyond the end of the
3868 : // highest layer they write: such gaps between layer data and the frozen LSN
3869 : // are only legal on sharded tenants.
3870 2327 : debug_assert!(
3871 2327 : self.shard_identity.count.count() > 1
3872 2327 : || flushed_to_lsn >= frozen_to_lsn
3873 56 : || !flushed_to_lsn.is_valid()
3874 : );
3875 :
3876 2327 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3877 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3878 : // to us via layer_flush_start_rx, then advance it here.
3879 : //
3880 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3881 : // never encounter a gap in the wal.
3882 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3883 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3884 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3885 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3886 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3887 0 : }
3888 0 : }
3889 2327 : }
3890 :
3891 : // Notify any listeners that we're done
3892 2327 : let _ = self
3893 2327 : .layer_flush_done_tx
3894 2327 : .send_replace((flush_counter, result));
3895 : }
3896 20 : }
3897 :
3898 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
3899 2240 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3900 2240 : let mut rx = self.layer_flush_done_tx.subscribe();
3901 : loop {
3902 : {
3903 4551 : let (last_result_counter, last_result) = &*rx.borrow();
3904 4551 : if *last_result_counter >= request {
3905 2240 : if let Err(err) = last_result {
3906 : // We already logged the original error in
3907 : // flush_loop. We cannot propagate it to the caller
3908 : // here, because it might not be Cloneable
3909 0 : return Err(err.clone());
3910 : } else {
3911 2240 : return Ok(());
3912 : }
3913 2311 : }
3914 2311 : }
3915 2311 : trace!("waiting for flush to complete");
3916 2311 : tokio::select! {
3917 2311 : rx_e = rx.changed() => {
3918 2311 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3919 : },
3920 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3921 : // the notification from [`flush_loop`] that it completed.
3922 2311 : _ = self.cancel.cancelled() => {
3923 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3924 0 : return Ok(())
3925 : }
3926 : };
3927 2311 : trace!("done")
3928 : }
3929 2240 : }
3930 :
3931 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3932 : ///
3933 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3934 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3935 : async fn flush_frozen_layer(
3936 : self: &Arc<Self>,
3937 : frozen_layer: Arc<InMemoryLayer>,
3938 : ctx: &RequestContext,
3939 : ) -> Result<Lsn, FlushLayerError> {
3940 : debug_assert_current_span_has_tenant_and_timeline_id();
3941 :
3942 : // As a special case, when we have just imported an image into the repository,
3943 : // instead of writing out a L0 delta layer, we directly write out image layer
3944 : // files instead. This is possible as long as *all* the data imported into the
3945 : // repository have the same LSN.
3946 : let lsn_range = frozen_layer.get_lsn_range();
3947 :
3948 : // Whether to directly create image layers for this flush, or flush them as delta layers
3949 : let create_image_layer =
3950 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3951 :
3952 : #[cfg(test)]
3953 : {
3954 : match &mut *self.flush_loop_state.lock().unwrap() {
3955 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3956 : panic!("flush loop not running")
3957 : }
3958 : FlushLoopState::Running {
3959 : expect_initdb_optimization,
3960 : initdb_optimization_count,
3961 : ..
3962 : } => {
3963 : if create_image_layer {
3964 : *initdb_optimization_count += 1;
3965 : } else {
3966 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3967 : }
3968 : }
3969 : }
3970 : }
3971 :
3972 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3973 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3974 : // require downloading anything during initial import.
3975 : let ((rel_partition, metadata_partition), _lsn) = self
3976 : .repartition(
3977 : self.initdb_lsn,
3978 : self.get_compaction_target_size(),
3979 : EnumSet::empty(),
3980 : ctx,
3981 : )
3982 : .await
3983 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
3984 :
3985 : if self.cancel.is_cancelled() {
3986 : return Err(FlushLayerError::Cancelled);
3987 : }
3988 :
3989 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
3990 : // So that the key ranges don't overlap.
3991 : let mut partitions = KeyPartitioning::default();
3992 : partitions.parts.extend(rel_partition.parts);
3993 : if !metadata_partition.parts.is_empty() {
3994 : assert_eq!(
3995 : metadata_partition.parts.len(),
3996 : 1,
3997 : "currently sparse keyspace should only contain a single metadata keyspace"
3998 : );
3999 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
4000 : // every single key within the keyspace, and therefore, it's safe to force converting it
4001 : // into a dense keyspace before calling this function.
4002 : partitions
4003 : .parts
4004 : .extend(metadata_partition.into_dense().parts);
4005 : }
4006 :
4007 : let mut layers_to_upload = Vec::new();
4008 : layers_to_upload.extend(
4009 : self.create_image_layers(
4010 : &partitions,
4011 : self.initdb_lsn,
4012 : ImageLayerCreationMode::Initial,
4013 : ctx,
4014 : )
4015 : .await?,
4016 : );
4017 :
4018 : (layers_to_upload, None)
4019 : } else {
4020 : // Normal case, write out a L0 delta layer file.
4021 : // `create_delta_layer` will not modify the layer map.
4022 : // We will remove frozen layer and add delta layer in one atomic operation later.
4023 : let Some(layer) = self
4024 : .create_delta_layer(&frozen_layer, None, ctx)
4025 : .await
4026 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
4027 : else {
4028 : panic!("delta layer cannot be empty if no filter is applied");
4029 : };
4030 : (
4031 : // FIXME: even though we have a single image and single delta layer assumption
4032 : // we push them to vec
4033 : vec![layer.clone()],
4034 : Some(layer),
4035 : )
4036 : };
4037 :
4038 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
4039 :
4040 : if self.cancel.is_cancelled() {
4041 : return Err(FlushLayerError::Cancelled);
4042 : }
4043 :
4044 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
4045 :
4046 : // The new on-disk layers are now in the layer map. We can remove the
4047 : // in-memory layer from the map now. The flushed layer is stored in
4048 : // the mapping in `create_delta_layer`.
4049 : {
4050 : let mut guard = self.layers.write().await;
4051 :
4052 : guard.open_mut()?.finish_flush_l0_layer(
4053 : delta_layer_to_add.as_ref(),
4054 : &frozen_layer,
4055 : &self.metrics,
4056 : );
4057 :
4058 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
4059 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
4060 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
4061 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
4062 : }
4063 : // release lock on 'layers'
4064 : };
4065 :
4066 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
4067 : // This makes us refuse ingest until the new layers have been persisted to the remote
4068 : // TODO: remove this, and rely on l0_flush_{delay,stall}_threshold instead.
4069 : if self.get_l0_flush_wait_upload() {
4070 : let start = Instant::now();
4071 : self.remote_client
4072 : .wait_completion()
4073 : .await
4074 0 : .map_err(|e| match e {
4075 : WaitCompletionError::UploadQueueShutDownOrStopped
4076 : | WaitCompletionError::NotInitialized(
4077 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
4078 0 : ) => FlushLayerError::Cancelled,
4079 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
4080 0 : FlushLayerError::Other(anyhow!(e).into())
4081 : }
4082 0 : })?;
4083 : let duration = start.elapsed().as_secs_f64();
4084 : self.metrics.flush_wait_upload_time_gauge_add(duration);
4085 : }
4086 :
4087 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
4088 : // a compaction can delete the file and then it won't be available for uploads any more.
4089 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
4090 : // race situation.
4091 : // See https://github.com/neondatabase/neon/issues/4526
4092 : pausable_failpoint!("flush-frozen-pausable");
4093 :
4094 : // This failpoint is used by another test case `test_pageserver_recovery`.
4095 : fail_point!("flush-frozen-exit");
4096 :
4097 : Ok(Lsn(lsn_range.end.0 - 1))
4098 : }
4099 :
4100 : /// Return true if the value changed
4101 : ///
4102 : /// This function must only be used from the layer flush task.
4103 2344 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
4104 2344 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
4105 2344 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
4106 :
4107 2344 : self.metrics
4108 2344 : .disk_consistent_lsn_gauge
4109 2344 : .set(new_value.0 as i64);
4110 2344 : new_value != old_value
4111 2344 : }
4112 :
4113 : /// Update metadata file
4114 2444 : fn schedule_uploads(
4115 2444 : &self,
4116 2444 : disk_consistent_lsn: Lsn,
4117 2444 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
4118 2444 : ) -> anyhow::Result<()> {
4119 2444 : // We can only save a valid 'prev_record_lsn' value on disk if we
4120 2444 : // flushed *all* in-memory changes to disk. We only track
4121 2444 : // 'prev_record_lsn' in memory for the latest processed record, so we
4122 2444 : // don't remember what the correct value that corresponds to some old
4123 2444 : // LSN is. But if we flush everything, then the value corresponding
4124 2444 : // current 'last_record_lsn' is correct and we can store it on disk.
4125 2444 : let RecordLsn {
4126 2444 : last: last_record_lsn,
4127 2444 : prev: prev_record_lsn,
4128 2444 : } = self.last_record_lsn.load();
4129 2444 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
4130 2188 : Some(prev_record_lsn)
4131 : } else {
4132 256 : None
4133 : };
4134 :
4135 2444 : let update = crate::tenant::metadata::MetadataUpdate::new(
4136 2444 : disk_consistent_lsn,
4137 2444 : ondisk_prev_record_lsn,
4138 2444 : *self.latest_gc_cutoff_lsn.read(),
4139 2444 : );
4140 2444 :
4141 2444 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
4142 0 : "{}",
4143 0 : x.unwrap()
4144 2444 : ));
4145 :
4146 4812 : for layer in layers_to_upload {
4147 2368 : self.remote_client.schedule_layer_file_upload(layer)?;
4148 : }
4149 2444 : self.remote_client
4150 2444 : .schedule_index_upload_for_metadata_update(&update)?;
4151 :
4152 2444 : Ok(())
4153 2444 : }
4154 :
4155 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
4156 0 : self.remote_client
4157 0 : .preserve_initdb_archive(
4158 0 : &self.tenant_shard_id.tenant_id,
4159 0 : &self.timeline_id,
4160 0 : &self.cancel,
4161 0 : )
4162 0 : .await
4163 0 : }
4164 :
4165 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
4166 : // in layer map immediately. The caller is responsible to put it into the layer map.
4167 1936 : async fn create_delta_layer(
4168 1936 : self: &Arc<Self>,
4169 1936 : frozen_layer: &Arc<InMemoryLayer>,
4170 1936 : key_range: Option<Range<Key>>,
4171 1936 : ctx: &RequestContext,
4172 1936 : ) -> anyhow::Result<Option<ResidentLayer>> {
4173 1936 : let self_clone = Arc::clone(self);
4174 1936 : let frozen_layer = Arc::clone(frozen_layer);
4175 1936 : let ctx = ctx.attached_child();
4176 1936 : let work = async move {
4177 1936 : let Some((desc, path)) = frozen_layer
4178 1936 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
4179 1936 : .await?
4180 : else {
4181 0 : return Ok(None);
4182 : };
4183 1936 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
4184 :
4185 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
4186 : // We just need to fsync the directory in which these inodes are linked,
4187 : // which we know to be the timeline directory.
4188 : //
4189 : // We use fatal_err() below because the after write_to_disk returns with success,
4190 : // the in-memory state of the filesystem already has the layer file in its final place,
4191 : // and subsequent pageserver code could think it's durable while it really isn't.
