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 compaction::CompactionOutcome;
22 : use enumset::EnumSet;
23 : use fail::fail_point;
24 : use futures::{stream::FuturesUnordered, StreamExt};
25 : use handle::ShardTimelineId;
26 : use layer_manager::Shutdown;
27 : use offload::OffloadError;
28 : use once_cell::sync::Lazy;
29 : use pageserver_api::models::PageTraceEvent;
30 : use pageserver_api::{
31 : key::{
32 : KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
33 : SPARSE_RANGE,
34 : },
35 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
36 : models::{
37 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
38 : DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy,
39 : InMemoryLayerInfo, LayerMapInfo, LsnLease, TimelineState,
40 : },
41 : reltag::BlockNumber,
42 : shard::{ShardIdentity, ShardNumber, TenantShardId},
43 : };
44 : use rand::Rng;
45 : use remote_storage::DownloadError;
46 : use serde_with::serde_as;
47 : use storage_broker::BrokerClientChannel;
48 : use tokio::runtime::Handle;
49 : use tokio::sync::mpsc::Sender;
50 : use tokio::sync::{oneshot, watch, Notify};
51 : use tokio_util::sync::CancellationToken;
52 : use tracing::*;
53 : use utils::critical;
54 : use utils::rate_limit::RateLimit;
55 : use utils::{
56 : fs_ext,
57 : guard_arc_swap::GuardArcSwap,
58 : pausable_failpoint,
59 : postgres_client::PostgresClientProtocol,
60 : sync::gate::{Gate, GateGuard},
61 : };
62 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
63 :
64 : use std::array;
65 : use std::cmp::{max, min};
66 : use std::collections::btree_map::Entry;
67 : use std::collections::{BTreeMap, HashMap, HashSet};
68 : use std::ops::{ControlFlow, Deref, Range};
69 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering as AtomicOrdering};
70 : use std::sync::{Arc, Mutex, OnceLock, RwLock, Weak};
71 : use std::time::{Duration, Instant, SystemTime};
72 :
73 : use crate::l0_flush::{self, L0FlushGlobalState};
74 : use crate::tenant::storage_layer::ImageLayerName;
75 : use crate::{
76 : aux_file::AuxFileSizeEstimator,
77 : page_service::TenantManagerTypes,
78 : tenant::{
79 : config::AttachmentMode,
80 : layer_map::{LayerMap, SearchResult},
81 : metadata::TimelineMetadata,
82 : storage_layer::{
83 : inmemory_layer::IndexEntry, BatchLayerWriter, IoConcurrency, PersistentLayerDesc,
84 : ValueReconstructSituation,
85 : },
86 : },
87 : walingest::WalLagCooldown,
88 : walredo,
89 : };
90 : use crate::{
91 : context::{DownloadBehavior, RequestContext},
92 : disk_usage_eviction_task::DiskUsageEvictionInfo,
93 : pgdatadir_mapping::CollectKeySpaceError,
94 : };
95 : use crate::{
96 : disk_usage_eviction_task::finite_f32,
97 : tenant::storage_layer::{
98 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
99 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
100 : ValuesReconstructState,
101 : },
102 : };
103 : use crate::{
104 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
105 : };
106 : use crate::{
107 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
108 : };
109 : use crate::{
110 : pgdatadir_mapping::DirectoryKind,
111 : virtual_file::{MaybeFatalIo, VirtualFile},
112 : };
113 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
114 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
115 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
116 :
117 : use crate::config::PageServerConf;
118 : use crate::keyspace::{KeyPartitioning, KeySpace};
119 : use crate::metrics::{TimelineMetrics, DELTAS_PER_READ_GLOBAL, LAYERS_PER_READ_GLOBAL};
120 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
121 : use crate::tenant::config::TenantConfOpt;
122 : use pageserver_api::reltag::RelTag;
123 : use pageserver_api::shard::ShardIndex;
124 :
125 : use postgres_connection::PgConnectionConfig;
126 : use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
127 : use utils::{
128 : completion,
129 : generation::Generation,
130 : id::TimelineId,
131 : lsn::{AtomicLsn, Lsn, RecordLsn},
132 : seqwait::SeqWait,
133 : simple_rcu::{Rcu, RcuReadGuard},
134 : };
135 :
136 : use crate::task_mgr;
137 : use crate::task_mgr::TaskKind;
138 : use crate::tenant::gc_result::GcResult;
139 : use crate::ZERO_PAGE;
140 : use pageserver_api::key::Key;
141 :
142 : use self::delete::DeleteTimelineFlow;
143 : pub(super) use self::eviction_task::EvictionTaskTenantState;
144 : use self::eviction_task::EvictionTaskTimelineState;
145 : use self::layer_manager::LayerManager;
146 : use self::logical_size::LogicalSize;
147 : use self::walreceiver::{WalReceiver, WalReceiverConf};
148 :
149 : use super::{
150 : config::TenantConf, storage_layer::LayerVisibilityHint, upload_queue::NotInitialized,
151 : MaybeOffloaded,
152 : };
153 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
154 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
155 : use super::{
156 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
157 : storage_layer::ReadableLayer,
158 : };
159 : use super::{
160 : secondary::heatmap::{HeatMapLayer, HeatMapTimeline},
161 : GcError,
162 : };
163 :
164 : #[cfg(test)]
165 : use pageserver_api::value::Value;
166 :
167 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
168 : pub(crate) enum FlushLoopState {
169 : NotStarted,
170 : Running {
171 : #[cfg(test)]
172 : expect_initdb_optimization: bool,
173 : #[cfg(test)]
174 : initdb_optimization_count: usize,
175 : },
176 : Exited,
177 : }
178 :
179 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
180 : pub enum ImageLayerCreationMode {
181 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
182 : Try,
183 : /// Force creating the image layers if possible. For now, no image layers will be created
184 : /// for metadata keys. Used in compaction code path with force flag enabled.
185 : Force,
186 : /// Initial ingestion of the data, and no data should be dropped in this function. This
187 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
188 : /// code path.
189 : Initial,
190 : }
191 :
192 : #[derive(Clone, Debug, Default)]
193 : pub enum LastImageLayerCreationStatus {
194 : Incomplete {
195 : /// The last key of the partition (exclusive) that was processed in the last
196 : /// image layer creation attempt. We will continue from this key in the next
197 : /// attempt.
198 : last_key: Key,
199 : },
200 : Complete,
201 : #[default]
202 : Initial,
203 : }
204 :
205 : impl std::fmt::Display for ImageLayerCreationMode {
206 1176 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
207 1176 : write!(f, "{:?}", self)
208 1176 : }
209 : }
210 :
211 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
212 : /// Can be removed after all refactors are done.
213 92 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
214 92 : drop(rlock)
215 92 : }
216 :
217 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
218 : /// Can be removed after all refactors are done.
219 1268 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
220 1268 : drop(rlock)
221 1268 : }
222 :
223 : /// The outward-facing resources required to build a Timeline
224 : pub struct TimelineResources {
225 : pub remote_client: RemoteTimelineClient,
226 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
227 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
228 : pub l0_compaction_trigger: Arc<Notify>,
229 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
230 : }
231 :
232 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
233 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
234 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
235 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
236 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
237 : pub(crate) struct RelSizeCache {
238 : pub(crate) complete_as_of: Lsn,
239 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
240 : }
241 :
242 : pub struct Timeline {
243 : pub(crate) conf: &'static PageServerConf,
244 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
245 :
246 : myself: Weak<Self>,
247 :
248 : pub(crate) tenant_shard_id: TenantShardId,
249 : pub timeline_id: TimelineId,
250 :
251 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
252 : /// Never changes for the lifetime of this [`Timeline`] object.
253 : ///
254 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
255 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
256 : pub(crate) generation: Generation,
257 :
258 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
259 : /// to shards, and is constant through the lifetime of this Timeline.
260 : shard_identity: ShardIdentity,
261 :
262 : pub pg_version: u32,
263 :
264 : /// The tuple has two elements.
265 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
266 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
267 : ///
268 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
269 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
270 : ///
271 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
272 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
273 : /// `PersistentLayerDesc`'s.
274 : ///
275 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
276 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
277 : /// runtime, e.g., during page reconstruction.
278 : ///
279 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
280 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
281 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
282 :
283 : last_freeze_at: AtomicLsn,
284 : // Atomic would be more appropriate here.
285 : last_freeze_ts: RwLock<Instant>,
286 :
287 : pub(crate) standby_horizon: AtomicLsn,
288 :
289 : // WAL redo manager. `None` only for broken tenants.
290 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
291 :
292 : /// Remote storage client.
293 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
294 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
295 :
296 : // What page versions do we hold in the repository? If we get a
297 : // request > last_record_lsn, we need to wait until we receive all
298 : // the WAL up to the request. The SeqWait provides functions for
299 : // that. TODO: If we get a request for an old LSN, such that the
300 : // versions have already been garbage collected away, we should
301 : // throw an error, but we don't track that currently.
302 : //
303 : // last_record_lsn.load().last points to the end of last processed WAL record.
304 : //
305 : // We also remember the starting point of the previous record in
306 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
307 : // first WAL record when the node is started up. But here, we just
308 : // keep track of it.
309 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
310 :
311 : // All WAL records have been processed and stored durably on files on
312 : // local disk, up to this LSN. On crash and restart, we need to re-process
313 : // the WAL starting from this point.
314 : //
315 : // Some later WAL records might have been processed and also flushed to disk
316 : // already, so don't be surprised to see some, but there's no guarantee on
317 : // them yet.
318 : disk_consistent_lsn: AtomicLsn,
319 :
320 : // Parent timeline that this timeline was branched from, and the LSN
321 : // of the branch point.
322 : ancestor_timeline: Option<Arc<Timeline>>,
323 : ancestor_lsn: Lsn,
324 :
325 : pub(super) metrics: TimelineMetrics,
326 :
327 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
328 : // in `crate::page_service` writes these metrics.
329 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
330 :
331 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
332 :
333 : /// Ensures layers aren't frozen by checkpointer between
334 : /// [`Timeline::get_layer_for_write`] and layer reads.
335 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
336 : /// Must always be acquired before the layer map/individual layer lock
337 : /// to avoid deadlock.
338 : ///
339 : /// The state is cleared upon freezing.
340 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
341 :
342 : /// Used to avoid multiple `flush_loop` tasks running
343 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
344 :
345 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
346 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
347 : /// The flush cycle counter is sent back on the layer_flush_done channel when
348 : /// the flush finishes. You can use that to wait for the flush to finish.
349 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
350 : /// read by whoever sends an update
351 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
352 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
353 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
354 :
355 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
356 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
357 :
358 : pub(crate) gc_compaction_layer_update_lock: tokio::sync::RwLock<()>,
359 :
360 : // List of child timelines and their branch points. This is needed to avoid
361 : // garbage collecting data that is still needed by the child timelines.
362 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
363 :
364 : pub(crate) last_image_layer_creation_status: ArcSwap<LastImageLayerCreationStatus>,
365 :
366 : // It may change across major versions so for simplicity
367 : // keep it after running initdb for a timeline.
368 : // It is needed in checks when we want to error on some operations
369 : // when they are requested for pre-initdb lsn.
370 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
371 : // though let's keep them both for better error visibility.
372 : pub initdb_lsn: Lsn,
373 :
374 : /// The repartitioning result. Allows a single writer and multiple readers.
375 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
376 :
377 : /// Configuration: how often should the partitioning be recalculated.
378 : repartition_threshold: u64,
379 :
380 : last_image_layer_creation_check_at: AtomicLsn,
381 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
382 :
383 : /// Current logical size of the "datadir", at the last LSN.
384 : current_logical_size: LogicalSize,
385 :
386 : /// Information about the last processed message by the WAL receiver,
387 : /// or None if WAL receiver has not received anything for this timeline
388 : /// yet.
389 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
390 : pub walreceiver: Mutex<Option<WalReceiver>>,
391 :
392 : /// Relation size cache
393 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
394 :
395 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
396 :
397 : state: watch::Sender<TimelineState>,
398 :
399 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
400 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
401 : pub delete_progress: TimelineDeleteProgress,
402 :
403 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
404 :
405 : /// Load or creation time information about the disk_consistent_lsn and when the loading
406 : /// happened. Used for consumption metrics.
407 : pub(crate) loaded_at: (Lsn, SystemTime),
408 :
409 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
410 : pub(crate) gate: Gate,
411 :
412 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
413 : /// to the timeline should drop out when this token fires.
414 : pub(crate) cancel: CancellationToken,
415 :
416 : /// Make sure we only have one running compaction at a time in tests.
417 : ///
418 : /// Must only be taken in two places:
419 : /// - [`Timeline::compact`] (this file)
420 : /// - [`delete::delete_local_timeline_directory`]
421 : ///
422 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
423 : compaction_lock: tokio::sync::Mutex<()>,
424 :
425 : /// If true, the last compaction failed.
426 : compaction_failed: AtomicBool,
427 :
428 : /// Notifies the tenant compaction loop that there is pending L0 compaction work.
429 : l0_compaction_trigger: Arc<Notify>,
430 :
431 : /// Make sure we only have one running gc at a time.
432 : ///
433 : /// Must only be taken in two places:
434 : /// - [`Timeline::gc`] (this file)
435 : /// - [`delete::delete_local_timeline_directory`]
436 : ///
437 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
438 : gc_lock: tokio::sync::Mutex<()>,
439 :
440 : /// Cloned from [`super::Tenant::pagestream_throttle`] on construction.
441 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
442 :
443 : /// Size estimator for aux file v2
444 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
445 :
446 : /// Some test cases directly place keys into the timeline without actually modifying the directory
447 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
448 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
449 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
450 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
451 : #[cfg(test)]
452 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
453 :
454 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
455 :
456 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
457 :
458 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
459 :
460 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
461 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
462 :
463 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
464 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
465 : }
466 :
467 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
468 :
469 : pub struct WalReceiverInfo {
470 : pub wal_source_connconf: PgConnectionConfig,
471 : pub last_received_msg_lsn: Lsn,
472 : pub last_received_msg_ts: u128,
473 : }
474 :
475 : /// Information about how much history needs to be retained, needed by
476 : /// Garbage Collection.
477 : #[derive(Default)]
478 : pub(crate) struct GcInfo {
479 : /// Specific LSNs that are needed.
480 : ///
481 : /// Currently, this includes all points where child branches have
482 : /// been forked off from. In the future, could also include
483 : /// explicit user-defined snapshot points.
484 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
485 :
486 : /// The cutoff coordinates, which are combined by selecting the minimum.
487 : pub(crate) cutoffs: GcCutoffs,
488 :
489 : /// Leases granted to particular LSNs.
490 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
491 :
492 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
493 : pub(crate) within_ancestor_pitr: bool,
494 : }
495 :
496 : impl GcInfo {
497 572 : pub(crate) fn min_cutoff(&self) -> Lsn {
498 572 : self.cutoffs.select_min()
499 572 : }
500 :
501 464 : pub(super) fn insert_child(
502 464 : &mut self,
503 464 : child_id: TimelineId,
504 464 : child_lsn: Lsn,
505 464 : is_offloaded: MaybeOffloaded,
506 464 : ) {
507 464 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
508 464 : self.retain_lsns.sort_by_key(|i| i.0);
509 464 : }
510 :
511 8 : pub(super) fn remove_child_maybe_offloaded(
512 8 : &mut self,
513 8 : child_id: TimelineId,
514 8 : maybe_offloaded: MaybeOffloaded,
515 8 : ) -> bool {
516 8 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
517 8 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
518 8 : let mut removed = false;
519 12 : self.retain_lsns.retain(|i| {
520 12 : if removed {
521 4 : return true;
522 8 : }
523 8 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
524 8 : removed |= remove;
525 8 : !remove
526 12 : });
527 8 : removed
528 8 : }
529 :
530 8 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
531 8 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
532 8 : }
533 :
534 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
535 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
536 0 : }
537 464 : pub(crate) fn lsn_covered_by_lease(&self, lsn: Lsn) -> bool {
538 464 : self.leases.contains_key(&lsn)
539 464 : }
540 : }
541 :
542 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
543 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
544 : /// between time-based and space-based retention for observability and consumption metrics purposes.
545 : #[derive(Debug, Clone)]
546 : pub(crate) struct GcCutoffs {
547 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
548 : /// history we must keep to retain a specified number of bytes of WAL.
549 : pub(crate) space: Lsn,
550 :
551 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
552 : /// history we must keep to enable reading back at least the PITR interval duration.
553 : pub(crate) time: Lsn,
554 : }
555 :
556 : impl Default for GcCutoffs {
557 892 : fn default() -> Self {
558 892 : Self {
559 892 : space: Lsn::INVALID,
560 892 : time: Lsn::INVALID,
561 892 : }
562 892 : }
563 : }
564 :
565 : impl GcCutoffs {
566 572 : fn select_min(&self) -> Lsn {
567 572 : std::cmp::min(self.space, self.time)
568 572 : }
569 : }
570 :
571 : pub(crate) struct TimelineVisitOutcome {
572 : completed_keyspace: KeySpace,
573 : image_covered_keyspace: KeySpace,
574 : }
575 :
576 : /// An error happened in a get() operation.
577 : #[derive(thiserror::Error, Debug)]
578 : pub(crate) enum PageReconstructError {
579 : #[error(transparent)]
580 : Other(anyhow::Error),
581 :
582 : #[error("Ancestor LSN wait error: {0}")]
583 : AncestorLsnTimeout(WaitLsnError),
584 :
585 : #[error("timeline shutting down")]
586 : Cancelled,
587 :
588 : /// An error happened replaying WAL records
589 : #[error(transparent)]
590 : WalRedo(anyhow::Error),
591 :
592 : #[error("{0}")]
593 : MissingKey(MissingKeyError),
594 : }
595 :
596 : impl From<anyhow::Error> for PageReconstructError {
597 0 : fn from(value: anyhow::Error) -> Self {
598 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
599 0 : match value.downcast::<PageReconstructError>() {
600 0 : Ok(pre) => pre,
601 0 : Err(other) => PageReconstructError::Other(other),
602 : }
603 0 : }
604 : }
605 :
606 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
607 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
608 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
609 0 : }
610 : }
611 :
612 : impl From<layer_manager::Shutdown> for PageReconstructError {
613 0 : fn from(_: layer_manager::Shutdown) -> Self {
614 0 : PageReconstructError::Cancelled
615 0 : }
616 : }
617 :
618 : impl GetVectoredError {
619 : #[cfg(test)]
620 12 : pub(crate) fn is_missing_key_error(&self) -> bool {
621 12 : matches!(self, Self::MissingKey(_))
622 12 : }
623 : }
624 :
625 : impl From<layer_manager::Shutdown> for GetVectoredError {
626 0 : fn from(_: layer_manager::Shutdown) -> Self {
627 0 : GetVectoredError::Cancelled
628 0 : }
629 : }
630 :
631 : /// A layer identifier when used in the [`ReadPath`] structure. This enum is for observability purposes
632 : /// only and not used by the "real read path".
633 : pub enum ReadPathLayerId {
634 : PersistentLayer(PersistentLayerKey),
635 : InMemoryLayer(Range<Lsn>),
636 : }
637 :
638 : impl std::fmt::Display for ReadPathLayerId {
639 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
640 0 : match self {
641 0 : ReadPathLayerId::PersistentLayer(key) => write!(f, "{}", key),
642 0 : ReadPathLayerId::InMemoryLayer(range) => {
643 0 : write!(f, "in-mem {}..{}", range.start, range.end)
644 : }
645 : }
646 0 : }
647 : }
648 : pub struct ReadPath {
649 : keyspace: KeySpace,
650 : lsn: Lsn,
651 : path: Vec<(ReadPathLayerId, KeySpace, Range<Lsn>)>,
652 : }
653 :
654 : impl ReadPath {
655 1255178 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
656 1255178 : Self {
657 1255178 : keyspace,
658 1255178 : lsn,
659 1255178 : path: Vec::new(),
660 1255178 : }
661 1255178 : }
662 :
663 1767919 : pub fn record_layer_visit(
664 1767919 : &mut self,
665 1767919 : layer_to_read: &ReadableLayer,
666 1767919 : keyspace_to_read: &KeySpace,
667 1767919 : lsn_range: &Range<Lsn>,
668 1767919 : ) {
669 1767919 : let id = match layer_to_read {
670 554724 : ReadableLayer::PersistentLayer(layer) => {
671 554724 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
672 : }
673 1213195 : ReadableLayer::InMemoryLayer(layer) => {
674 1213195 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
675 : }
676 : };
677 1767919 : self.path
678 1767919 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
679 1767919 : }
680 : }
681 :
682 : impl std::fmt::Display for ReadPath {
683 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
684 0 : writeln!(f, "Read path for {} at lsn {}:", self.keyspace, self.lsn)?;
685 0 : for (idx, (layer_id, keyspace, lsn_range)) in self.path.iter().enumerate() {
686 0 : writeln!(
687 0 : f,
688 0 : "{}: {} {}..{} {}",
689 0 : idx, layer_id, lsn_range.start, lsn_range.end, keyspace
690 0 : )?;
691 : }
692 0 : Ok(())
693 0 : }
694 : }
695 :
696 : #[derive(thiserror::Error)]
697 : pub struct MissingKeyError {
698 : key: Key,
699 : shard: ShardNumber,
700 : cont_lsn: Lsn,
701 : request_lsn: Lsn,
702 : ancestor_lsn: Option<Lsn>,
703 : /// Debug information about the read path if there's an error
704 : read_path: Option<ReadPath>,
705 : backtrace: Option<std::backtrace::Backtrace>,
706 : }
707 :
708 : impl std::fmt::Debug for MissingKeyError {
709 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
710 0 : write!(f, "{}", self)
711 0 : }
712 : }
713 :
714 : impl std::fmt::Display for MissingKeyError {
715 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
716 0 : write!(
717 0 : f,
718 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
719 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
720 0 : )?;
721 :
722 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
723 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
724 0 : }
725 :
726 0 : if let Some(ref read_path) = self.read_path {
727 0 : write!(f, "\n{}", read_path)?;
728 0 : }
729 :
730 0 : if let Some(ref backtrace) = self.backtrace {
731 0 : write!(f, "\n{}", backtrace)?;
732 0 : }
733 :
734 0 : Ok(())
735 0 : }
736 : }
737 :
738 : impl PageReconstructError {
739 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
740 0 : pub(crate) fn is_stopping(&self) -> bool {
741 : use PageReconstructError::*;
742 0 : match self {
743 0 : Cancelled => true,
744 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
745 : }
746 0 : }
747 : }
748 :
749 : #[derive(thiserror::Error, Debug)]
750 : pub(crate) enum CreateImageLayersError {
751 : #[error("timeline shutting down")]
752 : Cancelled,
753 :
754 : #[error("read failed")]
755 : GetVectoredError(#[source] GetVectoredError),
756 :
757 : #[error("reconstruction failed")]
758 : PageReconstructError(#[source] PageReconstructError),
759 :
760 : #[error(transparent)]
761 : Other(#[from] anyhow::Error),
762 : }
763 :
764 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
765 0 : fn from(_: layer_manager::Shutdown) -> Self {
766 0 : CreateImageLayersError::Cancelled
767 0 : }
768 : }
769 :
770 : #[derive(thiserror::Error, Debug, Clone)]
771 : pub(crate) enum FlushLayerError {
772 : /// Timeline cancellation token was cancelled
773 : #[error("timeline shutting down")]
774 : Cancelled,
775 :
776 : /// We tried to flush a layer while the Timeline is in an unexpected state
777 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
778 : NotRunning(FlushLoopState),
779 :
780 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
781 : // loop via a watch channel, where we can only borrow it.
782 : #[error("create image layers (shared)")]
783 : CreateImageLayersError(Arc<CreateImageLayersError>),
784 :
785 : #[error("other (shared)")]
786 : Other(#[from] Arc<anyhow::Error>),
787 : }
788 :
789 : impl FlushLayerError {
790 : // When crossing from generic anyhow errors to this error type, we explicitly check
791 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
792 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
793 0 : let cancelled = timeline.cancel.is_cancelled()
794 : // The upload queue might have been shut down before the official cancellation of the timeline.
795 0 : || err
796 0 : .downcast_ref::<NotInitialized>()
797 0 : .map(NotInitialized::is_stopping)
798 0 : .unwrap_or_default();
799 0 : if cancelled {
800 0 : Self::Cancelled
801 : } else {
802 0 : Self::Other(Arc::new(err))
803 : }
804 0 : }
805 : }
806 :
807 : impl From<layer_manager::Shutdown> for FlushLayerError {
808 0 : fn from(_: layer_manager::Shutdown) -> Self {
809 0 : FlushLayerError::Cancelled
810 0 : }
811 : }
812 :
813 : #[derive(thiserror::Error, Debug)]
814 : pub(crate) enum GetVectoredError {
815 : #[error("timeline shutting down")]
816 : Cancelled,
817 :
818 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
819 : Oversized(u64),
820 :
821 : #[error("requested at invalid LSN: {0}")]
822 : InvalidLsn(Lsn),
823 :
824 : #[error("requested key not found: {0}")]
825 : MissingKey(MissingKeyError),
826 :
827 : #[error("ancestry walk")]
828 : GetReadyAncestorError(#[source] GetReadyAncestorError),
829 :
830 : #[error(transparent)]
831 : Other(#[from] anyhow::Error),
832 : }
833 :
834 : impl From<GetReadyAncestorError> for GetVectoredError {
835 4 : fn from(value: GetReadyAncestorError) -> Self {
836 : use GetReadyAncestorError::*;
837 4 : match value {
838 0 : Cancelled => GetVectoredError::Cancelled,
839 : AncestorLsnTimeout(_) | BadState { .. } => {
840 4 : GetVectoredError::GetReadyAncestorError(value)
841 : }
842 : }
843 4 : }
844 : }
845 :
846 : #[derive(thiserror::Error, Debug)]
847 : pub(crate) enum GetReadyAncestorError {
848 : #[error("ancestor LSN wait error")]
849 : AncestorLsnTimeout(#[from] WaitLsnError),
850 :
851 : #[error("bad state on timeline {timeline_id}: {state:?}")]
852 : BadState {
853 : timeline_id: TimelineId,
854 : state: TimelineState,
855 : },
856 :
857 : #[error("cancelled")]
858 : Cancelled,
859 : }
860 :
861 : #[derive(Clone, Copy)]
862 : pub enum LogicalSizeCalculationCause {
863 : Initial,
864 : ConsumptionMetricsSyntheticSize,
865 : EvictionTaskImitation,
866 : TenantSizeHandler,
867 : }
868 :
869 : pub enum GetLogicalSizePriority {
870 : User,
871 : Background,
872 : }
873 :
874 0 : #[derive(Debug, enumset::EnumSetType)]
875 : pub(crate) enum CompactFlags {
876 : ForceRepartition,
877 : ForceImageLayerCreation,
878 : ForceL0Compaction,
879 : OnlyL0Compaction,
880 : EnhancedGcBottomMostCompaction,
881 : DryRun,
882 : /// Disables compaction yielding e.g. due to high L0 count. This is set e.g. when requesting
883 : /// compaction via HTTP API.
884 : NoYield,
885 : }
886 :
887 : #[serde_with::serde_as]
888 0 : #[derive(Debug, Clone, serde::Deserialize)]
889 : pub(crate) struct CompactRequest {
890 : pub compact_key_range: Option<CompactKeyRange>,
891 : pub compact_lsn_range: Option<CompactLsnRange>,
892 : /// Whether the compaction job should be scheduled.
893 : #[serde(default)]
894 : pub scheduled: bool,
895 : /// Whether the compaction job should be split across key ranges.
896 : #[serde(default)]
897 : pub sub_compaction: bool,
898 : /// Max job size for each subcompaction job.
899 : pub sub_compaction_max_job_size_mb: Option<u64>,
900 : }
901 :
902 : #[derive(Debug, Clone, Default)]
903 : pub(crate) struct CompactOptions {
904 : pub flags: EnumSet<CompactFlags>,
905 : /// If set, the compaction will only compact the key range specified by this option.
906 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
907 : pub compact_key_range: Option<CompactKeyRange>,
908 : /// If set, the compaction will only compact the LSN within this value.
909 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
910 : pub compact_lsn_range: Option<CompactLsnRange>,
911 : /// Enable sub-compaction (split compaction job across key ranges).
912 : /// This option is only used by GC compaction.
913 : pub sub_compaction: bool,
914 : /// Set job size for the GC compaction.
915 : /// This option is only used by GC compaction.
