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::{
154 : debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf, HeatMapTimeline,
155 : };
156 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
157 : use super::{
158 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
159 : storage_layer::ReadableLayer,
160 : };
161 : use super::{secondary::heatmap::HeatMapLayer, GcError};
162 :
163 : #[cfg(test)]
164 : use pageserver_api::value::Value;
165 :
166 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
167 : pub(crate) enum FlushLoopState {
168 : NotStarted,
169 : Running {
170 : #[cfg(test)]
171 : expect_initdb_optimization: bool,
172 : #[cfg(test)]
173 : initdb_optimization_count: usize,
174 : },
175 : Exited,
176 : }
177 :
178 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
179 : pub enum ImageLayerCreationMode {
180 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
181 : Try,
182 : /// Force creating the image layers if possible. For now, no image layers will be created
183 : /// for metadata keys. Used in compaction code path with force flag enabled.
184 : Force,
185 : /// Initial ingestion of the data, and no data should be dropped in this function. This
186 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
187 : /// code path.
188 : Initial,
189 : }
190 :
191 : #[derive(Clone, Debug, Default)]
192 : pub enum LastImageLayerCreationStatus {
193 : Incomplete {
194 : /// The last key of the partition (exclusive) that was processed in the last
195 : /// image layer creation attempt. We will continue from this key in the next
196 : /// attempt.
197 : last_key: Key,
198 : },
199 : Complete,
200 : #[default]
201 : Initial,
202 : }
203 :
204 : impl std::fmt::Display for ImageLayerCreationMode {
205 1181 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
206 1181 : write!(f, "{:?}", self)
207 1181 : }
208 : }
209 :
210 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
211 : /// Can be removed after all refactors are done.
212 92 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
213 92 : drop(rlock)
214 92 : }
215 :
216 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
217 : /// Can be removed after all refactors are done.
218 1273 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
219 1273 : drop(rlock)
220 1273 : }
221 :
222 : /// The outward-facing resources required to build a Timeline
223 : pub struct TimelineResources {
224 : pub remote_client: RemoteTimelineClient,
225 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
226 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
227 : pub l0_compaction_trigger: Arc<Notify>,
228 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
229 : }
230 :
231 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
232 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
233 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
234 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
235 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
236 : pub(crate) struct RelSizeCache {
237 : pub(crate) complete_as_of: Lsn,
238 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
239 : }
240 :
241 : pub struct Timeline {
242 : pub(crate) conf: &'static PageServerConf,
243 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
244 :
245 : myself: Weak<Self>,
246 :
247 : pub(crate) tenant_shard_id: TenantShardId,
248 : pub timeline_id: TimelineId,
249 :
250 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
251 : /// Never changes for the lifetime of this [`Timeline`] object.
252 : ///
253 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
254 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
255 : pub(crate) generation: Generation,
256 :
257 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
258 : /// to shards, and is constant through the lifetime of this Timeline.
259 : shard_identity: ShardIdentity,
260 :
261 : pub pg_version: u32,
262 :
263 : /// The tuple has two elements.
264 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
265 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
266 : ///
267 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
268 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
269 : ///
270 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
271 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
272 : /// `PersistentLayerDesc`'s.
273 : ///
274 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
275 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
276 : /// runtime, e.g., during page reconstruction.
277 : ///
278 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
279 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
280 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
281 :
282 : last_freeze_at: AtomicLsn,
283 : // Atomic would be more appropriate here.
284 : last_freeze_ts: RwLock<Instant>,
285 :
286 : pub(crate) standby_horizon: AtomicLsn,
287 :
288 : // WAL redo manager. `None` only for broken tenants.
289 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
290 :
291 : /// Remote storage client.
292 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
293 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
294 :
295 : // What page versions do we hold in the repository? If we get a
296 : // request > last_record_lsn, we need to wait until we receive all
297 : // the WAL up to the request. The SeqWait provides functions for
298 : // that. TODO: If we get a request for an old LSN, such that the
299 : // versions have already been garbage collected away, we should
300 : // throw an error, but we don't track that currently.
301 : //
302 : // last_record_lsn.load().last points to the end of last processed WAL record.
303 : //
304 : // We also remember the starting point of the previous record in
305 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
306 : // first WAL record when the node is started up. But here, we just
307 : // keep track of it.
308 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
309 :
310 : // All WAL records have been processed and stored durably on files on
311 : // local disk, up to this LSN. On crash and restart, we need to re-process
312 : // the WAL starting from this point.
313 : //
314 : // Some later WAL records might have been processed and also flushed to disk
315 : // already, so don't be surprised to see some, but there's no guarantee on
316 : // them yet.
317 : disk_consistent_lsn: AtomicLsn,
318 :
319 : // Parent timeline that this timeline was branched from, and the LSN
320 : // of the branch point.
321 : ancestor_timeline: Option<Arc<Timeline>>,
322 : ancestor_lsn: Lsn,
323 :
324 : pub(super) metrics: TimelineMetrics,
325 :
326 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
327 : // in `crate::page_service` writes these metrics.
328 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
329 :
330 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
331 :
332 : /// Ensures layers aren't frozen by checkpointer between
333 : /// [`Timeline::get_layer_for_write`] and layer reads.
334 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
335 : /// Must always be acquired before the layer map/individual layer lock
336 : /// to avoid deadlock.
337 : ///
338 : /// The state is cleared upon freezing.
339 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
340 :
341 : /// Used to avoid multiple `flush_loop` tasks running
342 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
343 :
344 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
345 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
346 : /// The flush cycle counter is sent back on the layer_flush_done channel when
347 : /// the flush finishes. You can use that to wait for the flush to finish.
348 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
349 : /// read by whoever sends an update
350 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
351 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
352 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
353 :
354 : // The LSN at which we have executed GC: whereas [`Self::gc_info`] records the LSN at which
355 : // we _intend_ to GC (i.e. the PITR cutoff), this LSN records where we actually last did it.
356 : // Because PITR interval is mutable, it's possible for this LSN to be earlier or later than
357 : // the planned GC cutoff.
358 : pub applied_gc_cutoff_lsn: Rcu<Lsn>,
359 :
360 : pub(crate) gc_compaction_layer_update_lock: tokio::sync::RwLock<()>,
361 :
362 : // List of child timelines and their branch points. This is needed to avoid
363 : // garbage collecting data that is still needed by the child timelines.
364 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
365 :
366 : pub(crate) last_image_layer_creation_status: ArcSwap<LastImageLayerCreationStatus>,
367 :
368 : // It may change across major versions so for simplicity
369 : // keep it after running initdb for a timeline.
370 : // It is needed in checks when we want to error on some operations
371 : // when they are requested for pre-initdb lsn.
372 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
373 : // though let's keep them both for better error visibility.
374 : pub initdb_lsn: Lsn,
375 :
376 : /// The repartitioning result. Allows a single writer and multiple readers.
377 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
378 :
379 : /// Configuration: how often should the partitioning be recalculated.