4192 1936 : let timeline_dir = VirtualFile::open(
4193 1936 : &self_clone
4194 1936 : .conf
4195 1936 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
4196 1936 : &ctx,
4197 1936 : )
4198 1936 : .await
4199 1936 : .fatal_err("VirtualFile::open for timeline dir fsync");
4200 1936 : timeline_dir
4201 1936 : .sync_all()
4202 1936 : .await
4203 1936 : .fatal_err("VirtualFile::sync_all timeline dir");
4204 1936 : anyhow::Ok(Some(new_delta))
4205 1936 : };
4206 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
4207 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
4208 : use crate::virtual_file::io_engine::IoEngine;
4209 1936 : match crate::virtual_file::io_engine::get() {
4210 0 : IoEngine::NotSet => panic!("io engine not set"),
4211 : IoEngine::StdFs => {
4212 968 : let span = tracing::info_span!("blocking");
4213 968 : tokio::task::spawn_blocking({
4214 968 : move || Handle::current().block_on(work.instrument(span))
4215 968 : })
4216 968 : .await
4217 968 : .context("spawn_blocking")
4218 968 : .and_then(|x| x)
4219 : }
4220 : #[cfg(target_os = "linux")]
4221 968 : IoEngine::TokioEpollUring => work.await,
4222 : }
4223 1936 : }
4224 :
4225 1136 : async fn repartition(
4226 1136 : &self,
4227 1136 : lsn: Lsn,
4228 1136 : partition_size: u64,
4229 1136 : flags: EnumSet<CompactFlags>,
4230 1136 : ctx: &RequestContext,
4231 1136 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
4232 1136 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
4233 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
4234 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
4235 : // and hence before the compaction task starts.
4236 0 : return Err(CompactionError::Other(anyhow!(
4237 0 : "repartition() called concurrently"
4238 0 : )));
4239 : };
4240 1136 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
4241 1136 : if lsn < *partition_lsn {
4242 0 : return Err(CompactionError::Other(anyhow!(
4243 0 : "repartition() called with LSN going backwards, this should not happen"
4244 0 : )));
4245 1136 : }
4246 1136 :
4247 1136 : let distance = lsn.0 - partition_lsn.0;
4248 1136 : if *partition_lsn != Lsn(0)
4249 524 : && distance <= self.repartition_threshold
4250 524 : && !flags.contains(CompactFlags::ForceRepartition)
4251 : {
4252 496 : debug!(
4253 : distance,
4254 : threshold = self.repartition_threshold,
4255 0 : "no repartitioning needed"
4256 : );
4257 496 : return Ok((
4258 496 : (dense_partition.clone(), sparse_partition.clone()),
4259 496 : *partition_lsn,
4260 496 : ));
4261 640 : }
4262 :
4263 640 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4264 640 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4265 640 : let sparse_partitioning = SparseKeyPartitioning {
4266 640 : parts: vec![sparse_ks],
4267 640 : }; // no partitioning for metadata keys for now
4268 640 : let result = ((dense_partitioning, sparse_partitioning), lsn);
4269 640 : guard.write(result.clone());
4270 640 : Ok(result)
4271 1136 : }
4272 :
4273 : // Is it time to create a new image layer for the given partition?
4274 28 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4275 28 : let threshold = self.get_image_creation_threshold();
4276 :
4277 28 : let guard = self.layers.read().await;
4278 28 : let Ok(layers) = guard.layer_map() else {
4279 0 : return false;
4280 : };
4281 :
4282 28 : let mut max_deltas = 0;
4283 56 : for part_range in &partition.ranges {
4284 28 : let image_coverage = layers.image_coverage(part_range, lsn);
4285 56 : for (img_range, last_img) in image_coverage {
4286 28 : let img_lsn = if let Some(last_img) = last_img {
4287 0 : last_img.get_lsn_range().end
4288 : } else {
4289 28 : Lsn(0)
4290 : };
4291 : // Let's consider an example:
4292 : //
4293 : // delta layer with LSN range 71-81
4294 : // delta layer with LSN range 81-91
4295 : // delta layer with LSN range 91-101
4296 : // image layer at LSN 100
4297 : //
4298 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4299 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4300 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4301 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4302 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4303 28 : if img_lsn < lsn {
4304 28 : let num_deltas =
4305 28 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4306 28 :
4307 28 : max_deltas = max_deltas.max(num_deltas);
4308 28 : if num_deltas >= threshold {
4309 0 : debug!(
4310 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4311 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4312 : );
4313 0 : return true;
4314 28 : }
4315 0 : }
4316 : }
4317 : }
4318 :
4319 28 : debug!(
4320 : max_deltas,
4321 0 : "none of the partitioned ranges had >= {threshold} deltas"
4322 : );
4323 28 : false
4324 28 : }
4325 :
4326 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4327 : /// so that at most one image layer will be produced from this function.
4328 : #[allow(clippy::too_many_arguments)]
4329 460 : async fn create_image_layer_for_rel_blocks(
4330 460 : self: &Arc<Self>,
4331 460 : partition: &KeySpace,
4332 460 : mut image_layer_writer: ImageLayerWriter,
4333 460 : lsn: Lsn,
4334 460 : ctx: &RequestContext,
4335 460 : img_range: Range<Key>,
4336 460 : start: Key,
4337 460 : io_concurrency: IoConcurrency,
4338 460 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4339 460 : let mut wrote_keys = false;
4340 460 :
4341 460 : let mut key_request_accum = KeySpaceAccum::new();
4342 3044 : for range in &partition.ranges {
4343 2584 : let mut key = range.start;
4344 5600 : while key < range.end {
4345 : // Decide whether to retain this key: usually we do, but sharded tenants may
4346 : // need to drop keys that don't belong to them. If we retain the key, add it
4347 : // to `key_request_accum` for later issuing a vectored get
4348 3016 : if self.shard_identity.is_key_disposable(&key) {
4349 0 : debug!(
4350 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4351 0 : key,
4352 0 : self.shard_identity.get_shard_number(&key)
4353 : );
4354 3016 : } else {
4355 3016 : key_request_accum.add_key(key);
4356 3016 : }
4357 :
4358 3016 : let last_key_in_range = key.next() == range.end;
4359 3016 : key = key.next();
4360 3016 :
4361 3016 : // Maybe flush `key_rest_accum`
4362 3016 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4363 3016 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4364 : {
4365 2584 : let results = self
4366 2584 : .get_vectored(
4367 2584 : key_request_accum.consume_keyspace(),
4368 2584 : lsn,
4369 2584 : io_concurrency.clone(),
4370 2584 : ctx,
4371 2584 : )
4372 2584 : .await?;
4373 :
4374 2584 : if self.cancel.is_cancelled() {
4375 0 : return Err(CreateImageLayersError::Cancelled);
4376 2584 : }
4377 :
4378 5600 : for (img_key, img) in results {
4379 3016 : let img = match img {
4380 3016 : Ok(img) => img,
4381 0 : Err(err) => {
4382 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4383 0 : // page without losing any actual user data. That seems better
4384 0 : // than failing repeatedly and getting stuck.
4385 0 : //
4386 0 : // We had a bug at one point, where we truncated the FSM and VM
4387 0 : // in the pageserver, but the Postgres didn't know about that
4388 0 : // and continued to generate incremental WAL records for pages
4389 0 : // that didn't exist in the pageserver. Trying to replay those
4390 0 : // WAL records failed to find the previous image of the page.
4391 0 : // This special case allows us to recover from that situation.
4392 0 : // See https://github.com/neondatabase/neon/issues/2601.
4393 0 : //
4394 0 : // Unfortunately we cannot do this for the main fork, or for
4395 0 : // any metadata keys, keys, as that would lead to actual data
4396 0 : // loss.
4397 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4398 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4399 0 : ZERO_PAGE.clone()
4400 : } else {
4401 0 : return Err(CreateImageLayersError::from(err));
4402 : }
4403 : }
4404 : };
4405 :
4406 : // Write all the keys we just read into our new image layer.
4407 3016 : image_layer_writer.put_image(img_key, img, ctx).await?;
4408 3016 : wrote_keys = true;
4409 : }
4410 432 : }
4411 : }
4412 : }
4413 :
4414 460 : if wrote_keys {
4415 : // Normal path: we have written some data into the new image layer for this
4416 : // partition, so flush it to disk.
4417 460 : info!(
4418 0 : "produced image layer for rel {}",
4419 0 : ImageLayerName {
4420 0 : key_range: img_range.clone(),
4421 0 : lsn
4422 0 : },
4423 : );
4424 460 : Ok(ImageLayerCreationOutcome {
4425 460 : unfinished_image_layer: Some(image_layer_writer),
4426 460 : next_start_key: img_range.end,
4427 460 : })
4428 : } else {
4429 : // Special case: the image layer may be empty if this is a sharded tenant and the
4430 : // partition does not cover any keys owned by this shard. In this case, to ensure
4431 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4432 : // layer we write will cover the key range that we just scanned.
4433 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4434 0 : Ok(ImageLayerCreationOutcome {
4435 0 : unfinished_image_layer: None,
4436 0 : next_start_key: start,
4437 0 : })
4438 : }
4439 460 : }
4440 :
4441 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4442 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4443 : /// would not be too large to fit in a single image layer.
4444 : #[allow(clippy::too_many_arguments)]
4445 440 : async fn create_image_layer_for_metadata_keys(
4446 440 : self: &Arc<Self>,
4447 440 : partition: &KeySpace,
4448 440 : mut image_layer_writer: ImageLayerWriter,
4449 440 : lsn: Lsn,
4450 440 : ctx: &RequestContext,
4451 440 : img_range: Range<Key>,
4452 440 : mode: ImageLayerCreationMode,
4453 440 : start: Key,
4454 440 : io_concurrency: IoConcurrency,
4455 440 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4456 440 : // Metadata keys image layer creation.
4457 440 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
4458 440 : let begin = Instant::now();
4459 440 : let data = self
4460 440 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4461 440 : .await?;
4462 440 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4463 440 : let mut new_data = BTreeMap::new();
4464 440 : let mut total_kb_retrieved = 0;
4465 440 : let mut total_keys_retrieved = 0;
4466 20464 : for (k, v) in data {
4467 20024 : let v = v?;
4468 20024 : total_kb_retrieved += KEY_SIZE + v.len();
4469 20024 : total_keys_retrieved += 1;
4470 20024 : new_data.insert(k, v);
4471 : }
4472 440 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4473 440 : };
4474 440 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4475 440 : let elapsed = begin.elapsed();
4476 440 :
4477 440 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4478 440 : info!(
4479 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()
4480 : );
4481 :
4482 440 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4483 4 : return Ok(ImageLayerCreationOutcome {
4484 4 : unfinished_image_layer: None,
4485 4 : next_start_key: img_range.end,
4486 4 : });
4487 436 : }
4488 436 : if self.cancel.is_cancelled() {
4489 0 : return Err(CreateImageLayersError::Cancelled);
4490 436 : }
4491 436 : let mut wrote_any_image = false;
4492 20460 : for (k, v) in data {
4493 20024 : if v.is_empty() {
4494 : // the key has been deleted, it does not need an image
4495 : // in metadata keyspace, an empty image == tombstone
4496 16 : continue;
4497 20008 : }
4498 20008 : wrote_any_image = true;
4499 20008 :
4500 20008 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4501 20008 :
4502 20008 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4503 20008 : // on the normal data path either.