916 : pub sub_compaction_max_job_size_mb: Option<u64>,
917 : }
918 :
919 : impl std::fmt::Debug for Timeline {
920 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
921 0 : write!(f, "Timeline<{}>", self.timeline_id)
922 0 : }
923 : }
924 :
925 : #[derive(thiserror::Error, Debug)]
926 : pub(crate) enum WaitLsnError {
927 : // Called on a timeline which is shutting down
928 : #[error("Shutdown")]
929 : Shutdown,
930 :
931 : // Called on an timeline not in active state or shutting down
932 : #[error("Bad timeline state: {0:?}")]
933 : BadState(TimelineState),
934 :
935 : // Timeout expired while waiting for LSN to catch up with goal.
936 : #[error("{0}")]
937 : Timeout(String),
938 : }
939 :
940 : // The impls below achieve cancellation mapping for errors.
941 : // Perhaps there's a way of achieving this with less cruft.
942 :
943 : impl From<CreateImageLayersError> for CompactionError {
944 0 : fn from(e: CreateImageLayersError) -> Self {
945 0 : match e {
946 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
947 0 : CreateImageLayersError::Other(e) => {
948 0 : CompactionError::Other(e.context("create image layers"))
949 : }
950 0 : _ => CompactionError::Other(e.into()),
951 : }
952 0 : }
953 : }
954 :
955 : impl From<CreateImageLayersError> for FlushLayerError {
956 0 : fn from(e: CreateImageLayersError) -> Self {
957 0 : match e {
958 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
959 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
960 : }
961 0 : }
962 : }
963 :
964 : impl From<PageReconstructError> for CreateImageLayersError {
965 0 : fn from(e: PageReconstructError) -> Self {
966 0 : match e {
967 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
968 0 : _ => CreateImageLayersError::PageReconstructError(e),
969 : }
970 0 : }
971 : }
972 :
973 : impl From<GetVectoredError> for CreateImageLayersError {
974 0 : fn from(e: GetVectoredError) -> Self {
975 0 : match e {
976 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
977 0 : _ => CreateImageLayersError::GetVectoredError(e),
978 : }
979 0 : }
980 : }
981 :
982 : impl From<GetVectoredError> for PageReconstructError {
983 12 : fn from(e: GetVectoredError) -> Self {
984 12 : match e {
985 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
986 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
987 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
988 8 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
989 4 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
990 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
991 : }
992 12 : }
993 : }
994 :
995 : impl From<GetReadyAncestorError> for PageReconstructError {
996 4 : fn from(e: GetReadyAncestorError) -> Self {
997 : use GetReadyAncestorError::*;
998 4 : match e {
999 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
1000 4 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
1001 0 : Cancelled => PageReconstructError::Cancelled,
1002 : }
1003 4 : }
1004 : }
1005 :
1006 : pub(crate) enum WaitLsnTimeout {
1007 : Custom(Duration),
1008 : // Use the [`PageServerConf::wait_lsn_timeout`] default
1009 : Default,
1010 : }
1011 :
1012 : pub(crate) enum WaitLsnWaiter<'a> {
1013 : Timeline(&'a Timeline),
1014 : Tenant,
1015 : PageService,
1016 : HttpEndpoint,
1017 : }
1018 :
1019 : /// Argument to [`Timeline::shutdown`].
1020 : #[derive(Debug, Clone, Copy)]
1021 : pub(crate) enum ShutdownMode {
1022 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
1023 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
1024 : ///
1025 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1026 : /// the call to [`Timeline::shutdown`].
1027 : FreezeAndFlush,
1028 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1029 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1030 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1031 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1032 : Reload,
1033 : /// Shut down immediately, without waiting for any open layers to flush.
1034 : Hard,
1035 : }
1036 :
1037 : enum ImageLayerCreationOutcome {
1038 : /// We generated an image layer
1039 : Generated {
1040 : unfinished_image_layer: ImageLayerWriter,
1041 : },
1042 : /// The key range is empty
1043 : Empty,
1044 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1045 : /// the image layer creation.
1046 : Skip,
1047 : }
1048 :
1049 : /// Public interface functions
1050 : impl Timeline {
1051 : /// Get the LSN where this branch was created
1052 776 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1053 776 : self.ancestor_lsn
1054 776 : }
1055 :
1056 : /// Get the ancestor's timeline id
1057 24 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1058 24 : self.ancestor_timeline
1059 24 : .as_ref()
1060 24 : .map(|ancestor| ancestor.timeline_id)
1061 24 : }
1062 :
1063 : /// Get the ancestor timeline
1064 4 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1065 4 : self.ancestor_timeline.as_ref()
1066 4 : }
1067 :
1068 : /// Get the bytes written since the PITR cutoff on this branch, and
1069 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1070 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1071 0 : let gc_info = self.gc_info.read().unwrap();
1072 0 : let history = self
1073 0 : .get_last_record_lsn()
1074 0 : .checked_sub(gc_info.cutoffs.time)
1075 0 : .unwrap_or(Lsn(0))
1076 0 : .0;
1077 0 : (history, gc_info.within_ancestor_pitr)
1078 0 : }
1079 :
1080 : /// Lock and get timeline's GC cutoff
1081 1705224 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1082 1705224 : self.latest_gc_cutoff_lsn.read()
1083 1705224 : }
1084 :
1085 : /// Look up given page version.
1086 : ///
1087 : /// If a remote layer file is needed, it is downloaded as part of this
1088 : /// call.
1089 : ///
1090 : /// This method enforces [`Self::pagestream_throttle`] internally.
1091 : ///
1092 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1093 : /// abstraction above this needs to store suitable metadata to track what
1094 : /// data exists with what keys, in separate metadata entries. If a
1095 : /// non-existent key is requested, we may incorrectly return a value from
1096 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1097 : /// non-existing key.
1098 : ///
1099 : /// # Cancel-Safety
1100 : ///
1101 : /// This method is cancellation-safe.
1102 : #[inline(always)]
1103 1215142 : pub(crate) async fn get(
1104 1215142 : &self,
1105 1215142 : key: Key,
1106 1215142 : lsn: Lsn,
1107 1215142 : ctx: &RequestContext,
1108 1215142 : ) -> Result<Bytes, PageReconstructError> {
1109 1215142 : if !lsn.is_valid() {
1110 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1111 1215142 : }
1112 1215142 :
1113 1215142 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1114 1215142 : // already checked the key against the shard_identity when looking up the Timeline from
1115 1215142 : // page_service.
1116 1215142 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1117 :
1118 1215142 : let keyspace = KeySpace {
1119 1215142 : ranges: vec![key..key.next()],
1120 1215142 : };
1121 1215142 :
1122 1215142 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1123 :
1124 1215142 : let vectored_res = self
1125 1215142 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
1126 1215142 : .await;
1127 :
1128 1215142 : let key_value = vectored_res?.pop_first();
1129 1215130 : match key_value {
1130 1215106 : Some((got_key, value)) => {
1131 1215106 : if got_key != key {
1132 0 : error!(
1133 0 : "Expected {}, but singular vectored get returned {}",
1134 : key, got_key
1135 : );
1136 0 : Err(PageReconstructError::Other(anyhow!(
1137 0 : "Singular vectored get returned wrong key"
1138 0 : )))
1139 : } else {
1140 1215106 : value
1141 : }
1142 : }
1143 24 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1144 24 : key,
1145 24 : shard: self.shard_identity.get_shard_number(&key),
1146 24 : cont_lsn: Lsn(0),
1147 24 : request_lsn: lsn,
1148 24 : ancestor_lsn: None,
1149 24 : backtrace: None,
1150 24 : read_path: None,
1151 24 : })),
1152 : }
1153 1215142 : }
1154 :
1155 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1156 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1157 :
1158 : /// Look up multiple page versions at a given LSN
1159 : ///
1160 : /// This naive implementation will be replaced with a more efficient one
1161 : /// which actually vectorizes the read path.
1162 39308 : pub(crate) async fn get_vectored(
1163 39308 : &self,
1164 39308 : keyspace: KeySpace,
1165 39308 : lsn: Lsn,
1166 39308 : io_concurrency: super::storage_layer::IoConcurrency,
1167 39308 : ctx: &RequestContext,
1168 39308 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1169 39308 : if !lsn.is_valid() {
1170 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1171 39308 : }
1172 39308 :
1173 39308 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1174 39308 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1175 0 : return Err(GetVectoredError::Oversized(key_count));
1176 39308 : }
1177 :
1178 78616 : for range in &keyspace.ranges {
1179 39308 : let mut key = range.start;
1180 79076 : while key != range.end {
1181 39768 : assert!(!self.shard_identity.is_key_disposable(&key));
1182 39768 : key = key.next();
1183 : }
1184 : }
1185 :
1186 39308 : trace!(
1187 0 : "get vectored request for {:?}@{} from task kind {:?}",
1188 0 : keyspace,
1189 0 : lsn,
1190 0 : ctx.task_kind(),
1191 : );
1192 :
1193 39308 : let start = crate::metrics::GET_VECTORED_LATENCY
1194 39308 : .for_task_kind(ctx.task_kind())
1195 39308 : .map(|metric| (metric, Instant::now()));
1196 :
1197 39308 : let res = self
1198 39308 : .get_vectored_impl(
1199 39308 : keyspace.clone(),
1200 39308 : lsn,
1201 39308 : &mut ValuesReconstructState::new(io_concurrency),
1202 39308 : ctx,
1203 39308 : )
1204 39308 : .await;
1205 :
1206 39308 : if let Some((metric, start)) = start {
1207 0 : let elapsed = start.elapsed();
1208 0 : metric.observe(elapsed.as_secs_f64());
1209 39308 : }
1210 :
1211 39308 : res
1212 39308 : }
1213 :
1214 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1215 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1216 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1217 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1218 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1219 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1220 : /// the scan operation will not cause OOM in the future.
1221 24 : pub(crate) async fn scan(
1222 24 : &self,
1223 24 : keyspace: KeySpace,
1224 24 : lsn: Lsn,
1225 24 : ctx: &RequestContext,
1226 24 : io_concurrency: super::storage_layer::IoConcurrency,
1227 24 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1228 24 : if !lsn.is_valid() {
1229 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1230 24 : }
1231 24 :
1232 24 : trace!(
1233 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1234 0 : keyspace,
1235 0 : lsn,
1236 0 : ctx.task_kind()
1237 : );
1238 :
1239 : // We should generalize this into Keyspace::contains in the future.
1240 48 : for range in &keyspace.ranges {
1241 24 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1242 24 : || range.end.field1 > METADATA_KEY_END_PREFIX
1243 : {
1244 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1245 0 : "only metadata keyspace can be scanned"
1246 0 : )));
1247 24 : }
1248 : }
1249 :
1250 24 : let start = crate::metrics::SCAN_LATENCY
1251 24 : .for_task_kind(ctx.task_kind())
1252 24 : .map(ScanLatencyOngoingRecording::start_recording);
1253 :
1254 24 : let vectored_res = self
1255 24 : .get_vectored_impl(
1256 24 : keyspace.clone(),
1257 24 : lsn,
1258 24 : &mut ValuesReconstructState::new(io_concurrency),
1259 24 : ctx,
1260 24 : )
1261 24 : .await;
1262 :
1263 24 : if let Some(recording) = start {
1264 0 : recording.observe();
1265 24 : }
1266 :
1267 24 : vectored_res
1268 24 : }
1269 :
1270 1255178 : pub(super) async fn get_vectored_impl(
1271 1255178 : &self,
1272 1255178 : keyspace: KeySpace,
1273 1255178 : lsn: Lsn,
1274 1255178 : reconstruct_state: &mut ValuesReconstructState,
1275 1255178 : ctx: &RequestContext,
1276 1255178 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1277 1255178 : let read_path = if self.conf.enable_read_path_debugging {
1278 1255178 : Some(ReadPath::new(keyspace.clone(), lsn))
1279 : } else {
1280 0 : None
1281 : };
1282 1255178 : reconstruct_state.read_path = read_path;
1283 :
1284 1255178 : let traversal_res: Result<(), _> = self
1285 1255178 : .get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1286 1255178 : .await;
1287 1255178 : if let Err(err) = traversal_res {
1288 : // Wait for all the spawned IOs to complete.
1289 : // See comments on `spawn_io` inside `storage_layer` for more details.
1290 32 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1291 32 : .into_values()
1292 32 : .map(|state| state.collect_pending_ios())
1293 32 : .collect::<FuturesUnordered<_>>();
1294 32 : while collect_futs.next().await.is_some() {}
1295 32 : return Err(err);
1296 1255146 : };
1297 1255146 :
1298 1255146 : let layers_visited = reconstruct_state.get_layers_visited();
1299 1255146 :
1300 1255146 : let futs = FuturesUnordered::new();
1301 1327810 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1302 1327810 : futs.push({
1303 1327810 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1304 1327810 : async move {
1305 1327810 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1306 :
1307 1327810 : let converted = match state.collect_pending_ios().await {
1308 1327810 : Ok(ok) => ok,
1309 0 : Err(err) => {
1310 0 : return (key, Err(err));
1311 : }
1312 : };
1313 1327810 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1314 1327810 :
1315 1327810 : // The walredo module expects the records to be descending in terms of Lsn.
1316 1327810 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1317 1327810 : debug_assert!(
1318 1327810 : converted
1319 1327810 : .records
1320 1327810 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1321 0 : "{converted:?}"
1322 : );
1323 :
1324 : (
1325 1327810 : key,
1326 1327810 : walredo_self.reconstruct_value(key, lsn, converted).await,
1327 : )
1328 1327810 : }
1329 1327810 : });
1330 1327810 : }
1331 :
1332 1255146 : let results = futs
1333 1255146 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1334 1255146 : .await;
1335 :
1336 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1337 : // when they're missing. Instead they are omitted from the resulting btree
1338 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1339 : // to avoid infinite results.
1340 1255146 : if !results.is_empty() {
1341 : // Record the total number of layers visited towards each key in the batch. While some
1342 : // layers may not intersect with a given read, and the cost of layer visits are
1343 : // amortized across the batch, each visited layer contributes directly to the observed
1344 : // latency for every read in the batch, which is what we care about.
1345 1254662 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1346 0 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1347 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1348 0 : LOG_PACER.lock().unwrap().call(|| {
1349 0 : let num_keys = keyspace.total_raw_size();
1350 0 : let num_pages = results.len();
1351 0 : tracing::info!(
1352 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1353 0 : lsn = %lsn,
1354 0 : "Vectored read for {keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1355 : );
1356 0 : });
1357 1254662 : }
1358 :
1359 2582472 : for _ in &results {
1360 1327810 : self.metrics.layers_per_read.observe(layers_visited as f64);
1361 1327810 : LAYERS_PER_READ_GLOBAL.observe(layers_visited as f64);
1362 1327810 : }
1363 484 : }
1364 :
1365 1255146 : Ok(results)
1366 1255178 : }
1367 :
1368 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1369 548192 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1370 548192 : self.last_record_lsn.load().last
1371 548192 : }
1372 :
1373 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1374 0 : self.last_record_lsn.load().prev
1375 0 : }
1376 :
1377 : /// Atomically get both last and prev.
1378 456 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1379 456 : self.last_record_lsn.load()
1380 456 : }
1381 :
1382 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1383 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1384 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1385 0 : self.last_record_lsn.status_receiver()
1386 0 : }
1387 :
1388 1652 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1389 1652 : self.disk_consistent_lsn.load()
1390 1652 : }
1391 :
1392 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1393 : /// not validated with control plane yet.
1394 : /// See [`Self::get_remote_consistent_lsn_visible`].
1395 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1396 0 : self.remote_client.remote_consistent_lsn_projected()
1397 0 : }
1398 :
1399 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1400 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1401 : /// generation validation in the deletion queue.
1402 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1403 0 : self.remote_client.remote_consistent_lsn_visible()
1404 0 : }
1405 :
1406 : /// The sum of the file size of all historic layers in the layer map.
1407 : /// This method makes no distinction between local and remote layers.
1408 : /// Hence, the result **does not represent local filesystem usage**.
1409 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1410 0 : let guard = self.layers.read().await;
1411 0 : guard.layer_size_sum()
1412 0 : }
1413 :
1414 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1415 0 : self.metrics.resident_physical_size_get()
1416 0 : }
1417 :
1418 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1419 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1420 0 : }
1421 :
1422 : ///
1423 : /// Wait until WAL has been received and processed up to this LSN.
1424 : ///
1425 : /// You should call this before any of the other get_* or list_* functions. Calling
1426 : /// those functions with an LSN that has been processed yet is an error.
1427 : ///
1428 449442 : pub(crate) async fn wait_lsn(
1429 449442 : &self,
1430 449442 : lsn: Lsn,
1431 449442 : who_is_waiting: WaitLsnWaiter<'_>,
1432 449442 : timeout: WaitLsnTimeout,
1433 449442 : ctx: &RequestContext, /* Prepare for use by cancellation */
1434 449442 : ) -> Result<(), WaitLsnError> {
1435 449442 : let state = self.current_state();
1436 449442 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1437 0 : return Err(WaitLsnError::Shutdown);
1438 449442 : } else if !matches!(state, TimelineState::Active) {
1439 0 : return Err(WaitLsnError::BadState(state));
1440 449442 : }
1441 449442 :
1442 449442 : if cfg!(debug_assertions) {
1443 449442 : match ctx.task_kind() {
1444 : TaskKind::WalReceiverManager
1445 : | TaskKind::WalReceiverConnectionHandler
1446 : | TaskKind::WalReceiverConnectionPoller => {
1447 0 : let is_myself = match who_is_waiting {
1448 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1449 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService | WaitLsnWaiter::HttpEndpoint => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1450 : };
1451 0 : if is_myself {
1452 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1453 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1454 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1455 0 : }
1456 0 : } else {
1457 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1458 0 : // our walreceiver task can make progress independent of theirs
1459 0 : }
1460 : }
1461 449442 : _ => {}
1462 : }
1463 0 : }
1464 :
1465 449442 : let timeout = match timeout {
1466 0 : WaitLsnTimeout::Custom(t) => t,
1467 449442 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1468 : };
1469 :
1470 449442 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1471 449442 :
1472 449442 : match self.last_record_lsn.wait_for_timeout(lsn, timeout).await {
1473 449442 : Ok(()) => Ok(()),
1474 0 : Err(e) => {
1475 : use utils::seqwait::SeqWaitError::*;
1476 0 : match e {
1477 0 : Shutdown => Err(WaitLsnError::Shutdown),
1478 : Timeout => {
1479 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1480 0 : drop(_timer);
1481 0 : let walreceiver_status = self.walreceiver_status();
1482 0 : Err(WaitLsnError::Timeout(format!(
1483 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1484 0 : lsn,
1485 0 : self.get_last_record_lsn(),
1486 0 : self.get_disk_consistent_lsn(),
1487 0 : walreceiver_status,
1488 0 : )))
1489 : }
1490 : }
1491 : }
1492 : }
1493 449442 : }
1494 :
1495 0 : pub(crate) fn walreceiver_status(&self) -> String {
1496 0 : match &*self.walreceiver.lock().unwrap() {
1497 0 : None => "stopping or stopped".to_string(),
1498 0 : Some(walreceiver) => match walreceiver.status() {
1499 0 : Some(status) => status.to_human_readable_string(),
1500 0 : None => "Not active".to_string(),
1501 : },
1502 : }
1503 0 : }
1504 :
1505 : /// Check that it is valid to request operations with that lsn.
1506 464 : pub(crate) fn check_lsn_is_in_scope(
1507 464 : &self,
1508 464 : lsn: Lsn,
1509 464 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1510 464 : ) -> anyhow::Result<()> {
1511 464 : ensure!(
1512 464 : lsn >= **latest_gc_cutoff_lsn,
1513 8 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1514 8 : lsn,
1515 8 : **latest_gc_cutoff_lsn,
1516 : );
1517 456 : Ok(())
1518 464 : }
1519 :
1520 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1521 20 : pub(crate) fn init_lsn_lease(
1522 20 : &self,
1523 20 : lsn: Lsn,
1524 20 : length: Duration,
1525 20 : ctx: &RequestContext,
1526 20 : ) -> anyhow::Result<LsnLease> {
1527 20 : self.make_lsn_lease(lsn, length, true, ctx)
1528 20 : }
1529 :
1530 : /// 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.
1531 8 : pub(crate) fn renew_lsn_lease(
1532 8 : &self,
1533 8 : lsn: Lsn,
1534 8 : length: Duration,
1535 8 : ctx: &RequestContext,
1536 8 : ) -> anyhow::Result<LsnLease> {
1537 8 : self.make_lsn_lease(lsn, length, false, ctx)
1538 8 : }
1539 :
1540 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1541 : ///
1542 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1543 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1544 : ///
1545 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1546 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1547 28 : fn make_lsn_lease(
1548 28 : &self,
1549 28 : lsn: Lsn,
1550 28 : length: Duration,
1551 28 : init: bool,
1552 28 : _ctx: &RequestContext,
1553 28 : ) -> anyhow::Result<LsnLease> {
1554 24 : let lease = {
1555 : // Normalize the requested LSN to be aligned, and move to the first record
1556 : // if it points to the beginning of the page (header).
1557 28 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1558 28 :
1559 28 : let mut gc_info = self.gc_info.write().unwrap();
1560 28 :
1561 28 : let valid_until = SystemTime::now() + length;
1562 28 :
1563 28 : let entry = gc_info.leases.entry(lsn);
1564 28 :
1565 28 : match entry {
1566 12 : Entry::Occupied(mut occupied) => {
1567 12 : let existing_lease = occupied.get_mut();
1568 12 : if valid_until > existing_lease.valid_until {
1569 4 : existing_lease.valid_until = valid_until;
1570 4 : let dt: DateTime<Utc> = valid_until.into();
1571 4 : info!("lease extended to {}", dt);
1572 : } else {
1573 8 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1574 8 : info!("existing lease covers greater length, valid until {}", dt);
1575 : }
1576 :
1577 12 : existing_lease.clone()
1578 : }
1579 16 : Entry::Vacant(vacant) => {
1580 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
1581 : // not blocked by the lsn lease deadline.
1582 16 : let validate = {
1583 16 : let conf = self.tenant_conf.load();
1584 16 : conf.location.attach_mode == AttachmentMode::Single
1585 16 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1586 : };
1587 :
1588 16 : if init || validate {
1589 16 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1590 16 : if lsn < *latest_gc_cutoff_lsn {
1591 4 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1592 12 : }
1593 0 : }
1594 :
1595 12 : let dt: DateTime<Utc> = valid_until.into();
1596 12 : info!("lease created, valid until {}", dt);
1597 12 : vacant.insert(LsnLease { valid_until }).clone()
1598 : }
1599 : }
1600 : };
1601 :
1602 24 : Ok(lease)
1603 28 : }
1604 :
1605 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1606 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1607 : #[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))]
1608 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1609 : self.freeze0().await
1610 : }
1611 :
1612 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1613 : #[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))]
1614 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1615 : self.freeze_and_flush0().await
1616 : }
1617 :
1618 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1619 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1620 2240 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1621 2240 : let mut g = self.write_lock.lock().await;
1622 2240 : let to_lsn = self.get_last_record_lsn();
1623 2240 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1624 2240 : }
1625 :
1626 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1627 : // polluting the span hierarchy.
1628 2240 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1629 2240 : let token = self.freeze0().await?;
1630 2240 : self.wait_flush_completion(token).await
1631 2240 : }
1632 :
1633 : // Check if an open ephemeral layer should be closed: this provides
1634 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1635 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1636 : // ephemeral layer bytes has been breached.
1637 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1638 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1639 : // If the write lock is held, there is an active wal receiver: rolling open layers
1640 : // is their responsibility while they hold this lock.
1641 0 : return;
1642 : };
1643 :
1644 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1645 : // time, and this was missed.
1646 : // if write_guard.is_none() { return; }
1647 :
1648 0 : let Ok(layers_guard) = self.layers.try_read() else {
1649 : // Don't block if the layer lock is busy
1650 0 : return;
1651 : };
1652 :
1653 0 : let Ok(lm) = layers_guard.layer_map() else {
1654 0 : return;
1655 : };
1656 :
1657 0 : let Some(open_layer) = &lm.open_layer else {
1658 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1659 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1660 : // that didn't result in writes to this shard.
1661 :
1662 : // Must not hold the layers lock while waiting for a flush.
1663 0 : drop(layers_guard);
1664 0 :
1665 0 : let last_record_lsn = self.get_last_record_lsn();
1666 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1667 0 : if last_record_lsn > disk_consistent_lsn {
1668 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1669 : // we are a sharded tenant and have skipped some WAL
1670 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1671 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1672 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1673 : // without any data ingested (yet) doesn't write a remote index as soon as it
1674 : // sees its LSN advance: we only do this if we've been layer-less
1675 : // for some time.
1676 0 : tracing::debug!(
1677 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1678 : disk_consistent_lsn,
1679 : last_record_lsn
1680 : );
1681 :
1682 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1683 : // We know there is no open layer, so we can request freezing without actually
1684 : // freezing anything. This is true even if we have dropped the layers_guard, we
1685 : // still hold the write_guard.
1686 0 : let _ = async {
1687 0 : let token = self
1688 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1689 0 : .await?;
1690 0 : self.wait_flush_completion(token).await
1691 0 : }
1692 0 : .await;
1693 0 : }
1694 0 : }
1695 :
1696 0 : return;
1697 : };
1698 :
1699 0 : let Some(current_size) = open_layer.try_len() else {
1700 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1701 : // read lock to get size should always succeed.
1702 0 : tracing::warn!("Lock conflict while reading size of open layer");
1703 0 : return;
1704 : };
1705 :
1706 0 : let current_lsn = self.get_last_record_lsn();
1707 :
1708 0 : let checkpoint_distance_override = open_layer.tick().await;
1709 :
1710 0 : if let Some(size_override) = checkpoint_distance_override {
1711 0 : if current_size > size_override {
1712 : // This is not harmful, but it only happens in relatively rare cases where
1713 : // time-based checkpoints are not happening fast enough to keep the amount of
1714 : // ephemeral data within configured limits. It's a sign of stress on the system.
1715 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1716 0 : }
1717 0 : }
1718 :
1719 0 : let checkpoint_distance =
1720 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1721 0 :
1722 0 : if self.should_roll(
1723 0 : current_size,
1724 0 : current_size,
1725 0 : checkpoint_distance,
1726 0 : self.get_last_record_lsn(),
1727 0 : self.last_freeze_at.load(),
1728 0 : open_layer.get_opened_at(),
1729 0 : ) {
1730 0 : match open_layer.info() {
1731 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1732 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1733 0 : // happens asynchronously in the background.
1734 0 : tracing::debug!(
1735 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1736 : );
1737 : }
1738 : InMemoryLayerInfo::Open { .. } => {
1739 : // Upgrade to a write lock and freeze the layer
1740 0 : drop(layers_guard);
1741 0 : let res = self
1742 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1743 0 : .await;
1744 :
1745 0 : if let Err(e) = res {
1746 0 : tracing::info!(
1747 0 : "failed to flush frozen layer after background freeze: {e:#}"
1748 : );
1749 0 : }
1750 : }
1751 : }
1752 0 : }
1753 0 : }
1754 :
1755 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1756 : ///
1757 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1758 : /// child timelines that are not offloaded yet.
1759 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1760 0 : if self.remote_client.is_archived() != Some(true) {
1761 0 : return (false, "the timeline is not archived");
1762 0 : }
1763 0 : if !self.remote_client.no_pending_work() {
1764 : // if the remote client is still processing some work, we can't offload
1765 0 : return (false, "the upload queue is not drained yet");
1766 0 : }
1767 0 :
1768 0 : (true, "ok")
1769 0 : }
1770 :
1771 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1772 : /// compaction tasks.
1773 768 : pub(crate) async fn compact(
1774 768 : self: &Arc<Self>,
1775 768 : cancel: &CancellationToken,
1776 768 : flags: EnumSet<CompactFlags>,
1777 768 : ctx: &RequestContext,
1778 768 : ) -> Result<CompactionOutcome, CompactionError> {
1779 768 : self.compact_with_options(
1780 768 : cancel,
1781 768 : CompactOptions {
1782 768 : flags,
1783 768 : compact_key_range: None,
1784 768 : compact_lsn_range: None,
1785 768 : sub_compaction: false,
1786 768 : sub_compaction_max_job_size_mb: None,
1787 768 : },
1788 768 : ctx,
1789 768 : )
1790 768 : .await
1791 768 : }
1792 :
1793 : /// Outermost timeline compaction operation; downloads needed layers.
1794 : ///
1795 : /// NB: the cancellation token is usually from a background task, but can also come from a
1796 : /// request task.
1797 768 : pub(crate) async fn compact_with_options(
1798 768 : self: &Arc<Self>,
1799 768 : cancel: &CancellationToken,
1800 768 : options: CompactOptions,
1801 768 : ctx: &RequestContext,
1802 768 : ) -> Result<CompactionOutcome, CompactionError> {
1803 768 : // Acquire the compaction lock and task semaphore.
1804 768 : //
1805 768 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
1806 768 : // out by other background tasks (including image compaction). We request this via
1807 768 : // `BackgroundLoopKind::L0Compaction`.