380 : repartition_threshold: u64,
381 :
382 : last_image_layer_creation_check_at: AtomicLsn,
383 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
384 :
385 : /// Current logical size of the "datadir", at the last LSN.
386 : current_logical_size: LogicalSize,
387 :
388 : /// Information about the last processed message by the WAL receiver,
389 : /// or None if WAL receiver has not received anything for this timeline
390 : /// yet.
391 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
392 : pub walreceiver: Mutex<Option<WalReceiver>>,
393 :
394 : /// Relation size cache
395 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
396 :
397 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
398 :
399 : state: watch::Sender<TimelineState>,
400 :
401 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
402 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
403 : pub delete_progress: TimelineDeleteProgress,
404 :
405 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
406 :
407 : /// Load or creation time information about the disk_consistent_lsn and when the loading
408 : /// happened. Used for consumption metrics.
409 : pub(crate) loaded_at: (Lsn, SystemTime),
410 :
411 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
412 : pub(crate) gate: Gate,
413 :
414 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
415 : /// to the timeline should drop out when this token fires.
416 : pub(crate) cancel: CancellationToken,
417 :
418 : /// Make sure we only have one running compaction at a time in tests.
419 : ///
420 : /// Must only be taken in two places:
421 : /// - [`Timeline::compact`] (this file)
422 : /// - [`delete::delete_local_timeline_directory`]
423 : ///
424 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
425 : compaction_lock: tokio::sync::Mutex<()>,
426 :
427 : /// If true, the last compaction failed.
428 : compaction_failed: AtomicBool,
429 :
430 : /// Notifies the tenant compaction loop that there is pending L0 compaction work.
431 : l0_compaction_trigger: Arc<Notify>,
432 :
433 : /// Make sure we only have one running gc at a time.
434 : ///
435 : /// Must only be taken in two places:
436 : /// - [`Timeline::gc`] (this file)
437 : /// - [`delete::delete_local_timeline_directory`]
438 : ///
439 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
440 : gc_lock: tokio::sync::Mutex<()>,
441 :
442 : /// Cloned from [`super::Tenant::pagestream_throttle`] on construction.
443 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
444 :
445 : /// Size estimator for aux file v2
446 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
447 :
448 : /// Some test cases directly place keys into the timeline without actually modifying the directory
449 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
450 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
451 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
452 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
453 : #[cfg(test)]
454 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
455 :
456 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
457 :
458 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
459 :
460 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
461 :
462 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
463 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
464 :
465 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
466 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
467 :
468 : previous_heatmap: ArcSwapOption<PreviousHeatmap>,
469 : }
470 :
471 : pub(crate) enum PreviousHeatmap {
472 : Active {
473 : heatmap: HeatMapTimeline,
474 : read_at: std::time::Instant,
475 : },
476 : Obsolete,
477 : }
478 :
479 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
480 :
481 : pub struct WalReceiverInfo {
482 : pub wal_source_connconf: PgConnectionConfig,
483 : pub last_received_msg_lsn: Lsn,
484 : pub last_received_msg_ts: u128,
485 : }
486 :
487 : /// Information about how much history needs to be retained, needed by
488 : /// Garbage Collection.
489 : #[derive(Default)]
490 : pub(crate) struct GcInfo {
491 : /// Specific LSNs that are needed.
492 : ///
493 : /// Currently, this includes all points where child branches have
494 : /// been forked off from. In the future, could also include
495 : /// explicit user-defined snapshot points.
496 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
497 :
498 : /// The cutoff coordinates, which are combined by selecting the minimum.
499 : pub(crate) cutoffs: GcCutoffs,
500 :
501 : /// Leases granted to particular LSNs.
502 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
503 :
504 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
505 : pub(crate) within_ancestor_pitr: bool,
506 : }
507 :
508 : impl GcInfo {
509 600 : pub(crate) fn min_cutoff(&self) -> Lsn {
510 600 : self.cutoffs.select_min()
511 600 : }
512 :
513 464 : pub(super) fn insert_child(
514 464 : &mut self,
515 464 : child_id: TimelineId,
516 464 : child_lsn: Lsn,
517 464 : is_offloaded: MaybeOffloaded,
518 464 : ) {
519 464 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
520 464 : self.retain_lsns.sort_by_key(|i| i.0);
521 464 : }
522 :
523 8 : pub(super) fn remove_child_maybe_offloaded(
524 8 : &mut self,
525 8 : child_id: TimelineId,
526 8 : maybe_offloaded: MaybeOffloaded,
527 8 : ) -> bool {
528 8 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
529 8 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
530 8 : let mut removed = false;
531 12 : self.retain_lsns.retain(|i| {
532 12 : if removed {
533 4 : return true;
534 8 : }
535 8 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
536 8 : removed |= remove;
537 8 : !remove
538 12 : });
539 8 : removed
540 8 : }
541 :
542 8 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
543 8 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
544 8 : }
545 :
546 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
547 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
548 0 : }
549 464 : pub(crate) fn lsn_covered_by_lease(&self, lsn: Lsn) -> bool {
550 464 : self.leases.contains_key(&lsn)
551 464 : }
552 : }
553 :
554 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
555 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
556 : /// between time-based and space-based retention for observability and consumption metrics purposes.
557 : #[derive(Debug, Clone)]
558 : pub(crate) struct GcCutoffs {
559 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
560 : /// history we must keep to retain a specified number of bytes of WAL.
561 : pub(crate) space: Lsn,
562 :
563 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
564 : /// history we must keep to enable reading back at least the PITR interval duration.
565 : pub(crate) time: Lsn,
566 : }
567 :
568 : impl Default for GcCutoffs {
569 897 : fn default() -> Self {
570 897 : Self {
571 897 : space: Lsn::INVALID,
572 897 : time: Lsn::INVALID,
573 897 : }
574 897 : }
575 : }
576 :
577 : impl GcCutoffs {
578 600 : fn select_min(&self) -> Lsn {
579 600 : std::cmp::min(self.space, self.time)
580 600 : }
581 : }
582 :
583 : pub(crate) struct TimelineVisitOutcome {
584 : completed_keyspace: KeySpace,
585 : image_covered_keyspace: KeySpace,
586 : }
587 :
588 : /// An error happened in a get() operation.