4504 20008 : image_layer_writer.put_image(k, v, ctx).await?;
4505 : }
4506 :
4507 436 : if wrote_any_image {
4508 : // Normal path: we have written some data into the new image layer for this
4509 : // partition, so flush it to disk.
4510 24 : info!(
4511 0 : "created image layer for metadata {}",
4512 0 : ImageLayerName {
4513 0 : key_range: img_range.clone(),
4514 0 : lsn
4515 0 : }
4516 : );
4517 24 : Ok(ImageLayerCreationOutcome {
4518 24 : unfinished_image_layer: Some(image_layer_writer),
4519 24 : next_start_key: img_range.end,
4520 24 : })
4521 : } else {
4522 : // Special case: the image layer may be empty if this is a sharded tenant and the
4523 : // partition does not cover any keys owned by this shard. In this case, to ensure
4524 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4525 : // layer we write will cover the key range that we just scanned.
4526 412 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4527 412 : Ok(ImageLayerCreationOutcome {
4528 412 : unfinished_image_layer: None,
4529 412 : next_start_key: start,
4530 412 : })
4531 : }
4532 440 : }
4533 :
4534 : /// Predicate function which indicates whether we should check if new image layers
4535 : /// are required. Since checking if new image layers are required is expensive in
4536 : /// terms of CPU, we only do it in the following cases:
4537 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4538 : /// 2. If enough time has passed since the last check:
4539 : /// 1. For large tenants, we wish to perform the check more often since they
4540 : /// suffer from the lack of image layers
4541 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4542 1136 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4543 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4544 :
4545 1136 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4546 1136 : let distance = lsn
4547 1136 : .checked_sub(last_checks_at)
4548 1136 : .expect("Attempt to compact with LSN going backwards");
4549 1136 : let min_distance =
4550 1136 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4551 1136 :
4552 1136 : let distance_based_decision = distance.0 >= min_distance;
4553 1136 :
4554 1136 : let mut time_based_decision = false;
4555 1136 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4556 1136 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4557 932 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4558 : {
4559 0 : self.get_checkpoint_timeout()
4560 : } else {
4561 932 : Duration::from_secs(3600 * 48)
4562 : };
4563 :
4564 932 : time_based_decision = match *last_check_instant {
4565 524 : Some(last_check) => {
4566 524 : let elapsed = last_check.elapsed();
4567 524 : elapsed >= check_required_after
4568 : }
4569 408 : None => true,
4570 : };
4571 204 : }
4572 :
4573 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4574 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4575 : // check.
4576 1136 : let decision = distance_based_decision || time_based_decision;
4577 :
4578 1136 : if decision {
4579 412 : self.last_image_layer_creation_check_at.store(lsn);
4580 412 : *last_check_instant = Some(Instant::now());
4581 724 : }
4582 :
4583 1136 : decision
4584 1136 : }
4585 :
4586 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4587 : async fn create_image_layers(
4588 : self: &Arc<Timeline>,
4589 : partitioning: &KeyPartitioning,
4590 : lsn: Lsn,
4591 : mode: ImageLayerCreationMode,
4592 : ctx: &RequestContext,
4593 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4594 : let timer = self.metrics.create_images_time_histo.start_timer();
4595 :
4596 : // We need to avoid holes between generated image layers.
4597 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4598 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4599 : //
4600 : // How such hole between partitions can appear?
4601 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4602 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4603 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4604 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4605 : let mut start = Key::MIN;
4606 :
4607 : let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
4608 :
4609 : let mut batch_image_writer = BatchLayerWriter::new(self.conf).await?;
4610 :
4611 : for partition in partitioning.parts.iter() {
4612 : if self.cancel.is_cancelled() {
4613 : return Err(CreateImageLayersError::Cancelled);
4614 : }
4615 :
4616 : let img_range = start..partition.ranges.last().unwrap().end;
4617 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4618 : if compact_metadata {
4619 : for range in &partition.ranges {
4620 : assert!(
4621 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4622 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4623 : "metadata keys must be partitioned separately"
4624 : );
4625 : }
4626 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4627 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4628 : // might mess up with evictions.
4629 : start = img_range.end;
4630 : continue;
4631 : }
4632 : // For initial and force modes, we always generate image layers for metadata keys.
4633 : } else if let ImageLayerCreationMode::Try = mode {
4634 : // check_for_image_layers = false -> skip
4635 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4636 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4637 : start = img_range.end;
4638 : continue;
4639 : }
4640 : }
4641 : if let ImageLayerCreationMode::Force = mode {
4642 : // When forced to create image layers, we might try and create them where they already
4643 : // exist. This mode is only used in tests/debug.
4644 : let layers = self.layers.read().await;
4645 : if layers.contains_key(&PersistentLayerKey {
4646 : key_range: img_range.clone(),
4647 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4648 : is_delta: false,
4649 : }) {
4650 : tracing::info!(
4651 : "Skipping image layer at {lsn} {}..{}, already exists",
4652 : img_range.start,
4653 : img_range.end
4654 : );
4655 : start = img_range.end;
4656 : continue;
4657 : }
4658 : }
4659 :
4660 : let image_layer_writer = ImageLayerWriter::new(
4661 : self.conf,
4662 : self.timeline_id,
4663 : self.tenant_shard_id,
4664 : &img_range,
4665 : lsn,
4666 : ctx,
4667 : )
4668 : .await?;
4669 :
4670 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4671 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4672 0 : "failpoint image-layer-writer-fail-before-finish"
4673 0 : )))
4674 0 : });
4675 :
4676 : let io_concurrency = IoConcurrency::spawn_from_conf(
4677 : self.conf,
4678 : self.gate
4679 : .enter()
4680 0 : .map_err(|_| CreateImageLayersError::Cancelled)?,
4681 : );
4682 :
4683 : let ImageLayerCreationOutcome {
4684 : unfinished_image_layer,
4685 : next_start_key,
4686 : } = if !compact_metadata {
4687 : self.create_image_layer_for_rel_blocks(
4688 : partition,
4689 : image_layer_writer,
4690 : lsn,
4691 : ctx,
4692 : img_range.clone(),
4693 : start,
4694 : io_concurrency,
4695 : )
4696 : .await?
4697 : } else {
4698 : self.create_image_layer_for_metadata_keys(
4699 : partition,
4700 : image_layer_writer,
4701 : lsn,
4702 : ctx,
4703 : img_range.clone(),
4704 : mode,
4705 : start,
4706 : io_concurrency,
4707 : )
4708 : .await?
4709 : };
4710 : start = next_start_key;
4711 : if let Some(unfinished_image_layer) = unfinished_image_layer {
4712 : batch_image_writer.add_unfinished_image_writer(
4713 : unfinished_image_layer,
4714 : img_range,
4715 : lsn,
4716 : );
4717 : }
4718 : }
4719 :
4720 : let image_layers = batch_image_writer.finish(self, ctx).await?;
4721 :
4722 : let mut guard = self.layers.write().await;
4723 :
4724 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4725 : // now they are being scheduled outside of write lock; current way is inconsistent with
4726 : // compaction lock order.
4727 : guard
4728 : .open_mut()?
4729 : .track_new_image_layers(&image_layers, &self.metrics);
4730 : drop_wlock(guard);
4731 : timer.stop_and_record();
4732 :
4733 : // Creating image layers may have caused some previously visible layers to be covered
4734 : if !image_layers.is_empty() {
4735 : self.update_layer_visibility().await?;
4736 : }
4737 :
4738 : Ok(image_layers)
4739 : }
4740 :
4741 : /// Wait until the background initial logical size calculation is complete, or
4742 : /// this Timeline is shut down. Calling this function will cause the initial
4743 : /// logical size calculation to skip waiting for the background jobs barrier.
4744 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4745 0 : if !self.shard_identity.is_shard_zero() {
4746 : // We don't populate logical size on shard >0: skip waiting for it.
4747 0 : return;
4748 0 : }
4749 0 :
4750 0 : if self.remote_client.is_deleting() {
4751 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4752 0 : return;
4753 0 : }
4754 0 :
4755 0 : if self.current_logical_size.current_size().is_exact() {
4756 : // root timelines are initialized with exact count, but never start the background
4757 : // calculation
4758 0 : return;
4759 0 : }
4760 :
4761 0 : if let Some(await_bg_cancel) = self
4762 0 : .current_logical_size
4763 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4764 0 : .get()
4765 0 : {
4766 0 : await_bg_cancel.cancel();
4767 0 : } else {
4768 : // We should not wait if we were not able to explicitly instruct
4769 : // the logical size cancellation to skip the concurrency limit semaphore.
4770 : // TODO: this is an unexpected case. We should restructure so that it
4771 : // can't happen.
4772 0 : tracing::warn!(
4773 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4774 : );
4775 0 : debug_assert!(false);
4776 : }
4777 :
4778 0 : tokio::select!(
4779 0 : _ = self.current_logical_size.initialized.acquire() => {},
4780 0 : _ = self.cancel.cancelled() => {}
4781 : )
4782 0 : }
4783 :
4784 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4785 : /// Timelines layers up to the ancestor_lsn.
4786 : ///
4787 : /// Requires a timeline that:
4788 : /// - has an ancestor to detach from
4789 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4790 : /// a technical requirement
4791 : ///
4792 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4793 : /// polled again until completion.
4794 : ///
4795 : /// During the operation all timelines sharing the data with this timeline will be reparented
4796 : /// from our ancestor to be branches of this timeline.
4797 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4798 0 : self: &Arc<Timeline>,
4799 0 : tenant: &crate::tenant::Tenant,
4800 0 : options: detach_ancestor::Options,
4801 0 : ctx: &RequestContext,
4802 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
4803 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4804 0 : }
4805 :
4806 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
4807 : /// reparents any reparentable children of previous ancestor.
4808 : ///
4809 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
4810 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
4811 : /// successfully, tenant must be reloaded.
4812 : ///
4813 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
4814 : /// resetting the tenant.
4815 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
4816 0 : self: &Arc<Timeline>,
4817 0 : tenant: &crate::tenant::Tenant,
4818 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4819 0 : ctx: &RequestContext,
4820 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
4821 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
4822 0 : }
4823 :
4824 : /// Final step which unblocks the GC.
4825 : ///
4826 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
4827 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4828 0 : self: &Arc<Timeline>,
4829 0 : tenant: &crate::tenant::Tenant,
4830 0 : attempt: detach_ancestor::Attempt,
4831 0 : ctx: &RequestContext,
4832 0 : ) -> Result<(), detach_ancestor::Error> {
4833 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
4834 0 : }
4835 : }
4836 :
4837 : impl Drop for Timeline {
4838 20 : fn drop(&mut self) {
4839 20 : if let Some(ancestor) = &self.ancestor_timeline {
4840 : // This lock should never be poisoned, but in case it is we do a .map() instead of
4841 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
4842 8 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
4843 8 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
4844 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
4845 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
4846 8 : }
4847 0 : }
4848 12 : }
4849 20 : info!(
4850 0 : "Timeline {} for tenant {} is being dropped",
4851 : self.timeline_id, self.tenant_shard_id.tenant_id
4852 : );
4853 20 : }
4854 : }
4855 :
4856 : /// Top-level failure to compact.
4857 : #[derive(Debug, thiserror::Error)]
4858 : pub(crate) enum CompactionError {
4859 : #[error("The timeline or pageserver is shutting down")]
4860 : ShuttingDown,
4861 : /// Compaction tried to offload a timeline and failed
4862 : #[error("Failed to offload timeline: {0}")]
4863 : Offload(OffloadError),
4864 : /// Compaction cannot be done right now; page reconstruction and so on.