1808 768 : //
1809 768 : // If this is a regular compaction pass, and L0-only compaction is enabled in the config,
1810 768 : // then we should yield for immediate L0 compaction if necessary while we're waiting for the
1811 768 : // background task semaphore. There's no point yielding otherwise, since we'd just end up
1812 768 : // right back here.
1813 768 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
1814 768 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
1815 0 : true => BackgroundLoopKind::L0Compaction,
1816 768 : false => BackgroundLoopKind::Compaction,
1817 : };
1818 768 : let yield_for_l0 = !is_l0_only
1819 768 : && self.get_compaction_l0_first()
1820 0 : && !options.flags.contains(CompactFlags::NoYield);
1821 :
1822 768 : let acquire = async move {
1823 768 : let guard = self.compaction_lock.lock().await;
1824 768 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
1825 768 : (guard, permit)
1826 768 : };
1827 :
1828 768 : let (_guard, _permit) = tokio::select! {
1829 768 : (guard, permit) = acquire => (guard, permit),
1830 768 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
1831 0 : return Ok(CompactionOutcome::YieldForL0);
1832 : }
1833 768 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
1834 768 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
1835 : };
1836 :
1837 768 : let last_record_lsn = self.get_last_record_lsn();
1838 768 :
1839 768 : // Last record Lsn could be zero in case the timeline was just created
1840 768 : if !last_record_lsn.is_valid() {
1841 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1842 0 : return Ok(CompactionOutcome::Skipped);
1843 768 : }
1844 :
1845 768 : let result = match self.get_compaction_algorithm_settings().kind {
1846 : CompactionAlgorithm::Tiered => {
1847 0 : self.compact_tiered(cancel, ctx).await?;
1848 0 : Ok(CompactionOutcome::Done)
1849 : }
1850 768 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1851 : };
1852 :
1853 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
1854 0 : match result {
1855 768 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
1856 : Err(CompactionError::Other(_)) => {
1857 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
1858 : }
1859 : // Don't change the current value on offload failure or shutdown. We don't want to
1860 : // abruptly stall nor resume L0 flushes in these cases.
1861 0 : Err(CompactionError::Offload(_)) => {}
1862 0 : Err(CompactionError::ShuttingDown) => {}
1863 : };
1864 :
1865 768 : result
1866 768 : }
1867 :
1868 : /// Mutate the timeline with a [`TimelineWriter`].
1869 10266380 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1870 10266380 : TimelineWriter {
1871 10266380 : tl: self,
1872 10266380 : write_guard: self.write_lock.lock().await,
1873 : }
1874 10266380 : }
1875 :
1876 0 : pub(crate) fn activate(
1877 0 : self: &Arc<Self>,
1878 0 : parent: Arc<crate::tenant::Tenant>,
1879 0 : broker_client: BrokerClientChannel,
1880 0 : background_jobs_can_start: Option<&completion::Barrier>,
1881 0 : ctx: &RequestContext,
1882 0 : ) {
1883 0 : if self.tenant_shard_id.is_shard_zero() {
1884 0 : // Logical size is only maintained accurately on shard zero.
1885 0 : self.spawn_initial_logical_size_computation_task(ctx);
1886 0 : }
1887 0 : self.launch_wal_receiver(ctx, broker_client);
1888 0 : self.set_state(TimelineState::Active);
1889 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1890 0 : }
1891 :
1892 : /// After this function returns, there are no timeline-scoped tasks are left running.
1893 : ///
1894 : /// The preferred pattern for is:
1895 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1896 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1897 : /// go the extra mile and keep track of JoinHandles
1898 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1899 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1900 : ///
1901 : /// For legacy reasons, we still have multiple tasks spawned using
1902 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1903 : /// We refer to these as "timeline-scoped task_mgr tasks".
1904 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1905 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1906 : /// or [`task_mgr::shutdown_watcher`].
1907 : /// We want to gradually convert the code base away from these.
1908 : ///
1909 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1910 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1911 : /// ones that aren't mentioned here):
1912 : /// - [`TaskKind::TimelineDeletionWorker`]
1913 : /// - NB: also used for tenant deletion
1914 : /// - [`TaskKind::RemoteUploadTask`]`
1915 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1916 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1917 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1918 : /// - [`TaskKind::Eviction`]
1919 : /// - [`TaskKind::LayerFlushTask`]
1920 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1921 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1922 20 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1923 20 : debug_assert_current_span_has_tenant_and_timeline_id();
1924 20 :
1925 20 : // Regardless of whether we're going to try_freeze_and_flush
1926 20 : // or not, stop ingesting any more data. Walreceiver only provides
1927 20 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1928 20 : // So, only after the self.gate.close() below will we know for sure that
1929 20 : // no walreceiver tasks are left.
1930 20 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1931 20 : // data during the call to `self.freeze_and_flush()` below.
1932 20 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1933 20 : // which is what we'd need to properly model early shutdown of the walreceiver
1934 20 : // task sub-tree before the other Timeline task sub-trees.
1935 20 : let walreceiver = self.walreceiver.lock().unwrap().take();
1936 20 : tracing::debug!(
1937 0 : is_some = walreceiver.is_some(),
1938 0 : "Waiting for WalReceiverManager..."
1939 : );
1940 20 : if let Some(walreceiver) = walreceiver {
1941 0 : walreceiver.cancel();
1942 20 : }
1943 : // ... and inform any waiters for newer LSNs that there won't be any.
1944 20 : self.last_record_lsn.shutdown();
1945 20 :
1946 20 : if let ShutdownMode::FreezeAndFlush = mode {
1947 12 : let do_flush = if let Some((open, frozen)) = self
1948 12 : .layers
1949 12 : .read()
1950 12 : .await
1951 12 : .layer_map()
1952 12 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1953 12 : .ok()
1954 12 : .filter(|(open, frozen)| *open || *frozen > 0)
1955 : {
1956 0 : if self.remote_client.is_archived() == Some(true) {
1957 : // No point flushing on shutdown for an archived timeline: it is not important
1958 : // to have it nice and fresh after our restart, and trying to flush here might
1959 : // race with trying to offload it (which also stops the flush loop)
1960 0 : false
1961 : } else {
1962 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1963 0 : true
1964 : }
1965 : } else {
1966 : // this is double-shutdown, it'll be a no-op
1967 12 : true
1968 : };
1969 :
1970 : // we shut down walreceiver above, so, we won't add anything more
1971 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1972 : // to reach the disk & upload queue, then shut the upload queue and
1973 : // wait for it to drain.
1974 12 : if do_flush {
1975 12 : match self.freeze_and_flush().await {
1976 : Ok(_) => {
1977 : // drain the upload queue
1978 : // if we did not wait for completion here, it might be our shutdown process
1979 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1980 : // be spawned.
1981 : //
1982 : // what is problematic is the shutting down of RemoteTimelineClient, because
1983 : // obviously it does not make sense to stop while we wait for it, but what
1984 : // about corner cases like s3 suddenly hanging up?
1985 12 : self.remote_client.shutdown().await;
1986 : }
1987 : Err(FlushLayerError::Cancelled) => {
1988 : // this is likely the second shutdown, ignore silently.
1989 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1990 0 : debug_assert!(self.cancel.is_cancelled());
1991 : }
1992 0 : Err(e) => {
1993 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1994 0 : // we have some extra WAL replay to do next time the timeline starts.
1995 0 : warn!("failed to freeze and flush: {e:#}");
1996 : }
1997 : }
1998 :
1999 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2000 : // we also do a final check here to ensure that the queue is empty.
2001 12 : if !self.remote_client.no_pending_work() {
2002 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
2003 12 : }
2004 0 : }
2005 8 : }
2006 :
2007 20 : if let ShutdownMode::Reload = mode {
2008 : // drain the upload queue
2009 4 : self.remote_client.shutdown().await;
2010 4 : if !self.remote_client.no_pending_work() {
2011 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
2012 4 : }
2013 16 : }
2014 :
2015 : // Signal any subscribers to our cancellation token to drop out
2016 20 : tracing::debug!("Cancelling CancellationToken");
2017 20 : self.cancel.cancel();
2018 20 :
2019 20 : // Ensure Prevent new page service requests from starting.
2020 20 : self.handles.shutdown();
2021 20 :
2022 20 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2023 20 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2024 20 : self.remote_client.stop();
2025 20 :
2026 20 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2027 20 : // to shut down the upload queue tasks.
2028 20 : // TODO: fix that, task management should be encapsulated inside remote_client.
2029 20 : task_mgr::shutdown_tasks(
2030 20 : Some(TaskKind::RemoteUploadTask),
2031 20 : Some(self.tenant_shard_id),
2032 20 : Some(self.timeline_id),
2033 20 : )
2034 20 : .await;
2035 :
2036 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2037 20 : tracing::debug!("Waiting for tasks...");
2038 20 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2039 :
2040 : {
2041 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2042 : // open.
2043 20 : let mut write_guard = self.write_lock.lock().await;
2044 20 : self.layers.write().await.shutdown(&mut write_guard);
2045 20 : }
2046 20 :
2047 20 : // Finally wait until any gate-holders are complete.
2048 20 : //
2049 20 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2050 20 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2051 20 : self.gate.close().await;
2052 :
2053 20 : self.metrics.shutdown();
2054 20 : }
2055 :
2056 896 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2057 896 : match (self.current_state(), new_state) {
2058 896 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2059 4 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2060 : }
2061 0 : (st, TimelineState::Loading) => {
2062 0 : error!("ignoring transition from {st:?} into Loading state");
2063 : }
2064 0 : (TimelineState::Broken { .. }, new_state) => {
2065 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2066 : }
2067 : (TimelineState::Stopping, TimelineState::Active) => {
2068 0 : error!("Not activating a Stopping timeline");
2069 : }
2070 892 : (_, new_state) => {
2071 892 : self.state.send_replace(new_state);
2072 892 : }
2073 : }
2074 896 : }
2075 :
2076 4 : pub(crate) fn set_broken(&self, reason: String) {
2077 4 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2078 4 : let broken_state = TimelineState::Broken {
2079 4 : reason,
2080 4 : backtrace: backtrace_str,
2081 4 : };
2082 4 : self.set_state(broken_state);
2083 4 :
2084 4 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2085 4 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2086 4 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2087 4 : self.cancel.cancel();
2088 4 : }
2089 :
2090 451558 : pub(crate) fn current_state(&self) -> TimelineState {
2091 451558 : self.state.borrow().clone()
2092 451558 : }
2093 :
2094 12 : pub(crate) fn is_broken(&self) -> bool {
2095 12 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2096 12 : }
2097 :
2098 444 : pub(crate) fn is_active(&self) -> bool {
2099 444 : self.current_state() == TimelineState::Active
2100 444 : }
2101 :
2102 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
2103 0 : self.remote_client.is_archived()
2104 0 : }
2105 :
2106 776 : pub(crate) fn is_stopping(&self) -> bool {
2107 776 : self.current_state() == TimelineState::Stopping
2108 776 : }
2109 :
2110 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2111 0 : self.state.subscribe()
2112 0 : }
2113 :
2114 449446 : pub(crate) async fn wait_to_become_active(
2115 449446 : &self,
2116 449446 : _ctx: &RequestContext, // Prepare for use by cancellation
2117 449446 : ) -> Result<(), TimelineState> {
2118 449446 : let mut receiver = self.state.subscribe();
2119 : loop {
2120 449446 : let current_state = receiver.borrow().clone();
2121 449446 : match current_state {
2122 : TimelineState::Loading => {
2123 0 : receiver
2124 0 : .changed()
2125 0 : .await
2126 0 : .expect("holding a reference to self");
2127 : }
2128 : TimelineState::Active { .. } => {
2129 449442 : return Ok(());
2130 : }
2131 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2132 : // There's no chance the timeline can transition back into ::Active
2133 4 : return Err(current_state);
2134 : }
2135 : }
2136 : }
2137 449446 : }
2138 :
2139 0 : pub(crate) async fn layer_map_info(
2140 0 : &self,
2141 0 : reset: LayerAccessStatsReset,
2142 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2143 0 : let guard = self.layers.read().await;
2144 0 : let layer_map = guard.layer_map()?;
2145 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2146 0 : if let Some(open_layer) = &layer_map.open_layer {
2147 0 : in_memory_layers.push(open_layer.info());
2148 0 : }
2149 0 : for frozen_layer in &layer_map.frozen_layers {
2150 0 : in_memory_layers.push(frozen_layer.info());
2151 0 : }
2152 :
2153 0 : let historic_layers = layer_map
2154 0 : .iter_historic_layers()
2155 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2156 0 : .collect();
2157 0 :
2158 0 : Ok(LayerMapInfo {
2159 0 : in_memory_layers,
2160 0 : historic_layers,
2161 0 : })
2162 0 : }
2163 :
2164 : #[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))]
2165 : pub(crate) async fn download_layer(
2166 : &self,
2167 : layer_file_name: &LayerName,
2168 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2169 : let Some(layer) = self
2170 : .find_layer(layer_file_name)
2171 : .await
2172 0 : .map_err(|e| match e {
2173 0 : layer_manager::Shutdown => {
2174 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2175 0 : }
2176 0 : })?
2177 : else {
2178 : return Ok(None);
2179 : };
2180 :
2181 : layer.download().await?;
2182 :
2183 : Ok(Some(true))
2184 : }
2185 :
2186 : /// Evict just one layer.
2187 : ///
2188 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2189 0 : pub(crate) async fn evict_layer(
2190 0 : &self,
2191 0 : layer_file_name: &LayerName,
2192 0 : ) -> anyhow::Result<Option<bool>> {
2193 0 : let _gate = self
2194 0 : .gate
2195 0 : .enter()
2196 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2197 :
2198 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2199 0 : return Ok(None);
2200 : };
2201 :
2202 : // curl has this by default
2203 0 : let timeout = std::time::Duration::from_secs(120);
2204 0 :
2205 0 : match local_layer.evict_and_wait(timeout).await {
2206 0 : Ok(()) => Ok(Some(true)),
2207 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2208 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2209 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2210 : }
2211 0 : }
2212 :
2213 9606020 : fn should_roll(
2214 9606020 : &self,
2215 9606020 : layer_size: u64,
2216 9606020 : projected_layer_size: u64,
2217 9606020 : checkpoint_distance: u64,
2218 9606020 : projected_lsn: Lsn,
2219 9606020 : last_freeze_at: Lsn,
2220 9606020 : opened_at: Instant,
2221 9606020 : ) -> bool {
2222 9606020 : let distance = projected_lsn.widening_sub(last_freeze_at);
2223 9606020 :
2224 9606020 : // Rolling the open layer can be triggered by:
2225 9606020 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2226 9606020 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2227 9606020 : // account for how writes are distributed across shards: we expect each node to consume
2228 9606020 : // 1/count of the LSN on average.
2229 9606020 : // 2. The size of the currently open layer.
2230 9606020 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2231 9606020 : // up and suspend activity.
2232 9606020 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2233 0 : info!(
2234 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2235 : projected_lsn, layer_size, distance
2236 : );
2237 :
2238 0 : true
2239 9606020 : } else if projected_layer_size >= checkpoint_distance {
2240 : // NB: this check is relied upon by:
2241 160 : let _ = IndexEntry::validate_checkpoint_distance;
2242 160 : info!(
2243 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2244 : projected_lsn, layer_size, projected_layer_size
2245 : );
2246 :
2247 160 : true
2248 9605860 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2249 0 : info!(
2250 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2251 0 : projected_lsn,
2252 0 : layer_size,
2253 0 : opened_at.elapsed()
2254 : );
2255 :
2256 0 : true
2257 : } else {
2258 9605860 : false
2259 : }
2260 9606020 : }
2261 : }
2262 :
2263 : /// Number of times we will compute partition within a checkpoint distance.
2264 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2265 :
2266 : // Private functions
2267 : impl Timeline {
2268 24 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2269 24 : let tenant_conf = self.tenant_conf.load();
2270 24 : tenant_conf
2271 24 : .tenant_conf
2272 24 : .lsn_lease_length
2273 24 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2274 24 : }
2275 :
2276 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2277 0 : let tenant_conf = self.tenant_conf.load();
2278 0 : tenant_conf
2279 0 : .tenant_conf
2280 0 : .lsn_lease_length_for_ts
2281 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2282 0 : }
2283 :
2284 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2285 0 : let tenant_conf = self.tenant_conf.load();
2286 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2287 0 : }
2288 :
2289 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2290 0 : let tenant_conf = self.tenant_conf.load();
2291 0 : tenant_conf
2292 0 : .tenant_conf
2293 0 : .lazy_slru_download
2294 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2295 0 : }
2296 :
2297 9609108 : fn get_checkpoint_distance(&self) -> u64 {
2298 9609108 : let tenant_conf = self.tenant_conf.load();
2299 9609108 : tenant_conf
2300 9609108 : .tenant_conf
2301 9609108 : .checkpoint_distance
2302 9609108 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2303 9609108 : }
2304 :
2305 9605860 : fn get_checkpoint_timeout(&self) -> Duration {
2306 9605860 : let tenant_conf = self.tenant_conf.load();
2307 9605860 : tenant_conf
2308 9605860 : .tenant_conf
2309 9605860 : .checkpoint_timeout
2310 9605860 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2311 9605860 : }
2312 :
2313 5008 : fn get_compaction_period(&self) -> Duration {
2314 5008 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2315 5008 : tenant_conf
2316 5008 : .compaction_period
2317 5008 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2318 5008 : }
2319 :
2320 1364 : fn get_compaction_target_size(&self) -> u64 {
2321 1364 : let tenant_conf = self.tenant_conf.load();
2322 1364 : tenant_conf
2323 1364 : .tenant_conf
2324 1364 : .compaction_target_size
2325 1364 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2326 1364 : }
2327 :
2328 3204 : fn get_compaction_threshold(&self) -> usize {
2329 3204 : let tenant_conf = self.tenant_conf.load();
2330 3204 : tenant_conf
2331 3204 : .tenant_conf
2332 3204 : .compaction_threshold
2333 3204 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2334 3204 : }
2335 :
2336 92 : fn get_compaction_upper_limit(&self) -> usize {
2337 92 : let tenant_conf = self.tenant_conf.load();
2338 92 : tenant_conf
2339 92 : .tenant_conf
2340 92 : .compaction_upper_limit
2341 92 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2342 92 : }
2343 :
2344 768 : pub fn get_compaction_l0_first(&self) -> bool {
2345 768 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2346 768 : tenant_conf
2347 768 : .compaction_l0_first
2348 768 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2349 768 : }
2350 :
2351 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2352 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2353 0 : tenant_conf
2354 0 : .compaction_l0_semaphore
2355 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2356 0 : }
2357 :
2358 2504 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2359 : // Disable L0 flushes by default. This and compaction needs further tuning.
2360 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 0; // TODO: default to e.g. 3
2361 :
2362 : // If compaction is disabled, don't delay.
2363 2504 : if self.get_compaction_period() == Duration::ZERO {
2364 2504 : return None;
2365 0 : }
2366 0 :
2367 0 : let compaction_threshold = self.get_compaction_threshold();
2368 0 : let tenant_conf = self.tenant_conf.load();
2369 0 : let l0_flush_delay_threshold = tenant_conf
2370 0 : .tenant_conf
2371 0 : .l0_flush_delay_threshold
2372 0 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2373 0 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2374 0 :
2375 0 : // 0 disables backpressure.
2376 0 : if l0_flush_delay_threshold == 0 {
2377 0 : return None;
2378 0 : }
2379 0 :
2380 0 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2381 0 : // backpressure flushes below this.
2382 0 : // TODO: the tenant config should have validation to prevent this instead.
2383 0 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2384 0 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2385 2504 : }
2386 :
2387 2504 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2388 : // Disable L0 stalls by default. In ingest benchmarks, we see image compaction take >10
2389 : // minutes, blocking L0 compaction, and we can't stall L0 flushes for that long.
2390 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2391 :
2392 : // If compaction is disabled, don't stall.
2393 2504 : if self.get_compaction_period() == Duration::ZERO {
2394 2504 : return None;
2395 0 : }
2396 0 :
2397 0 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2398 0 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2399 0 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2400 0 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2401 0 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2402 0 : return None;
2403 0 : }
2404 0 :
2405 0 : let compaction_threshold = self.get_compaction_threshold();
2406 0 : let tenant_conf = self.tenant_conf.load();
2407 0 : let l0_flush_stall_threshold = tenant_conf
2408 0 : .tenant_conf
2409 0 : .l0_flush_stall_threshold
2410 0 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2411 0 :
2412 0 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2413 0 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2414 0 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2415 0 : if cfg!(feature = "testing")
2416 0 : && compaction_threshold == 1
2417 0 : && l0_flush_stall_threshold.is_none()
2418 : {
2419 0 : return None;
2420 0 : }
2421 0 :
2422 0 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2423 0 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2424 0 :
2425 0 : // 0 disables backpressure.
2426 0 : if l0_flush_stall_threshold == 0 {
2427 0 : return None;
2428 0 : }
2429 0 :
2430 0 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2431 0 : // backpressure flushes below this.
2432 0 : // TODO: the tenant config should have validation to prevent this instead.
2433 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2434 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2435 2504 : }
2436 :
2437 2344 : fn get_l0_flush_wait_upload(&self) -> bool {
2438 2344 : let tenant_conf = self.tenant_conf.load();
2439 2344 : tenant_conf
2440 2344 : .tenant_conf
2441 2344 : .l0_flush_wait_upload
2442 2344 : .unwrap_or(self.conf.default_tenant_conf.l0_flush_wait_upload)
2443 2344 : }
2444 :
2445 0 : fn get_image_creation_threshold(&self) -> usize {
2446 0 : let tenant_conf = self.tenant_conf.load();
2447 0 : tenant_conf
2448 0 : .tenant_conf
2449 0 : .image_creation_threshold
2450 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2451 0 : }
2452 :
2453 768 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2454 768 : let tenant_conf = &self.tenant_conf.load();
2455 768 : tenant_conf
2456 768 : .tenant_conf
2457 768 : .compaction_algorithm
2458 768 : .as_ref()
2459 768 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2460 768 : .clone()
2461 768 : }
2462 :
2463 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2464 0 : let tenant_conf = self.tenant_conf.load();
2465 0 : tenant_conf
2466 0 : .tenant_conf
2467 0 : .eviction_policy
2468 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2469 0 : }
2470 :
2471 892 : fn get_evictions_low_residence_duration_metric_threshold(
2472 892 : tenant_conf: &TenantConfOpt,
2473 892 : default_tenant_conf: &TenantConf,
2474 892 : ) -> Duration {
2475 892 : tenant_conf
2476 892 : .evictions_low_residence_duration_metric_threshold
2477 892 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2478 892 : }
2479 :
2480 1176 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2481 1176 : let tenant_conf = self.tenant_conf.load();
2482 1176 : tenant_conf
2483 1176 : .tenant_conf
2484 1176 : .image_layer_creation_check_threshold
2485 1176 : .unwrap_or(
2486 1176 : self.conf
2487 1176 : .default_tenant_conf
2488 1176 : .image_layer_creation_check_threshold,
2489 1176 : )
2490 1176 : }
2491 :
2492 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
2493 0 : let tenant_conf = self.tenant_conf.load();
2494 0 : tenant_conf
2495 0 : .tenant_conf
2496 0 : .image_creation_preempt_threshold
2497 0 : .unwrap_or(
2498 0 : self.conf
2499 0 : .default_tenant_conf
2500 0 : .image_creation_preempt_threshold,
2501 0 : )
2502 0 : }
2503 :
2504 : /// Resolve the effective WAL receiver protocol to use for this tenant.
2505 : ///
2506 : /// Priority order is:
2507 : /// 1. Tenant config override
2508 : /// 2. Default value for tenant config override
2509 : /// 3. Pageserver config override
2510 : /// 4. Pageserver config default
2511 0 : pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
2512 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2513 0 : tenant_conf
2514 0 : .wal_receiver_protocol_override
2515 0 : .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
2516 0 : .unwrap_or(self.conf.wal_receiver_protocol)
2517 0 : }
2518 :
2519 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2520 0 : // NB: Most tenant conf options are read by background loops, so,
2521 0 : // changes will automatically be picked up.
2522 0 :
2523 0 : // The threshold is embedded in the metric. So, we need to update it.
2524 0 : {
2525 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2526 0 : &new_conf.tenant_conf,
2527 0 : &self.conf.default_tenant_conf,
2528 0 : );
2529 0 :
2530 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2531 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2532 0 :
2533 0 : let timeline_id_str = self.timeline_id.to_string();
2534 0 :
2535 0 : self.remote_client.update_config(&new_conf.location);
2536 0 :
2537 0 : self.metrics
2538 0 : .evictions_with_low_residence_duration
2539 0 : .write()
2540 0 : .unwrap()
2541 0 : .change_threshold(
2542 0 : &tenant_id_str,
2543 0 : &shard_id_str,
2544 0 : &timeline_id_str,
2545 0 : new_threshold,
2546 0 : );
2547 0 : }
2548 0 : }
2549 :
2550 : /// Open a Timeline handle.
2551 : ///
2552 : /// Loads the metadata for the timeline into memory, but not the layer map.
2553 : #[allow(clippy::too_many_arguments)]
2554 892 : pub(super) fn new(
2555 892 : conf: &'static PageServerConf,
2556 892 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2557 892 : metadata: &TimelineMetadata,
2558 892 : ancestor: Option<Arc<Timeline>>,
2559 892 : timeline_id: TimelineId,
2560 892 : tenant_shard_id: TenantShardId,
2561 892 : generation: Generation,
2562 892 : shard_identity: ShardIdentity,
2563 892 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2564 892 : resources: TimelineResources,
2565 892 : pg_version: u32,
2566 892 : state: TimelineState,
2567 892 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2568 892 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2569 892 : cancel: CancellationToken,
2570 892 : ) -> Arc<Self> {
2571 892 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2572 892 : let (state, _) = watch::channel(state);
2573 892 :
2574 892 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2575 892 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2576 892 :
2577 892 : let evictions_low_residence_duration_metric_threshold = {
2578 892 : let loaded_tenant_conf = tenant_conf.load();
2579 892 : Self::get_evictions_low_residence_duration_metric_threshold(
2580 892 : &loaded_tenant_conf.tenant_conf,
2581 892 : &conf.default_tenant_conf,
2582 892 : )
2583 : };
2584 :
2585 892 : if let Some(ancestor) = &ancestor {
2586 460 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2587 460 : // If we construct an explicit timeline object, it's obviously not offloaded
2588 460 : let is_offloaded = MaybeOffloaded::No;
2589 460 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2590 460 : }
2591 :
2592 892 : Arc::new_cyclic(|myself| {
2593 892 : let metrics = TimelineMetrics::new(
2594 892 : &tenant_shard_id,
2595 892 : &timeline_id,
2596 892 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2597 892 : "mtime",
2598 892 : evictions_low_residence_duration_metric_threshold,
2599 892 : ),
2600 892 : );
2601 892 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2602 :
2603 892 : let mut result = Timeline {
2604 892 : conf,
2605 892 : tenant_conf,
2606 892 : myself: myself.clone(),
2607 892 : timeline_id,
2608 892 : tenant_shard_id,
2609 892 : generation,
2610 892 : shard_identity,
2611 892 : pg_version,
2612 892 : layers: Default::default(),
2613 892 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
2614 892 :
2615 892 : walredo_mgr,
2616 892 : walreceiver: Mutex::new(None),
2617 892 :
2618 892 : remote_client: Arc::new(resources.remote_client),
2619 892 :
2620 892 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2621 892 : last_record_lsn: SeqWait::new(RecordLsn {
2622 892 : last: disk_consistent_lsn,
2623 892 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2624 892 : }),
2625 892 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2626 892 :
2627 892 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2628 892 : last_freeze_ts: RwLock::new(Instant::now()),
2629 892 :
2630 892 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2631 892 :
2632 892 : ancestor_timeline: ancestor,
2633 892 : ancestor_lsn: metadata.ancestor_lsn(),
2634 892 :
2635 892 : metrics,
2636 892 :
2637 892 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2638 892 : &tenant_shard_id,
2639 892 : &timeline_id,
2640 892 : resources.pagestream_throttle_metrics,
2641 892 : ),
2642 892 :
2643 6244 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2644 892 :
2645 892 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2646 892 :
2647 892 : layer_flush_start_tx,
2648 892 : layer_flush_done_tx,
2649 892 :
2650 892 : write_lock: tokio::sync::Mutex::new(None),
2651 892 :
2652 892 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2653 892 :
2654 892 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
2655 892 : LastImageLayerCreationStatus::default(),
2656 892 : )),
2657 892 :
2658 892 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2659 892 : initdb_lsn: metadata.initdb_lsn(),
2660 892 :
2661 892 : current_logical_size: if disk_consistent_lsn.is_valid() {
2662 : // we're creating timeline data with some layer files existing locally,
2663 : // need to recalculate timeline's logical size based on data in the layers.