589 : #[derive(thiserror::Error, Debug)]
590 : pub(crate) enum PageReconstructError {
591 : #[error(transparent)]
592 : Other(anyhow::Error),
593 :
594 : #[error("Ancestor LSN wait error: {0}")]
595 : AncestorLsnTimeout(WaitLsnError),
596 :
597 : #[error("timeline shutting down")]
598 : Cancelled,
599 :
600 : /// An error happened replaying WAL records
601 : #[error(transparent)]
602 : WalRedo(anyhow::Error),
603 :
604 : #[error("{0}")]
605 : MissingKey(MissingKeyError),
606 : }
607 :
608 : impl From<anyhow::Error> for PageReconstructError {
609 0 : fn from(value: anyhow::Error) -> Self {
610 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
611 0 : match value.downcast::<PageReconstructError>() {
612 0 : Ok(pre) => pre,
613 0 : Err(other) => PageReconstructError::Other(other),
614 : }
615 0 : }
616 : }
617 :
618 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
619 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
620 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
621 0 : }
622 : }
623 :
624 : impl From<layer_manager::Shutdown> for PageReconstructError {
625 0 : fn from(_: layer_manager::Shutdown) -> Self {
626 0 : PageReconstructError::Cancelled
627 0 : }
628 : }
629 :
630 : impl GetVectoredError {
631 : #[cfg(test)]
632 12 : pub(crate) fn is_missing_key_error(&self) -> bool {
633 12 : matches!(self, Self::MissingKey(_))
634 12 : }
635 : }
636 :
637 : impl From<layer_manager::Shutdown> for GetVectoredError {
638 0 : fn from(_: layer_manager::Shutdown) -> Self {
639 0 : GetVectoredError::Cancelled
640 0 : }
641 : }
642 :
643 : /// A layer identifier when used in the [`ReadPath`] structure. This enum is for observability purposes
644 : /// only and not used by the "real read path".
645 : pub enum ReadPathLayerId {
646 : PersistentLayer(PersistentLayerKey),
647 : InMemoryLayer(Range<Lsn>),
648 : }
649 :
650 : impl std::fmt::Display for ReadPathLayerId {
651 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
652 0 : match self {
653 0 : ReadPathLayerId::PersistentLayer(key) => write!(f, "{}", key),
654 0 : ReadPathLayerId::InMemoryLayer(range) => {
655 0 : write!(f, "in-mem {}..{}", range.start, range.end)
656 : }
657 : }
658 0 : }
659 : }
660 : pub struct ReadPath {
661 : keyspace: KeySpace,
662 : lsn: Lsn,
663 : path: Vec<(ReadPathLayerId, KeySpace, Range<Lsn>)>,
664 : }
665 :
666 : impl ReadPath {
667 1255313 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
668 1255313 : Self {
669 1255313 : keyspace,
670 1255313 : lsn,
671 1255313 : path: Vec::new(),
672 1255313 : }
673 1255313 : }
674 :
675 1767933 : pub fn record_layer_visit(
676 1767933 : &mut self,
677 1767933 : layer_to_read: &ReadableLayer,
678 1767933 : keyspace_to_read: &KeySpace,
679 1767933 : lsn_range: &Range<Lsn>,
680 1767933 : ) {
681 1767933 : let id = match layer_to_read {
682 554613 : ReadableLayer::PersistentLayer(layer) => {
683 554613 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
684 : }
685 1213320 : ReadableLayer::InMemoryLayer(layer) => {
686 1213320 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
687 : }
688 : };
689 1767933 : self.path
690 1767933 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
691 1767933 : }
692 : }
693 :
694 : impl std::fmt::Display for ReadPath {
695 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
696 0 : writeln!(f, "Read path for {} at lsn {}:", self.keyspace, self.lsn)?;
697 0 : for (idx, (layer_id, keyspace, lsn_range)) in self.path.iter().enumerate() {
698 0 : writeln!(
699 0 : f,
700 0 : "{}: {} {}..{} {}",
701 0 : idx, layer_id, lsn_range.start, lsn_range.end, keyspace
702 0 : )?;
703 : }
704 0 : Ok(())
705 0 : }
706 : }
707 :
708 : #[derive(thiserror::Error)]
709 : pub struct MissingKeyError {
710 : key: Key,
711 : shard: ShardNumber,
712 : cont_lsn: Lsn,
713 : request_lsn: Lsn,
714 : ancestor_lsn: Option<Lsn>,
715 : /// Debug information about the read path if there's an error
716 : read_path: Option<ReadPath>,
717 : backtrace: Option<std::backtrace::Backtrace>,
718 : }
719 :
720 : impl std::fmt::Debug for MissingKeyError {
721 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
722 0 : write!(f, "{}", self)
723 0 : }
724 : }
725 :
726 : impl std::fmt::Display for MissingKeyError {
727 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
728 0 : write!(
729 0 : f,
730 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
731 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
732 0 : )?;
733 :
734 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
735 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
736 0 : }
737 :
738 0 : if let Some(ref read_path) = self.read_path {
739 0 : write!(f, "\n{}", read_path)?;
740 0 : }
741 :
742 0 : if let Some(ref backtrace) = self.backtrace {
743 0 : write!(f, "\n{}", backtrace)?;
744 0 : }
745 :
746 0 : Ok(())
747 0 : }
748 : }
749 :
750 : impl PageReconstructError {
751 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
752 0 : pub(crate) fn is_stopping(&self) -> bool {
753 : use PageReconstructError::*;
754 0 : match self {
755 0 : Cancelled => true,
756 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
757 : }
758 0 : }
759 : }
760 :
761 : #[derive(thiserror::Error, Debug)]
762 : pub(crate) enum CreateImageLayersError {
763 : #[error("timeline shutting down")]
764 : Cancelled,
765 :
766 : #[error("read failed")]
767 : GetVectoredError(#[source] GetVectoredError),
768 :
769 : #[error("reconstruction failed")]
770 : PageReconstructError(#[source] PageReconstructError),
771 :
772 : #[error(transparent)]
773 : Other(#[from] anyhow::Error),
774 : }
775 :
776 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
777 0 : fn from(_: layer_manager::Shutdown) -> Self {
778 0 : CreateImageLayersError::Cancelled
779 0 : }
780 : }
781 :
782 : #[derive(thiserror::Error, Debug, Clone)]
783 : pub(crate) enum FlushLayerError {
784 : /// Timeline cancellation token was cancelled
785 : #[error("timeline shutting down")]
786 : Cancelled,
787 :
788 : /// We tried to flush a layer while the Timeline is in an unexpected state
789 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
790 : NotRunning(FlushLoopState),
791 :
792 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
793 : // loop via a watch channel, where we can only borrow it.
794 : #[error("create image layers (shared)")]
795 : CreateImageLayersError(Arc<CreateImageLayersError>),
796 :
797 : #[error("other (shared)")]
798 : Other(#[from] Arc<anyhow::Error>),
799 : }
800 :
801 : impl FlushLayerError {
802 : // When crossing from generic anyhow errors to this error type, we explicitly check
803 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
804 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
805 0 : let cancelled = timeline.cancel.is_cancelled()
806 : // The upload queue might have been shut down before the official cancellation of the timeline.