4865 : #[error(transparent)]
4866 : Other(anyhow::Error),
4867 : }
4868 :
4869 : impl From<OffloadError> for CompactionError {
4870 0 : fn from(e: OffloadError) -> Self {
4871 0 : match e {
4872 0 : OffloadError::Cancelled => Self::ShuttingDown,
4873 0 : _ => Self::Offload(e),
4874 : }
4875 0 : }
4876 : }
4877 :
4878 : impl CompactionError {
4879 0 : pub fn is_cancelled(&self) -> bool {
4880 0 : matches!(self, CompactionError::ShuttingDown)
4881 0 : }
4882 : }
4883 :
4884 : impl From<CollectKeySpaceError> for CompactionError {
4885 0 : fn from(err: CollectKeySpaceError) -> Self {
4886 0 : match err {
4887 : CollectKeySpaceError::Cancelled
4888 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4889 0 : CompactionError::ShuttingDown
4890 : }
4891 0 : e => CompactionError::Other(e.into()),
4892 : }
4893 0 : }
4894 : }
4895 :
4896 : impl From<super::upload_queue::NotInitialized> for CompactionError {
4897 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
4898 0 : match value {
4899 : super::upload_queue::NotInitialized::Uninitialized => {
4900 0 : CompactionError::Other(anyhow::anyhow!(value))
4901 : }
4902 : super::upload_queue::NotInitialized::ShuttingDown
4903 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
4904 : }
4905 0 : }
4906 : }
4907 :
4908 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
4909 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
4910 0 : match e {
4911 : super::storage_layer::layer::DownloadError::TimelineShutdown
4912 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
4913 0 : CompactionError::ShuttingDown
4914 : }
4915 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
4916 : | super::storage_layer::layer::DownloadError::DownloadRequired
4917 : | super::storage_layer::layer::DownloadError::NotFile(_)
4918 : | super::storage_layer::layer::DownloadError::DownloadFailed
4919 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
4920 0 : CompactionError::Other(anyhow::anyhow!(e))
4921 : }
4922 : #[cfg(test)]
4923 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
4924 0 : CompactionError::Other(anyhow::anyhow!(e))
4925 : }
4926 : }
4927 0 : }
4928 : }
4929 :
4930 : impl From<layer_manager::Shutdown> for CompactionError {
4931 0 : fn from(_: layer_manager::Shutdown) -> Self {
4932 0 : CompactionError::ShuttingDown
4933 0 : }
4934 : }
4935 :
4936 : #[serde_as]
4937 392 : #[derive(serde::Serialize)]
4938 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4939 :
4940 : #[derive(Default)]
4941 : enum DurationRecorder {
4942 : #[default]
4943 : NotStarted,
4944 : Recorded(RecordedDuration, tokio::time::Instant),
4945 : }
4946 :
4947 : impl DurationRecorder {
4948 1008 : fn till_now(&self) -> DurationRecorder {
4949 1008 : match self {
4950 : DurationRecorder::NotStarted => {
4951 0 : panic!("must only call on recorded measurements")
4952 : }
4953 1008 : DurationRecorder::Recorded(_, ended) => {
4954 1008 : let now = tokio::time::Instant::now();
4955 1008 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4956 1008 : }
4957 1008 : }
4958 1008 : }
4959 392 : fn into_recorded(self) -> Option<RecordedDuration> {
4960 392 : match self {
4961 0 : DurationRecorder::NotStarted => None,
4962 392 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4963 : }
4964 392 : }
4965 : }
4966 :
4967 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
4968 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
4969 : /// the layer descriptor requires the user to provide the ranges, which should cover all
4970 : /// keys specified in the `data` field.
4971 : #[cfg(test)]
4972 : #[derive(Clone)]
4973 : pub struct DeltaLayerTestDesc {
4974 : pub lsn_range: Range<Lsn>,
4975 : pub key_range: Range<Key>,
4976 : pub data: Vec<(Key, Lsn, Value)>,
4977 : }
4978 :
4979 : #[cfg(test)]
4980 : impl DeltaLayerTestDesc {
4981 4 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
4982 4 : Self {
4983 4 : lsn_range,
4984 4 : key_range,
4985 4 : data,
4986 4 : }
4987 4 : }
4988 :
4989 176 : pub fn new_with_inferred_key_range(
4990 176 : lsn_range: Range<Lsn>,
4991 176 : data: Vec<(Key, Lsn, Value)>,
4992 176 : ) -> Self {
4993 440 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
4994 440 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
4995 176 : Self {
4996 176 : key_range: (*key_min)..(key_max.next()),
4997 176 : lsn_range,
4998 176 : data,
4999 176 : }
5000 176 : }
5001 :
5002 20 : pub(crate) fn layer_name(&self) -> LayerName {
5003 20 : LayerName::Delta(super::storage_layer::DeltaLayerName {
5004 20 : key_range: self.key_range.clone(),
5005 20 : lsn_range: self.lsn_range.clone(),
5006 20 : })
5007 20 : }
5008 : }
5009 :
5010 : impl Timeline {
5011 56 : async fn finish_compact_batch(
5012 56 : self: &Arc<Self>,
5013 56 : new_deltas: &[ResidentLayer],
5014 56 : new_images: &[ResidentLayer],
5015 56 : layers_to_remove: &[Layer],
5016 56 : ) -> Result<(), CompactionError> {
5017 56 : let mut guard = tokio::select! {
5018 56 : guard = self.layers.write() => guard,
5019 56 : _ = self.cancel.cancelled() => {
5020 0 : return Err(CompactionError::ShuttingDown);
5021 : }
5022 : };
5023 :
5024 56 : let mut duplicated_layers = HashSet::new();
5025 56 :
5026 56 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
5027 :
5028 672 : for l in new_deltas {
5029 616 : if guard.contains(l.as_ref()) {
5030 : // expected in tests
5031 0 : tracing::error!(layer=%l, "duplicated L1 layer");
5032 :
5033 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
5034 : // `guard` on self.layers. as of writing this, there are no error returns except
5035 : // for compact_level0_phase1 creating an L0, which does not happen in practice
5036 : // because we have not implemented L0 => L0 compaction.
5037 0 : duplicated_layers.insert(l.layer_desc().key());
5038 616 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
5039 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
5040 616 : } else {
5041 616 : insert_layers.push(l.clone());
5042 616 : }
5043 : }
5044 :
5045 : // only remove those inputs which were not outputs
5046 56 : let remove_layers: Vec<Layer> = layers_to_remove
5047 56 : .iter()
5048 804 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
5049 56 : .cloned()
5050 56 : .collect();
5051 56 :
5052 56 : if !new_images.is_empty() {
5053 0 : guard
5054 0 : .open_mut()?
5055 0 : .track_new_image_layers(new_images, &self.metrics);
5056 56 : }
5057 :
5058 56 : guard
5059 56 : .open_mut()?
5060 56 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
5061 56 :
5062 56 : self.remote_client
5063 56 : .schedule_compaction_update(&remove_layers, new_deltas)?;
5064 :
5065 56 : drop_wlock(guard);
5066 56 :
5067 56 : Ok(())
5068 56 : }
5069 :
5070 0 : async fn rewrite_layers(
5071 0 : self: &Arc<Self>,
5072 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
5073 0 : mut drop_layers: Vec<Layer>,
5074 0 : ) -> Result<(), CompactionError> {
5075 0 : let mut guard = self.layers.write().await;
5076 :
5077 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
5078 : // to avoid double-removing, and avoid rewriting something that was removed.
5079 0 : replace_layers.retain(|(l, _)| guard.contains(l));
5080 0 : drop_layers.retain(|l| guard.contains(l));
5081 0 :
5082 0 : guard
5083 0 : .open_mut()?
5084 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
5085 0 :
5086 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
5087 0 :
5088 0 : self.remote_client
5089 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
5090 :
5091 0 : Ok(())
5092 0 : }
5093 :
5094 : /// Schedules the uploads of the given image layers
5095 728 : fn upload_new_image_layers(
5096 728 : self: &Arc<Self>,
5097 728 : new_images: impl IntoIterator<Item = ResidentLayer>,
5098 728 : ) -> Result<(), super::upload_queue::NotInitialized> {
5099 780 : for layer in new_images {
5100 52 : self.remote_client.schedule_layer_file_upload(layer)?;
5101 : }
5102 : // should any new image layer been created, not uploading index_part will
5103 : // result in a mismatch between remote_physical_size and layermap calculated
5104 : // size, which will fail some tests, but should not be an issue otherwise.
5105 728 : self.remote_client
5106 728 : .schedule_index_upload_for_file_changes()?;
5107 728 : Ok(())
5108 728 : }
5109 :
5110 0 : async fn find_gc_time_cutoff(
5111 0 : &self,
5112 0 : now: SystemTime,
5113 0 : pitr: Duration,
5114 0 : cancel: &CancellationToken,
5115 0 : ctx: &RequestContext,
5116 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
5117 0 : debug_assert_current_span_has_tenant_and_timeline_id();
5118 0 : if self.shard_identity.is_shard_zero() {
5119 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
5120 0 : let time_range = if pitr == Duration::ZERO {
5121 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
5122 : } else {
5123 0 : pitr
5124 : };
5125 :
5126 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
5127 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
5128 0 : let timestamp = to_pg_timestamp(time_cutoff);
5129 :
5130 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
5131 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
5132 0 : LsnForTimestamp::Future(lsn) => {
5133 0 : // The timestamp is in the future. That sounds impossible,
5134 0 : // but what it really means is that there hasn't been
5135 0 : // any commits since the cutoff timestamp.
5136 0 : //
5137 0 : // In this case we should use the LSN of the most recent commit,
5138 0 : // which is implicitly the last LSN in the log.
5139 0 : debug!("future({})", lsn);
5140 0 : Some(self.get_last_record_lsn())
5141 : }
5142 0 : LsnForTimestamp::Past(lsn) => {
5143 0 : debug!("past({})", lsn);
5144 0 : None
5145 : }
5146 0 : LsnForTimestamp::NoData(lsn) => {
5147 0 : debug!("nodata({})", lsn);
5148 0 : None
5149 : }
5150 : };
5151 0 : Ok(time_cutoff)
5152 : } else {
5153 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
5154 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
5155 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
5156 : // space cutoff that we would also have respected ourselves.
5157 0 : match self
5158 0 : .remote_client
5159 0 : .download_foreign_index(ShardNumber(0), cancel)
5160 0 : .await
5161 : {
5162 0 : Ok((index_part, index_generation, _index_mtime)) => {
5163 0 : tracing::info!("GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
5164 0 : index_part.metadata.latest_gc_cutoff_lsn());
5165 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
5166 : }
5167 : Err(DownloadError::NotFound) => {
5168 : // This is unexpected, because during timeline creations shard zero persists to remote
5169 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
5170 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
5171 : // state, then shard zero should _eventually_ write an index when it GCs.
5172 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
5173 0 : Ok(None)
5174 : }
5175 0 : Err(e) => {
5176 0 : // TODO: this function should return a different error type than page reconstruct error
5177 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
5178 : }
5179 : }
5180 :
5181 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
5182 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
5183 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
5184 : // new location. This is legal in principle, but problematic in practice because it might result
5185 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
5186 : // because they have GC'd past the branch point.