2664 468 : LogicalSize::deferred_initial(disk_consistent_lsn)
2665 : } else {
2666 : // we're creating timeline data without any layers existing locally,
2667 : // initial logical size is 0.
2668 424 : LogicalSize::empty_initial()
2669 : },
2670 :
2671 892 : partitioning: GuardArcSwap::new((
2672 892 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2673 892 : Lsn(0),
2674 892 : )),
2675 892 : repartition_threshold: 0,
2676 892 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2677 892 : last_image_layer_creation_check_instant: Mutex::new(None),
2678 892 :
2679 892 : last_received_wal: Mutex::new(None),
2680 892 : rel_size_cache: RwLock::new(RelSizeCache {
2681 892 : complete_as_of: disk_consistent_lsn,
2682 892 : map: HashMap::new(),
2683 892 : }),
2684 892 :
2685 892 : download_all_remote_layers_task_info: RwLock::new(None),
2686 892 :
2687 892 : state,
2688 892 :
2689 892 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2690 892 : EvictionTaskTimelineState::default(),
2691 892 : ),
2692 892 : delete_progress: TimelineDeleteProgress::default(),
2693 892 :
2694 892 : cancel,
2695 892 : gate: Gate::default(),
2696 892 :
2697 892 : compaction_lock: tokio::sync::Mutex::default(),
2698 892 : compaction_failed: AtomicBool::default(),
2699 892 : l0_compaction_trigger: resources.l0_compaction_trigger,
2700 892 : gc_lock: tokio::sync::Mutex::default(),
2701 892 :
2702 892 : standby_horizon: AtomicLsn::new(0),
2703 892 :
2704 892 : pagestream_throttle: resources.pagestream_throttle,
2705 892 :
2706 892 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2707 892 :
2708 892 : #[cfg(test)]
2709 892 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2710 892 :
2711 892 : l0_flush_global_state: resources.l0_flush_global_state,
2712 892 :
2713 892 : handles: Default::default(),
2714 892 :
2715 892 : attach_wal_lag_cooldown,
2716 892 :
2717 892 : create_idempotency,
2718 892 :
2719 892 : page_trace: Default::default(),
2720 892 : };
2721 892 :
2722 892 : result.repartition_threshold =
2723 892 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2724 892 :
2725 892 : result
2726 892 : .metrics
2727 892 : .last_record_lsn_gauge
2728 892 : .set(disk_consistent_lsn.0 as i64);
2729 892 : result
2730 892 : })
2731 892 : }
2732 :
2733 1288 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2734 1288 : let Ok(guard) = self.gate.enter() else {
2735 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2736 0 : return;
2737 : };
2738 1288 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2739 1288 : match *flush_loop_state {
2740 880 : FlushLoopState::NotStarted => (),
2741 : FlushLoopState::Running { .. } => {
2742 408 : info!(
2743 0 : "skipping attempt to start flush_loop twice {}/{}",
2744 0 : self.tenant_shard_id, self.timeline_id
2745 : );
2746 408 : return;
2747 : }
2748 : FlushLoopState::Exited => {
2749 0 : info!(
2750 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2751 0 : self.tenant_shard_id, self.timeline_id
2752 : );
2753 0 : return;
2754 : }
2755 : }
2756 :
2757 880 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2758 880 : let self_clone = Arc::clone(self);
2759 880 :
2760 880 : debug!("spawning flush loop");
2761 880 : *flush_loop_state = FlushLoopState::Running {
2762 880 : #[cfg(test)]
2763 880 : expect_initdb_optimization: false,
2764 880 : #[cfg(test)]
2765 880 : initdb_optimization_count: 0,
2766 880 : };
2767 880 : task_mgr::spawn(
2768 880 : task_mgr::BACKGROUND_RUNTIME.handle(),
2769 880 : task_mgr::TaskKind::LayerFlushTask,
2770 880 : self.tenant_shard_id,
2771 880 : Some(self.timeline_id),
2772 880 : "layer flush task",
2773 880 : async move {
2774 880 : let _guard = guard;
2775 880 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2776 880 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2777 20 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2778 20 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2779 20 : *flush_loop_state = FlushLoopState::Exited;
2780 20 : Ok(())
2781 20 : }
2782 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))
2783 : );
2784 1288 : }
2785 :
2786 : /// Creates and starts the wal receiver.
2787 : ///
2788 : /// This function is expected to be called at most once per Timeline's lifecycle
2789 : /// when the timeline is activated.
2790 0 : fn launch_wal_receiver(
2791 0 : self: &Arc<Self>,
2792 0 : ctx: &RequestContext,
2793 0 : broker_client: BrokerClientChannel,
2794 0 : ) {
2795 0 : info!(
2796 0 : "launching WAL receiver for timeline {} of tenant {}",
2797 0 : self.timeline_id, self.tenant_shard_id
2798 : );
2799 :
2800 0 : let tenant_conf = self.tenant_conf.load();
2801 0 : let wal_connect_timeout = tenant_conf
2802 0 : .tenant_conf
2803 0 : .walreceiver_connect_timeout
2804 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2805 0 : let lagging_wal_timeout = tenant_conf
2806 0 : .tenant_conf
2807 0 : .lagging_wal_timeout
2808 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2809 0 : let max_lsn_wal_lag = tenant_conf
2810 0 : .tenant_conf
2811 0 : .max_lsn_wal_lag
2812 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2813 0 :
2814 0 : let mut guard = self.walreceiver.lock().unwrap();
2815 0 : assert!(
2816 0 : guard.is_none(),
2817 0 : "multiple launches / re-launches of WAL receiver are not supported"
2818 : );
2819 0 : *guard = Some(WalReceiver::start(
2820 0 : Arc::clone(self),
2821 0 : WalReceiverConf {
2822 0 : protocol: self.resolve_wal_receiver_protocol(),
2823 0 : wal_connect_timeout,
2824 0 : lagging_wal_timeout,
2825 0 : max_lsn_wal_lag,
2826 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2827 0 : availability_zone: self.conf.availability_zone.clone(),
2828 0 : ingest_batch_size: self.conf.ingest_batch_size,
2829 0 : },
2830 0 : broker_client,
2831 0 : ctx,
2832 0 : ));
2833 0 : }
2834 :
2835 : /// Initialize with an empty layer map. Used when creating a new timeline.
2836 880 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2837 880 : let mut layers = self.layers.try_write().expect(
2838 880 : "in the context where we call this function, no other task has access to the object",
2839 880 : );
2840 880 : layers
2841 880 : .open_mut()
2842 880 : .expect("in this context the LayerManager must still be open")
2843 880 : .initialize_empty(Lsn(start_lsn.0));
2844 880 : }
2845 :
2846 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2847 : /// files.
2848 12 : pub(super) async fn load_layer_map(
2849 12 : &self,
2850 12 : disk_consistent_lsn: Lsn,
2851 12 : index_part: IndexPart,
2852 12 : ) -> anyhow::Result<()> {
2853 : use init::{Decision::*, Discovered, DismissedLayer};
2854 : use LayerName::*;
2855 :
2856 12 : let mut guard = self.layers.write().await;
2857 :
2858 12 : let timer = self.metrics.load_layer_map_histo.start_timer();
2859 12 :
2860 12 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2861 12 : // structs representing all files on disk
2862 12 : let timeline_path = self
2863 12 : .conf
2864 12 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2865 12 : let conf = self.conf;
2866 12 : let span = tracing::Span::current();
2867 12 :
2868 12 : // Copy to move into the task we're about to spawn
2869 12 : let this = self.myself.upgrade().expect("&self method holds the arc");
2870 :
2871 12 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2872 12 : move || {
2873 12 : let _g = span.entered();
2874 12 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2875 12 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2876 12 : let mut unrecognized_files = Vec::new();
2877 12 :
2878 12 : let mut path = timeline_path;
2879 :
2880 44 : for discovered in discovered {
2881 32 : let (name, kind) = match discovered {
2882 32 : Discovered::Layer(layer_file_name, local_metadata) => {
2883 32 : discovered_layers.push((layer_file_name, local_metadata));
2884 32 : continue;
2885 : }
2886 0 : Discovered::IgnoredBackup(path) => {
2887 0 : std::fs::remove_file(path)
2888 0 : .or_else(fs_ext::ignore_not_found)
2889 0 : .fatal_err("Removing .old file");
2890 0 : continue;
2891 : }
2892 0 : Discovered::Unknown(file_name) => {
2893 0 : // we will later error if there are any
2894 0 : unrecognized_files.push(file_name);
2895 0 : continue;
2896 : }
2897 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2898 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2899 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2900 : };
2901 0 : path.push(Utf8Path::new(&name));
2902 0 : init::cleanup(&path, kind)?;
2903 0 : path.pop();
2904 : }
2905 :
2906 12 : if !unrecognized_files.is_empty() {
2907 : // assume that if there are any there are many many.
2908 0 : let n = unrecognized_files.len();
2909 0 : let first = &unrecognized_files[..n.min(10)];
2910 0 : anyhow::bail!(
2911 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2912 0 : );
2913 12 : }
2914 12 :
2915 12 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2916 12 :
2917 12 : let mut loaded_layers = Vec::new();
2918 12 : let mut needs_cleanup = Vec::new();
2919 12 : let mut total_physical_size = 0;
2920 :
2921 44 : for (name, decision) in decided {
2922 32 : let decision = match decision {
2923 32 : Ok(decision) => decision,
2924 0 : Err(DismissedLayer::Future { local }) => {
2925 0 : if let Some(local) = local {
2926 0 : init::cleanup_future_layer(
2927 0 : &local.local_path,
2928 0 : &name,
2929 0 : disk_consistent_lsn,
2930 0 : )?;
2931 0 : }
2932 0 : needs_cleanup.push(name);
2933 0 : continue;
2934 : }
2935 0 : Err(DismissedLayer::LocalOnly(local)) => {
2936 0 : init::cleanup_local_only_file(&name, &local)?;
2937 : // this file never existed remotely, we will have to do rework
2938 0 : continue;
2939 : }
2940 0 : Err(DismissedLayer::BadMetadata(local)) => {
2941 0 : init::cleanup_local_file_for_remote(&local)?;
2942 : // this file never existed remotely, we will have to do rework
2943 0 : continue;
2944 : }
2945 : };
2946 :
2947 32 : match &name {
2948 24 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2949 8 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2950 : }
2951 :
2952 32 : tracing::debug!(layer=%name, ?decision, "applied");
2953 :
2954 32 : let layer = match decision {
2955 32 : Resident { local, remote } => {
2956 32 : total_physical_size += local.file_size;
2957 32 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2958 32 : .drop_eviction_guard()
2959 : }
2960 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2961 : };
2962 :
2963 32 : loaded_layers.push(layer);
2964 : }
2965 12 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2966 12 : }
2967 12 : })
2968 12 : .await
2969 12 : .map_err(anyhow::Error::new)
2970 12 : .and_then(|x| x)?;
2971 :
2972 12 : let num_layers = loaded_layers.len();
2973 12 :
2974 12 : guard
2975 12 : .open_mut()
2976 12 : .expect("layermanager must be open during init")
2977 12 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2978 12 :
2979 12 : self.remote_client
2980 12 : .schedule_layer_file_deletion(&needs_cleanup)?;
2981 12 : self.remote_client
2982 12 : .schedule_index_upload_for_file_changes()?;
2983 : // This barrier orders above DELETEs before any later operations.
2984 : // This is critical because code executing after the barrier might
2985 : // create again objects with the same key that we just scheduled for deletion.
2986 : // For example, if we just scheduled deletion of an image layer "from the future",
2987 : // later compaction might run again and re-create the same image layer.
2988 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2989 : // "same" here means same key range and LSN.
2990 : //
2991 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2992 : // the upload queue may execute the PUT first, then the DELETE.
2993 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2994 : //
2995 : // 1. a future image layer is created and uploaded
2996 : // 2. ps restart
2997 : // 3. the future layer from (1) is deleted during load layer map
2998 : // 4. image layer is re-created and uploaded
2999 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3000 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3001 : //
3002 : // See https://github.com/neondatabase/neon/issues/5878
3003 : //
3004 : // NB: generation numbers naturally protect against this because they disambiguate
3005 : // (1) and (4)
3006 : // TODO: this is basically a no-op now, should we remove it?
3007 12 : self.remote_client.schedule_barrier()?;
3008 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
3009 : // on retry.
3010 :
3011 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
3012 12 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
3013 12 : self.update_layer_visibility().await?;
3014 :
3015 12 : info!(
3016 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3017 : num_layers, disk_consistent_lsn, total_physical_size
3018 : );
3019 :
3020 12 : timer.stop_and_record();
3021 12 : Ok(())
3022 12 : }
3023 :
3024 : /// Retrieve current logical size of the timeline.
3025 : ///
3026 : /// The size could be lagging behind the actual number, in case
3027 : /// the initial size calculation has not been run (gets triggered on the first size access).
3028 : ///
3029 : /// return size and boolean flag that shows if the size is exact
3030 0 : pub(crate) fn get_current_logical_size(
3031 0 : self: &Arc<Self>,
3032 0 : priority: GetLogicalSizePriority,
3033 0 : ctx: &RequestContext,
3034 0 : ) -> logical_size::CurrentLogicalSize {
3035 0 : if !self.tenant_shard_id.is_shard_zero() {
3036 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3037 : // when HTTP API is serving a GET for timeline zero, return zero
3038 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3039 0 : }
3040 0 :
3041 0 : let current_size = self.current_logical_size.current_size();
3042 0 : debug!("Current size: {current_size:?}");
3043 :
3044 0 : match (current_size.accuracy(), priority) {
3045 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3046 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3047 0 : // background task will eventually deliver an exact value, we're in no rush
3048 0 : }
3049 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3050 : // background task is not ready, but user is asking for it now;
3051 : // => make the background task skip the line
3052 : // (The alternative would be to calculate the size here, but,
3053 : // it can actually take a long time if the user has a lot of rels.
3054 : // And we'll inevitable need it again; So, let the background task do the work.)
3055 0 : match self
3056 0 : .current_logical_size
3057 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3058 0 : .get()
3059 : {
3060 0 : Some(cancel) => cancel.cancel(),
3061 : None => {
3062 0 : match self.current_state() {
3063 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3064 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3065 0 : // Don't make noise.
3066 0 : }
3067 : TimelineState::Loading => {
3068 : // Import does not return an activated timeline.
3069 0 : info!("discarding priority boost for logical size calculation because timeline is not yet active");
3070 : }
3071 : TimelineState::Active => {
3072 : // activation should be setting the once cell
3073 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
3074 0 : debug_assert!(false);
3075 : }
3076 : }
3077 : }
3078 : }
3079 : }
3080 : }
3081 :
3082 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3083 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3084 0 : let first = self
3085 0 : .current_logical_size
3086 0 : .did_return_approximate_to_walreceiver
3087 0 : .compare_exchange(
3088 0 : false,
3089 0 : true,
3090 0 : AtomicOrdering::Relaxed,
3091 0 : AtomicOrdering::Relaxed,
3092 0 : )
3093 0 : .is_ok();
3094 0 : if first {
3095 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3096 0 : }
3097 0 : }
3098 0 : }
3099 :
3100 0 : current_size
3101 0 : }
3102 :
3103 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3104 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3105 : // nothing to do for freshly created timelines;
3106 0 : assert_eq!(
3107 0 : self.current_logical_size.current_size().accuracy(),
3108 0 : logical_size::Accuracy::Exact,
3109 0 : );
3110 0 : self.current_logical_size.initialized.add_permits(1);
3111 0 : return;
3112 : };
3113 :
3114 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3115 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3116 0 : self.current_logical_size
3117 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3118 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3119 0 :
3120 0 : let self_clone = Arc::clone(self);
3121 0 : let background_ctx = ctx.detached_child(
3122 0 : TaskKind::InitialLogicalSizeCalculation,
3123 0 : DownloadBehavior::Download,
3124 0 : );
3125 0 : task_mgr::spawn(
3126 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3127 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3128 0 : self.tenant_shard_id,
3129 0 : Some(self.timeline_id),
3130 0 : "initial size calculation",
3131 : // NB: don't log errors here, task_mgr will do that.
3132 0 : async move {
3133 0 : self_clone
3134 0 : .initial_logical_size_calculation_task(
3135 0 : initial_part_end,
3136 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3137 0 : background_ctx,
3138 0 : )
3139 0 : .await;
3140 0 : Ok(())
3141 0 : }
3142 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)),
3143 : );
3144 0 : }
3145 :
3146 : /// # Cancellation
3147 : ///
3148 : /// This method is sensitive to `Timeline::cancel`.
3149 : ///
3150 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3151 : ///
3152 : /// # Cancel-Safety
3153 : ///
3154 : /// It does Timeline IO, hence this should be polled to completion because
3155 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3156 : /// to reason about.
3157 0 : async fn initial_logical_size_calculation_task(
3158 0 : self: Arc<Self>,
3159 0 : initial_part_end: Lsn,
3160 0 : skip_concurrency_limiter: CancellationToken,
3161 0 : background_ctx: RequestContext,
3162 0 : ) {
3163 0 : scopeguard::defer! {
3164 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3165 0 : self.current_logical_size.initialized.add_permits(1);
3166 0 : }
3167 0 :
3168 0 : let try_once = |attempt: usize| {
3169 0 : let background_ctx = &background_ctx;
3170 0 : let self_ref = &self;
3171 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3172 0 : async move {
3173 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3174 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3175 0 : background_ctx,
3176 0 : );
3177 :
3178 : use crate::metrics::initial_logical_size::StartCircumstances;
3179 0 : let (_maybe_permit, circumstances) = tokio::select! {
3180 0 : permit = wait_for_permit => {
3181 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3182 : }
3183 0 : _ = self_ref.cancel.cancelled() => {
3184 0 : return Err(CalculateLogicalSizeError::Cancelled);
3185 : }
3186 0 : () = skip_concurrency_limiter.cancelled() => {
3187 : // Some action that is part of a end user interaction requested logical size
3188 : // => break out of the rate limit
3189 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3190 : // but then again what happens if they cancel; also, we should just be using
3191 : // one runtime across the entire process, so, let's leave this for now.
3192 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3193 : }
3194 : };
3195 :
3196 0 : let metrics_guard = if attempt == 1 {
3197 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3198 : } else {
3199 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3200 : };
3201 :
3202 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3203 0 : self_ref.conf,
3204 0 : self_ref
3205 0 : .gate
3206 0 : .enter()
3207 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3208 : );
3209 :
3210 0 : let calculated_size = self_ref
3211 0 : .logical_size_calculation_task(
3212 0 : initial_part_end,
3213 0 : LogicalSizeCalculationCause::Initial,
3214 0 : background_ctx,
3215 0 : )
3216 0 : .await?;
3217 :
3218 0 : self_ref
3219 0 : .trigger_aux_file_size_computation(
3220 0 : initial_part_end,
3221 0 : background_ctx,
3222 0 : io_concurrency,
3223 0 : )
3224 0 : .await?;
3225 :
3226 : // TODO: add aux file size to logical size
3227 :
3228 0 : Ok((calculated_size, metrics_guard))
3229 0 : }
3230 0 : };
3231 :
3232 0 : let retrying = async {
3233 0 : let mut attempt = 0;
3234 : loop {
3235 0 : attempt += 1;
3236 0 :
3237 0 : match try_once(attempt).await {
3238 0 : Ok(res) => return ControlFlow::Continue(res),
3239 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3240 : Err(
3241 0 : e @ (CalculateLogicalSizeError::Decode(_)
3242 0 : | CalculateLogicalSizeError::PageRead(_)),
3243 0 : ) => {
3244 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3245 : // exponential back-off doesn't make sense at these long intervals;
3246 : // use fixed retry interval with generous jitter instead
3247 0 : let sleep_duration = Duration::from_secs(
3248 0 : u64::try_from(
3249 0 : // 1hour base
3250 0 : (60_i64 * 60_i64)
3251 0 : // 10min jitter
3252 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3253 0 : )
3254 0 : .expect("10min < 1hour"),
3255 0 : );
3256 0 : tokio::select! {
3257 0 : _ = tokio::time::sleep(sleep_duration) => {}
3258 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3259 : }
3260 : }
3261 : }
3262 : }
3263 0 : };
3264 :
3265 0 : let (calculated_size, metrics_guard) = match retrying.await {
3266 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3267 0 : ControlFlow::Break(()) => return,
3268 : };
3269 :
3270 : // we cannot query current_logical_size.current_size() to know the current
3271 : // *negative* value, only truncated to u64.
3272 0 : let added = self
3273 0 : .current_logical_size
3274 0 : .size_added_after_initial
3275 0 : .load(AtomicOrdering::Relaxed);
3276 0 :
3277 0 : let sum = calculated_size.saturating_add_signed(added);
3278 0 :
3279 0 : // set the gauge value before it can be set in `update_current_logical_size`.
3280 0 : self.metrics.current_logical_size_gauge.set(sum);
3281 0 :
3282 0 : self.current_logical_size
3283 0 : .initial_logical_size
3284 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3285 0 : .ok()
3286 0 : .expect("only this task sets it");
3287 0 : }
3288 :
3289 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3290 0 : self: &Arc<Self>,
3291 0 : lsn: Lsn,
3292 0 : cause: LogicalSizeCalculationCause,
3293 0 : ctx: RequestContext,
3294 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3295 0 : let (sender, receiver) = oneshot::channel();
3296 0 : let self_clone = Arc::clone(self);
3297 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3298 0 : // we should stop the size calculation work and return an error.
3299 0 : // That would require restructuring this function's API to
3300 0 : // return the result directly, instead of a Receiver for the result.
3301 0 : let ctx = ctx.detached_child(
3302 0 : TaskKind::OndemandLogicalSizeCalculation,
3303 0 : DownloadBehavior::Download,
3304 0 : );
3305 0 : task_mgr::spawn(
3306 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3307 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3308 0 : self.tenant_shard_id,
3309 0 : Some(self.timeline_id),
3310 0 : "ondemand logical size calculation",
3311 0 : async move {
3312 0 : let res = self_clone
3313 0 : .logical_size_calculation_task(lsn, cause, &ctx)
3314 0 : .await;
3315 0 : let _ = sender.send(res).ok();
3316 0 : Ok(()) // Receiver is responsible for handling errors
3317 0 : }
3318 0 : .in_current_span(),
3319 0 : );
3320 0 : receiver
3321 0 : }
3322 :
3323 : #[instrument(skip_all)]
3324 : async fn logical_size_calculation_task(
3325 : self: &Arc<Self>,
3326 : lsn: Lsn,
3327 : cause: LogicalSizeCalculationCause,
3328 : ctx: &RequestContext,
3329 : ) -> Result<u64, CalculateLogicalSizeError> {
3330 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3331 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3332 : // accurate relation sizes, and they do not emit consumption metrics.
3333 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3334 :
3335 : let guard = self
3336 : .gate
3337 : .enter()
3338 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3339 :
3340 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3341 : }
3342 :
3343 : /// Calculate the logical size of the database at the latest LSN.
3344 : ///
3345 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3346 : /// especially if we need to download remote layers.
3347 0 : async fn calculate_logical_size(
3348 0 : &self,
3349 0 : up_to_lsn: Lsn,
3350 0 : cause: LogicalSizeCalculationCause,
3351 0 : _guard: &GateGuard,
3352 0 : ctx: &RequestContext,
3353 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3354 0 : info!(
3355 0 : "Calculating logical size for timeline {} at {}",
3356 : self.timeline_id, up_to_lsn
3357 : );
3358 :
3359 0 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
3360 : {
3361 0 : return Err(CalculateLogicalSizeError::Cancelled);
3362 0 : }
3363 :
3364 : // See if we've already done the work for initial size calculation.
3365 : // This is a short-cut for timelines that are mostly unused.
3366 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3367 0 : return Ok(size);
3368 0 : }
3369 0 : let storage_time_metrics = match cause {
3370 : LogicalSizeCalculationCause::Initial
3371 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3372 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3373 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3374 0 : &self.metrics.imitate_logical_size_histo
3375 : }
3376 : };
3377 0 : let timer = storage_time_metrics.start_timer();
3378 0 : let logical_size = self
3379 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3380 0 : .await?;
3381 0 : debug!("calculated logical size: {logical_size}");
3382 0 : timer.stop_and_record();
3383 0 : Ok(logical_size)
3384 0 : }
3385 :
3386 : /// Update current logical size, adding `delta' to the old value.
3387 541140 : fn update_current_logical_size(&self, delta: i64) {
3388 541140 : let logical_size = &self.current_logical_size;
3389 541140 : logical_size.increment_size(delta);
3390 541140 :
3391 541140 : // Also set the value in the prometheus gauge. Note that
3392 541140 : // there is a race condition here: if this is is called by two
3393 541140 : // threads concurrently, the prometheus gauge might be set to
3394 541140 : // one value while current_logical_size is set to the
3395 541140 : // other.
3396 541140 : match logical_size.current_size() {
3397 541140 : CurrentLogicalSize::Exact(ref new_current_size) => self
3398 541140 : .metrics
3399 541140 : .current_logical_size_gauge
3400 541140 : .set(new_current_size.into()),
3401 0 : CurrentLogicalSize::Approximate(_) => {
3402 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3403 0 : // forth between the initial size calculation task.
3404 0 : }
3405 : }
3406 541140 : }
3407 :
3408 5912 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3409 5912 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3410 5912 : let aux_metric =
3411 5912 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3412 5912 :
3413 5912 : let sum_of_entries = self
3414 5912 : .directory_metrics
3415 5912 : .iter()
3416 41384 : .map(|v| v.load(AtomicOrdering::Relaxed))
3417 5912 : .sum();
3418 : // Set a high general threshold and a lower threshold for the auxiliary files,
3419 : // as we can have large numbers of relations in the db directory.
3420 : const SUM_THRESHOLD: u64 = 5000;
3421 : const AUX_THRESHOLD: u64 = 1000;
3422 5912 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3423 0 : self.metrics
3424 0 : .directory_entries_count_gauge
3425 0 : .set(sum_of_entries);
3426 5912 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3427 0 : metric.set(sum_of_entries);
3428 5912 : }
3429 5912 : }
3430 :
3431 0 : async fn find_layer(
3432 0 : &self,
3433 0 : layer_name: &LayerName,
3434 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3435 0 : let guard = self.layers.read().await;
3436 0 : let layer = guard
3437 0 : .layer_map()?
3438 0 : .iter_historic_layers()
3439 0 : .find(|l| &l.layer_name() == layer_name)
3440 0 : .map(|found| guard.get_from_desc(&found));
3441 0 : Ok(layer)
3442 0 : }
3443 :
3444 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3445 : /// indicating which layers are currently on-disk on the primary.
3446 : ///
3447 : /// None is returned if the Timeline is in a state where uploading a heatmap
3448 : /// doesn't make sense, such as shutting down or initializing. The caller
3449 : /// should treat this as a cue to simply skip doing any heatmap uploading
3450 : /// for this timeline.
3451 4 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3452 4 : if !self.is_active() {
3453 0 : return None;
3454 4 : }
3455 :
3456 4 : let guard = self.layers.read().await;
3457 :
3458 20 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3459 20 : match layer.visibility() {
3460 : LayerVisibilityHint::Visible => {
3461 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3462 16 : let last_activity_ts = layer.latest_activity();
3463 16 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
3464 : }
3465 : LayerVisibilityHint::Covered => {
3466 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3467 4 : None
3468 : }
3469 : }
3470 20 : });
3471 4 :
3472 4 : let mut layers = resident.collect::<Vec<_>>();
3473 4 :
3474 4 : // Sort layers in order of which to download first. For a large set of layers to download, we
3475 4 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3476 4 : // or hours later:
3477 4 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3478 4 : // only exist for a few minutes before being compacted into L1s.
3479 4 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3480 4 : // the layer is likely to be covered by an image layer during compaction.