807 0 : || err
808 0 : .downcast_ref::<NotInitialized>()
809 0 : .map(NotInitialized::is_stopping)
810 0 : .unwrap_or_default();
811 0 : if cancelled {
812 0 : Self::Cancelled
813 : } else {
814 0 : Self::Other(Arc::new(err))
815 : }
816 0 : }
817 : }
818 :
819 : impl From<layer_manager::Shutdown> for FlushLayerError {
820 0 : fn from(_: layer_manager::Shutdown) -> Self {
821 0 : FlushLayerError::Cancelled
822 0 : }
823 : }
824 :
825 : #[derive(thiserror::Error, Debug)]
826 : pub(crate) enum GetVectoredError {
827 : #[error("timeline shutting down")]
828 : Cancelled,
829 :
830 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
831 : Oversized(u64),
832 :
833 : #[error("requested at invalid LSN: {0}")]
834 : InvalidLsn(Lsn),
835 :
836 : #[error("requested key not found: {0}")]
837 : MissingKey(MissingKeyError),
838 :
839 : #[error("ancestry walk")]
840 : GetReadyAncestorError(#[source] GetReadyAncestorError),
841 :
842 : #[error(transparent)]
843 : Other(#[from] anyhow::Error),
844 : }
845 :
846 : impl From<GetReadyAncestorError> for GetVectoredError {
847 4 : fn from(value: GetReadyAncestorError) -> Self {
848 : use GetReadyAncestorError::*;
849 4 : match value {
850 0 : Cancelled => GetVectoredError::Cancelled,
851 : AncestorLsnTimeout(_) | BadState { .. } => {
852 4 : GetVectoredError::GetReadyAncestorError(value)
853 : }
854 : }
855 4 : }
856 : }
857 :
858 : #[derive(thiserror::Error, Debug)]
859 : pub(crate) enum GetReadyAncestorError {
860 : #[error("ancestor LSN wait error")]
861 : AncestorLsnTimeout(#[from] WaitLsnError),
862 :
863 : #[error("bad state on timeline {timeline_id}: {state:?}")]
864 : BadState {
865 : timeline_id: TimelineId,
866 : state: TimelineState,
867 : },
868 :
869 : #[error("cancelled")]
870 : Cancelled,
871 : }
872 :
873 : #[derive(Clone, Copy)]
874 : pub enum LogicalSizeCalculationCause {
875 : Initial,
876 : ConsumptionMetricsSyntheticSize,
877 : EvictionTaskImitation,
878 : TenantSizeHandler,
879 : }
880 :
881 : pub enum GetLogicalSizePriority {
882 : User,
883 : Background,
884 : }
885 :
886 0 : #[derive(Debug, enumset::EnumSetType)]
887 : pub(crate) enum CompactFlags {
888 : ForceRepartition,
889 : ForceImageLayerCreation,
890 : ForceL0Compaction,
891 : OnlyL0Compaction,
892 : EnhancedGcBottomMostCompaction,
893 : DryRun,
894 : /// Disables compaction yielding e.g. due to high L0 count. This is set e.g. when requesting
895 : /// compaction via HTTP API.
896 : NoYield,
897 : }
898 :
899 : #[serde_with::serde_as]
900 0 : #[derive(Debug, Clone, serde::Deserialize)]
901 : pub(crate) struct CompactRequest {
902 : pub compact_key_range: Option<CompactKeyRange>,
903 : pub compact_lsn_range: Option<CompactLsnRange>,
904 : /// Whether the compaction job should be scheduled.
905 : #[serde(default)]
906 : pub scheduled: bool,
907 : /// Whether the compaction job should be split across key ranges.
908 : #[serde(default)]
909 : pub sub_compaction: bool,
910 : /// Max job size for each subcompaction job.
911 : pub sub_compaction_max_job_size_mb: Option<u64>,
912 : }
913 :
914 : #[derive(Debug, Clone, Default)]
915 : pub(crate) struct CompactOptions {
916 : pub flags: EnumSet<CompactFlags>,
917 : /// If set, the compaction will only compact the key range specified by this option.
918 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
919 : pub compact_key_range: Option<CompactKeyRange>,
920 : /// If set, the compaction will only compact the LSN within this value.
921 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
922 : pub compact_lsn_range: Option<CompactLsnRange>,
923 : /// Enable sub-compaction (split compaction job across key ranges).
924 : /// This option is only used by GC compaction.
925 : pub sub_compaction: bool,
926 : /// Set job size for the GC compaction.
927 : /// This option is only used by GC compaction.
928 : pub sub_compaction_max_job_size_mb: Option<u64>,
929 : }
930 :
931 : impl std::fmt::Debug for Timeline {
932 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
933 0 : write!(f, "Timeline<{}>", self.timeline_id)
934 0 : }
935 : }
936 :
937 : #[derive(thiserror::Error, Debug)]
938 : pub(crate) enum WaitLsnError {
939 : // Called on a timeline which is shutting down
940 : #[error("Shutdown")]
941 : Shutdown,
942 :
943 : // Called on an timeline not in active state or shutting down
944 : #[error("Bad timeline state: {0:?}")]
945 : BadState(TimelineState),
946 :
947 : // Timeout expired while waiting for LSN to catch up with goal.
948 : #[error("{0}")]
949 : Timeout(String),
950 : }
951 :
952 : // The impls below achieve cancellation mapping for errors.
953 : // Perhaps there's a way of achieving this with less cruft.
954 :
955 : impl From<CreateImageLayersError> for CompactionError {
956 0 : fn from(e: CreateImageLayersError) -> Self {
957 0 : match e {
958 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
959 0 : CreateImageLayersError::Other(e) => {
960 0 : CompactionError::Other(e.context("create image layers"))
961 : }
962 0 : _ => CompactionError::Other(e.into()),
963 : }
964 0 : }
965 : }
966 :
967 : impl From<CreateImageLayersError> for FlushLayerError {
968 0 : fn from(e: CreateImageLayersError) -> Self {
969 0 : match e {
970 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
971 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
972 : }
973 0 : }
974 : }
975 :
976 : impl From<PageReconstructError> for CreateImageLayersError {
977 0 : fn from(e: PageReconstructError) -> Self {
978 0 : match e {
979 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
980 0 : _ => CreateImageLayersError::PageReconstructError(e),
981 : }
982 0 : }
983 : }
984 :
985 : impl From<GetVectoredError> for CreateImageLayersError {
986 0 : fn from(e: GetVectoredError) -> Self {
987 0 : match e {
988 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
989 0 : _ => CreateImageLayersError::GetVectoredError(e),
990 : }
991 0 : }
992 : }
993 :
994 : impl From<GetVectoredError> for PageReconstructError {
995 12 : fn from(e: GetVectoredError) -> Self {
996 12 : match e {
997 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
998 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
999 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
1000 8 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
1001 4 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
1002 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
1003 : }
1004 12 : }
1005 : }
1006 :
1007 : impl From<GetReadyAncestorError> for PageReconstructError {
1008 4 : fn from(e: GetReadyAncestorError) -> Self {
1009 : use GetReadyAncestorError::*;
1010 4 : match e {
1011 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
1012 4 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
1013 0 : Cancelled => PageReconstructError::Cancelled,
1014 : }
1015 4 : }
1016 : }
1017 :
1018 : pub(crate) enum WaitLsnTimeout {
1019 : Custom(Duration),
1020 : // Use the [`PageServerConf::wait_lsn_timeout`] default
1021 : Default,
1022 : }
1023 :
1024 : pub(crate) enum WaitLsnWaiter<'a> {
1025 : Timeline(&'a Timeline),
1026 : Tenant,
1027 : PageService,
1028 : HttpEndpoint,
1029 : }
1030 :
1031 : /// Argument to [`Timeline::shutdown`].