5187 : }
5188 0 : }
5189 :
5190 : /// Find the Lsns above which layer files need to be retained on
5191 : /// garbage collection.
5192 : ///
5193 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
5194 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
5195 : /// the space-based retention.
5196 : ///
5197 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
5198 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
5199 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
5200 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
5201 : /// in the response as the GC cutoff point for the timeline.
5202 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
5203 : pub(super) async fn find_gc_cutoffs(
5204 : &self,
5205 : now: SystemTime,
5206 : space_cutoff: Lsn,
5207 : pitr: Duration,
5208 : cancel: &CancellationToken,
5209 : ctx: &RequestContext,
5210 : ) -> Result<GcCutoffs, PageReconstructError> {
5211 : let _timer = self
5212 : .metrics
5213 : .find_gc_cutoffs_histo
5214 : .start_timer()
5215 : .record_on_drop();
5216 :
5217 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
5218 :
5219 : if cfg!(test) {
5220 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
5221 : if pitr == Duration::ZERO {
5222 : return Ok(GcCutoffs {
5223 : time: self.get_last_record_lsn(),
5224 : space: space_cutoff,
5225 : });
5226 : }
5227 : }
5228 :
5229 : // Calculate a time-based limit on how much to retain:
5230 : // - if PITR interval is set, then this is our cutoff.
5231 : // - if PITR interval is not set, then we do a lookup
5232 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
5233 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
5234 :
5235 : Ok(match (pitr, time_cutoff) {
5236 : (Duration::ZERO, Some(time_cutoff)) => {
5237 : // PITR is not set. Retain the size-based limit, or the default time retention,
5238 : // whichever requires less data.
5239 : GcCutoffs {
5240 : time: self.get_last_record_lsn(),
5241 : space: std::cmp::max(time_cutoff, space_cutoff),
5242 : }
5243 : }
5244 : (Duration::ZERO, None) => {
5245 : // PITR is not set, and time lookup failed
5246 : GcCutoffs {
5247 : time: self.get_last_record_lsn(),
5248 : space: space_cutoff,
5249 : }
5250 : }
5251 : (_, None) => {
5252 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
5253 : // cannot advance beyond what was already GC'd, and respect space-based retention
5254 : GcCutoffs {
5255 : time: *self.get_latest_gc_cutoff_lsn(),
5256 : space: space_cutoff,
5257 : }
5258 : }
5259 : (_, Some(time_cutoff)) => {
5260 : // PITR interval is set and we looked up timestamp successfully. Ignore
5261 : // size based retention and make time cutoff authoritative
5262 : GcCutoffs {
5263 : time: time_cutoff,
5264 : space: time_cutoff,
5265 : }
5266 : }
5267 : })
5268 : }
5269 :
5270 : /// Garbage collect layer files on a timeline that are no longer needed.
5271 : ///
5272 : /// Currently, we don't make any attempt at removing unneeded page versions
5273 : /// within a layer file. We can only remove the whole file if it's fully
5274 : /// obsolete.
5275 8 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
5276 : // this is most likely the background tasks, but it might be the spawned task from
5277 : // immediate_gc
5278 8 : let _g = tokio::select! {
5279 8 : guard = self.gc_lock.lock() => guard,
5280 8 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
5281 : };
5282 8 : let timer = self.metrics.garbage_collect_histo.start_timer();
5283 8 :
5284 8 : fail_point!("before-timeline-gc");
5285 8 :
5286 8 : // Is the timeline being deleted?
5287 8 : if self.is_stopping() {
5288 0 : return Err(GcError::TimelineCancelled);
5289 8 : }
5290 8 :
5291 8 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
5292 8 : let gc_info = self.gc_info.read().unwrap();
5293 8 :
5294 8 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
5295 8 : let time_cutoff = gc_info.cutoffs.time;
5296 8 : let retain_lsns = gc_info
5297 8 : .retain_lsns
5298 8 : .iter()
5299 8 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
5300 8 : .collect();
5301 8 :
5302 8 : // Gets the maximum LSN that holds the valid lease.
5303 8 : //
5304 8 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
5305 8 : // Here, we do not check for stale leases again.
5306 8 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
5307 8 :
5308 8 : (
5309 8 : space_cutoff,
5310 8 : time_cutoff,
5311 8 : retain_lsns,
5312 8 : max_lsn_with_valid_lease,
5313 8 : )
5314 8 : };
5315 8 :
5316 8 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
5317 8 : let standby_horizon = self.standby_horizon.load();
5318 8 : // Hold GC for the standby, but as a safety guard do it only within some
5319 8 : // reasonable lag.
5320 8 : if standby_horizon != Lsn::INVALID {
5321 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
5322 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
5323 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
5324 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
5325 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
5326 : } else {
5327 0 : warn!(
5328 0 : "standby is lagging for more than {}MB, not holding gc for it",
5329 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
5330 : )
5331 : }
5332 0 : }
5333 8 : }
5334 :
5335 : // Reset standby horizon to ignore it if it is not updated till next GC.
5336 : // It is an easy way to unset it when standby disappears without adding
5337 : // more conf options.
5338 8 : self.standby_horizon.store(Lsn::INVALID);
5339 8 : self.metrics
5340 8 : .standby_horizon_gauge
5341 8 : .set(Lsn::INVALID.0 as i64);
5342 :
5343 8 : let res = self
5344 8 : .gc_timeline(
5345 8 : space_cutoff,
5346 8 : time_cutoff,
5347 8 : retain_lsns,
5348 8 : max_lsn_with_valid_lease,
5349 8 : new_gc_cutoff,
5350 8 : )
5351 8 : .instrument(
5352 8 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5353 : )
5354 8 : .await?;
5355 :
5356 : // only record successes
5357 8 : timer.stop_and_record();
5358 8 :
5359 8 : Ok(res)
5360 8 : }
5361 :
5362 8 : async fn gc_timeline(
5363 8 : &self,
5364 8 : space_cutoff: Lsn,
5365 8 : time_cutoff: Lsn,
5366 8 : retain_lsns: Vec<Lsn>,
5367 8 : max_lsn_with_valid_lease: Option<Lsn>,
5368 8 : new_gc_cutoff: Lsn,
5369 8 : ) -> Result<GcResult, GcError> {
5370 8 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5371 8 :
5372 8 : let now = SystemTime::now();
5373 8 : let mut result: GcResult = GcResult::default();
5374 8 :
5375 8 : // Nothing to GC. Return early.
5376 8 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
5377 8 : if latest_gc_cutoff >= new_gc_cutoff {
5378 0 : info!(
5379 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5380 : );
5381 0 : return Ok(result);
5382 8 : }
5383 :
5384 : // We need to ensure that no one tries to read page versions or create
5385 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5386 : // for details. This will block until the old value is no longer in use.
5387 : //
5388 : // The GC cutoff should only ever move forwards.
5389 8 : let waitlist = {
5390 8 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
5391 8 : if *write_guard > new_gc_cutoff {
5392 0 : return Err(GcError::BadLsn {
5393 0 : why: format!(
5394 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5395 0 : *write_guard, new_gc_cutoff
5396 0 : ),
5397 0 : });
5398 8 : }
5399 8 :
5400 8 : write_guard.store_and_unlock(new_gc_cutoff)
5401 8 : };
5402 8 : waitlist.wait().await;
5403 :
5404 8 : info!("GC starting");
5405 :
5406 8 : debug!("retain_lsns: {:?}", retain_lsns);
5407 :
5408 8 : let mut layers_to_remove = Vec::new();
5409 :
5410 : // Scan all layers in the timeline (remote or on-disk).
5411 : //
5412 : // Garbage collect the layer if all conditions are satisfied:
5413 : // 1. it is older than cutoff LSN;
5414 : // 2. it is older than PITR interval;
5415 : // 3. it doesn't need to be retained for 'retain_lsns';
5416 : // 4. it does not need to be kept for LSNs holding valid leases.
5417 : // 5. newer on-disk image layers cover the layer's whole key range
5418 : //
5419 : // TODO holding a write lock is too agressive and avoidable
5420 8 : let mut guard = self.layers.write().await;
5421 8 : let layers = guard.layer_map()?;
5422 48 : 'outer: for l in layers.iter_historic_layers() {
5423 48 : result.layers_total += 1;
5424 48 :
5425 48 : // 1. Is it newer than GC horizon cutoff point?
5426 48 : if l.get_lsn_range().end > space_cutoff {
5427 4 : info!(
5428 0 : "keeping {} because it's newer than space_cutoff {}",
5429 0 : l.layer_name(),
5430 : space_cutoff,
5431 : );
5432 4 : result.layers_needed_by_cutoff += 1;
5433 4 : continue 'outer;
5434 44 : }
5435 44 :
5436 44 : // 2. It is newer than PiTR cutoff point?
5437 44 : if l.get_lsn_range().end > time_cutoff {
5438 0 : info!(
5439 0 : "keeping {} because it's newer than time_cutoff {}",
5440 0 : l.layer_name(),
5441 : time_cutoff,
5442 : );
5443 0 : result.layers_needed_by_pitr += 1;
5444 0 : continue 'outer;
5445 44 : }
5446 :
5447 : // 3. Is it needed by a child branch?
5448 : // NOTE With that we would keep data that
5449 : // might be referenced by child branches forever.
5450 : // We can track this in child timeline GC and delete parent layers when
5451 : // they are no longer needed. This might be complicated with long inheritance chains.
5452 : //
5453 : // TODO Vec is not a great choice for `retain_lsns`
5454 44 : for retain_lsn in &retain_lsns {
5455 : // start_lsn is inclusive
5456 0 : if &l.get_lsn_range().start <= retain_lsn {
5457 0 : info!(
5458 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5459 0 : l.layer_name(),
5460 0 : retain_lsn,
5461 0 : l.is_incremental(),
5462 : );
5463 0 : result.layers_needed_by_branches += 1;
5464 0 : continue 'outer;
5465 0 : }
5466 : }
5467 :
5468 : // 4. Is there a valid lease that requires us to keep this layer?
5469 44 : if let Some(lsn) = &max_lsn_with_valid_lease {
5470 : // keep if layer start <= any of the lease
5471 36 : if &l.get_lsn_range().start <= lsn {
5472 28 : info!(
5473 0 : "keeping {} because there is a valid lease preventing GC at {}",
5474 0 : l.layer_name(),
5475 : lsn,
5476 : );
5477 28 : result.layers_needed_by_leases += 1;
5478 28 : continue 'outer;
5479 8 : }
5480 8 : }
5481 :
5482 : // 5. Is there a later on-disk layer for this relation?
5483 : //
5484 : // The end-LSN is exclusive, while disk_consistent_lsn is
5485 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5486 : // OK for a delta layer to have end LSN 101, but if the end LSN
5487 : // is 102, then it might not have been fully flushed to disk
5488 : // before crash.
5489 : //
5490 : // For example, imagine that the following layers exist:
5491 : //
5492 : // 1000 - image (A)
5493 : // 1000-2000 - delta (B)
5494 : // 2000 - image (C)
5495 : // 2000-3000 - delta (D)
5496 : // 3000 - image (E)
5497 : //
5498 : // If GC horizon is at 2500, we can remove layers A and B, but
5499 : // we cannot remove C, even though it's older than 2500, because
5500 : // the delta layer 2000-3000 depends on it.