3481 40 : layers.sort_by_key(|(desc, _meta, _atime)| {
3482 40 : std::cmp::Reverse((
3483 40 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3484 40 : desc.lsn_range.end,
3485 40 : ))
3486 40 : });
3487 4 :
3488 4 : let layers = layers
3489 4 : .into_iter()
3490 16 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3491 4 : .collect();
3492 4 :
3493 4 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3494 4 : }
3495 :
3496 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3497 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3498 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3499 0 : // branchpoint in the value in IndexPart::lineage
3500 0 : self.ancestor_lsn == lsn
3501 0 : || (self.ancestor_lsn == Lsn::INVALID
3502 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3503 0 : }
3504 : }
3505 :
3506 : impl Timeline {
3507 : #[allow(clippy::doc_lazy_continuation)]
3508 : /// Get the data needed to reconstruct all keys in the provided keyspace
3509 : ///
3510 : /// The algorithm is as follows:
3511 : /// 1. While some keys are still not done and there's a timeline to visit:
3512 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3513 : /// 2.1: Build the fringe for the current keyspace
3514 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3515 : /// intersects
3516 : /// 2.3. Pop the timeline from the fringe
3517 : /// 2.4. If the fringe is empty, go back to 1
3518 1255178 : async fn get_vectored_reconstruct_data(
3519 1255178 : &self,
3520 1255178 : mut keyspace: KeySpace,
3521 1255178 : request_lsn: Lsn,
3522 1255178 : reconstruct_state: &mut ValuesReconstructState,
3523 1255178 : ctx: &RequestContext,
3524 1255178 : ) -> Result<(), GetVectoredError> {
3525 1255178 : let mut timeline_owned: Arc<Timeline>;
3526 1255178 : let mut timeline = self;
3527 1255178 :
3528 1255178 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3529 :
3530 1255174 : let missing_keyspace = loop {
3531 1704620 : if self.cancel.is_cancelled() {
3532 0 : return Err(GetVectoredError::Cancelled);
3533 1704620 : }
3534 :
3535 : let TimelineVisitOutcome {
3536 1704620 : completed_keyspace: completed,
3537 1704620 : image_covered_keyspace,
3538 1704620 : } = Self::get_vectored_reconstruct_data_timeline(
3539 1704620 : timeline,
3540 1704620 : keyspace.clone(),
3541 1704620 : cont_lsn,
3542 1704620 : reconstruct_state,
3543 1704620 : &self.cancel,
3544 1704620 : ctx,
3545 1704620 : )
3546 1704620 : .await?;
3547 :
3548 1704620 : keyspace.remove_overlapping_with(&completed);
3549 1704620 :
3550 1704620 : // Do not descend into the ancestor timeline for aux files.
3551 1704620 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3552 1704620 : // stalling compaction.
3553 1704620 : keyspace.remove_overlapping_with(&KeySpace {
3554 1704620 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
3555 1704620 : });
3556 1704620 :
3557 1704620 : // Keyspace is fully retrieved
3558 1704620 : if keyspace.is_empty() {
3559 1255118 : break None;
3560 449502 : }
3561 :
3562 449502 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3563 : // Not fully retrieved but no ancestor timeline.
3564 56 : break Some(keyspace);
3565 : };
3566 :
3567 : // Now we see if there are keys covered by the image layer but does not exist in the
3568 : // image layer, which means that the key does not exist.
3569 :
3570 : // The block below will stop the vectored search if any of the keys encountered an image layer
3571 : // which did not contain a snapshot for said key. Since we have already removed all completed
3572 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3573 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3574 : // and stop the search as a result of that.
3575 449446 : let mut removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3576 449446 : // Do not fire missing key error for sparse keys.
3577 449446 : removed.remove_overlapping_with(&KeySpace {
3578 449446 : ranges: vec![SPARSE_RANGE],
3579 449446 : });
3580 449446 : if !removed.is_empty() {
3581 0 : break Some(removed);
3582 449446 : }
3583 449446 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3584 449446 : // keyspace.
3585 449446 :
3586 449446 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3587 449446 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3588 449446 : timeline_owned = timeline
3589 449446 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3590 449446 : .await?;
3591 449442 : timeline = &*timeline_owned;
3592 : };
3593 :
3594 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
3595 1255174 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
3596 56 : let mut missing_keyspace = missing_keyspace;
3597 56 : missing_keyspace.remove_overlapping_with(&KeySpace {
3598 56 : ranges: vec![SPARSE_RANGE],
3599 56 : });
3600 56 : if missing_keyspace.is_empty() {
3601 28 : None
3602 : } else {
3603 28 : Some(missing_keyspace)
3604 : }
3605 : } else {
3606 1255118 : None
3607 : };
3608 :
3609 1255174 : if let Some(missing_keyspace) = missing_keyspace {
3610 28 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3611 28 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3612 28 : shard: self
3613 28 : .shard_identity
3614 28 : .get_shard_number(&missing_keyspace.start().unwrap()),
3615 28 : cont_lsn,
3616 28 : request_lsn,
3617 28 : ancestor_lsn: Some(timeline.ancestor_lsn),
3618 28 : backtrace: None,
3619 28 : read_path: std::mem::take(&mut reconstruct_state.read_path),
3620 28 : }));
3621 1255146 : }
3622 1255146 :
3623 1255146 : Ok(())
3624 1255178 : }
3625 :
3626 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3627 : ///
3628 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3629 : /// the current keyspace. The current keyspace of the search at any given timeline
3630 : /// is the original keyspace minus all the keys that have been completed minus
3631 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3632 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3633 : ///
3634 : /// This is basically a depth-first search visitor implementation where a vertex
3635 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3636 : ///
3637 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3638 : /// and get all the required reconstruct data from the layer in one go.
3639 : ///
3640 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3641 : /// decides how to deal with these two keyspaces.
3642 1704620 : async fn get_vectored_reconstruct_data_timeline(
3643 1704620 : timeline: &Timeline,
3644 1704620 : keyspace: KeySpace,
3645 1704620 : mut cont_lsn: Lsn,
3646 1704620 : reconstruct_state: &mut ValuesReconstructState,
3647 1704620 : cancel: &CancellationToken,
3648 1704620 : ctx: &RequestContext,
3649 1704620 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3650 1704620 : let mut unmapped_keyspace = keyspace.clone();
3651 1704620 : let mut fringe = LayerFringe::new();
3652 1704620 :
3653 1704620 : let mut completed_keyspace = KeySpace::default();
3654 1704620 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3655 1704620 :
3656 1704620 : // Prevent GC from progressing while visiting the current timeline.
3657 1704620 : // If we are GC-ing because a new image layer was added while traversing
3658 1704620 : // the timeline, then it will remove layers that are required for fulfilling
3659 1704620 : // the current get request (read-path cannot "look back" and notice the new
3660 1704620 : // image layer).
3661 1704620 : let _gc_cutoff_holder = timeline.get_latest_gc_cutoff_lsn();
3662 :
3663 : // See `compaction::compact_with_gc` for why we need this.
3664 1704620 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
3665 :
3666 : loop {
3667 3472539 : if cancel.is_cancelled() {
3668 0 : return Err(GetVectoredError::Cancelled);
3669 3472539 : }
3670 3472539 :
3671 3472539 : let (keys_done_last_step, keys_with_image_coverage) =
3672 3472539 : reconstruct_state.consume_done_keys();
3673 3472539 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3674 3472539 : completed_keyspace.merge(&keys_done_last_step);
3675 3472539 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3676 4996 : unmapped_keyspace
3677 4996 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3678 4996 : image_covered_keyspace.add_range(keys_with_image_coverage);
3679 3467543 : }
3680 :
3681 : // Do not descent any further if the last layer we visited
3682 : // completed all keys in the keyspace it inspected. This is not
3683 : // required for correctness, but avoids visiting extra layers
3684 : // which turns out to be a perf bottleneck in some cases.
3685 3472539 : if !unmapped_keyspace.is_empty() {
3686 2261949 : let guard = timeline.layers.read().await;
3687 2261949 : let layers = guard.layer_map()?;
3688 :
3689 2261949 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3690 1826443 : let start_lsn = l.get_lsn_range().start;
3691 1826443 : cont_lsn > start_lsn
3692 2261949 : });
3693 2261949 :
3694 2261949 : match in_memory_layer {
3695 1213195 : Some(l) => {
3696 1213195 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3697 1213195 : fringe.update(
3698 1213195 : ReadableLayer::InMemoryLayer(l),
3699 1213195 : unmapped_keyspace.clone(),
3700 1213195 : lsn_range,
3701 1213195 : );
3702 1213195 : }
3703 : None => {
3704 1048802 : for range in unmapped_keyspace.ranges.iter() {
3705 1048802 : let results = layers.range_search(range.clone(), cont_lsn);
3706 1048802 :
3707 1048802 : results
3708 1048802 : .found
3709 1048802 : .into_iter()
3710 1048802 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3711 554782 : (
3712 554782 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3713 554782 : keyspace_accum.to_keyspace(),
3714 554782 : lsn_floor..cont_lsn,
3715 554782 : )
3716 1048802 : })
3717 1048802 : .for_each(|(layer, keyspace, lsn_range)| {
3718 554782 : fringe.update(layer, keyspace, lsn_range)
3719 1048802 : });
3720 1048802 : }
3721 : }
3722 : }
3723 :
3724 : // It's safe to drop the layer map lock after planning the next round of reads.
3725 : // The fringe keeps readable handles for the layers which are safe to read even
3726 : // if layers were compacted or flushed.
3727 : //
3728 : // The more interesting consideration is: "Why is the read algorithm still correct
3729 : // if the layer map changes while it is operating?". Doing a vectored read on a
3730 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3731 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3732 : // range at *a particular point in time*. It is fine for the answer to be different
3733 : // at two different time points.
3734 2261949 : drop(guard);
3735 1210590 : }
3736 :
3737 3472539 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3738 1767919 : if let Some(ref mut read_path) = reconstruct_state.read_path {
3739 1767919 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
3740 1767919 : }
3741 1767919 : let next_cont_lsn = lsn_range.start;
3742 1767919 : layer_to_read
3743 1767919 : .get_values_reconstruct_data(
3744 1767919 : keyspace_to_read.clone(),
3745 1767919 : lsn_range,
3746 1767919 : reconstruct_state,
3747 1767919 : ctx,
3748 1767919 : )
3749 1767919 : .await?;
3750 :
3751 1767919 : unmapped_keyspace = keyspace_to_read;
3752 1767919 : cont_lsn = next_cont_lsn;
3753 1767919 :
3754 1767919 : reconstruct_state.on_layer_visited(&layer_to_read);
3755 : } else {
3756 1704620 : break;
3757 1704620 : }
3758 1704620 : }
3759 1704620 :
3760 1704620 : Ok(TimelineVisitOutcome {
3761 1704620 : completed_keyspace,
3762 1704620 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3763 1704620 : })
3764 1704620 : }
3765 :
3766 449446 : async fn get_ready_ancestor_timeline(
3767 449446 : &self,
3768 449446 : ancestor: &Arc<Timeline>,
3769 449446 : ctx: &RequestContext,
3770 449446 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3771 449446 : // It's possible that the ancestor timeline isn't active yet, or
3772 449446 : // is active but hasn't yet caught up to the branch point. Wait
3773 449446 : // for it.
3774 449446 : //
3775 449446 : // This cannot happen while the pageserver is running normally,
3776 449446 : // because you cannot create a branch from a point that isn't
3777 449446 : // present in the pageserver yet. However, we don't wait for the
3778 449446 : // branch point to be uploaded to cloud storage before creating
3779 449446 : // a branch. I.e., the branch LSN need not be remote consistent
3780 449446 : // for the branching operation to succeed.
3781 449446 : //
3782 449446 : // Hence, if we try to load a tenant in such a state where
3783 449446 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3784 449446 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3785 449446 : // then we will need to wait for the ancestor timeline to
3786 449446 : // re-stream WAL up to branch_lsn before we access it.
3787 449446 : //
3788 449446 : // How can a tenant get in such a state?
3789 449446 : // - ungraceful pageserver process exit
3790 449446 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3791 449446 : //
3792 449446 : // NB: this could be avoided by requiring
3793 449446 : // branch_lsn >= remote_consistent_lsn
3794 449446 : // during branch creation.
3795 449446 : match ancestor.wait_to_become_active(ctx).await {
3796 449442 : Ok(()) => {}
3797 : Err(TimelineState::Stopping) => {
3798 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3799 0 : return Err(GetReadyAncestorError::Cancelled);
3800 : }
3801 4 : Err(state) => {
3802 4 : return Err(GetReadyAncestorError::BadState {
3803 4 : timeline_id: ancestor.timeline_id,
3804 4 : state,
3805 4 : });
3806 : }
3807 : }
3808 449442 : ancestor
3809 449442 : .wait_lsn(
3810 449442 : self.ancestor_lsn,
3811 449442 : WaitLsnWaiter::Timeline(self),
3812 449442 : WaitLsnTimeout::Default,
3813 449442 : ctx,
3814 449442 : )
3815 449442 : .await
3816 449442 : .map_err(|e| match e {
3817 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3818 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3819 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3820 0 : timeline_id: ancestor.timeline_id,
3821 0 : state,
3822 0 : },
3823 449442 : })?;
3824 :
3825 449442 : Ok(ancestor.clone())
3826 449446 : }
3827 :
3828 594312 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3829 594312 : &self.shard_identity
3830 594312 : }
3831 :
3832 : #[inline(always)]
3833 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3834 0 : ShardTimelineId {
3835 0 : shard_index: ShardIndex {
3836 0 : shard_number: self.shard_identity.number,
3837 0 : shard_count: self.shard_identity.count,
3838 0 : },
3839 0 : timeline_id: self.timeline_id,
3840 0 : }
3841 0 : }
3842 :
3843 : /// Returns a non-frozen open in-memory layer for ingestion.
3844 : ///
3845 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3846 : /// this function without holding the mutex.
3847 2592 : async fn get_layer_for_write(
3848 2592 : &self,
3849 2592 : lsn: Lsn,
3850 2592 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3851 2592 : ctx: &RequestContext,
3852 2592 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3853 2592 : let mut guard = self.layers.write().await;
3854 :
3855 2592 : let last_record_lsn = self.get_last_record_lsn();
3856 2592 : ensure!(
3857 2592 : lsn > last_record_lsn,
3858 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3859 : lsn,
3860 : last_record_lsn,
3861 : );
3862 :
3863 2592 : let layer = guard
3864 2592 : .open_mut()?
3865 2592 : .get_layer_for_write(
3866 2592 : lsn,
3867 2592 : self.conf,
3868 2592 : self.timeline_id,
3869 2592 : self.tenant_shard_id,
3870 2592 : &self.gate,
3871 2592 : ctx,
3872 2592 : )
3873 2592 : .await?;
3874 2592 : Ok(layer)
3875 2592 : }
3876 :
3877 10558184 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3878 10558184 : assert!(new_lsn.is_aligned());
3879 :
3880 10558184 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
3881 10558184 : self.last_record_lsn.advance(new_lsn);
3882 10558184 : }
3883 :
3884 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3885 : ///
3886 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3887 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3888 2400 : async fn freeze_inmem_layer_at(
3889 2400 : &self,
3890 2400 : at: Lsn,
3891 2400 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3892 2400 : ) -> Result<u64, FlushLayerError> {
3893 2400 : let frozen = {
3894 2400 : let mut guard = self.layers.write().await;
3895 2400 : guard
3896 2400 : .open_mut()?
3897 2400 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
3898 2400 : .await
3899 : };
3900 :
3901 2400 : if frozen {
3902 2344 : let now = Instant::now();
3903 2344 : *(self.last_freeze_ts.write().unwrap()) = now;
3904 2344 : }
3905 :
3906 : // Increment the flush cycle counter and wake up the flush task.
3907 : // Remember the new value, so that when we listen for the flush
3908 : // to finish, we know when the flush that we initiated has
3909 : // finished, instead of some other flush that was started earlier.
3910 2400 : let mut my_flush_request = 0;
3911 2400 :
3912 2400 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3913 2400 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3914 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3915 2400 : }
3916 2400 :
3917 2400 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3918 2400 : my_flush_request = *counter + 1;
3919 2400 : *counter = my_flush_request;
3920 2400 : *lsn = std::cmp::max(at, *lsn);
3921 2400 : });
3922 2400 :
3923 2400 : assert_ne!(my_flush_request, 0);
3924 :
3925 2400 : Ok(my_flush_request)
3926 2400 : }
3927 :
3928 : /// Layer flusher task's main loop.
3929 880 : async fn flush_loop(
3930 880 : self: &Arc<Self>,
3931 880 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3932 880 : ctx: &RequestContext,
3933 880 : ) {
3934 : // Subscribe to L0 delta layer updates, for compaction backpressure.
3935 880 : let mut watch_l0 = match self.layers.read().await.layer_map() {
3936 880 : Ok(lm) => lm.watch_level0_deltas(),
3937 0 : Err(Shutdown) => return,
3938 : };
3939 :
3940 880 : info!("started flush loop");
3941 : loop {
3942 3206 : tokio::select! {
3943 3206 : _ = self.cancel.cancelled() => {
3944 20 : info!("shutting down layer flush task due to Timeline::cancel");
3945 20 : break;
3946 : },
3947 3206 : _ = layer_flush_start_rx.changed() => {}
3948 2326 : }
3949 2326 : trace!("waking up");
3950 2326 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3951 2326 :
3952 2326 : // The highest LSN to which we flushed in the loop over frozen layers
3953 2326 : let mut flushed_to_lsn = Lsn(0);
3954 :
3955 2326 : let result = loop {
3956 4670 : if self.cancel.is_cancelled() {
3957 0 : info!("dropping out of flush loop for timeline shutdown");
3958 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3959 : // anyone waiting on that will respect self.cancel as well: they will stop
3960 : // waiting at the same time we as drop out of this loop.
3961 0 : return;
3962 4670 : }
3963 4670 :
3964 4670 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
3965 4670 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
3966 4670 : if flushed_to_lsn >= frozen_to_lsn {
3967 2270 : break Ok(());
3968 2400 : }
3969 :
3970 : // Fetch the next layer to flush, if any.
3971 2400 : let (layer, l0_count, frozen_count, frozen_size) = {
3972 2400 : let layers = self.layers.read().await;
3973 2400 : let Ok(lm) = layers.layer_map() else {
3974 0 : info!("dropping out of flush loop for timeline shutdown");
3975 0 : return;
3976 : };
3977 2400 : let l0_count = lm.level0_deltas().len();
3978 2400 : let frozen_count = lm.frozen_layers.len();
3979 2400 : let frozen_size: u64 = lm
3980 2400 : .frozen_layers
3981 2400 : .iter()
3982 2420 : .map(|l| l.estimated_in_mem_size())
3983 2400 : .sum();
3984 2400 : let layer = lm.frozen_layers.front().cloned();
3985 2400 : (layer, l0_count, frozen_count, frozen_size)
3986 2400 : // drop 'layers' lock
3987 2400 : };
3988 2400 : let Some(layer) = layer else {
3989 56 : break Ok(());
3990 : };
3991 :
3992 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
3993 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
3994 : //
3995 : // NB: the compaction loop only checks `compaction_threshold` every 20 seconds, so
3996 : // we can end up stalling before compaction even starts. Consider making it more
3997 : // responsive (e.g. via `watch_level0_deltas`).
3998 2344 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
3999 0 : if l0_count >= stall_threshold {
4000 0 : warn!(
4001 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
4002 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4003 : );
4004 0 : let stall_timer = self
4005 0 : .metrics
4006 0 : .flush_delay_histo
4007 0 : .start_timer()
4008 0 : .record_on_drop();
4009 0 : tokio::select! {
4010 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
4011 0 : if let Ok(l0) = result.as_deref() {
4012 0 : let delay = stall_timer.elapsed().as_secs_f64();
4013 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
4014 0 : }
4015 : },
4016 0 : _ = self.cancel.cancelled() => {},
4017 : }
4018 0 : continue; // check again
4019 0 : }
4020 2344 : }
4021 :
4022 : // Flush the layer.
4023 2344 : let flush_timer = self.metrics.flush_time_histo.start_timer();
4024 2344 : match self.flush_frozen_layer(layer, ctx).await {
4025 2344 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
4026 : Err(FlushLayerError::Cancelled) => {
4027 0 : info!("dropping out of flush loop for timeline shutdown");
4028 0 : return;
4029 : }
4030 0 : err @ Err(
4031 0 : FlushLayerError::NotRunning(_)
4032 0 : | FlushLayerError::Other(_)
4033 0 : | FlushLayerError::CreateImageLayersError(_),
4034 0 : ) => {
4035 0 : error!("could not flush frozen layer: {err:?}");
4036 0 : break err.map(|_| ());
4037 : }
4038 : }
4039 2344 : let flush_duration = flush_timer.stop_and_record();
4040 2344 :
4041 2344 : // Notify the tenant compaction loop if L0 compaction is needed.
4042 2344 : let l0_count = *watch_l0.borrow();
4043 2344 : if l0_count >= self.get_compaction_threshold() {
4044 952 : self.l0_compaction_trigger.notify_one();
4045 1392 : }
4046 :
4047 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
4048 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
4049 : // ingestion by having ephemeral layer rolls wait for flushes.
4050 2344 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
4051 0 : if l0_count >= delay_threshold {
4052 0 : let delay = flush_duration.as_secs_f64();
4053 0 : info!(
4054 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
4055 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4056 : );
4057 0 : let _delay_timer = self
4058 0 : .metrics
4059 0 : .flush_delay_histo
4060 0 : .start_timer()
4061 0 : .record_on_drop();
4062 0 : tokio::select! {
4063 0 : _ = tokio::time::sleep(flush_duration) => {},
4064 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
4065 0 : _ = self.cancel.cancelled() => {},
4066 : }
4067 0 : }
4068 2344 : }
4069 : };
4070 :
4071 : // Unsharded tenants should never advance their LSN beyond the end of the
4072 : // highest layer they write: such gaps between layer data and the frozen LSN
4073 : // are only legal on sharded tenants.
4074 2326 : debug_assert!(
4075 2326 : self.shard_identity.count.count() > 1
4076 2326 : || flushed_to_lsn >= frozen_to_lsn
4077 56 : || !flushed_to_lsn.is_valid()
4078 : );
4079 :
4080 2326 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
4081 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
4082 : // to us via layer_flush_start_rx, then advance it here.
4083 : //
4084 : // This path is only taken for tenants with multiple shards: single sharded tenants should
4085 : // never encounter a gap in the wal.
4086 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
4087 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
4088 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
4089 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
4090 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
4091 0 : }
4092 0 : }
4093 2326 : }
4094 :
4095 : // Notify any listeners that we're done
4096 2326 : let _ = self
4097 2326 : .layer_flush_done_tx
4098 2326 : .send_replace((flush_counter, result));
4099 : }
4100 20 : }
4101 :
4102 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
4103 2240 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
4104 2240 : let mut rx = self.layer_flush_done_tx.subscribe();
4105 : loop {
4106 : {
4107 4559 : let (last_result_counter, last_result) = &*rx.borrow();
4108 4559 : if *last_result_counter >= request {
4109 2240 : if let Err(err) = last_result {
4110 : // We already logged the original error in
4111 : // flush_loop. We cannot propagate it to the caller
4112 : // here, because it might not be Cloneable
4113 0 : return Err(err.clone());
4114 : } else {
4115 2240 : return Ok(());
4116 : }
4117 2319 : }
4118 2319 : }
4119 2319 : trace!("waiting for flush to complete");
4120 2319 : tokio::select! {
4121 2319 : rx_e = rx.changed() => {
4122 2319 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
4123 : },
4124 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
4125 : // the notification from [`flush_loop`] that it completed.
4126 2319 : _ = self.cancel.cancelled() => {
4127 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
4128 0 : return Ok(())
4129 : }
4130 : };
4131 2319 : trace!("done")
4132 : }
4133 2240 : }
4134 :
4135 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
4136 : ///
4137 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
4138 : #[instrument(skip_all, fields(layer=%frozen_layer))]
4139 : async fn flush_frozen_layer(
4140 : self: &Arc<Self>,
4141 : frozen_layer: Arc<InMemoryLayer>,
4142 : ctx: &RequestContext,
4143 : ) -> Result<Lsn, FlushLayerError> {
4144 : debug_assert_current_span_has_tenant_and_timeline_id();
4145 :
4146 : // As a special case, when we have just imported an image into the repository,
4147 : // instead of writing out a L0 delta layer, we directly write out image layer
4148 : // files instead. This is possible as long as *all* the data imported into the
4149 : // repository have the same LSN.
4150 : let lsn_range = frozen_layer.get_lsn_range();
4151 :
4152 : // Whether to directly create image layers for this flush, or flush them as delta layers
4153 : let create_image_layer =
4154 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
4155 :
4156 : #[cfg(test)]
4157 : {
4158 : match &mut *self.flush_loop_state.lock().unwrap() {
4159 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
4160 : panic!("flush loop not running")
4161 : }
4162 : FlushLoopState::Running {
4163 : expect_initdb_optimization,
4164 : initdb_optimization_count,
4165 : ..
4166 : } => {
4167 : if create_image_layer {
4168 : *initdb_optimization_count += 1;
4169 : } else {
4170 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
4171 : }
4172 : }
4173 : }
4174 : }
4175 :
4176 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
4177 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
4178 : // require downloading anything during initial import.
4179 : let ((rel_partition, metadata_partition), _lsn) = self
4180 : .repartition(
4181 : self.initdb_lsn,
4182 : self.get_compaction_target_size(),
4183 : EnumSet::empty(),
4184 : ctx,
4185 : )
4186 : .await
4187 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
4188 :
4189 : if self.cancel.is_cancelled() {
4190 : return Err(FlushLayerError::Cancelled);
4191 : }
4192 :
4193 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
4194 : // So that the key ranges don't overlap.
4195 : let mut partitions = KeyPartitioning::default();
4196 : partitions.parts.extend(rel_partition.parts);
4197 : if !metadata_partition.parts.is_empty() {
4198 : assert_eq!(
4199 : metadata_partition.parts.len(),
4200 : 1,
4201 : "currently sparse keyspace should only contain a single metadata keyspace"
4202 : );
4203 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
4204 : // every single key within the keyspace, and therefore, it's safe to force converting it
4205 : // into a dense keyspace before calling this function.
4206 : partitions
4207 : .parts
4208 : .extend(metadata_partition.into_dense().parts);
4209 : }
4210 :
4211 : let mut layers_to_upload = Vec::new();
4212 : let (generated_image_layers, is_complete) = self
4213 : .create_image_layers(
4214 : &partitions,
4215 : self.initdb_lsn,
4216 : ImageLayerCreationMode::Initial,
4217 : ctx,
4218 : LastImageLayerCreationStatus::Initial,
4219 : false, // don't yield for L0, we're flushing L0
4220 : )
4221 : .await?;
4222 : debug_assert!(
4223 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
4224 : "init image generation mode must fully cover the keyspace"
4225 : );
4226 : layers_to_upload.extend(generated_image_layers);
4227 :
4228 : (layers_to_upload, None)
4229 : } else {
4230 : // Normal case, write out a L0 delta layer file.
4231 : // `create_delta_layer` will not modify the layer map.
4232 : // We will remove frozen layer and add delta layer in one atomic operation later.
4233 : let Some(layer) = self
4234 : .create_delta_layer(&frozen_layer, None, ctx)
4235 : .await
4236 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
4237 : else {
4238 : panic!("delta layer cannot be empty if no filter is applied");
4239 : };
4240 : (
4241 : // FIXME: even though we have a single image and single delta layer assumption
4242 : // we push them to vec
4243 : vec![layer.clone()],
4244 : Some(layer),
4245 : )
4246 : };
4247 :
4248 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
4249 :
4250 : if self.cancel.is_cancelled() {
4251 : return Err(FlushLayerError::Cancelled);
4252 : }
4253 :
4254 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
4255 :
4256 : // The new on-disk layers are now in the layer map. We can remove the
4257 : // in-memory layer from the map now. The flushed layer is stored in
4258 : // the mapping in `create_delta_layer`.
4259 : {
4260 : let mut guard = self.layers.write().await;
4261 :
4262 : guard.open_mut()?.finish_flush_l0_layer(
4263 : delta_layer_to_add.as_ref(),
4264 : &frozen_layer,
4265 : &self.metrics,
4266 : );
4267 :
4268 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
4269 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
4270 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
4271 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
4272 : }
4273 : // release lock on 'layers'
4274 : };
4275 :
4276 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
4277 : // This makes us refuse ingest until the new layers have been persisted to the remote
4278 : // TODO: remove this, and rely on l0_flush_{delay,stall}_threshold instead.
4279 : if self.get_l0_flush_wait_upload() {
4280 : let start = Instant::now();
4281 : self.remote_client
4282 : .wait_completion()
4283 : .await
4284 0 : .map_err(|e| match e {
4285 : WaitCompletionError::UploadQueueShutDownOrStopped
4286 : | WaitCompletionError::NotInitialized(
4287 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
4288 0 : ) => FlushLayerError::Cancelled,
4289 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
4290 0 : FlushLayerError::Other(anyhow!(e).into())
4291 : }
4292 0 : })?;
4293 : let duration = start.elapsed().as_secs_f64();
4294 : self.metrics.flush_wait_upload_time_gauge_add(duration);
4295 : }
4296 :
4297 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
4298 : // a compaction can delete the file and then it won't be available for uploads any more.
4299 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
4300 : // race situation.
4301 : // See https://github.com/neondatabase/neon/issues/4526
4302 : pausable_failpoint!("flush-frozen-pausable");
4303 :
4304 : // This failpoint is used by another test case `test_pageserver_recovery`.