1032 : #[derive(Debug, Clone, Copy)]
1033 : pub(crate) enum ShutdownMode {
1034 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
1035 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
1036 : ///
1037 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1038 : /// the call to [`Timeline::shutdown`].
1039 : FreezeAndFlush,
1040 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1041 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1042 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1043 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1044 : Reload,
1045 : /// Shut down immediately, without waiting for any open layers to flush.
1046 : Hard,
1047 : }
1048 :
1049 : enum ImageLayerCreationOutcome {
1050 : /// We generated an image layer
1051 : Generated {
1052 : unfinished_image_layer: ImageLayerWriter,
1053 : },
1054 : /// The key range is empty
1055 : Empty,
1056 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1057 : /// the image layer creation.
1058 : Skip,
1059 : }
1060 :
1061 : /// Public interface functions
1062 : impl Timeline {
1063 : /// Get the LSN where this branch was created
1064 776 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1065 776 : self.ancestor_lsn
1066 776 : }
1067 :
1068 : /// Get the ancestor's timeline id
1069 24 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1070 24 : self.ancestor_timeline
1071 24 : .as_ref()
1072 24 : .map(|ancestor| ancestor.timeline_id)
1073 24 : }
1074 :
1075 : /// Get the ancestor timeline
1076 4 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1077 4 : self.ancestor_timeline.as_ref()
1078 4 : }
1079 :
1080 : /// Get the bytes written since the PITR cutoff on this branch, and
1081 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1082 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1083 0 : let gc_info = self.gc_info.read().unwrap();
1084 0 : let history = self
1085 0 : .get_last_record_lsn()
1086 0 : .checked_sub(gc_info.cutoffs.time)
1087 0 : .unwrap_or(Lsn(0))
1088 0 : .0;
1089 0 : (history, gc_info.within_ancestor_pitr)
1090 0 : }
1091 :
1092 : /// Read timeline's GC cutoff: this is the LSN at which GC has started to happen
1093 1707589 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1094 1707589 : self.applied_gc_cutoff_lsn.read()
1095 1707589 : }
1096 :
1097 : /// Read timeline's planned GC cutoff: this is the logical end of history that users
1098 : /// are allowed to read (based on configured PITR), even if physically we have more history.
1099 0 : pub(crate) fn get_gc_cutoff_lsn(&self) -> Lsn {
1100 0 : self.gc_info.read().unwrap().cutoffs.time
1101 0 : }
1102 :
1103 : /// Look up given page version.
1104 : ///
1105 : /// If a remote layer file is needed, it is downloaded as part of this
1106 : /// call.
1107 : ///
1108 : /// This method enforces [`Self::pagestream_throttle`] internally.
1109 : ///
1110 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1111 : /// abstraction above this needs to store suitable metadata to track what
1112 : /// data exists with what keys, in separate metadata entries. If a
1113 : /// non-existent key is requested, we may incorrectly return a value from
1114 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1115 : /// non-existing key.
1116 : ///
1117 : /// # Cancel-Safety
1118 : ///
1119 : /// This method is cancellation-safe.
1120 : #[inline(always)]
1121 1215242 : pub(crate) async fn get(
1122 1215242 : &self,
1123 1215242 : key: Key,
1124 1215242 : lsn: Lsn,
1125 1215242 : ctx: &RequestContext,
1126 1215242 : ) -> Result<Bytes, PageReconstructError> {
1127 1215242 : if !lsn.is_valid() {
1128 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1129 1215242 : }
1130 1215242 :
1131 1215242 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1132 1215242 : // already checked the key against the shard_identity when looking up the Timeline from
1133 1215242 : // page_service.
1134 1215242 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1135 :
1136 1215242 : let keyspace = KeySpace {
1137 1215242 : ranges: vec![key..key.next()],
1138 1215242 : };
1139 1215242 :
1140 1215242 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1141 :
1142 1215242 : let vectored_res = self
1143 1215242 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
1144 1215242 : .await;
1145 :
1146 1215242 : let key_value = vectored_res?.pop_first();
1147 1215230 : match key_value {
1148 1215206 : Some((got_key, value)) => {
1149 1215206 : if got_key != key {
1150 0 : error!(
1151 0 : "Expected {}, but singular vectored get returned {}",
1152 : key, got_key
1153 : );
1154 0 : Err(PageReconstructError::Other(anyhow!(
1155 0 : "Singular vectored get returned wrong key"
1156 0 : )))
1157 : } else {
1158 1215206 : value
1159 : }
1160 : }
1161 24 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1162 24 : key,
1163 24 : shard: self.shard_identity.get_shard_number(&key),
1164 24 : cont_lsn: Lsn(0),
1165 24 : request_lsn: lsn,
1166 24 : ancestor_lsn: None,
1167 24 : backtrace: None,
1168 24 : read_path: None,
1169 24 : })),
1170 : }
1171 1215242 : }
1172 :
1173 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1174 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1175 :
1176 : /// Look up multiple page versions at a given LSN
1177 : ///
1178 : /// This naive implementation will be replaced with a more efficient one
1179 : /// which actually vectorizes the read path.
1180 39338 : pub(crate) async fn get_vectored(
1181 39338 : &self,
1182 39338 : keyspace: KeySpace,
1183 39338 : lsn: Lsn,
1184 39338 : io_concurrency: super::storage_layer::IoConcurrency,
1185 39338 : ctx: &RequestContext,
1186 39338 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1187 39338 : if !lsn.is_valid() {
1188 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1189 39338 : }
1190 39338 :
1191 39338 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1192 39338 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1193 0 : return Err(GetVectoredError::Oversized(key_count));
1194 39338 : }
1195 :
1196 78676 : for range in &keyspace.ranges {
1197 39338 : let mut key = range.start;
1198 79141 : while key != range.end {
1199 39803 : assert!(!self.shard_identity.is_key_disposable(&key));
1200 39803 : key = key.next();
1201 : }
1202 : }
1203 :
1204 39338 : trace!(
1205 0 : "get vectored request for {:?}@{} from task kind {:?}",
1206 0 : keyspace,
1207 0 : lsn,
1208 0 : ctx.task_kind(),
1209 : );
1210 :
1211 39338 : let start = crate::metrics::GET_VECTORED_LATENCY
1212 39338 : .for_task_kind(ctx.task_kind())
1213 39338 : .map(|metric| (metric, Instant::now()));
1214 :
1215 39338 : let res = self
1216 39338 : .get_vectored_impl(
1217 39338 : keyspace.clone(),
1218 39338 : lsn,
1219 39338 : &mut ValuesReconstructState::new(io_concurrency),
1220 39338 : ctx,
1221 39338 : )
1222 39338 : .await;
1223 :
1224 39338 : if let Some((metric, start)) = start {
1225 0 : let elapsed = start.elapsed();
1226 0 : metric.observe(elapsed.as_secs_f64());
1227 39338 : }
1228 :
1229 39338 : res
1230 39338 : }
1231 :
1232 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1233 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1234 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1235 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1236 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1237 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1238 : /// the scan operation will not cause OOM in the future.