5501 16 : if !layers
5502 16 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5503 : {
5504 12 : info!("keeping {} because it is the latest layer", l.layer_name());
5505 12 : result.layers_not_updated += 1;
5506 12 : continue 'outer;
5507 4 : }
5508 4 :
5509 4 : // We didn't find any reason to keep this file, so remove it.
5510 4 : info!(
5511 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5512 0 : l.layer_name(),
5513 0 : l.is_incremental(),
5514 : );
5515 4 : layers_to_remove.push(l);
5516 : }
5517 :
5518 8 : if !layers_to_remove.is_empty() {
5519 : // Persist the new GC cutoff value before we actually remove anything.
5520 : // This unconditionally schedules also an index_part.json update, even though, we will
5521 : // be doing one a bit later with the unlinked gc'd layers.
5522 4 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5523 4 : self.schedule_uploads(disk_consistent_lsn, None)
5524 4 : .map_err(|e| {
5525 0 : if self.cancel.is_cancelled() {
5526 0 : GcError::TimelineCancelled
5527 : } else {
5528 0 : GcError::Remote(e)
5529 : }
5530 4 : })?;
5531 :
5532 4 : let gc_layers = layers_to_remove
5533 4 : .iter()
5534 4 : .map(|x| guard.get_from_desc(x))
5535 4 : .collect::<Vec<Layer>>();
5536 4 :
5537 4 : result.layers_removed = gc_layers.len() as u64;
5538 4 :
5539 4 : self.remote_client.schedule_gc_update(&gc_layers)?;
5540 :
5541 4 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5542 4 :
5543 4 : #[cfg(feature = "testing")]
5544 4 : {
5545 4 : result.doomed_layers = gc_layers;
5546 4 : }
5547 4 : }
5548 :
5549 8 : info!(
5550 0 : "GC completed removing {} layers, cutoff {}",
5551 : result.layers_removed, new_gc_cutoff
5552 : );
5553 :
5554 8 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5555 8 : Ok(result)
5556 8 : }
5557 :
5558 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5559 1340006 : async fn reconstruct_value(
5560 1340006 : &self,
5561 1340006 : key: Key,
5562 1340006 : request_lsn: Lsn,
5563 1340006 : mut data: ValueReconstructState,
5564 1340006 : ) -> Result<Bytes, PageReconstructError> {
5565 1340006 : // Perform WAL redo if needed
5566 1340006 : data.records.reverse();
5567 1340006 :
5568 1340006 : // If we have a page image, and no WAL, we're all set
5569 1340006 : if data.records.is_empty() {
5570 1338370 : if let Some((img_lsn, img)) = &data.img {
5571 1338370 : trace!(
5572 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5573 : key,
5574 : img_lsn,
5575 : request_lsn,
5576 : );
5577 1338370 : Ok(img.clone())
5578 : } else {
5579 0 : Err(PageReconstructError::from(anyhow!(
5580 0 : "base image for {key} at {request_lsn} not found"
5581 0 : )))
5582 : }
5583 : } else {
5584 : // We need to do WAL redo.
5585 : //
5586 : // If we don't have a base image, then the oldest WAL record better initialize
5587 : // the page
5588 1636 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5589 0 : Err(PageReconstructError::from(anyhow!(
5590 0 : "Base image for {} at {} not found, but got {} WAL records",
5591 0 : key,
5592 0 : request_lsn,
5593 0 : data.records.len()
5594 0 : )))
5595 : } else {
5596 1636 : if data.img.is_some() {
5597 1504 : trace!(
5598 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5599 0 : data.records.len(),
5600 : key,
5601 : request_lsn
5602 : );
5603 : } else {
5604 132 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5605 : };
5606 1636 : let res = self
5607 1636 : .walredo_mgr
5608 1636 : .as_ref()
5609 1636 : .context("timeline has no walredo manager")
5610 1636 : .map_err(PageReconstructError::WalRedo)?
5611 1636 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5612 1636 : .await;
5613 1636 : let img = match res {
5614 1636 : Ok(img) => img,
5615 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5616 0 : Err(walredo::Error::Other(e)) => {
5617 0 : return Err(PageReconstructError::WalRedo(
5618 0 : e.context("reconstruct a page image"),
5619 0 : ))
5620 : }
5621 : };
5622 1636 : Ok(img)
5623 : }
5624 : }
5625 1340006 : }
5626 :
5627 0 : pub(crate) async fn spawn_download_all_remote_layers(
5628 0 : self: Arc<Self>,
5629 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5630 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5631 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5632 :
5633 : // this is not really needed anymore; it has tests which really check the return value from
5634 : // http api. it would be better not to maintain this anymore.
5635 :
5636 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5637 0 : if let Some(st) = &*status_guard {
5638 0 : match &st.state {
5639 : DownloadRemoteLayersTaskState::Running => {
5640 0 : return Err(st.clone());
5641 : }
5642 : DownloadRemoteLayersTaskState::ShutDown
5643 0 : | DownloadRemoteLayersTaskState::Completed => {
5644 0 : *status_guard = None;
5645 0 : }
5646 : }
5647 0 : }
5648 :
5649 0 : let self_clone = Arc::clone(&self);
5650 0 : let task_id = task_mgr::spawn(
5651 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5652 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5653 0 : self.tenant_shard_id,
5654 0 : Some(self.timeline_id),
5655 0 : "download all remote layers task",
5656 0 : async move {
5657 0 : self_clone.download_all_remote_layers(request).await;
5658 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5659 0 : match &mut *status_guard {
5660 : None => {
5661 0 : warn!("tasks status is supposed to be Some(), since we are running");
5662 : }
5663 0 : Some(st) => {
5664 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5665 0 : if st.task_id != exp_task_id {
5666 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5667 0 : } else {
5668 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5669 0 : }
5670 : }
5671 : };
5672 0 : Ok(())
5673 0 : }
5674 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))
5675 : );
5676 :
5677 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5678 0 : task_id: format!("{task_id}"),
5679 0 : state: DownloadRemoteLayersTaskState::Running,
5680 0 : total_layer_count: 0,
5681 0 : successful_download_count: 0,
5682 0 : failed_download_count: 0,
5683 0 : };
5684 0 : *status_guard = Some(initial_info.clone());
5685 0 :
5686 0 : Ok(initial_info)
5687 0 : }
5688 :
5689 0 : async fn download_all_remote_layers(
5690 0 : self: &Arc<Self>,
5691 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5692 0 : ) {
5693 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5694 :
5695 0 : let remaining = {
5696 0 : let guard = self.layers.read().await;
5697 0 : let Ok(lm) = guard.layer_map() else {
5698 : // technically here we could look into iterating accessible layers, but downloading
5699 : // all layers of a shutdown timeline makes no sense regardless.
5700 0 : tracing::info!("attempted to download all layers of shutdown timeline");
5701 0 : return;
5702 : };
5703 0 : lm.iter_historic_layers()
5704 0 : .map(|desc| guard.get_from_desc(&desc))
5705 0 : .collect::<Vec<_>>()
5706 0 : };
5707 0 : let total_layer_count = remaining.len();
5708 :
5709 : macro_rules! lock_status {
5710 : ($st:ident) => {
5711 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5712 : let st = st
5713 : .as_mut()
5714 : .expect("this function is only called after the task has been spawned");
5715 : assert_eq!(
5716 : st.task_id,
5717 : format!(
5718 : "{}",
5719 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5720 : )
5721 : );
5722 : let $st = st;
5723 : };
5724 : }
5725 :
5726 : {
5727 0 : lock_status!(st);
5728 0 : st.total_layer_count = total_layer_count as u64;
5729 0 : }
5730 0 :
5731 0 : let mut remaining = remaining.into_iter();
5732 0 : let mut have_remaining = true;
5733 0 : let mut js = tokio::task::JoinSet::new();
5734 0 :
5735 0 : let cancel = task_mgr::shutdown_token();
5736 0 :
5737 0 : let limit = request.max_concurrent_downloads;
5738 :
5739 : loop {
5740 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5741 0 : let Some(next) = remaining.next() else {
5742 0 : have_remaining = false;
5743 0 : break;
5744 : };
5745 :
5746 0 : let span = tracing::info_span!("download", layer = %next);
5747 :
5748 0 : js.spawn(
5749 0 : async move {
5750 0 : let res = next.download().await;
5751 0 : (next, res)
5752 0 : }
5753 0 : .instrument(span),
5754 0 : );
5755 0 : }
5756 :
5757 0 : while let Some(res) = js.join_next().await {
5758 0 : match res {
5759 : Ok((_, Ok(_))) => {
5760 0 : lock_status!(st);
5761 0 : st.successful_download_count += 1;
5762 : }
5763 0 : Ok((layer, Err(e))) => {
5764 0 : tracing::error!(%layer, "download failed: {e:#}");
5765 0 : lock_status!(st);
5766 0 : st.failed_download_count += 1;
5767 : }
5768 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5769 0 : Err(je) if je.is_panic() => {
5770 0 : lock_status!(st);
5771 0 : st.failed_download_count += 1;
5772 : }
5773 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5774 : }
5775 : }
5776 :
5777 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5778 0 : break;
5779 0 : }
5780 : }
5781 :
5782 : {
5783 0 : lock_status!(st);
5784 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5785 : }
5786 0 : }
5787 :
5788 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5789 0 : &self,
5790 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5791 0 : self.download_all_remote_layers_task_info
5792 0 : .read()
5793 0 : .unwrap()
5794 0 : .clone()
5795 0 : }
5796 : }
5797 :
5798 : impl Timeline {
5799 : /// Returns non-remote layers for eviction.
5800 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5801 0 : let guard = self.layers.read().await;
5802 0 : let mut max_layer_size: Option<u64> = None;
5803 0 :
5804 0 : let resident_layers = guard
5805 0 : .likely_resident_layers()
5806 0 : .map(|layer| {
5807 0 : let file_size = layer.layer_desc().file_size;
5808 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5809 0 :
5810 0 : let last_activity_ts = layer.latest_activity();
5811 0 :
5812 0 : EvictionCandidate {
5813 0 : layer: layer.to_owned().into(),
5814 0 : last_activity_ts,
5815 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5816 0 : visibility: layer.visibility(),
5817 0 : }
5818 0 : })
5819 0 : .collect();
5820 0 :
5821 0 : DiskUsageEvictionInfo {
5822 0 : max_layer_size,
5823 0 : resident_layers,
5824 0 : }
5825 0 : }
5826 :
5827 3720 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5828 3720 : ShardIndex {
5829 3720 : shard_number: self.tenant_shard_id.shard_number,
5830 3720 : shard_count: self.tenant_shard_id.shard_count,
5831 3720 : }
5832 3720 : }
5833 :
5834 : /// Persistently blocks gc for `Manual` reason.
5835 : ///
5836 : /// Returns true if no such block existed before, false otherwise.
5837 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
5838 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5839 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5840 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
5841 0 : }
5842 :
5843 : /// Persistently unblocks gc for `Manual` reason.
5844 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
5845 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5846 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5847 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
5848 0 : }
5849 :
5850 : #[cfg(test)]
5851 92 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5852 92 : self.last_record_lsn.advance(new_lsn);
5853 92 : }
5854 :
5855 : #[cfg(test)]
5856 4 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
5857 4 : self.disk_consistent_lsn.store(new_value);
5858 4 : }
5859 :
5860 : /// Force create an image layer and place it into the layer map.