4305 : fail_point!("flush-frozen-exit");
4306 :
4307 : Ok(Lsn(lsn_range.end.0 - 1))
4308 : }
4309 :
4310 : /// Return true if the value changed
4311 : ///
4312 : /// This function must only be used from the layer flush task.
4313 2344 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
4314 2344 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
4315 2344 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
4316 :
4317 2344 : self.metrics
4318 2344 : .disk_consistent_lsn_gauge
4319 2344 : .set(new_value.0 as i64);
4320 2344 : new_value != old_value
4321 2344 : }
4322 :
4323 : /// Update metadata file
4324 2444 : fn schedule_uploads(
4325 2444 : &self,
4326 2444 : disk_consistent_lsn: Lsn,
4327 2444 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
4328 2444 : ) -> anyhow::Result<()> {
4329 2444 : // We can only save a valid 'prev_record_lsn' value on disk if we
4330 2444 : // flushed *all* in-memory changes to disk. We only track
4331 2444 : // 'prev_record_lsn' in memory for the latest processed record, so we
4332 2444 : // don't remember what the correct value that corresponds to some old
4333 2444 : // LSN is. But if we flush everything, then the value corresponding
4334 2444 : // current 'last_record_lsn' is correct and we can store it on disk.
4335 2444 : let RecordLsn {
4336 2444 : last: last_record_lsn,
4337 2444 : prev: prev_record_lsn,
4338 2444 : } = self.last_record_lsn.load();
4339 2444 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
4340 2188 : Some(prev_record_lsn)
4341 : } else {
4342 256 : None
4343 : };
4344 :
4345 2444 : let update = crate::tenant::metadata::MetadataUpdate::new(
4346 2444 : disk_consistent_lsn,
4347 2444 : ondisk_prev_record_lsn,
4348 2444 : *self.latest_gc_cutoff_lsn.read(),
4349 2444 : );
4350 2444 :
4351 2444 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
4352 0 : "{}",
4353 0 : x.unwrap()
4354 2444 : ));
4355 :
4356 4812 : for layer in layers_to_upload {
4357 2368 : self.remote_client.schedule_layer_file_upload(layer)?;
4358 : }
4359 2444 : self.remote_client
4360 2444 : .schedule_index_upload_for_metadata_update(&update)?;
4361 :
4362 2444 : Ok(())
4363 2444 : }
4364 :
4365 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
4366 0 : self.remote_client
4367 0 : .preserve_initdb_archive(
4368 0 : &self.tenant_shard_id.tenant_id,
4369 0 : &self.timeline_id,
4370 0 : &self.cancel,
4371 0 : )
4372 0 : .await
4373 0 : }
4374 :
4375 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
4376 : // in layer map immediately. The caller is responsible to put it into the layer map.
4377 1936 : async fn create_delta_layer(
4378 1936 : self: &Arc<Self>,
4379 1936 : frozen_layer: &Arc<InMemoryLayer>,
4380 1936 : key_range: Option<Range<Key>>,
4381 1936 : ctx: &RequestContext,
4382 1936 : ) -> anyhow::Result<Option<ResidentLayer>> {
4383 1936 : let self_clone = Arc::clone(self);
4384 1936 : let frozen_layer = Arc::clone(frozen_layer);
4385 1936 : let ctx = ctx.attached_child();
4386 1936 : let work = async move {
4387 1936 : let Some((desc, path)) = frozen_layer
4388 1936 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
4389 1936 : .await?
4390 : else {
4391 0 : return Ok(None);
4392 : };
4393 1936 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
4394 :
4395 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
4396 : // We just need to fsync the directory in which these inodes are linked,
4397 : // which we know to be the timeline directory.
4398 : //
4399 : // We use fatal_err() below because the after write_to_disk returns with success,
4400 : // the in-memory state of the filesystem already has the layer file in its final place,
4401 : // and subsequent pageserver code could think it's durable while it really isn't.
4402 1936 : let timeline_dir = VirtualFile::open(
4403 1936 : &self_clone
4404 1936 : .conf
4405 1936 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
4406 1936 : &ctx,
4407 1936 : )
4408 1936 : .await
4409 1936 : .fatal_err("VirtualFile::open for timeline dir fsync");
4410 1936 : timeline_dir
4411 1936 : .sync_all()
4412 1936 : .await
4413 1936 : .fatal_err("VirtualFile::sync_all timeline dir");
4414 1936 : anyhow::Ok(Some(new_delta))
4415 1936 : };
4416 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
4417 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
4418 : use crate::virtual_file::io_engine::IoEngine;
4419 1936 : match crate::virtual_file::io_engine::get() {
4420 0 : IoEngine::NotSet => panic!("io engine not set"),
4421 : IoEngine::StdFs => {
4422 968 : let span = tracing::info_span!("blocking");
4423 968 : tokio::task::spawn_blocking({
4424 968 : move || Handle::current().block_on(work.instrument(span))
4425 968 : })
4426 968 : .await
4427 968 : .context("spawn_blocking")
4428 968 : .and_then(|x| x)
4429 : }
4430 : #[cfg(target_os = "linux")]
4431 968 : IoEngine::TokioEpollUring => work.await,
4432 : }
4433 1936 : }
4434 :
4435 1176 : async fn repartition(
4436 1176 : &self,
4437 1176 : lsn: Lsn,
4438 1176 : partition_size: u64,
4439 1176 : flags: EnumSet<CompactFlags>,
4440 1176 : ctx: &RequestContext,
4441 1176 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
4442 1176 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
4443 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
4444 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
4445 : // and hence before the compaction task starts.
4446 0 : return Err(CompactionError::Other(anyhow!(
4447 0 : "repartition() called concurrently"
4448 0 : )));
4449 : };
4450 1176 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
4451 1176 : if lsn < *partition_lsn {
4452 0 : return Err(CompactionError::Other(anyhow!(
4453 0 : "repartition() called with LSN going backwards, this should not happen"
4454 0 : )));
4455 1176 : }
4456 1176 :
4457 1176 : let distance = lsn.0 - partition_lsn.0;
4458 1176 : if *partition_lsn != Lsn(0)
4459 564 : && distance <= self.repartition_threshold
4460 564 : && !flags.contains(CompactFlags::ForceRepartition)
4461 : {
4462 536 : debug!(
4463 : distance,
4464 : threshold = self.repartition_threshold,
4465 0 : "no repartitioning needed"
4466 : );
4467 536 : return Ok((
4468 536 : (dense_partition.clone(), sparse_partition.clone()),
4469 536 : *partition_lsn,
4470 536 : ));
4471 640 : }
4472 :
4473 640 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4474 640 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4475 640 : let sparse_partitioning = SparseKeyPartitioning {
4476 640 : parts: vec![sparse_ks],
4477 640 : }; // no partitioning for metadata keys for now
4478 640 : let result = ((dense_partitioning, sparse_partitioning), lsn);
4479 640 : guard.write(result.clone());
4480 640 : Ok(result)
4481 1176 : }
4482 :
4483 : // Is it time to create a new image layer for the given partition? True if we want to generate.
4484 0 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4485 0 : let threshold = self.get_image_creation_threshold();
4486 :
4487 0 : let guard = self.layers.read().await;
4488 0 : let Ok(layers) = guard.layer_map() else {
4489 0 : return false;
4490 : };
4491 :
4492 0 : let mut max_deltas = 0;
4493 0 : for part_range in &partition.ranges {
4494 0 : let image_coverage = layers.image_coverage(part_range, lsn);
4495 0 : for (img_range, last_img) in image_coverage {
4496 0 : let img_lsn = if let Some(last_img) = last_img {
4497 0 : last_img.get_lsn_range().end
4498 : } else {
4499 0 : Lsn(0)
4500 : };
4501 : // Let's consider an example:
4502 : //
4503 : // delta layer with LSN range 71-81
4504 : // delta layer with LSN range 81-91
4505 : // delta layer with LSN range 91-101
4506 : // image layer at LSN 100
4507 : //
4508 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4509 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4510 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4511 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4512 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4513 0 : if img_lsn < lsn {
4514 0 : let num_deltas =
4515 0 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4516 0 :
4517 0 : max_deltas = max_deltas.max(num_deltas);
4518 0 : if num_deltas >= threshold {
4519 0 : debug!(
4520 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4521 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4522 : );
4523 0 : return true;
4524 0 : }
4525 0 : }
4526 : }
4527 : }
4528 :
4529 0 : debug!(
4530 : max_deltas,
4531 0 : "none of the partitioned ranges had >= {threshold} deltas"
4532 : );
4533 0 : false
4534 0 : }
4535 :
4536 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4537 : /// so that at most one image layer will be produced from this function.
4538 : #[allow(clippy::too_many_arguments)]
4539 452 : async fn create_image_layer_for_rel_blocks(
4540 452 : self: &Arc<Self>,
4541 452 : partition: &KeySpace,
4542 452 : mut image_layer_writer: ImageLayerWriter,
4543 452 : lsn: Lsn,
4544 452 : ctx: &RequestContext,
4545 452 : img_range: Range<Key>,
4546 452 : io_concurrency: IoConcurrency,
4547 452 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4548 452 : let mut wrote_keys = false;
4549 452 :
4550 452 : let mut key_request_accum = KeySpaceAccum::new();
4551 2988 : for range in &partition.ranges {
4552 2536 : let mut key = range.start;
4553 5496 : while key < range.end {
4554 : // Decide whether to retain this key: usually we do, but sharded tenants may
4555 : // need to drop keys that don't belong to them. If we retain the key, add it
4556 : // to `key_request_accum` for later issuing a vectored get
4557 2960 : if self.shard_identity.is_key_disposable(&key) {
4558 0 : debug!(
4559 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4560 0 : key,
4561 0 : self.shard_identity.get_shard_number(&key)
4562 : );
4563 2960 : } else {
4564 2960 : key_request_accum.add_key(key);
4565 2960 : }
4566 :
4567 2960 : let last_key_in_range = key.next() == range.end;
4568 2960 : key = key.next();
4569 2960 :
4570 2960 : // Maybe flush `key_rest_accum`
4571 2960 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4572 2960 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4573 : {
4574 2536 : let results = self
4575 2536 : .get_vectored(
4576 2536 : key_request_accum.consume_keyspace(),
4577 2536 : lsn,
4578 2536 : io_concurrency.clone(),
4579 2536 : ctx,
4580 2536 : )
4581 2536 : .await?;
4582 :
4583 2536 : if self.cancel.is_cancelled() {
4584 0 : return Err(CreateImageLayersError::Cancelled);
4585 2536 : }
4586 :
4587 5496 : for (img_key, img) in results {
4588 2960 : let img = match img {
4589 2960 : Ok(img) => img,
4590 0 : Err(err) => {
4591 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4592 0 : // page without losing any actual user data. That seems better
4593 0 : // than failing repeatedly and getting stuck.
4594 0 : //
4595 0 : // We had a bug at one point, where we truncated the FSM and VM
4596 0 : // in the pageserver, but the Postgres didn't know about that
4597 0 : // and continued to generate incremental WAL records for pages
4598 0 : // that didn't exist in the pageserver. Trying to replay those
4599 0 : // WAL records failed to find the previous image of the page.
4600 0 : // This special case allows us to recover from that situation.
4601 0 : // See https://github.com/neondatabase/neon/issues/2601.
4602 0 : //
4603 0 : // Unfortunately we cannot do this for the main fork, or for
4604 0 : // any metadata keys, keys, as that would lead to actual data
4605 0 : // loss.
4606 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4607 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4608 0 : ZERO_PAGE.clone()
4609 : } else {
4610 0 : return Err(CreateImageLayersError::from(err));
4611 : }
4612 : }
4613 : };
4614 :
4615 : // Write all the keys we just read into our new image layer.
4616 2960 : image_layer_writer.put_image(img_key, img, ctx).await?;
4617 2960 : wrote_keys = true;
4618 : }
4619 424 : }
4620 : }
4621 : }
4622 :
4623 452 : if wrote_keys {
4624 : // Normal path: we have written some data into the new image layer for this
4625 : // partition, so flush it to disk.
4626 452 : info!(
4627 0 : "produced image layer for rel {}",
4628 0 : ImageLayerName {
4629 0 : key_range: img_range.clone(),
4630 0 : lsn
4631 0 : },
4632 : );
4633 452 : Ok(ImageLayerCreationOutcome::Generated {
4634 452 : unfinished_image_layer: image_layer_writer,
4635 452 : })
4636 : } else {
4637 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4638 0 : Ok(ImageLayerCreationOutcome::Empty)
4639 : }
4640 452 : }
4641 :
4642 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4643 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4644 : /// would not be too large to fit in a single image layer.
4645 : ///
4646 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
4647 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
4648 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
4649 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
4650 : #[allow(clippy::too_many_arguments)]
4651 436 : async fn create_image_layer_for_metadata_keys(
4652 436 : self: &Arc<Self>,
4653 436 : partition: &KeySpace,
4654 436 : mut image_layer_writer: ImageLayerWriter,
4655 436 : lsn: Lsn,
4656 436 : ctx: &RequestContext,
4657 436 : img_range: Range<Key>,
4658 436 : mode: ImageLayerCreationMode,
4659 436 : io_concurrency: IoConcurrency,
4660 436 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4661 436 : // Metadata keys image layer creation.
4662 436 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
4663 436 : let begin = Instant::now();
4664 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
4665 : // not contain too many keys, otherwise this takes a lot of memory.
4666 436 : let data = self
4667 436 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4668 436 : .await?;
4669 436 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4670 436 : let mut new_data = BTreeMap::new();
4671 436 : let mut total_kb_retrieved = 0;
4672 436 : let mut total_keys_retrieved = 0;
4673 12460 : for (k, v) in data {
4674 12024 : let v = v?;
4675 12024 : total_kb_retrieved += KEY_SIZE + v.len();
4676 12024 : total_keys_retrieved += 1;
4677 12024 : new_data.insert(k, v);
4678 : }
4679 436 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4680 436 : };
4681 436 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4682 436 : let elapsed = begin.elapsed();
4683 436 :
4684 436 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4685 436 : info!(
4686 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()
4687 : );
4688 :
4689 436 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4690 0 : return Ok(ImageLayerCreationOutcome::Skip);
4691 436 : }
4692 436 : if self.cancel.is_cancelled() {
4693 0 : return Err(CreateImageLayersError::Cancelled);
4694 436 : }
4695 436 : let mut wrote_any_image = false;
4696 12460 : for (k, v) in data {
4697 12024 : if v.is_empty() {
4698 : // the key has been deleted, it does not need an image
4699 : // in metadata keyspace, an empty image == tombstone
4700 8 : continue;
4701 12016 : }
4702 12016 : wrote_any_image = true;
4703 12016 :
4704 12016 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4705 12016 :
4706 12016 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4707 12016 : // on the normal data path either.
4708 12016 : image_layer_writer.put_image(k, v, ctx).await?;
4709 : }
4710 :
4711 436 : if wrote_any_image {
4712 : // Normal path: we have written some data into the new image layer for this
4713 : // partition, so flush it to disk.
4714 20 : info!(
4715 0 : "created image layer for metadata {}",
4716 0 : ImageLayerName {
4717 0 : key_range: img_range.clone(),
4718 0 : lsn
4719 0 : }
4720 : );
4721 20 : Ok(ImageLayerCreationOutcome::Generated {
4722 20 : unfinished_image_layer: image_layer_writer,
4723 20 : })
4724 : } else {
4725 416 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4726 416 : Ok(ImageLayerCreationOutcome::Empty)
4727 : }
4728 436 : }
4729 :
4730 : /// Predicate function which indicates whether we should check if new image layers
4731 : /// are required. Since checking if new image layers are required is expensive in
4732 : /// terms of CPU, we only do it in the following cases:
4733 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4734 : /// 2. If enough time has passed since the last check:
4735 : /// 1. For large tenants, we wish to perform the check more often since they
4736 : /// suffer from the lack of image layers
4737 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4738 1176 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4739 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4740 :
4741 1176 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4742 1176 : let distance = lsn
4743 1176 : .checked_sub(last_checks_at)
4744 1176 : .expect("Attempt to compact with LSN going backwards");
4745 1176 : let min_distance =
4746 1176 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4747 1176 :
4748 1176 : let distance_based_decision = distance.0 >= min_distance;
4749 1176 :
4750 1176 : let mut time_based_decision = false;
4751 1176 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4752 1176 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4753 972 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4754 : {
4755 0 : self.get_checkpoint_timeout()
4756 : } else {
4757 972 : Duration::from_secs(3600 * 48)
4758 : };
4759 :
4760 972 : time_based_decision = match *last_check_instant {
4761 564 : Some(last_check) => {
4762 564 : let elapsed = last_check.elapsed();
4763 564 : elapsed >= check_required_after
4764 : }
4765 408 : None => true,
4766 : };
4767 204 : }
4768 :
4769 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4770 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4771 : // check.
4772 1176 : let decision = distance_based_decision || time_based_decision;
4773 :
4774 1176 : if decision {
4775 408 : self.last_image_layer_creation_check_at.store(lsn);
4776 408 : *last_check_instant = Some(Instant::now());
4777 768 : }
4778 :
4779 1176 : decision
4780 1176 : }
4781 :
4782 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
4783 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
4784 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4785 : async fn create_image_layers(
4786 : self: &Arc<Timeline>,
4787 : partitioning: &KeyPartitioning,
4788 : lsn: Lsn,
4789 : mode: ImageLayerCreationMode,
4790 : ctx: &RequestContext,
4791 : last_status: LastImageLayerCreationStatus,
4792 : yield_for_l0: bool,
4793 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
4794 : let timer = self.metrics.create_images_time_histo.start_timer();
4795 :
4796 : if partitioning.parts.is_empty() {
4797 : warn!("no partitions to create image layers for");
4798 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
4799 : }
4800 :
4801 : // We need to avoid holes between generated image layers.
4802 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4803 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4804 : //
4805 : // How such hole between partitions can appear?
4806 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4807 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4808 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4809 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4810 : let mut start = Key::MIN;
4811 :
4812 : let check_for_image_layers =
4813 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
4814 : info!(
4815 : "resuming image layer creation: last_status=incomplete, continue from {}",
4816 : last_key
4817 : );
4818 : true
4819 : } else {
4820 : self.should_check_if_image_layers_required(lsn)
4821 : };
4822 :
4823 : let mut batch_image_writer = BatchLayerWriter::new(self.conf).await?;
4824 :
4825 : let mut all_generated = true;
4826 :
4827 : let mut partition_processed = 0;
4828 : let mut total_partitions = partitioning.parts.len();
4829 : let mut last_partition_processed = None;
4830 : let mut partition_parts = partitioning.parts.clone();
4831 :
4832 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
4833 : // We need to skip the partitions that have already been processed.
4834 : let mut found = false;
4835 : for (i, partition) in partition_parts.iter().enumerate() {
4836 : if last_key <= partition.end().unwrap() {
4837 : // ```plain
4838 : // |------|--------|----------|------|
4839 : // ^last_key
4840 : // ^start from this partition
4841 : // ```
4842 : // Why `i+1` instead of `i`?
4843 : // It is possible that the user did some writes after the previous image layer creation attempt so that
4844 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
4845 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
4846 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
4847 : if i + 1 >= total_partitions {
4848 : // In general, this case should not happen -- if last_key is on the last partition, the previous
4849 : // iteration of image layer creation should return a complete status.
4850 : break; // with found=false
4851 : }
4852 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
4853 : total_partitions = partition_parts.len();
4854 : // Update the start key to the partition start.
4855 : start = partition_parts[0].start().unwrap();
4856 : found = true;
4857 : break;
4858 : }
4859 : }
4860 : if !found {
4861 : // Last key is within the last partition, or larger than all partitions.
4862 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
4863 : }
4864 : }
4865 :
4866 : for partition in partition_parts.iter() {
4867 : if self.cancel.is_cancelled() {
4868 : return Err(CreateImageLayersError::Cancelled);
4869 : }
4870 : partition_processed += 1;
4871 : let img_range = start..partition.ranges.last().unwrap().end;
4872 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4873 : if compact_metadata {
4874 : for range in &partition.ranges {
4875 : assert!(
4876 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4877 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4878 : "metadata keys must be partitioned separately"
4879 : );
4880 : }
4881 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4882 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4883 : // might mess up with evictions.
4884 : start = img_range.end;
4885 : continue;
4886 : }
4887 : // For initial and force modes, we always generate image layers for metadata keys.
4888 : } else if let ImageLayerCreationMode::Try = mode {
4889 : // check_for_image_layers = false -> skip
4890 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4891 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4892 : start = img_range.end;
4893 : continue;
4894 : }
4895 : }
4896 : if let ImageLayerCreationMode::Force = mode {
4897 : // When forced to create image layers, we might try and create them where they already
4898 : // exist. This mode is only used in tests/debug.
4899 : let layers = self.layers.read().await;
4900 : if layers.contains_key(&PersistentLayerKey {
4901 : key_range: img_range.clone(),
4902 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4903 : is_delta: false,
4904 : }) {
4905 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
4906 : // in the future.
4907 : tracing::info!(
4908 : "Skipping image layer at {lsn} {}..{}, already exists",
4909 : img_range.start,
4910 : img_range.end
4911 : );
4912 : start = img_range.end;
4913 : continue;
4914 : }
4915 : }
4916 :
4917 : let image_layer_writer = ImageLayerWriter::new(
4918 : self.conf,
4919 : self.timeline_id,
4920 : self.tenant_shard_id,
4921 : &img_range,
4922 : lsn,
4923 : ctx,
4924 : )
4925 : .await?;
4926 :
4927 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4928 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4929 0 : "failpoint image-layer-writer-fail-before-finish"
4930 0 : )))
4931 0 : });
4932 :
4933 : let io_concurrency = IoConcurrency::spawn_from_conf(
4934 : self.conf,
4935 : self.gate
4936 : .enter()
4937 0 : .map_err(|_| CreateImageLayersError::Cancelled)?,
4938 : );
4939 :
4940 : let outcome = if !compact_metadata {
4941 : self.create_image_layer_for_rel_blocks(
4942 : partition,
4943 : image_layer_writer,
4944 : lsn,
4945 : ctx,
4946 : img_range.clone(),
4947 : io_concurrency,
4948 : )
4949 : .await?
4950 : } else {
4951 : self.create_image_layer_for_metadata_keys(
4952 : partition,
4953 : image_layer_writer,
4954 : lsn,
4955 : ctx,
4956 : img_range.clone(),
4957 : mode,
4958 : io_concurrency,
4959 : )
4960 : .await?
4961 : };
4962 : match outcome {
4963 : ImageLayerCreationOutcome::Empty => {
4964 : // No data in this partition, so we don't need to create an image layer (for now).
4965 : // The next image layer should cover this key range, so we don't advance the `start`
4966 : // key.
4967 : }
4968 : ImageLayerCreationOutcome::Generated {
4969 : unfinished_image_layer,
4970 : } => {
4971 : batch_image_writer.add_unfinished_image_writer(
4972 : unfinished_image_layer,
4973 : img_range.clone(),
4974 : lsn,
4975 : );
4976 : // The next image layer should be generated right after this one.
4977 : start = img_range.end;
4978 : }
4979 : ImageLayerCreationOutcome::Skip => {
4980 : // We don't need to create an image layer for this partition.
4981 : // The next image layer should NOT cover this range, otherwise
4982 : // the keyspace becomes empty (reads don't go past image layers).
4983 : start = img_range.end;
4984 : }
4985 : }
4986 :
4987 : if let ImageLayerCreationMode::Try = mode {
4988 : // We have at least made some progress
4989 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
4990 : // The `Try` mode is currently only used on the compaction path. We want to avoid
4991 : // image layer generation taking too long time and blocking L0 compaction. So in this
4992 : // mode, we also inspect the current number of L0 layers and skip image layer generation
4993 : // if there are too many of them.
4994 : let num_of_l0_layers = {
4995 : let layers = self.layers.read().await;
4996 : layers.layer_map()?.level0_deltas().len()
4997 : };
4998 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
4999 : * self.get_compaction_threshold();
5000 : if image_preempt_threshold != 0 && num_of_l0_layers >= image_preempt_threshold {
5001 : tracing::info!(
5002 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers {}",
5003 : partition.start().unwrap(), partition.end().unwrap(), num_of_l0_layers
5004 : );
5005 : last_partition_processed = Some(partition.clone());
5006 : all_generated = false;
5007 : break;
5008 : }
5009 : }
5010 : }
5011 : }
5012 :
5013 : let image_layers = batch_image_writer.finish(self, ctx).await?;
5014 :
5015 : let mut guard = self.layers.write().await;
5016 :
5017 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
5018 : // now they are being scheduled outside of write lock; current way is inconsistent with
5019 : // compaction lock order.
5020 : guard
5021 : .open_mut()?
5022 : .track_new_image_layers(&image_layers, &self.metrics);
5023 : drop_wlock(guard);
5024 : let duration = timer.stop_and_record();
5025 :
5026 : // Creating image layers may have caused some previously visible layers to be covered
5027 : if !image_layers.is_empty() {
5028 : self.update_layer_visibility().await?;
5029 : }
5030 :
5031 : let total_layer_size = image_layers
5032 : .iter()
5033 472 : .map(|l| l.metadata().file_size)
5034 : .sum::<u64>();
5035 :
5036 : info!(
5037 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
5038 : image_layers.len(),
5039 : total_layer_size,
5040 : duration.as_secs_f64(),
5041 : partition_processed,
5042 : total_partitions
5043 : );
5044 :
5045 : Ok((
5046 : image_layers,
5047 : if all_generated {
5048 : LastImageLayerCreationStatus::Complete
5049 : } else {
5050 : LastImageLayerCreationStatus::Incomplete {
5051 : last_key: if let Some(last_partition_processed) = last_partition_processed {
5052 : last_partition_processed.end().unwrap_or(Key::MIN)
5053 : } else {
5054 : // This branch should be unreachable, but in case it happens, we can just return the start key.
5055 : Key::MIN
5056 : },
5057 : }
5058 : },
5059 : ))
5060 : }
5061 :
5062 : /// Wait until the background initial logical size calculation is complete, or
5063 : /// this Timeline is shut down. Calling this function will cause the initial
5064 : /// logical size calculation to skip waiting for the background jobs barrier.
5065 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
5066 0 : if !self.shard_identity.is_shard_zero() {
5067 : // We don't populate logical size on shard >0: skip waiting for it.
5068 0 : return;
5069 0 : }
5070 0 :
5071 0 : if self.remote_client.is_deleting() {
5072 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
5073 0 : return;
5074 0 : }
5075 0 :
5076 0 : if self.current_logical_size.current_size().is_exact() {
5077 : // root timelines are initialized with exact count, but never start the background
5078 : // calculation
5079 0 : return;
5080 0 : }
5081 :
5082 0 : if let Some(await_bg_cancel) = self
5083 0 : .current_logical_size
5084 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
5085 0 : .get()
5086 0 : {
5087 0 : await_bg_cancel.cancel();
5088 0 : } else {
5089 : // We should not wait if we were not able to explicitly instruct
5090 : // the logical size cancellation to skip the concurrency limit semaphore.
5091 : // TODO: this is an unexpected case. We should restructure so that it
5092 : // can't happen.
5093 0 : tracing::warn!(
5094 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
5095 : );
5096 0 : debug_assert!(false);
5097 : }
5098 :
5099 0 : tokio::select!(
5100 0 : _ = self.current_logical_size.initialized.acquire() => {},
5101 0 : _ = self.cancel.cancelled() => {}
5102 : )
5103 0 : }
5104 :
5105 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
5106 : /// Timelines layers up to the ancestor_lsn.
5107 : ///
5108 : /// Requires a timeline that:
5109 : /// - has an ancestor to detach from
5110 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
5111 : /// a technical requirement
5112 : ///
5113 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
5114 : /// polled again until completion.
5115 : ///
5116 : /// During the operation all timelines sharing the data with this timeline will be reparented
5117 : /// from our ancestor to be branches of this timeline.
5118 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
5119 0 : self: &Arc<Timeline>,
5120 0 : tenant: &crate::tenant::Tenant,
5121 0 : options: detach_ancestor::Options,
5122 0 : ctx: &RequestContext,
5123 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
5124 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
5125 0 : }
5126 :
5127 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
5128 : /// reparents any reparentable children of previous ancestor.
5129 : ///
5130 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
5131 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
5132 : /// successfully, tenant must be reloaded.
5133 : ///
5134 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
5135 : /// resetting the tenant.
5136 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
5137 0 : self: &Arc<Timeline>,
5138 0 : tenant: &crate::tenant::Tenant,
5139 0 : prepared: detach_ancestor::PreparedTimelineDetach,
5140 0 : ctx: &RequestContext,
5141 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
5142 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
5143 0 : }
5144 :
5145 : /// Final step which unblocks the GC.