1239 24 : pub(crate) async fn scan(
1240 24 : &self,
1241 24 : keyspace: KeySpace,
1242 24 : lsn: Lsn,
1243 24 : ctx: &RequestContext,
1244 24 : io_concurrency: super::storage_layer::IoConcurrency,
1245 24 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1246 24 : if !lsn.is_valid() {
1247 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1248 24 : }
1249 24 :
1250 24 : trace!(
1251 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1252 0 : keyspace,
1253 0 : lsn,
1254 0 : ctx.task_kind()
1255 : );
1256 :
1257 : // We should generalize this into Keyspace::contains in the future.
1258 48 : for range in &keyspace.ranges {
1259 24 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1260 24 : || range.end.field1 > METADATA_KEY_END_PREFIX
1261 : {
1262 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1263 0 : "only metadata keyspace can be scanned"
1264 0 : )));
1265 24 : }
1266 : }
1267 :
1268 24 : let start = crate::metrics::SCAN_LATENCY
1269 24 : .for_task_kind(ctx.task_kind())
1270 24 : .map(ScanLatencyOngoingRecording::start_recording);
1271 :
1272 24 : let vectored_res = self
1273 24 : .get_vectored_impl(
1274 24 : keyspace.clone(),
1275 24 : lsn,
1276 24 : &mut ValuesReconstructState::new(io_concurrency),
1277 24 : ctx,
1278 24 : )
1279 24 : .await;
1280 :
1281 24 : if let Some(recording) = start {
1282 0 : recording.observe();
1283 24 : }
1284 :
1285 24 : vectored_res
1286 24 : }
1287 :
1288 1255313 : pub(super) async fn get_vectored_impl(
1289 1255313 : &self,
1290 1255313 : keyspace: KeySpace,
1291 1255313 : lsn: Lsn,
1292 1255313 : reconstruct_state: &mut ValuesReconstructState,
1293 1255313 : ctx: &RequestContext,
1294 1255313 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1295 1255313 : let read_path = if self.conf.enable_read_path_debugging {
1296 1255313 : Some(ReadPath::new(keyspace.clone(), lsn))
1297 : } else {
1298 0 : None
1299 : };
1300 1255313 : reconstruct_state.read_path = read_path;
1301 :
1302 1255313 : let traversal_res: Result<(), _> = self
1303 1255313 : .get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1304 1255313 : .await;
1305 1255313 : if let Err(err) = traversal_res {
1306 : // Wait for all the spawned IOs to complete.
1307 : // See comments on `spawn_io` inside `storage_layer` for more details.
1308 32 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1309 32 : .into_values()
1310 32 : .map(|state| state.collect_pending_ios())
1311 32 : .collect::<FuturesUnordered<_>>();
1312 32 : while collect_futs.next().await.is_some() {}
1313 32 : return Err(err);
1314 1255281 : };
1315 1255281 :
1316 1255281 : let layers_visited = reconstruct_state.get_layers_visited();
1317 1255281 :
1318 1255281 : let futs = FuturesUnordered::new();
1319 1327945 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1320 1327945 : futs.push({
1321 1327945 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1322 1327945 : async move {
1323 1327945 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1324 :
1325 1327945 : let converted = match state.collect_pending_ios().await {
1326 1327945 : Ok(ok) => ok,
1327 0 : Err(err) => {
1328 0 : return (key, Err(err));
1329 : }
1330 : };
1331 1327945 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1332 1327945 :
1333 1327945 : // The walredo module expects the records to be descending in terms of Lsn.
1334 1327945 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1335 1327945 : debug_assert!(
1336 1327945 : converted
1337 1327945 : .records
1338 1327945 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1339 0 : "{converted:?}"
1340 : );
1341 :
1342 : (
1343 1327945 : key,
1344 1327945 : walredo_self.reconstruct_value(key, lsn, converted).await,
1345 : )
1346 1327945 : }
1347 1327945 : });
1348 1327945 : }
1349 :
1350 1255281 : let results = futs
1351 1255281 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1352 1255281 : .await;
1353 :
1354 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1355 : // when they're missing. Instead they are omitted from the resulting btree
1356 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1357 : // to avoid infinite results.
1358 1255281 : if !results.is_empty() {
1359 : // Record the total number of layers visited towards each key in the batch. While some
1360 : // layers may not intersect with a given read, and the cost of layer visits are
1361 : // amortized across the batch, each visited layer contributes directly to the observed
1362 : // latency for every read in the batch, which is what we care about.
1363 1254792 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1364 0 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1365 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1366 0 : LOG_PACER.lock().unwrap().call(|| {
1367 0 : let num_keys = keyspace.total_raw_size();
1368 0 : let num_pages = results.len();
1369 0 : tracing::info!(
1370 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1371 0 : lsn = %lsn,
1372 0 : "Vectored read for {keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1373 : );
1374 0 : });
1375 1254792 : }
1376 :
1377 2582737 : for _ in &results {
1378 1327945 : self.metrics.layers_per_read.observe(layers_visited as f64);
1379 1327945 : LAYERS_PER_READ_GLOBAL.observe(layers_visited as f64);
1380 1327945 : }
1381 489 : }
1382 :
1383 1255281 : Ok(results)
1384 1255313 : }
1385 :
1386 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1387 548229 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1388 548229 : self.last_record_lsn.load().last
1389 548229 : }
1390 :
1391 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1392 0 : self.last_record_lsn.load().prev
1393 0 : }
1394 :
1395 : /// Atomically get both last and prev.
1396 456 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1397 456 : self.last_record_lsn.load()
1398 456 : }
1399 :
1400 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1401 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1402 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1403 0 : self.last_record_lsn.status_receiver()
1404 0 : }
1405 :
1406 1657 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1407 1657 : self.disk_consistent_lsn.load()
1408 1657 : }
1409 :
1410 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1411 : /// not validated with control plane yet.
1412 : /// See [`Self::get_remote_consistent_lsn_visible`].
1413 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1414 0 : self.remote_client.remote_consistent_lsn_projected()
1415 0 : }
1416 :
1417 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1418 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1419 : /// generation validation in the deletion queue.
1420 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1421 0 : self.remote_client.remote_consistent_lsn_visible()
1422 0 : }
1423 :
1424 : /// The sum of the file size of all historic layers in the layer map.
1425 : /// This method makes no distinction between local and remote layers.