5861 : ///
5862 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5863 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5864 : /// placed into the layer map in one run AND be validated.
5865 : #[cfg(test)]
5866 116 : pub(super) async fn force_create_image_layer(
5867 116 : self: &Arc<Timeline>,
5868 116 : lsn: Lsn,
5869 116 : mut images: Vec<(Key, Bytes)>,
5870 116 : check_start_lsn: Option<Lsn>,
5871 116 : ctx: &RequestContext,
5872 116 : ) -> anyhow::Result<()> {
5873 116 : let last_record_lsn = self.get_last_record_lsn();
5874 116 : assert!(
5875 116 : lsn <= last_record_lsn,
5876 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5877 : );
5878 116 : if let Some(check_start_lsn) = check_start_lsn {
5879 116 : assert!(lsn >= check_start_lsn);
5880 0 : }
5881 348 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5882 116 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5883 116 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5884 116 : let mut image_layer_writer = ImageLayerWriter::new(
5885 116 : self.conf,
5886 116 : self.timeline_id,
5887 116 : self.tenant_shard_id,
5888 116 : &(min_key..end_key),
5889 116 : lsn,
5890 116 : ctx,
5891 116 : )
5892 116 : .await?;
5893 580 : for (key, img) in images {
5894 464 : image_layer_writer.put_image(key, img, ctx).await?;
5895 : }
5896 116 : let (desc, path) = image_layer_writer.finish(ctx).await?;
5897 116 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5898 116 : info!("force created image layer {}", image_layer.local_path());
5899 : {
5900 116 : let mut guard = self.layers.write().await;
5901 116 : guard
5902 116 : .open_mut()
5903 116 : .unwrap()
5904 116 : .force_insert_layer(image_layer.clone());
5905 116 : }
5906 116 :
5907 116 : // Update remote_timeline_client state to reflect existence of this layer
5908 116 : self.remote_client
5909 116 : .schedule_layer_file_upload(image_layer)
5910 116 : .unwrap();
5911 116 :
5912 116 : Ok(())
5913 116 : }
5914 :
5915 : /// Force create a delta layer and place it into the layer map.
5916 : ///
5917 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5918 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5919 : /// placed into the layer map in one run AND be validated.
5920 : #[cfg(test)]
5921 180 : pub(super) async fn force_create_delta_layer(
5922 180 : self: &Arc<Timeline>,
5923 180 : mut deltas: DeltaLayerTestDesc,
5924 180 : check_start_lsn: Option<Lsn>,
5925 180 : ctx: &RequestContext,
5926 180 : ) -> anyhow::Result<()> {
5927 180 : let last_record_lsn = self.get_last_record_lsn();
5928 180 : deltas
5929 180 : .data
5930 264 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5931 180 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5932 180 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5933 624 : for (_, lsn, _) in &deltas.data {
5934 444 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5935 : }
5936 180 : assert!(
5937 180 : deltas.lsn_range.end <= last_record_lsn,
5938 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
5939 : deltas.lsn_range.end,
5940 : last_record_lsn
5941 : );
5942 180 : if let Some(check_start_lsn) = check_start_lsn {
5943 180 : assert!(deltas.lsn_range.start >= check_start_lsn);
5944 0 : }
5945 180 : let mut delta_layer_writer = DeltaLayerWriter::new(
5946 180 : self.conf,
5947 180 : self.timeline_id,
5948 180 : self.tenant_shard_id,
5949 180 : deltas.key_range.start,
5950 180 : deltas.lsn_range,
5951 180 : ctx,
5952 180 : )
5953 180 : .await?;
5954 624 : for (key, lsn, val) in deltas.data {
5955 444 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5956 : }
5957 180 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
5958 180 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5959 180 : info!("force created delta layer {}", delta_layer.local_path());
5960 : {
5961 180 : let mut guard = self.layers.write().await;
5962 180 : guard
5963 180 : .open_mut()
5964 180 : .unwrap()
5965 180 : .force_insert_layer(delta_layer.clone());
5966 180 : }
5967 180 :
5968 180 : // Update remote_timeline_client state to reflect existence of this layer
5969 180 : self.remote_client
5970 180 : .schedule_layer_file_upload(delta_layer)
5971 180 : .unwrap();
5972 180 :
5973 180 : Ok(())
5974 180 : }
5975 :
5976 : /// Return all keys at the LSN in the image layers
5977 : #[cfg(test)]
5978 12 : pub(crate) async fn inspect_image_layers(
5979 12 : self: &Arc<Timeline>,
5980 12 : lsn: Lsn,
5981 12 : ctx: &RequestContext,
5982 12 : io_concurrency: IoConcurrency,
5983 12 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
5984 12 : let mut all_data = Vec::new();
5985 12 : let guard = self.layers.read().await;
5986 68 : for layer in guard.layer_map()?.iter_historic_layers() {
5987 68 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
5988 16 : let layer = guard.get_from_desc(&layer);
5989 16 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
5990 16 : layer
5991 16 : .get_values_reconstruct_data(
5992 16 : KeySpace::single(Key::MIN..Key::MAX),
5993 16 : lsn..Lsn(lsn.0 + 1),
5994 16 : &mut reconstruct_data,
5995 16 : ctx,
5996 16 : )
5997 16 : .await?;
5998 132 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
5999 132 : let v = v.collect_pending_ios().await?;
6000 132 : all_data.push((k, v.img.unwrap().1));
6001 : }
6002 52 : }
6003 : }
6004 12 : all_data.sort();
6005 12 : Ok(all_data)
6006 12 : }
6007 :
6008 : /// Get all historic layer descriptors in the layer map
6009 : #[cfg(test)]
6010 48 : pub(crate) async fn inspect_historic_layers(
6011 48 : self: &Arc<Timeline>,
6012 48 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
6013 48 : let mut layers = Vec::new();
6014 48 : let guard = self.layers.read().await;
6015 228 : for layer in guard.layer_map()?.iter_historic_layers() {
6016 228 : layers.push(layer.key());
6017 228 : }
6018 48 : Ok(layers)
6019 48 : }
6020 :
6021 : #[cfg(test)]
6022 20 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
6023 20 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
6024 20 : keyspace.merge(&ks);
6025 20 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
6026 20 : }
6027 : }
6028 :
6029 : /// Tracking writes ingestion does to a particular in-memory layer.
6030 : ///
6031 : /// Cleared upon freezing a layer.
6032 : pub(crate) struct TimelineWriterState {
6033 : open_layer: Arc<InMemoryLayer>,
6034 : current_size: u64,
6035 : // Previous Lsn which passed through
6036 : prev_lsn: Option<Lsn>,
6037 : // Largest Lsn which passed through the current writer
6038 : max_lsn: Option<Lsn>,
6039 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
6040 : cached_last_freeze_at: Lsn,
6041 : }
6042 :
6043 : impl TimelineWriterState {
6044 2592 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
6045 2592 : Self {
6046 2592 : open_layer,
6047 2592 : current_size,
6048 2592 : prev_lsn: None,
6049 2592 : max_lsn: None,
6050 2592 : cached_last_freeze_at: last_freeze_at,
6051 2592 : }
6052 2592 : }
6053 : }
6054 :
6055 : /// Various functions to mutate the timeline.
6056 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
6057 : // This is probably considered a bad practice in Rust and should be fixed eventually,
6058 : // but will cause large code changes.
6059 : pub(crate) struct TimelineWriter<'a> {
6060 : tl: &'a Timeline,
6061 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
6062 : }
6063 :
6064 : impl Deref for TimelineWriter<'_> {
6065 : type Target = Timeline;
6066 :
6067 19796336 : fn deref(&self) -> &Self::Target {
6068 19796336 : self.tl
6069 19796336 : }
6070 : }
6071 :
6072 : #[derive(PartialEq)]
6073 : enum OpenLayerAction {
6074 : Roll,
6075 : Open,
6076 : None,
6077 : }
6078 :
6079 : impl TimelineWriter<'_> {
6080 9608464 : async fn handle_open_layer_action(
6081 9608464 : &mut self,
6082 9608464 : at: Lsn,
6083 9608464 : action: OpenLayerAction,
6084 9608464 : ctx: &RequestContext,
6085 9608464 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
6086 9608464 : match action {
6087 : OpenLayerAction::Roll => {
6088 160 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
6089 160 : self.roll_layer(freeze_at).await?;
6090 160 : self.open_layer(at, ctx).await?;
6091 : }
6092 2432 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
6093 : OpenLayerAction::None => {
6094 9605872 : assert!(self.write_guard.is_some());
6095 : }
6096 : }
6097 :
6098 9608464 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
6099 9608464 : }
6100 :
6101 2592 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
6102 2592 : let layer = self
6103 2592 : .tl
6104 2592 : .get_layer_for_write(at, &self.write_guard, ctx)
6105 2592 : .await?;
6106 2592 : let initial_size = layer.size().await?;
6107 :
6108 2592 : let last_freeze_at = self.last_freeze_at.load();
6109 2592 : self.write_guard.replace(TimelineWriterState::new(
6110 2592 : layer,
6111 2592 : initial_size,
6112 2592 : last_freeze_at,
6113 2592 : ));
6114 2592 :
6115 2592 : Ok(())
6116 2592 : }
6117 :
6118 160 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
6119 160 : let current_size = self.write_guard.as_ref().unwrap().current_size;
6120 :
6121 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
6122 : // to propagate the backpressure up into WAL ingestion.
6123 160 : let l0_count = self
6124 160 : .tl
6125 160 : .layers
6126 160 : .read()
6127 160 : .await
6128 160 : .layer_map()?
6129 160 : .level0_deltas()
6130 160 : .len();
6131 160 : let wait_thresholds = [
6132 160 : self.get_l0_flush_delay_threshold(),
6133 160 : self.get_l0_flush_stall_threshold(),
6134 160 : ];
6135 160 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
6136 :
6137 : // self.write_guard will be taken by the freezing
6138 160 : let flush_id = self
6139 160 : .tl
6140 160 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
6141 160 : .await?;
6142 :
6143 160 : assert!(self.write_guard.is_none());
6144 :
6145 160 : if let Some(wait_threshold) = wait_threshold {
6146 0 : if l0_count >= wait_threshold {
6147 0 : info!("layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers");
6148 0 : self.tl.wait_flush_completion(flush_id).await?;
6149 0 : }
6150 160 : }
6151 :
6152 160 : if current_size >= self.get_checkpoint_distance() * 2 {
6153 0 : warn!("Flushed oversized open layer with size {}", current_size)
6154 160 : }
6155 :
6156 160 : Ok(())
6157 160 : }
6158 :
6159 9608464 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
6160 9608464 : let state = &*self.write_guard;
6161 9608464 : let Some(state) = &state else {
6162 2432 : return OpenLayerAction::Open;
6163 : };
6164 :
6165 : #[cfg(feature = "testing")]
6166 9606032 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
6167 : // this check and assertion are not really needed because
6168 : // LayerManager::try_freeze_in_memory_layer will always clear out the
6169 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
6170 : // is no TimelineWriterState.
6171 0 : assert!(
6172 0 : state.open_layer.end_lsn.get().is_some(),
6173 0 : "our open_layer must be outdated"
6174 : );
6175 :
6176 : // this would be a memory leak waiting to happen because the in-memory layer always has
6177 : // an index
6178 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
6179 9606032 : }
6180 9606032 :
6181 9606032 : if state.prev_lsn == Some(lsn) {
6182 : // Rolling mid LSN is not supported by [downstream code].