5146 : ///
5147 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
5148 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
5149 0 : self: &Arc<Timeline>,
5150 0 : tenant: &crate::tenant::Tenant,
5151 0 : attempt: detach_ancestor::Attempt,
5152 0 : ctx: &RequestContext,
5153 0 : ) -> Result<(), detach_ancestor::Error> {
5154 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
5155 0 : }
5156 : }
5157 :
5158 : impl Drop for Timeline {
5159 20 : fn drop(&mut self) {
5160 20 : if let Some(ancestor) = &self.ancestor_timeline {
5161 : // This lock should never be poisoned, but in case it is we do a .map() instead of
5162 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
5163 8 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
5164 8 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
5165 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
5166 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
5167 8 : }
5168 0 : }
5169 12 : }
5170 20 : info!(
5171 0 : "Timeline {} for tenant {} is being dropped",
5172 : self.timeline_id, self.tenant_shard_id.tenant_id
5173 : );
5174 20 : }
5175 : }
5176 :
5177 : /// Top-level failure to compact.
5178 : #[derive(Debug, thiserror::Error)]
5179 : pub(crate) enum CompactionError {
5180 : #[error("The timeline or pageserver is shutting down")]
5181 : ShuttingDown,
5182 : /// Compaction tried to offload a timeline and failed
5183 : #[error("Failed to offload timeline: {0}")]
5184 : Offload(OffloadError),
5185 : /// Compaction cannot be done right now; page reconstruction and so on.
5186 : #[error(transparent)]
5187 : Other(anyhow::Error),
5188 : }
5189 :
5190 : impl From<OffloadError> for CompactionError {
5191 0 : fn from(e: OffloadError) -> Self {
5192 0 : match e {
5193 0 : OffloadError::Cancelled => Self::ShuttingDown,
5194 0 : _ => Self::Offload(e),
5195 : }
5196 0 : }
5197 : }
5198 :
5199 : impl CompactionError {
5200 0 : pub fn is_cancelled(&self) -> bool {
5201 0 : matches!(self, CompactionError::ShuttingDown)
5202 0 : }
5203 : }
5204 :
5205 : impl From<CollectKeySpaceError> for CompactionError {
5206 0 : fn from(err: CollectKeySpaceError) -> Self {
5207 0 : match err {
5208 : CollectKeySpaceError::Cancelled
5209 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
5210 0 : CompactionError::ShuttingDown
5211 : }
5212 0 : e => CompactionError::Other(e.into()),
5213 : }
5214 0 : }
5215 : }
5216 :
5217 : impl From<super::upload_queue::NotInitialized> for CompactionError {
5218 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
5219 0 : match value {
5220 : super::upload_queue::NotInitialized::Uninitialized => {
5221 0 : CompactionError::Other(anyhow::anyhow!(value))
5222 : }
5223 : super::upload_queue::NotInitialized::ShuttingDown
5224 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
5225 : }
5226 0 : }
5227 : }
5228 :
5229 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
5230 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
5231 0 : match e {
5232 : super::storage_layer::layer::DownloadError::TimelineShutdown
5233 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
5234 0 : CompactionError::ShuttingDown
5235 : }
5236 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
5237 : | super::storage_layer::layer::DownloadError::DownloadRequired
5238 : | super::storage_layer::layer::DownloadError::NotFile(_)
5239 : | super::storage_layer::layer::DownloadError::DownloadFailed
5240 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
5241 0 : CompactionError::Other(anyhow::anyhow!(e))
5242 : }
5243 : #[cfg(test)]
5244 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
5245 0 : CompactionError::Other(anyhow::anyhow!(e))
5246 : }
5247 : }
5248 0 : }
5249 : }
5250 :
5251 : impl From<layer_manager::Shutdown> for CompactionError {
5252 0 : fn from(_: layer_manager::Shutdown) -> Self {
5253 0 : CompactionError::ShuttingDown
5254 0 : }
5255 : }
5256 :
5257 : #[serde_as]
5258 644 : #[derive(serde::Serialize)]
5259 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
5260 :
5261 : #[derive(Default)]
5262 : enum DurationRecorder {
5263 : #[default]
5264 : NotStarted,
5265 : Recorded(RecordedDuration, tokio::time::Instant),
5266 : }
5267 :
5268 : impl DurationRecorder {
5269 1228 : fn till_now(&self) -> DurationRecorder {
5270 1228 : match self {
5271 : DurationRecorder::NotStarted => {
5272 0 : panic!("must only call on recorded measurements")
5273 : }
5274 1228 : DurationRecorder::Recorded(_, ended) => {
5275 1228 : let now = tokio::time::Instant::now();
5276 1228 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
5277 1228 : }
5278 1228 : }
5279 1228 : }
5280 644 : fn into_recorded(self) -> Option<RecordedDuration> {
5281 644 : match self {
5282 0 : DurationRecorder::NotStarted => None,
5283 644 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
5284 : }
5285 644 : }
5286 : }
5287 :
5288 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
5289 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
5290 : /// the layer descriptor requires the user to provide the ranges, which should cover all
5291 : /// keys specified in the `data` field.
5292 : #[cfg(test)]
5293 : #[derive(Clone)]
5294 : pub struct DeltaLayerTestDesc {
5295 : pub lsn_range: Range<Lsn>,
5296 : pub key_range: Range<Key>,
5297 : pub data: Vec<(Key, Lsn, Value)>,
5298 : }
5299 :
5300 : #[cfg(test)]
5301 : impl DeltaLayerTestDesc {
5302 4 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
5303 4 : Self {
5304 4 : lsn_range,
5305 4 : key_range,
5306 4 : data,
5307 4 : }
5308 4 : }
5309 :
5310 176 : pub fn new_with_inferred_key_range(
5311 176 : lsn_range: Range<Lsn>,
5312 176 : data: Vec<(Key, Lsn, Value)>,
5313 176 : ) -> Self {
5314 440 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
5315 440 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
5316 176 : Self {
5317 176 : key_range: (*key_min)..(key_max.next()),
5318 176 : lsn_range,
5319 176 : data,
5320 176 : }
5321 176 : }
5322 :
5323 20 : pub(crate) fn layer_name(&self) -> LayerName {
5324 20 : LayerName::Delta(super::storage_layer::DeltaLayerName {
5325 20 : key_range: self.key_range.clone(),
5326 20 : lsn_range: self.lsn_range.clone(),
5327 20 : })
5328 20 : }
5329 : }
5330 :
5331 : impl Timeline {
5332 92 : async fn finish_compact_batch(
5333 92 : self: &Arc<Self>,
5334 92 : new_deltas: &[ResidentLayer],
5335 92 : new_images: &[ResidentLayer],
5336 92 : layers_to_remove: &[Layer],
5337 92 : ) -> Result<(), CompactionError> {
5338 92 : let mut guard = tokio::select! {
5339 92 : guard = self.layers.write() => guard,
5340 92 : _ = self.cancel.cancelled() => {
5341 0 : return Err(CompactionError::ShuttingDown);
5342 : }
5343 : };
5344 :
5345 92 : let mut duplicated_layers = HashSet::new();
5346 92 :
5347 92 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
5348 :
5349 744 : for l in new_deltas {
5350 652 : if guard.contains(l.as_ref()) {
5351 : // expected in tests
5352 0 : tracing::error!(layer=%l, "duplicated L1 layer");
5353 :
5354 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
5355 : // `guard` on self.layers. as of writing this, there are no error returns except
5356 : // for compact_level0_phase1 creating an L0, which does not happen in practice
5357 : // because we have not implemented L0 => L0 compaction.
5358 0 : duplicated_layers.insert(l.layer_desc().key());
5359 652 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
5360 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
5361 652 : } else {
5362 652 : insert_layers.push(l.clone());
5363 652 : }
5364 : }
5365 :
5366 : // only remove those inputs which were not outputs
5367 92 : let remove_layers: Vec<Layer> = layers_to_remove
5368 92 : .iter()
5369 804 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
5370 92 : .cloned()
5371 92 : .collect();
5372 92 :
5373 92 : if !new_images.is_empty() {
5374 0 : guard
5375 0 : .open_mut()?
5376 0 : .track_new_image_layers(new_images, &self.metrics);
5377 92 : }
5378 :
5379 92 : guard
5380 92 : .open_mut()?
5381 92 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
5382 92 :
5383 92 : self.remote_client
5384 92 : .schedule_compaction_update(&remove_layers, new_deltas)?;
5385 :
5386 92 : drop_wlock(guard);
5387 92 :
5388 92 : Ok(())
5389 92 : }
5390 :
5391 0 : async fn rewrite_layers(
5392 0 : self: &Arc<Self>,
5393 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
5394 0 : mut drop_layers: Vec<Layer>,
5395 0 : ) -> Result<(), CompactionError> {
5396 0 : let mut guard = self.layers.write().await;
5397 :
5398 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
5399 : // to avoid double-removing, and avoid rewriting something that was removed.
5400 0 : replace_layers.retain(|(l, _)| guard.contains(l));
5401 0 : drop_layers.retain(|l| guard.contains(l));
5402 0 :
5403 0 : guard
5404 0 : .open_mut()?
5405 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
5406 0 :
5407 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
5408 0 :
5409 0 : self.remote_client
5410 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
5411 :
5412 0 : Ok(())
5413 0 : }
5414 :
5415 : /// Schedules the uploads of the given image layers
5416 768 : fn upload_new_image_layers(
5417 768 : self: &Arc<Self>,
5418 768 : new_images: impl IntoIterator<Item = ResidentLayer>,
5419 768 : ) -> Result<(), super::upload_queue::NotInitialized> {
5420 808 : for layer in new_images {
5421 40 : self.remote_client.schedule_layer_file_upload(layer)?;
5422 : }
5423 : // should any new image layer been created, not uploading index_part will
5424 : // result in a mismatch between remote_physical_size and layermap calculated
5425 : // size, which will fail some tests, but should not be an issue otherwise.
5426 768 : self.remote_client
5427 768 : .schedule_index_upload_for_file_changes()?;
5428 768 : Ok(())
5429 768 : }
5430 :
5431 0 : async fn find_gc_time_cutoff(
5432 0 : &self,
5433 0 : now: SystemTime,
5434 0 : pitr: Duration,
5435 0 : cancel: &CancellationToken,
5436 0 : ctx: &RequestContext,
5437 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
5438 0 : debug_assert_current_span_has_tenant_and_timeline_id();
5439 0 : if self.shard_identity.is_shard_zero() {
5440 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
5441 0 : let time_range = if pitr == Duration::ZERO {
5442 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
5443 : } else {
5444 0 : pitr
5445 : };
5446 :
5447 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
5448 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
5449 0 : let timestamp = to_pg_timestamp(time_cutoff);
5450 :
5451 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
5452 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
5453 0 : LsnForTimestamp::Future(lsn) => {
5454 0 : // The timestamp is in the future. That sounds impossible,
5455 0 : // but what it really means is that there hasn't been
5456 0 : // any commits since the cutoff timestamp.
5457 0 : //
5458 0 : // In this case we should use the LSN of the most recent commit,
5459 0 : // which is implicitly the last LSN in the log.
5460 0 : debug!("future({})", lsn);
5461 0 : Some(self.get_last_record_lsn())
5462 : }
5463 0 : LsnForTimestamp::Past(lsn) => {
5464 0 : debug!("past({})", lsn);
5465 0 : None
5466 : }
5467 0 : LsnForTimestamp::NoData(lsn) => {
5468 0 : debug!("nodata({})", lsn);
5469 0 : None
5470 : }
5471 : };
5472 0 : Ok(time_cutoff)
5473 : } else {
5474 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
5475 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
5476 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
5477 : // space cutoff that we would also have respected ourselves.
5478 0 : match self
5479 0 : .remote_client
5480 0 : .download_foreign_index(ShardNumber(0), cancel)
5481 0 : .await
5482 : {
5483 0 : Ok((index_part, index_generation, _index_mtime)) => {
5484 0 : tracing::info!("GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
5485 0 : index_part.metadata.latest_gc_cutoff_lsn());
5486 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
5487 : }
5488 : Err(DownloadError::NotFound) => {
5489 : // This is unexpected, because during timeline creations shard zero persists to remote
5490 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
5491 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
5492 : // state, then shard zero should _eventually_ write an index when it GCs.
5493 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
5494 0 : Ok(None)
5495 : }
5496 0 : Err(e) => {
5497 0 : // TODO: this function should return a different error type than page reconstruct error
5498 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
5499 : }
5500 : }
5501 :
5502 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
5503 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
5504 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
5505 : // new location. This is legal in principle, but problematic in practice because it might result
5506 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
5507 : // because they have GC'd past the branch point.
5508 : }
5509 0 : }
5510 :
5511 : /// Find the Lsns above which layer files need to be retained on
5512 : /// garbage collection.
5513 : ///
5514 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
5515 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
5516 : /// the space-based retention.
5517 : ///
5518 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
5519 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
5520 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
5521 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
5522 : /// in the response as the GC cutoff point for the timeline.
5523 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
5524 : pub(super) async fn find_gc_cutoffs(
5525 : &self,
5526 : now: SystemTime,
5527 : space_cutoff: Lsn,
5528 : pitr: Duration,
5529 : cancel: &CancellationToken,
5530 : ctx: &RequestContext,
5531 : ) -> Result<GcCutoffs, PageReconstructError> {
5532 : let _timer = self
5533 : .metrics
5534 : .find_gc_cutoffs_histo
5535 : .start_timer()
5536 : .record_on_drop();
5537 :
5538 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
5539 :
5540 : if cfg!(test) {
5541 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
5542 : if pitr == Duration::ZERO {
5543 : return Ok(GcCutoffs {
5544 : time: self.get_last_record_lsn(),
5545 : space: space_cutoff,
5546 : });
5547 : }
5548 : }
5549 :
5550 : // Calculate a time-based limit on how much to retain:
5551 : // - if PITR interval is set, then this is our cutoff.
5552 : // - if PITR interval is not set, then we do a lookup
5553 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
5554 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
5555 :
5556 : Ok(match (pitr, time_cutoff) {
5557 : (Duration::ZERO, Some(time_cutoff)) => {
5558 : // PITR is not set. Retain the size-based limit, or the default time retention,
5559 : // whichever requires less data.
5560 : GcCutoffs {
5561 : time: self.get_last_record_lsn(),
5562 : space: std::cmp::max(time_cutoff, space_cutoff),
5563 : }
5564 : }
5565 : (Duration::ZERO, None) => {
5566 : // PITR is not set, and time lookup failed
5567 : GcCutoffs {
5568 : time: self.get_last_record_lsn(),
5569 : space: space_cutoff,
5570 : }
5571 : }
5572 : (_, None) => {
5573 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
5574 : // cannot advance beyond what was already GC'd, and respect space-based retention
5575 : GcCutoffs {
5576 : time: *self.get_latest_gc_cutoff_lsn(),
5577 : space: space_cutoff,
5578 : }
5579 : }
5580 : (_, Some(time_cutoff)) => {
5581 : // PITR interval is set and we looked up timestamp successfully. Ignore
5582 : // size based retention and make time cutoff authoritative
5583 : GcCutoffs {
5584 : time: time_cutoff,
5585 : space: time_cutoff,
5586 : }
5587 : }
5588 : })
5589 : }
5590 :
5591 : /// Garbage collect layer files on a timeline that are no longer needed.
5592 : ///
5593 : /// Currently, we don't make any attempt at removing unneeded page versions
5594 : /// within a layer file. We can only remove the whole file if it's fully
5595 : /// obsolete.
5596 8 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
5597 : // this is most likely the background tasks, but it might be the spawned task from
5598 : // immediate_gc
5599 8 : let _g = tokio::select! {
5600 8 : guard = self.gc_lock.lock() => guard,
5601 8 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
5602 : };
5603 8 : let timer = self.metrics.garbage_collect_histo.start_timer();
5604 8 :
5605 8 : fail_point!("before-timeline-gc");
5606 8 :
5607 8 : // Is the timeline being deleted?
5608 8 : if self.is_stopping() {
5609 0 : return Err(GcError::TimelineCancelled);
5610 8 : }
5611 8 :
5612 8 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
5613 8 : let gc_info = self.gc_info.read().unwrap();
5614 8 :
5615 8 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
5616 8 : let time_cutoff = gc_info.cutoffs.time;
5617 8 : let retain_lsns = gc_info
5618 8 : .retain_lsns
5619 8 : .iter()
5620 8 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
5621 8 : .collect();
5622 8 :
5623 8 : // Gets the maximum LSN that holds the valid lease.
5624 8 : //
5625 8 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
5626 8 : // Here, we do not check for stale leases again.
5627 8 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
5628 8 :
5629 8 : (
5630 8 : space_cutoff,
5631 8 : time_cutoff,
5632 8 : retain_lsns,
5633 8 : max_lsn_with_valid_lease,
5634 8 : )
5635 8 : };
5636 8 :
5637 8 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
5638 8 : let standby_horizon = self.standby_horizon.load();
5639 8 : // Hold GC for the standby, but as a safety guard do it only within some
5640 8 : // reasonable lag.
5641 8 : if standby_horizon != Lsn::INVALID {
5642 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
5643 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
5644 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
5645 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
5646 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
5647 : } else {
5648 0 : warn!(
5649 0 : "standby is lagging for more than {}MB, not holding gc for it",
5650 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
5651 : )
5652 : }
5653 0 : }
5654 8 : }
5655 :
5656 : // Reset standby horizon to ignore it if it is not updated till next GC.
5657 : // It is an easy way to unset it when standby disappears without adding
5658 : // more conf options.
5659 8 : self.standby_horizon.store(Lsn::INVALID);
5660 8 : self.metrics
5661 8 : .standby_horizon_gauge
5662 8 : .set(Lsn::INVALID.0 as i64);
5663 :
5664 8 : let res = self
5665 8 : .gc_timeline(
5666 8 : space_cutoff,
5667 8 : time_cutoff,
5668 8 : retain_lsns,
5669 8 : max_lsn_with_valid_lease,
5670 8 : new_gc_cutoff,
5671 8 : )
5672 8 : .instrument(
5673 8 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5674 : )
5675 8 : .await?;
5676 :
5677 : // only record successes
5678 8 : timer.stop_and_record();
5679 8 :
5680 8 : Ok(res)
5681 8 : }
5682 :
5683 8 : async fn gc_timeline(
5684 8 : &self,
5685 8 : space_cutoff: Lsn,
5686 8 : time_cutoff: Lsn,
5687 8 : retain_lsns: Vec<Lsn>,
5688 8 : max_lsn_with_valid_lease: Option<Lsn>,
5689 8 : new_gc_cutoff: Lsn,
5690 8 : ) -> Result<GcResult, GcError> {
5691 8 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5692 8 :
5693 8 : let now = SystemTime::now();
5694 8 : let mut result: GcResult = GcResult::default();
5695 8 :
5696 8 : // Nothing to GC. Return early.
5697 8 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
5698 8 : if latest_gc_cutoff >= new_gc_cutoff {
5699 0 : info!(
5700 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5701 : );
5702 0 : return Ok(result);
5703 8 : }
5704 :
5705 : // We need to ensure that no one tries to read page versions or create
5706 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5707 : // for details. This will block until the old value is no longer in use.
5708 : //
5709 : // The GC cutoff should only ever move forwards.
5710 8 : let waitlist = {
5711 8 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
5712 8 : if *write_guard > new_gc_cutoff {
5713 0 : return Err(GcError::BadLsn {
5714 0 : why: format!(
5715 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5716 0 : *write_guard, new_gc_cutoff
5717 0 : ),
5718 0 : });
5719 8 : }
5720 8 :
5721 8 : write_guard.store_and_unlock(new_gc_cutoff)
5722 8 : };
5723 8 : waitlist.wait().await;
5724 :
5725 8 : info!("GC starting");
5726 :
5727 8 : debug!("retain_lsns: {:?}", retain_lsns);
5728 :
5729 8 : let mut layers_to_remove = Vec::new();
5730 :
5731 : // Scan all layers in the timeline (remote or on-disk).
5732 : //
5733 : // Garbage collect the layer if all conditions are satisfied:
5734 : // 1. it is older than cutoff LSN;
5735 : // 2. it is older than PITR interval;
5736 : // 3. it doesn't need to be retained for 'retain_lsns';
5737 : // 4. it does not need to be kept for LSNs holding valid leases.
5738 : // 5. newer on-disk image layers cover the layer's whole key range
5739 : //
5740 : // TODO holding a write lock is too agressive and avoidable
5741 8 : let mut guard = self.layers.write().await;
5742 8 : let layers = guard.layer_map()?;
5743 48 : 'outer: for l in layers.iter_historic_layers() {
5744 48 : result.layers_total += 1;
5745 48 :
5746 48 : // 1. Is it newer than GC horizon cutoff point?
5747 48 : if l.get_lsn_range().end > space_cutoff {
5748 4 : info!(
5749 0 : "keeping {} because it's newer than space_cutoff {}",
5750 0 : l.layer_name(),
5751 : space_cutoff,
5752 : );
5753 4 : result.layers_needed_by_cutoff += 1;
5754 4 : continue 'outer;
5755 44 : }
5756 44 :
5757 44 : // 2. It is newer than PiTR cutoff point?
5758 44 : if l.get_lsn_range().end > time_cutoff {
5759 0 : info!(
5760 0 : "keeping {} because it's newer than time_cutoff {}",
5761 0 : l.layer_name(),
5762 : time_cutoff,
5763 : );
5764 0 : result.layers_needed_by_pitr += 1;
5765 0 : continue 'outer;
5766 44 : }
5767 :
5768 : // 3. Is it needed by a child branch?
5769 : // NOTE With that we would keep data that
5770 : // might be referenced by child branches forever.
5771 : // We can track this in child timeline GC and delete parent layers when
5772 : // they are no longer needed. This might be complicated with long inheritance chains.
5773 : //
5774 : // TODO Vec is not a great choice for `retain_lsns`
5775 44 : for retain_lsn in &retain_lsns {
5776 : // start_lsn is inclusive
5777 0 : if &l.get_lsn_range().start <= retain_lsn {
5778 0 : info!(
5779 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5780 0 : l.layer_name(),
5781 0 : retain_lsn,
5782 0 : l.is_incremental(),
5783 : );
5784 0 : result.layers_needed_by_branches += 1;
5785 0 : continue 'outer;
5786 0 : }
5787 : }
5788 :
5789 : // 4. Is there a valid lease that requires us to keep this layer?
5790 44 : if let Some(lsn) = &max_lsn_with_valid_lease {
5791 : // keep if layer start <= any of the lease
5792 36 : if &l.get_lsn_range().start <= lsn {
5793 28 : info!(
5794 0 : "keeping {} because there is a valid lease preventing GC at {}",
5795 0 : l.layer_name(),
5796 : lsn,
5797 : );
5798 28 : result.layers_needed_by_leases += 1;
5799 28 : continue 'outer;
5800 8 : }
5801 8 : }
5802 :
5803 : // 5. Is there a later on-disk layer for this relation?
5804 : //
5805 : // The end-LSN is exclusive, while disk_consistent_lsn is
5806 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5807 : // OK for a delta layer to have end LSN 101, but if the end LSN
5808 : // is 102, then it might not have been fully flushed to disk
5809 : // before crash.
5810 : //
5811 : // For example, imagine that the following layers exist:
5812 : //
5813 : // 1000 - image (A)
5814 : // 1000-2000 - delta (B)
5815 : // 2000 - image (C)
5816 : // 2000-3000 - delta (D)
5817 : // 3000 - image (E)
5818 : //
5819 : // If GC horizon is at 2500, we can remove layers A and B, but
5820 : // we cannot remove C, even though it's older than 2500, because
5821 : // the delta layer 2000-3000 depends on it.
5822 16 : if !layers
5823 16 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5824 : {
5825 12 : info!("keeping {} because it is the latest layer", l.layer_name());
5826 12 : result.layers_not_updated += 1;
5827 12 : continue 'outer;
5828 4 : }
5829 4 :
5830 4 : // We didn't find any reason to keep this file, so remove it.
5831 4 : info!(
5832 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5833 0 : l.layer_name(),
5834 0 : l.is_incremental(),
5835 : );
5836 4 : layers_to_remove.push(l);
5837 : }
5838 :
5839 8 : if !layers_to_remove.is_empty() {
5840 : // Persist the new GC cutoff value before we actually remove anything.
5841 : // This unconditionally schedules also an index_part.json update, even though, we will
5842 : // be doing one a bit later with the unlinked gc'd layers.
5843 4 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5844 4 : self.schedule_uploads(disk_consistent_lsn, None)
5845 4 : .map_err(|e| {
5846 0 : if self.cancel.is_cancelled() {
5847 0 : GcError::TimelineCancelled
5848 : } else {
5849 0 : GcError::Remote(e)
5850 : }
5851 4 : })?;
5852 :
5853 4 : let gc_layers = layers_to_remove
5854 4 : .iter()
5855 4 : .map(|x| guard.get_from_desc(x))
5856 4 : .collect::<Vec<Layer>>();
5857 4 :
5858 4 : result.layers_removed = gc_layers.len() as u64;
5859 4 :
5860 4 : self.remote_client.schedule_gc_update(&gc_layers)?;
5861 :
5862 4 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5863 4 :
5864 4 : #[cfg(feature = "testing")]
5865 4 : {
5866 4 : result.doomed_layers = gc_layers;
5867 4 : }
5868 4 : }
5869 :
5870 8 : info!(
5871 0 : "GC completed removing {} layers, cutoff {}",
5872 : result.layers_removed, new_gc_cutoff
5873 : );
5874 :
5875 8 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5876 8 : Ok(result)
5877 8 : }
5878 :
5879 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5880 1331890 : async fn reconstruct_value(
5881 1331890 : &self,
5882 1331890 : key: Key,
5883 1331890 : request_lsn: Lsn,
5884 1331890 : mut data: ValueReconstructState,
5885 1331890 : ) -> Result<Bytes, PageReconstructError> {
5886 1331890 : // Perform WAL redo if needed
5887 1331890 : data.records.reverse();
5888 1331890 :
5889 1331890 : // If we have a page image, and no WAL, we're all set
5890 1331890 : if data.records.is_empty() {
5891 1330254 : if let Some((img_lsn, img)) = &data.img {
5892 1330254 : trace!(
5893 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5894 : key,
5895 : img_lsn,
5896 : request_lsn,
5897 : );
5898 1330254 : Ok(img.clone())
5899 : } else {
5900 0 : Err(PageReconstructError::from(anyhow!(
5901 0 : "base image for {key} at {request_lsn} not found"
5902 0 : )))
5903 : }
5904 : } else {
5905 : // We need to do WAL redo.
5906 : //
5907 : // If we don't have a base image, then the oldest WAL record better initialize
5908 : // the page
5909 1636 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5910 0 : Err(PageReconstructError::from(anyhow!(
5911 0 : "Base image for {} at {} not found, but got {} WAL records",
5912 0 : key,
5913 0 : request_lsn,
5914 0 : data.records.len()
5915 0 : )))
5916 : } else {
5917 1636 : if data.img.is_some() {
5918 1504 : trace!(
5919 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5920 0 : data.records.len(),
5921 : key,
5922 : request_lsn
5923 : );
5924 : } else {
5925 132 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5926 : };
5927 1636 : let res = self
5928 1636 : .walredo_mgr
5929 1636 : .as_ref()
5930 1636 : .context("timeline has no walredo manager")
5931 1636 : .map_err(PageReconstructError::WalRedo)?
5932 1636 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5933 1636 : .await;
5934 1636 : let img = match res {
5935 1636 : Ok(img) => img,
5936 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5937 0 : Err(walredo::Error::Other(err)) => {
5938 0 : critical!("walredo failure during page reconstruction: {err:?}");
5939 0 : return Err(PageReconstructError::WalRedo(
5940 0 : err.context("reconstruct a page image"),
5941 0 : ));
5942 : }
5943 : };
5944 1636 : Ok(img)
5945 : }
5946 : }
5947 1331890 : }
5948 :
5949 0 : pub(crate) async fn spawn_download_all_remote_layers(
5950 0 : self: Arc<Self>,
5951 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5952 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5953 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5954 :
5955 : // this is not really needed anymore; it has tests which really check the return value from
5956 : // http api. it would be better not to maintain this anymore.