1426 : /// Hence, the result **does not represent local filesystem usage**.
1427 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1428 0 : let guard = self.layers.read().await;
1429 0 : guard.layer_size_sum()
1430 0 : }
1431 :
1432 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1433 0 : self.metrics.resident_physical_size_get()
1434 0 : }
1435 :
1436 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1437 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1438 0 : }
1439 :
1440 : ///
1441 : /// Wait until WAL has been received and processed up to this LSN.
1442 : ///
1443 : /// You should call this before any of the other get_* or list_* functions. Calling
1444 : /// those functions with an LSN that has been processed yet is an error.
1445 : ///
1446 451672 : pub(crate) async fn wait_lsn(
1447 451672 : &self,
1448 451672 : lsn: Lsn,
1449 451672 : who_is_waiting: WaitLsnWaiter<'_>,
1450 451672 : timeout: WaitLsnTimeout,
1451 451672 : ctx: &RequestContext, /* Prepare for use by cancellation */
1452 451672 : ) -> Result<(), WaitLsnError> {
1453 451672 : let state = self.current_state();
1454 451672 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1455 0 : return Err(WaitLsnError::Shutdown);
1456 451672 : } else if !matches!(state, TimelineState::Active) {
1457 0 : return Err(WaitLsnError::BadState(state));
1458 451672 : }
1459 451672 :
1460 451672 : if cfg!(debug_assertions) {
1461 451672 : match ctx.task_kind() {
1462 : TaskKind::WalReceiverManager
1463 : | TaskKind::WalReceiverConnectionHandler
1464 : | TaskKind::WalReceiverConnectionPoller => {
1465 0 : let is_myself = match who_is_waiting {
1466 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1467 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService | WaitLsnWaiter::HttpEndpoint => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1468 : };
1469 0 : if is_myself {
1470 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1471 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1472 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1473 0 : }
1474 0 : } else {
1475 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1476 0 : // our walreceiver task can make progress independent of theirs
1477 0 : }
1478 : }
1479 451672 : _ => {}
1480 : }
1481 0 : }
1482 :
1483 451672 : let timeout = match timeout {
1484 0 : WaitLsnTimeout::Custom(t) => t,
1485 451672 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1486 : };
1487 :
1488 451672 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1489 451672 :
1490 451672 : match self.last_record_lsn.wait_for_timeout(lsn, timeout).await {
1491 451672 : Ok(()) => Ok(()),
1492 0 : Err(e) => {
1493 : use utils::seqwait::SeqWaitError::*;
1494 0 : match e {
1495 0 : Shutdown => Err(WaitLsnError::Shutdown),
1496 : Timeout => {
1497 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1498 0 : drop(_timer);
1499 0 : let walreceiver_status = self.walreceiver_status();
1500 0 : Err(WaitLsnError::Timeout(format!(
1501 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1502 0 : lsn,
1503 0 : self.get_last_record_lsn(),
1504 0 : self.get_disk_consistent_lsn(),
1505 0 : walreceiver_status,
1506 0 : )))
1507 : }
1508 : }
1509 : }
1510 : }
1511 451672 : }
1512 :
1513 0 : pub(crate) fn walreceiver_status(&self) -> String {
1514 0 : match &*self.walreceiver.lock().unwrap() {
1515 0 : None => "stopping or stopped".to_string(),
1516 0 : Some(walreceiver) => match walreceiver.status() {
1517 0 : Some(status) => status.to_human_readable_string(),
1518 0 : None => "Not active".to_string(),
1519 : },
1520 : }
1521 0 : }
1522 :
1523 : /// Check that it is valid to request operations with that lsn.
1524 464 : pub(crate) fn check_lsn_is_in_scope(
1525 464 : &self,
1526 464 : lsn: Lsn,
1527 464 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1528 464 : ) -> anyhow::Result<()> {
1529 464 : ensure!(
1530 464 : lsn >= **latest_gc_cutoff_lsn,
1531 8 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1532 8 : lsn,
1533 8 : **latest_gc_cutoff_lsn,
1534 : );
1535 456 : Ok(())
1536 464 : }
1537 :
1538 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1539 20 : pub(crate) fn init_lsn_lease(
1540 20 : &self,
1541 20 : lsn: Lsn,
1542 20 : length: Duration,
1543 20 : ctx: &RequestContext,
1544 20 : ) -> anyhow::Result<LsnLease> {
1545 20 : self.make_lsn_lease(lsn, length, true, ctx)
1546 20 : }
1547 :
1548 : /// 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.
1549 8 : pub(crate) fn renew_lsn_lease(
1550 8 : &self,
1551 8 : lsn: Lsn,
1552 8 : length: Duration,
1553 8 : ctx: &RequestContext,
1554 8 : ) -> anyhow::Result<LsnLease> {
1555 8 : self.make_lsn_lease(lsn, length, false, ctx)
1556 8 : }
1557 :
1558 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1559 : ///
1560 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1561 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1562 : ///
1563 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1564 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1565 28 : fn make_lsn_lease(
1566 28 : &self,
1567 28 : lsn: Lsn,
1568 28 : length: Duration,
1569 28 : init: bool,
1570 28 : _ctx: &RequestContext,
1571 28 : ) -> anyhow::Result<LsnLease> {
1572 24 : let lease = {
1573 : // Normalize the requested LSN to be aligned, and move to the first record
1574 : // if it points to the beginning of the page (header).
1575 28 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1576 28 :
1577 28 : let mut gc_info = self.gc_info.write().unwrap();
1578 28 : let planned_cutoff = gc_info.min_cutoff();
1579 28 :
1580 28 : let valid_until = SystemTime::now() + length;
1581 28 :
1582 28 : let entry = gc_info.leases.entry(lsn);
1583 28 :
1584 28 : match entry {
1585 12 : Entry::Occupied(mut occupied) => {
1586 12 : let existing_lease = occupied.get_mut();
1587 12 : if valid_until > existing_lease.valid_until {
1588 4 : existing_lease.valid_until = valid_until;
1589 4 : let dt: DateTime<Utc> = valid_until.into();
1590 4 : info!("lease extended to {}", dt);
1591 : } else {
1592 8 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1593 8 : info!("existing lease covers greater length, valid until {}", dt);
1594 : }
1595 :
1596 12 : existing_lease.clone()
1597 : }
1598 16 : Entry::Vacant(vacant) => {
1599 : // Reject already GC-ed LSN if we are in AttachedSingle and
1600 : // not blocked by the lsn lease deadline.