6183 : // Hence, only roll at LSN boundaries.
6184 : //
6185 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
6186 12 : return OpenLayerAction::None;
6187 9606020 : }
6188 9606020 :
6189 9606020 : if state.current_size == 0 {
6190 : // Don't roll empty layers
6191 0 : return OpenLayerAction::None;
6192 9606020 : }
6193 9606020 :
6194 9606020 : if self.tl.should_roll(
6195 9606020 : state.current_size,
6196 9606020 : state.current_size + new_value_size,
6197 9606020 : self.get_checkpoint_distance(),
6198 9606020 : lsn,
6199 9606020 : state.cached_last_freeze_at,
6200 9606020 : state.open_layer.get_opened_at(),
6201 9606020 : ) {
6202 160 : OpenLayerAction::Roll
6203 : } else {
6204 9605860 : OpenLayerAction::None
6205 : }
6206 9608464 : }
6207 :
6208 : /// Put a batch of keys at the specified Lsns.
6209 9608460 : pub(crate) async fn put_batch(
6210 9608460 : &mut self,
6211 9608460 : batch: SerializedValueBatch,
6212 9608460 : ctx: &RequestContext,
6213 9608460 : ) -> anyhow::Result<()> {
6214 9608460 : if !batch.has_data() {
6215 0 : return Ok(());
6216 9608460 : }
6217 9608460 :
6218 9608460 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
6219 9608460 : // We don't assert this in release builds, since key ownership policies may change over
6220 9608460 : // time. Stray keys will be removed during compaction.
6221 9608460 : if cfg!(debug_assertions) {
6222 19789792 : for metadata in &batch.metadata {
6223 10181332 : if let ValueMeta::Serialized(metadata) = metadata {
6224 10181332 : let key = Key::from_compact(metadata.key);
6225 10181332 : assert!(
6226 10181332 : self.shard_identity.is_key_local(&key)
6227 0 : || self.shard_identity.is_key_global(&key),
6228 0 : "key {key} does not belong on shard {}",
6229 0 : self.shard_identity.shard_index()
6230 : );
6231 0 : }
6232 : }
6233 0 : }
6234 :
6235 9608460 : let batch_max_lsn = batch.max_lsn;
6236 9608460 : let buf_size: u64 = batch.buffer_size() as u64;
6237 9608460 :
6238 9608460 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
6239 9608460 : let layer = self
6240 9608460 : .handle_open_layer_action(batch_max_lsn, action, ctx)
6241 9608460 : .await?;
6242 :
6243 9608460 : let res = layer.put_batch(batch, ctx).await;
6244 :
6245 9608460 : if res.is_ok() {
6246 9608460 : // Update the current size only when the entire write was ok.
6247 9608460 : // In case of failures, we may have had partial writes which
6248 9608460 : // render the size tracking out of sync. That's ok because
6249 9608460 : // the checkpoint distance should be significantly smaller
6250 9608460 : // than the S3 single shot upload limit of 5GiB.
6251 9608460 : let state = self.write_guard.as_mut().unwrap();
6252 9608460 :
6253 9608460 : state.current_size += buf_size;
6254 9608460 : state.prev_lsn = Some(batch_max_lsn);
6255 9608460 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
6256 9608460 : }
6257 :
6258 9608460 : res
6259 9608460 : }
6260 :
6261 : #[cfg(test)]
6262 : /// Test helper, for tests that would like to poke individual values without composing a batch
6263 8780308 : pub(crate) async fn put(
6264 8780308 : &mut self,
6265 8780308 : key: Key,
6266 8780308 : lsn: Lsn,
6267 8780308 : value: &Value,
6268 8780308 : ctx: &RequestContext,
6269 8780308 : ) -> anyhow::Result<()> {
6270 : use utils::bin_ser::BeSer;
6271 8780308 : if !key.is_valid_key_on_write_path() {
6272 0 : bail!(
6273 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
6274 0 : key
6275 0 : );
6276 8780308 : }
6277 8780308 : let val_ser_size = value.serialized_size().unwrap() as usize;
6278 8780308 : let batch = SerializedValueBatch::from_values(vec![(
6279 8780308 : key.to_compact(),
6280 8780308 : lsn,
6281 8780308 : val_ser_size,
6282 8780308 : value.clone(),
6283 8780308 : )]);
6284 8780308 :
6285 8780308 : self.put_batch(batch, ctx).await
6286 8780308 : }
6287 :
6288 4 : pub(crate) async fn delete_batch(
6289 4 : &mut self,
6290 4 : batch: &[(Range<Key>, Lsn)],
6291 4 : ctx: &RequestContext,
6292 4 : ) -> anyhow::Result<()> {
6293 4 : if let Some((_, lsn)) = batch.first() {
6294 4 : let action = self.get_open_layer_action(*lsn, 0);
6295 4 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
6296 4 : layer.put_tombstones(batch).await?;
6297 0 : }
6298 :
6299 4 : Ok(())
6300 4 : }
6301 :
6302 : /// Track the end of the latest digested WAL record.
6303 : /// Remember the (end of) last valid WAL record remembered in the timeline.
6304 : ///
6305 : /// Call this after you have finished writing all the WAL up to 'lsn'.
6306 : ///
6307 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
6308 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
6309 : /// the latest and can be read.
6310 10558184 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
6311 10558184 : self.tl.finish_write(new_lsn);
6312 10558184 : }
6313 :
6314 541140 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
6315 541140 : self.tl.update_current_logical_size(delta)
6316 541140 : }
6317 : }
6318 :
6319 : // We need TimelineWriter to be send in upcoming conversion of
6320 : // Timeline::layers to tokio::sync::RwLock.
6321 : #[test]
6322 4 : fn is_send() {
6323 4 : fn _assert_send<T: Send>() {}
6324 4 : _assert_send::<TimelineWriter<'_>>();
6325 4 : }
6326 :
6327 : #[cfg(test)]
6328 : mod tests {
6329 : use pageserver_api::key::Key;
6330 : use pageserver_api::value::Value;
6331 : use utils::{id::TimelineId, lsn::Lsn};
6332 :
6333 : use crate::tenant::{
6334 : harness::{test_img, TenantHarness},
6335 : layer_map::LayerMap,
6336 : storage_layer::{Layer, LayerName},
6337 : timeline::{DeltaLayerTestDesc, EvictionError},
6338 : Timeline,
6339 : };
6340 :
6341 : #[tokio::test]
6342 4 : async fn test_heatmap_generation() {
6343 4 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
6344 4 :
6345 4 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6346 4 : Lsn(0x10)..Lsn(0x20),
6347 4 : vec![(
6348 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6349 4 : Lsn(0x11),
6350 4 : Value::Image(test_img("foo")),
6351 4 : )],
6352 4 : );
6353 4 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6354 4 : Lsn(0x10)..Lsn(0x20),
6355 4 : vec![(
6356 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6357 4 : Lsn(0x11),
6358 4 : Value::Image(test_img("foo")),
6359 4 : )],
6360 4 : );
6361 4 : let l0_delta = DeltaLayerTestDesc::new(
6362 4 : Lsn(0x20)..Lsn(0x30),
6363 4 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6364 4 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6365 4 : vec![(
6366 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6367 4 : Lsn(0x25),
6368 4 : Value::Image(test_img("foo")),
6369 4 : )],
6370 4 : );
6371 4 : let delta_layers = vec![
6372 4 : covered_delta.clone(),
6373 4 : visible_delta.clone(),
6374 4 : l0_delta.clone(),
6375 4 : ];
6376 4 :
6377 4 : let image_layer = (
6378 4 : Lsn(0x40),
6379 4 : vec![(
6380 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6381 4 : test_img("bar"),
6382 4 : )],
6383 4 : );
6384 4 : let image_layers = vec![image_layer];
6385 4 :
6386 4 : let (tenant, ctx) = harness.load().await;
6387 4 : let timeline = tenant
6388 4 : .create_test_timeline_with_layers(
6389 4 : TimelineId::generate(),
6390 4 : Lsn(0x10),
6391 4 : 14,
6392 4 : &ctx,
6393 4 : delta_layers,
6394 4 : image_layers,
6395 4 : Lsn(0x100),
6396 4 : )
6397 4 : .await
6398 4 : .unwrap();
6399 4 :
6400 4 : // Layer visibility is an input to heatmap generation, so refresh it first
6401 4 : timeline.update_layer_visibility().await.unwrap();
6402 4 :
6403 4 : let heatmap = timeline
6404 4 : .generate_heatmap()
6405 4 : .await
6406 4 : .expect("Infallible while timeline is not shut down");
6407 4 :
6408 4 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
6409 4 :
6410 4 : // L0 should come last
6411 4 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
6412 4 :
6413 4 : let mut last_lsn = Lsn::MAX;
6414 20 : for layer in heatmap.layers {
6415 4 : // Covered layer should be omitted
6416 16 : assert!(layer.name != covered_delta.layer_name());
6417 4 :
6418 16 : let layer_lsn = match &layer.name {
6419 8 : LayerName::Delta(d) => d.lsn_range.end,
6420 8 : LayerName::Image(i) => i.lsn,
6421 4 : };
6422 4 :
6423 4 : // Apart from L0s, newest Layers should come first
6424 16 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
6425 12 : assert!(layer_lsn <= last_lsn);
6426 12 : last_lsn = layer_lsn;
6427 4 : }
6428 4 : }
6429 4 : }
6430 :
6431 : #[tokio::test]
6432 4 : async fn two_layer_eviction_attempts_at_the_same_time() {
6433 4 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6434 4 : .await
6435 4 : .unwrap();
6436 4 :
6437 4 : let (tenant, ctx) = harness.load().await;
6438 4 : let timeline = tenant
6439 4 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6440 4 : .await
6441 4 : .unwrap();
6442 4 :
6443 4 : let layer = find_some_layer(&timeline).await;
6444 4 : let layer = layer
6445 4 : .keep_resident()
6446 4 : .await
6447 4 : .expect("no download => no downloading errors")
6448 4 : .drop_eviction_guard();
6449 4 :
6450 4 : let forever = std::time::Duration::from_secs(120);
6451 4 :
6452 4 : let first = layer.evict_and_wait(forever);
6453 4 : let second = layer.evict_and_wait(forever);
6454 4 :
6455 4 : let (first, second) = tokio::join!(first, second);
6456 4 :
6457 4 : let res = layer.keep_resident().await;
6458 4 : assert!(res.is_none(), "{res:?}");
6459 4 :
6460 4 : match (first, second) {
6461 4 : (Ok(()), Ok(())) => {
6462 4 : // because there are no more timeline locks being taken on eviction path, we can
6463 4 : // witness all three outcomes here.
6464 4 : }
6465 4 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
6466 0 : // if one completes before the other, this is fine just as well.
6467 0 : }
6468 4 : other => unreachable!("unexpected {:?}", other),
6469 4 : }
6470 4 : }
6471 :
6472 4 : async fn find_some_layer(timeline: &Timeline) -> Layer {
6473 4 : let layers = timeline.layers.read().await;
6474 4 : let desc = layers
6475 4 : .layer_map()
6476 4 : .unwrap()
6477 4 : .iter_historic_layers()
6478 4 : .next()
6479 4 : .expect("must find one layer to evict");
6480 4 :
6481 4 : layers.get_from_desc(&desc)
6482 4 : }
6483 : }
|