5957 :
5958 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5959 0 : if let Some(st) = &*status_guard {
5960 0 : match &st.state {
5961 : DownloadRemoteLayersTaskState::Running => {
5962 0 : return Err(st.clone());
5963 : }
5964 : DownloadRemoteLayersTaskState::ShutDown
5965 0 : | DownloadRemoteLayersTaskState::Completed => {
5966 0 : *status_guard = None;
5967 0 : }
5968 : }
5969 0 : }
5970 :
5971 0 : let self_clone = Arc::clone(&self);
5972 0 : let task_id = task_mgr::spawn(
5973 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5974 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5975 0 : self.tenant_shard_id,
5976 0 : Some(self.timeline_id),
5977 0 : "download all remote layers task",
5978 0 : async move {
5979 0 : self_clone.download_all_remote_layers(request).await;
5980 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5981 0 : match &mut *status_guard {
5982 : None => {
5983 0 : warn!("tasks status is supposed to be Some(), since we are running");
5984 : }
5985 0 : Some(st) => {
5986 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5987 0 : if st.task_id != exp_task_id {
5988 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5989 0 : } else {
5990 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5991 0 : }
5992 : }
5993 : };
5994 0 : Ok(())
5995 0 : }
5996 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))
5997 : );
5998 :
5999 0 : let initial_info = DownloadRemoteLayersTaskInfo {
6000 0 : task_id: format!("{task_id}"),
6001 0 : state: DownloadRemoteLayersTaskState::Running,
6002 0 : total_layer_count: 0,
6003 0 : successful_download_count: 0,
6004 0 : failed_download_count: 0,
6005 0 : };
6006 0 : *status_guard = Some(initial_info.clone());
6007 0 :
6008 0 : Ok(initial_info)
6009 0 : }
6010 :
6011 0 : async fn download_all_remote_layers(
6012 0 : self: &Arc<Self>,
6013 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6014 0 : ) {
6015 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6016 :
6017 0 : let remaining = {
6018 0 : let guard = self.layers.read().await;
6019 0 : let Ok(lm) = guard.layer_map() else {
6020 : // technically here we could look into iterating accessible layers, but downloading
6021 : // all layers of a shutdown timeline makes no sense regardless.
6022 0 : tracing::info!("attempted to download all layers of shutdown timeline");
6023 0 : return;
6024 : };
6025 0 : lm.iter_historic_layers()
6026 0 : .map(|desc| guard.get_from_desc(&desc))
6027 0 : .collect::<Vec<_>>()
6028 0 : };
6029 0 : let total_layer_count = remaining.len();
6030 :
6031 : macro_rules! lock_status {
6032 : ($st:ident) => {
6033 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
6034 : let st = st
6035 : .as_mut()
6036 : .expect("this function is only called after the task has been spawned");
6037 : assert_eq!(
6038 : st.task_id,
6039 : format!(
6040 : "{}",
6041 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
6042 : )
6043 : );
6044 : let $st = st;
6045 : };
6046 : }
6047 :
6048 : {
6049 0 : lock_status!(st);
6050 0 : st.total_layer_count = total_layer_count as u64;
6051 0 : }
6052 0 :
6053 0 : let mut remaining = remaining.into_iter();
6054 0 : let mut have_remaining = true;
6055 0 : let mut js = tokio::task::JoinSet::new();
6056 0 :
6057 0 : let cancel = task_mgr::shutdown_token();
6058 0 :
6059 0 : let limit = request.max_concurrent_downloads;
6060 :
6061 : loop {
6062 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
6063 0 : let Some(next) = remaining.next() else {
6064 0 : have_remaining = false;
6065 0 : break;
6066 : };
6067 :
6068 0 : let span = tracing::info_span!("download", layer = %next);
6069 :
6070 0 : js.spawn(
6071 0 : async move {
6072 0 : let res = next.download().await;
6073 0 : (next, res)
6074 0 : }
6075 0 : .instrument(span),
6076 0 : );
6077 0 : }
6078 :
6079 0 : while let Some(res) = js.join_next().await {
6080 0 : match res {
6081 : Ok((_, Ok(_))) => {
6082 0 : lock_status!(st);
6083 0 : st.successful_download_count += 1;
6084 : }
6085 0 : Ok((layer, Err(e))) => {
6086 0 : tracing::error!(%layer, "download failed: {e:#}");
6087 0 : lock_status!(st);
6088 0 : st.failed_download_count += 1;
6089 : }
6090 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
6091 0 : Err(je) if je.is_panic() => {
6092 0 : lock_status!(st);
6093 0 : st.failed_download_count += 1;
6094 : }
6095 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
6096 : }
6097 : }
6098 :
6099 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
6100 0 : break;
6101 0 : }
6102 : }
6103 :
6104 : {
6105 0 : lock_status!(st);
6106 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6107 : }
6108 0 : }
6109 :
6110 0 : pub(crate) fn get_download_all_remote_layers_task_info(
6111 0 : &self,
6112 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
6113 0 : self.download_all_remote_layers_task_info
6114 0 : .read()
6115 0 : .unwrap()
6116 0 : .clone()
6117 0 : }
6118 : }
6119 :
6120 : impl Timeline {
6121 : /// Returns non-remote layers for eviction.
6122 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
6123 0 : let guard = self.layers.read().await;
6124 0 : let mut max_layer_size: Option<u64> = None;
6125 0 :
6126 0 : let resident_layers = guard
6127 0 : .likely_resident_layers()
6128 0 : .map(|layer| {
6129 0 : let file_size = layer.layer_desc().file_size;
6130 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
6131 0 :
6132 0 : let last_activity_ts = layer.latest_activity();
6133 0 :
6134 0 : EvictionCandidate {
6135 0 : layer: layer.to_owned().into(),
6136 0 : last_activity_ts,
6137 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
6138 0 : visibility: layer.visibility(),
6139 0 : }
6140 0 : })
6141 0 : .collect();
6142 0 :
6143 0 : DiskUsageEvictionInfo {
6144 0 : max_layer_size,
6145 0 : resident_layers,
6146 0 : }
6147 0 : }
6148 :
6149 3744 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
6150 3744 : ShardIndex {
6151 3744 : shard_number: self.tenant_shard_id.shard_number,
6152 3744 : shard_count: self.tenant_shard_id.shard_count,
6153 3744 : }
6154 3744 : }
6155 :
6156 : /// Persistently blocks gc for `Manual` reason.
6157 : ///
6158 : /// Returns true if no such block existed before, false otherwise.
6159 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
6160 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6161 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6162 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
6163 0 : }
6164 :
6165 : /// Persistently unblocks gc for `Manual` reason.
6166 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
6167 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6168 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6169 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
6170 0 : }
6171 :
6172 : #[cfg(test)]
6173 92 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
6174 92 : self.last_record_lsn.advance(new_lsn);
6175 92 : }
6176 :
6177 : #[cfg(test)]
6178 8 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
6179 8 : self.disk_consistent_lsn.store(new_value);
6180 8 : }
6181 :
6182 : /// Force create an image layer and place it into the layer map.
6183 : ///
6184 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
6185 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
6186 : /// placed into the layer map in one run AND be validated.
6187 : #[cfg(test)]
6188 116 : pub(super) async fn force_create_image_layer(
6189 116 : self: &Arc<Timeline>,
6190 116 : lsn: Lsn,
6191 116 : mut images: Vec<(Key, Bytes)>,
6192 116 : check_start_lsn: Option<Lsn>,
6193 116 : ctx: &RequestContext,
6194 116 : ) -> anyhow::Result<()> {
6195 116 : let last_record_lsn = self.get_last_record_lsn();
6196 116 : assert!(
6197 116 : lsn <= last_record_lsn,
6198 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
6199 : );
6200 116 : if let Some(check_start_lsn) = check_start_lsn {
6201 116 : assert!(lsn >= check_start_lsn);
6202 0 : }
6203 348 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
6204 116 : let min_key = *images.first().map(|(k, _)| k).unwrap();
6205 116 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
6206 116 : let mut image_layer_writer = ImageLayerWriter::new(
6207 116 : self.conf,
6208 116 : self.timeline_id,
6209 116 : self.tenant_shard_id,
6210 116 : &(min_key..end_key),
6211 116 : lsn,
6212 116 : ctx,
6213 116 : )
6214 116 : .await?;
6215 580 : for (key, img) in images {
6216 464 : image_layer_writer.put_image(key, img, ctx).await?;
6217 : }
6218 116 : let (desc, path) = image_layer_writer.finish(ctx).await?;
6219 116 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6220 116 : info!("force created image layer {}", image_layer.local_path());
6221 : {
6222 116 : let mut guard = self.layers.write().await;
6223 116 : guard
6224 116 : .open_mut()
6225 116 : .unwrap()
6226 116 : .force_insert_layer(image_layer.clone());
6227 116 : }
6228 116 :
6229 116 : // Update remote_timeline_client state to reflect existence of this layer
6230 116 : self.remote_client
6231 116 : .schedule_layer_file_upload(image_layer)
6232 116 : .unwrap();
6233 116 :
6234 116 : Ok(())
6235 116 : }
6236 :
6237 : /// Force create a delta layer and place it into the layer map.
6238 : ///
6239 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
6240 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
6241 : /// placed into the layer map in one run AND be validated.
6242 : #[cfg(test)]
6243 180 : pub(super) async fn force_create_delta_layer(
6244 180 : self: &Arc<Timeline>,
6245 180 : mut deltas: DeltaLayerTestDesc,
6246 180 : check_start_lsn: Option<Lsn>,
6247 180 : ctx: &RequestContext,
6248 180 : ) -> anyhow::Result<()> {
6249 180 : let last_record_lsn = self.get_last_record_lsn();
6250 180 : deltas
6251 180 : .data
6252 264 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
6253 180 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
6254 180 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
6255 624 : for (_, lsn, _) in &deltas.data {
6256 444 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
6257 : }
6258 180 : assert!(
6259 180 : deltas.lsn_range.end <= last_record_lsn,
6260 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
6261 : deltas.lsn_range.end,
6262 : last_record_lsn
6263 : );
6264 180 : if let Some(check_start_lsn) = check_start_lsn {
6265 180 : assert!(deltas.lsn_range.start >= check_start_lsn);
6266 0 : }
6267 180 : let mut delta_layer_writer = DeltaLayerWriter::new(
6268 180 : self.conf,
6269 180 : self.timeline_id,
6270 180 : self.tenant_shard_id,
6271 180 : deltas.key_range.start,
6272 180 : deltas.lsn_range,
6273 180 : ctx,
6274 180 : )
6275 180 : .await?;
6276 624 : for (key, lsn, val) in deltas.data {
6277 444 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
6278 : }
6279 180 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
6280 180 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6281 180 : info!("force created delta layer {}", delta_layer.local_path());
6282 : {
6283 180 : let mut guard = self.layers.write().await;
6284 180 : guard
6285 180 : .open_mut()
6286 180 : .unwrap()
6287 180 : .force_insert_layer(delta_layer.clone());
6288 180 : }
6289 180 :
6290 180 : // Update remote_timeline_client state to reflect existence of this layer
6291 180 : self.remote_client
6292 180 : .schedule_layer_file_upload(delta_layer)
6293 180 : .unwrap();
6294 180 :
6295 180 : Ok(())
6296 180 : }
6297 :
6298 : /// Return all keys at the LSN in the image layers
6299 : #[cfg(test)]
6300 12 : pub(crate) async fn inspect_image_layers(
6301 12 : self: &Arc<Timeline>,
6302 12 : lsn: Lsn,
6303 12 : ctx: &RequestContext,
6304 12 : io_concurrency: IoConcurrency,
6305 12 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
6306 12 : let mut all_data = Vec::new();
6307 12 : let guard = self.layers.read().await;
6308 68 : for layer in guard.layer_map()?.iter_historic_layers() {
6309 68 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
6310 12 : let layer = guard.get_from_desc(&layer);
6311 12 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
6312 12 : layer
6313 12 : .get_values_reconstruct_data(
6314 12 : KeySpace::single(Key::MIN..Key::MAX),
6315 12 : lsn..Lsn(lsn.0 + 1),
6316 12 : &mut reconstruct_data,
6317 12 : ctx,
6318 12 : )
6319 12 : .await?;
6320 104 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
6321 104 : let v = v.collect_pending_ios().await?;
6322 104 : all_data.push((k, v.img.unwrap().1));
6323 : }
6324 56 : }
6325 : }
6326 12 : all_data.sort();
6327 12 : Ok(all_data)
6328 12 : }
6329 :
6330 : /// Get all historic layer descriptors in the layer map
6331 : #[cfg(test)]
6332 48 : pub(crate) async fn inspect_historic_layers(
6333 48 : self: &Arc<Timeline>,
6334 48 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
6335 48 : let mut layers = Vec::new();
6336 48 : let guard = self.layers.read().await;
6337 228 : for layer in guard.layer_map()?.iter_historic_layers() {
6338 228 : layers.push(layer.key());
6339 228 : }
6340 48 : Ok(layers)
6341 48 : }
6342 :
6343 : #[cfg(test)]
6344 20 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
6345 20 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
6346 20 : keyspace.merge(&ks);
6347 20 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
6348 20 : }
6349 : }
6350 :
6351 : /// Tracking writes ingestion does to a particular in-memory layer.
6352 : ///
6353 : /// Cleared upon freezing a layer.
6354 : pub(crate) struct TimelineWriterState {
6355 : open_layer: Arc<InMemoryLayer>,
6356 : current_size: u64,
6357 : // Previous Lsn which passed through
6358 : prev_lsn: Option<Lsn>,
6359 : // Largest Lsn which passed through the current writer
6360 : max_lsn: Option<Lsn>,
6361 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
6362 : cached_last_freeze_at: Lsn,
6363 : }
6364 :
6365 : impl TimelineWriterState {
6366 2592 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
6367 2592 : Self {
6368 2592 : open_layer,
6369 2592 : current_size,
6370 2592 : prev_lsn: None,
6371 2592 : max_lsn: None,
6372 2592 : cached_last_freeze_at: last_freeze_at,
6373 2592 : }
6374 2592 : }
6375 : }
6376 :
6377 : /// Various functions to mutate the timeline.
6378 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
6379 : // This is probably considered a bad practice in Rust and should be fixed eventually,
6380 : // but will cause large code changes.
6381 : pub(crate) struct TimelineWriter<'a> {
6382 : tl: &'a Timeline,
6383 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
6384 : }
6385 :
6386 : impl Deref for TimelineWriter<'_> {
6387 : type Target = Timeline;
6388 :
6389 19796336 : fn deref(&self) -> &Self::Target {
6390 19796336 : self.tl
6391 19796336 : }
6392 : }
6393 :
6394 : #[derive(PartialEq)]
6395 : enum OpenLayerAction {
6396 : Roll,
6397 : Open,
6398 : None,
6399 : }
6400 :
6401 : impl TimelineWriter<'_> {
6402 9608464 : async fn handle_open_layer_action(
6403 9608464 : &mut self,
6404 9608464 : at: Lsn,
6405 9608464 : action: OpenLayerAction,
6406 9608464 : ctx: &RequestContext,
6407 9608464 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
6408 9608464 : match action {
6409 : OpenLayerAction::Roll => {
6410 160 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
6411 160 : self.roll_layer(freeze_at).await?;
6412 160 : self.open_layer(at, ctx).await?;
6413 : }
6414 2432 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
6415 : OpenLayerAction::None => {
6416 9605872 : assert!(self.write_guard.is_some());
6417 : }
6418 : }
6419 :
6420 9608464 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
6421 9608464 : }
6422 :
6423 2592 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
6424 2592 : let layer = self
6425 2592 : .tl
6426 2592 : .get_layer_for_write(at, &self.write_guard, ctx)
6427 2592 : .await?;
6428 2592 : let initial_size = layer.size().await?;
6429 :
6430 2592 : let last_freeze_at = self.last_freeze_at.load();
6431 2592 : self.write_guard.replace(TimelineWriterState::new(
6432 2592 : layer,
6433 2592 : initial_size,
6434 2592 : last_freeze_at,
6435 2592 : ));
6436 2592 :
6437 2592 : Ok(())
6438 2592 : }
6439 :
6440 160 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
6441 160 : let current_size = self.write_guard.as_ref().unwrap().current_size;
6442 :
6443 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
6444 : // to propagate the backpressure up into WAL ingestion.
6445 160 : let l0_count = self
6446 160 : .tl
6447 160 : .layers
6448 160 : .read()
6449 160 : .await
6450 160 : .layer_map()?
6451 160 : .level0_deltas()
6452 160 : .len();
6453 160 : let wait_thresholds = [
6454 160 : self.get_l0_flush_delay_threshold(),
6455 160 : self.get_l0_flush_stall_threshold(),
6456 160 : ];
6457 160 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
6458 :
6459 : // self.write_guard will be taken by the freezing
6460 160 : let flush_id = self
6461 160 : .tl
6462 160 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
6463 160 : .await?;
6464 :
6465 160 : assert!(self.write_guard.is_none());
6466 :
6467 160 : if let Some(wait_threshold) = wait_threshold {
6468 0 : if l0_count >= wait_threshold {
6469 0 : info!("layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers");
6470 0 : self.tl.wait_flush_completion(flush_id).await?;
6471 0 : }
6472 160 : }
6473 :
6474 160 : if current_size >= self.get_checkpoint_distance() * 2 {
6475 0 : warn!("Flushed oversized open layer with size {}", current_size)
6476 160 : }
6477 :
6478 160 : Ok(())
6479 160 : }
6480 :
6481 9608464 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
6482 9608464 : let state = &*self.write_guard;
6483 9608464 : let Some(state) = &state else {
6484 2432 : return OpenLayerAction::Open;
6485 : };
6486 :
6487 : #[cfg(feature = "testing")]
6488 9606032 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
6489 : // this check and assertion are not really needed because
6490 : // LayerManager::try_freeze_in_memory_layer will always clear out the
6491 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
6492 : // is no TimelineWriterState.
6493 0 : assert!(
6494 0 : state.open_layer.end_lsn.get().is_some(),
6495 0 : "our open_layer must be outdated"
6496 : );
6497 :
6498 : // this would be a memory leak waiting to happen because the in-memory layer always has
6499 : // an index
6500 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
6501 9606032 : }
6502 9606032 :
6503 9606032 : if state.prev_lsn == Some(lsn) {
6504 : // Rolling mid LSN is not supported by [downstream code].
6505 : // Hence, only roll at LSN boundaries.
6506 : //
6507 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
6508 12 : return OpenLayerAction::None;
6509 9606020 : }
6510 9606020 :
6511 9606020 : if state.current_size == 0 {
6512 : // Don't roll empty layers
6513 0 : return OpenLayerAction::None;
6514 9606020 : }
6515 9606020 :
6516 9606020 : if self.tl.should_roll(
6517 9606020 : state.current_size,
6518 9606020 : state.current_size + new_value_size,
6519 9606020 : self.get_checkpoint_distance(),
6520 9606020 : lsn,
6521 9606020 : state.cached_last_freeze_at,
6522 9606020 : state.open_layer.get_opened_at(),
6523 9606020 : ) {
6524 160 : OpenLayerAction::Roll
6525 : } else {
6526 9605860 : OpenLayerAction::None
6527 : }
6528 9608464 : }
6529 :
6530 : /// Put a batch of keys at the specified Lsns.
6531 9608460 : pub(crate) async fn put_batch(
6532 9608460 : &mut self,
6533 9608460 : batch: SerializedValueBatch,
6534 9608460 : ctx: &RequestContext,
6535 9608460 : ) -> anyhow::Result<()> {
6536 9608460 : if !batch.has_data() {
6537 0 : return Ok(());
6538 9608460 : }
6539 9608460 :
6540 9608460 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
6541 9608460 : // We don't assert this in release builds, since key ownership policies may change over
6542 9608460 : // time. Stray keys will be removed during compaction.
6543 9608460 : if cfg!(debug_assertions) {
6544 19789792 : for metadata in &batch.metadata {
6545 10181332 : if let ValueMeta::Serialized(metadata) = metadata {
6546 10181332 : let key = Key::from_compact(metadata.key);
6547 10181332 : assert!(
6548 10181332 : self.shard_identity.is_key_local(&key)
6549 0 : || self.shard_identity.is_key_global(&key),
6550 0 : "key {key} does not belong on shard {}",
6551 0 : self.shard_identity.shard_index()
6552 : );
6553 0 : }
6554 : }
6555 0 : }
6556 :
6557 9608460 : let batch_max_lsn = batch.max_lsn;
6558 9608460 : let buf_size: u64 = batch.buffer_size() as u64;
6559 9608460 :
6560 9608460 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
6561 9608460 : let layer = self
6562 9608460 : .handle_open_layer_action(batch_max_lsn, action, ctx)
6563 9608460 : .await?;
6564 :
6565 9608460 : let res = layer.put_batch(batch, ctx).await;
6566 :
6567 9608460 : if res.is_ok() {
6568 9608460 : // Update the current size only when the entire write was ok.
6569 9608460 : // In case of failures, we may have had partial writes which
6570 9608460 : // render the size tracking out of sync. That's ok because
6571 9608460 : // the checkpoint distance should be significantly smaller
6572 9608460 : // than the S3 single shot upload limit of 5GiB.
6573 9608460 : let state = self.write_guard.as_mut().unwrap();
6574 9608460 :
6575 9608460 : state.current_size += buf_size;
6576 9608460 : state.prev_lsn = Some(batch_max_lsn);
6577 9608460 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
6578 9608460 : }
6579 :
6580 9608460 : res
6581 9608460 : }
6582 :
6583 : #[cfg(test)]
6584 : /// Test helper, for tests that would like to poke individual values without composing a batch
6585 8780308 : pub(crate) async fn put(
6586 8780308 : &mut self,
6587 8780308 : key: Key,
6588 8780308 : lsn: Lsn,
6589 8780308 : value: &Value,
6590 8780308 : ctx: &RequestContext,
6591 8780308 : ) -> anyhow::Result<()> {
6592 : use utils::bin_ser::BeSer;
6593 8780308 : if !key.is_valid_key_on_write_path() {
6594 0 : bail!(
6595 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
6596 0 : key
6597 0 : );
6598 8780308 : }
6599 8780308 : let val_ser_size = value.serialized_size().unwrap() as usize;
6600 8780308 : let batch = SerializedValueBatch::from_values(vec![(
6601 8780308 : key.to_compact(),
6602 8780308 : lsn,
6603 8780308 : val_ser_size,
6604 8780308 : value.clone(),
6605 8780308 : )]);
6606 8780308 :
6607 8780308 : self.put_batch(batch, ctx).await
6608 8780308 : }
6609 :
6610 4 : pub(crate) async fn delete_batch(
6611 4 : &mut self,
6612 4 : batch: &[(Range<Key>, Lsn)],
6613 4 : ctx: &RequestContext,
6614 4 : ) -> anyhow::Result<()> {
6615 4 : if let Some((_, lsn)) = batch.first() {
6616 4 : let action = self.get_open_layer_action(*lsn, 0);
6617 4 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
6618 4 : layer.put_tombstones(batch).await?;
6619 0 : }
6620 :
6621 4 : Ok(())
6622 4 : }
6623 :
6624 : /// Track the end of the latest digested WAL record.
6625 : /// Remember the (end of) last valid WAL record remembered in the timeline.
6626 : ///
6627 : /// Call this after you have finished writing all the WAL up to 'lsn'.
6628 : ///
6629 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
6630 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
6631 : /// the latest and can be read.
6632 10558184 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
6633 10558184 : self.tl.finish_write(new_lsn);
6634 10558184 : }
6635 :
6636 541140 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
6637 541140 : self.tl.update_current_logical_size(delta)
6638 541140 : }
6639 : }
6640 :
6641 : // We need TimelineWriter to be send in upcoming conversion of
6642 : // Timeline::layers to tokio::sync::RwLock.
6643 : #[test]
6644 4 : fn is_send() {
6645 4 : fn _assert_send<T: Send>() {}
6646 4 : _assert_send::<TimelineWriter<'_>>();
6647 4 : }
6648 :
6649 : #[cfg(test)]
6650 : mod tests {
6651 : use pageserver_api::key::Key;
6652 : use pageserver_api::value::Value;
6653 : use utils::{id::TimelineId, lsn::Lsn};
6654 :
6655 : use crate::tenant::{
6656 : harness::{test_img, TenantHarness},
6657 : layer_map::LayerMap,
6658 : storage_layer::{Layer, LayerName},
6659 : timeline::{DeltaLayerTestDesc, EvictionError},
6660 : Timeline,
6661 : };
6662 :
6663 : #[tokio::test]
6664 4 : async fn test_heatmap_generation() {
6665 4 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
6666 4 :
6667 4 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6668 4 : Lsn(0x10)..Lsn(0x20),
6669 4 : vec![(
6670 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6671 4 : Lsn(0x11),
6672 4 : Value::Image(test_img("foo")),
6673 4 : )],
6674 4 : );
6675 4 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6676 4 : Lsn(0x10)..Lsn(0x20),
6677 4 : vec![(
6678 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6679 4 : Lsn(0x11),
6680 4 : Value::Image(test_img("foo")),
6681 4 : )],
6682 4 : );
6683 4 : let l0_delta = DeltaLayerTestDesc::new(
6684 4 : Lsn(0x20)..Lsn(0x30),
6685 4 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6686 4 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6687 4 : vec![(
6688 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6689 4 : Lsn(0x25),
6690 4 : Value::Image(test_img("foo")),
6691 4 : )],
6692 4 : );
6693 4 : let delta_layers = vec![
6694 4 : covered_delta.clone(),
6695 4 : visible_delta.clone(),
6696 4 : l0_delta.clone(),
6697 4 : ];
6698 4 :
6699 4 : let image_layer = (
6700 4 : Lsn(0x40),
6701 4 : vec![(
6702 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6703 4 : test_img("bar"),
6704 4 : )],
6705 4 : );
6706 4 : let image_layers = vec![image_layer];
6707 4 :
6708 4 : let (tenant, ctx) = harness.load().await;
6709 4 : let timeline = tenant
6710 4 : .create_test_timeline_with_layers(
6711 4 : TimelineId::generate(),
6712 4 : Lsn(0x10),
6713 4 : 14,
6714 4 : &ctx,
6715 4 : delta_layers,
6716 4 : image_layers,
6717 4 : Lsn(0x100),
6718 4 : )
6719 4 : .await
6720 4 : .unwrap();
6721 4 :
6722 4 : // Layer visibility is an input to heatmap generation, so refresh it first
6723 4 : timeline.update_layer_visibility().await.unwrap();
6724 4 :
6725 4 : let heatmap = timeline
6726 4 : .generate_heatmap()
6727 4 : .await
6728 4 : .expect("Infallible while timeline is not shut down");
6729 4 :
6730 4 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
6731 4 :
6732 4 : // L0 should come last
6733 4 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
6734 4 :
6735 4 : let mut last_lsn = Lsn::MAX;
6736 20 : for layer in heatmap.layers {
6737 4 : // Covered layer should be omitted
6738 16 : assert!(layer.name != covered_delta.layer_name());
6739 4 :
6740 16 : let layer_lsn = match &layer.name {
6741 8 : LayerName::Delta(d) => d.lsn_range.end,
6742 8 : LayerName::Image(i) => i.lsn,
6743 4 : };
6744 4 :
6745 4 : // Apart from L0s, newest Layers should come first
6746 16 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
6747 12 : assert!(layer_lsn <= last_lsn);
6748 12 : last_lsn = layer_lsn;
6749 4 : }
6750 4 : }
6751 4 : }
6752 :
6753 : #[tokio::test]
6754 4 : async fn two_layer_eviction_attempts_at_the_same_time() {
6755 4 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6756 4 : .await
6757 4 : .unwrap();
6758 4 :
6759 4 : let (tenant, ctx) = harness.load().await;
6760 4 : let timeline = tenant
6761 4 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6762 4 : .await
6763 4 : .unwrap();
6764 4 :
6765 4 : let layer = find_some_layer(&timeline).await;
6766 4 : let layer = layer
6767 4 : .keep_resident()
6768 4 : .await
6769 4 : .expect("no download => no downloading errors")
6770 4 : .drop_eviction_guard();
6771 4 :
6772 4 : let forever = std::time::Duration::from_secs(120);
6773 4 :
6774 4 : let first = layer.evict_and_wait(forever);
6775 4 : let second = layer.evict_and_wait(forever);
6776 4 :
6777 4 : let (first, second) = tokio::join!(first, second);
6778 4 :
6779 4 : let res = layer.keep_resident().await;
6780 4 : assert!(res.is_none(), "{res:?}");
6781 4 :
6782 4 : match (first, second) {
6783 4 : (Ok(()), Ok(())) => {
6784 4 : // because there are no more timeline locks being taken on eviction path, we can
6785 4 : // witness all three outcomes here.
6786 4 : }
6787 4 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
6788 0 : // if one completes before the other, this is fine just as well.
6789 0 : }
6790 4 : other => unreachable!("unexpected {:?}", other),
6791 4 : }
6792 4 : }
6793 :
6794 4 : async fn find_some_layer(timeline: &Timeline) -> Layer {
6795 4 : let layers = timeline.layers.read().await;
6796 4 : let desc = layers
6797 4 : .layer_map()
6798 4 : .unwrap()
6799 4 : .iter_historic_layers()
6800 4 : .next()
6801 4 : .expect("must find one layer to evict");
6802 4 :
6803 4 : layers.get_from_desc(&desc)
6804 4 : }
6805 : }
|