1601 16 : let validate = {
1602 16 : let conf = self.tenant_conf.load();
1603 16 : conf.location.attach_mode == AttachmentMode::Single
1604 16 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1605 : };
1606 :
1607 16 : if init || validate {
1608 16 : let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
1609 16 : if lsn < *latest_gc_cutoff_lsn {
1610 4 : bail!("tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1611 12 : }
1612 12 : if lsn < planned_cutoff {
1613 0 : bail!("tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}", lsn, planned_cutoff);
1614 12 : }
1615 0 : }
1616 :
1617 12 : let dt: DateTime<Utc> = valid_until.into();
1618 12 : info!("lease created, valid until {}", dt);
1619 12 : vacant.insert(LsnLease { valid_until }).clone()
1620 : }
1621 : }
1622 : };
1623 :
1624 24 : Ok(lease)
1625 28 : }
1626 :
1627 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1628 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1629 : #[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))]
1630 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1631 : self.freeze0().await
1632 : }
1633 :
1634 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1635 : #[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))]
1636 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1637 : self.freeze_and_flush0().await
1638 : }
1639 :
1640 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1641 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1642 2245 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1643 2245 : let mut g = self.write_lock.lock().await;
1644 2245 : let to_lsn = self.get_last_record_lsn();
1645 2245 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1646 2245 : }
1647 :
1648 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1649 : // polluting the span hierarchy.
1650 2245 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1651 2245 : let token = self.freeze0().await?;
1652 2245 : self.wait_flush_completion(token).await
1653 2245 : }
1654 :
1655 : // Check if an open ephemeral layer should be closed: this provides
1656 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1657 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1658 : // ephemeral layer bytes has been breached.
1659 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1660 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1661 : // If the write lock is held, there is an active wal receiver: rolling open layers
1662 : // is their responsibility while they hold this lock.
1663 0 : return;
1664 : };
1665 :
1666 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1667 : // time, and this was missed.
1668 : // if write_guard.is_none() { return; }
1669 :
1670 0 : let Ok(layers_guard) = self.layers.try_read() else {
1671 : // Don't block if the layer lock is busy
1672 0 : return;
1673 : };
1674 :
1675 0 : let Ok(lm) = layers_guard.layer_map() else {
1676 0 : return;
1677 : };
1678 :
1679 0 : let Some(open_layer) = &lm.open_layer else {
1680 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1681 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1682 : // that didn't result in writes to this shard.
1683 :
1684 : // Must not hold the layers lock while waiting for a flush.
1685 0 : drop(layers_guard);
1686 0 :
1687 0 : let last_record_lsn = self.get_last_record_lsn();
1688 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1689 0 : if last_record_lsn > disk_consistent_lsn {
1690 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1691 : // we are a sharded tenant and have skipped some WAL
1692 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1693 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1694 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1695 : // without any data ingested (yet) doesn't write a remote index as soon as it
1696 : // sees its LSN advance: we only do this if we've been layer-less
1697 : // for some time.
1698 0 : tracing::debug!(
1699 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1700 : disk_consistent_lsn,
1701 : last_record_lsn
1702 : );
1703 :
1704 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1705 : // We know there is no open layer, so we can request freezing without actually
1706 : // freezing anything. This is true even if we have dropped the layers_guard, we
1707 : // still hold the write_guard.
1708 0 : let _ = async {
1709 0 : let token = self
1710 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1711 0 : .await?;
1712 0 : self.wait_flush_completion(token).await
1713 0 : }
1714 0 : .await;
1715 0 : }
1716 0 : }
1717 :
1718 0 : return;
1719 : };
1720 :
1721 0 : let Some(current_size) = open_layer.try_len() else {
1722 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1723 : // read lock to get size should always succeed.
1724 0 : tracing::warn!("Lock conflict while reading size of open layer");
1725 0 : return;
1726 : };
1727 :
1728 0 : let current_lsn = self.get_last_record_lsn();
1729 :
1730 0 : let checkpoint_distance_override = open_layer.tick().await;
1731 :
1732 0 : if let Some(size_override) = checkpoint_distance_override {
1733 0 : if current_size > size_override {
1734 : // This is not harmful, but it only happens in relatively rare cases where
1735 : // time-based checkpoints are not happening fast enough to keep the amount of
1736 : // ephemeral data within configured limits. It's a sign of stress on the system.
1737 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1738 0 : }
1739 0 : }
1740 :
1741 0 : let checkpoint_distance =
1742 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1743 0 :
1744 0 : if self.should_roll(
1745 0 : current_size,
1746 0 : current_size,
1747 0 : checkpoint_distance,
1748 0 : self.get_last_record_lsn(),
1749 0 : self.last_freeze_at.load(),
1750 0 : open_layer.get_opened_at(),
1751 0 : ) {
1752 0 : match open_layer.info() {
1753 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1754 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1755 0 : // happens asynchronously in the background.
1756 0 : tracing::debug!(
1757 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1758 : );
1759 : }
1760 : InMemoryLayerInfo::Open { .. } => {
1761 : // Upgrade to a write lock and freeze the layer
1762 0 : drop(layers_guard);
1763 0 : let res = self
1764 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1765 0 : .await;
1766 :
1767 0 : if let Err(e) = res {
1768 0 : tracing::info!(
1769 0 : "failed to flush frozen layer after background freeze: {e:#}"
1770 : );
1771 0 : }
1772 : }
1773 : }
1774 0 : }
1775 0 : }
1776 :
1777 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1778 : ///
1779 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1780 : /// child timelines that are not offloaded yet.
1781 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1782 0 : if self.remote_client.is_archived() != Some(true) {
1783 0 : return (false, "the timeline is not archived");
1784 0 : }
1785 0 : if !self.remote_client.no_pending_work() {
1786 : // if the remote client is still processing some work, we can't offload
1787 0 : return (false, "the upload queue is not drained yet");
1788 0 : }
1789 0 :
1790 0 : (true, "ok")
1791 0 : }
1792 :
1793 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1794 : /// compaction tasks.
1795 768 : pub(crate) async fn compact(
1796 768 : self: &Arc<Self>,
1797 768 : cancel: &CancellationToken,
1798 768 : flags: EnumSet<CompactFlags>,
1799 768 : ctx: &RequestContext,
1800 768 : ) -> Result<CompactionOutcome, CompactionError> {
1801 768 : self.compact_with_options(
1802 768 : cancel,
1803 768 : CompactOptions {
1804 768 : flags,
1805 768 : compact_key_range: None,
1806 768 : compact_lsn_range: None,
1807 768 : sub_compaction: false,
1808 768 : sub_compaction_max_job_size_mb: None,
1809 768 : },
1810 768 : ctx,
1811 768 : )
1812 768 : .await
1813 768 : }
1814 :
1815 : /// Outermost timeline compaction operation; downloads needed layers.
1816 : ///
1817 : /// NB: the cancellation token is usually from a background task, but can also come from a
1818 : /// request task.
1819 768 : pub(crate) async fn compact_with_options(
1820 768 : self: &Arc<Self>,
1821 768 : cancel: &CancellationToken,
1822 768 : options: CompactOptions,
1823 768 : ctx: &RequestContext,
1824 768 : ) -> Result<CompactionOutcome, CompactionError> {
1825 768 : // Acquire the compaction lock and task semaphore.
1826 768 : //
1827 768 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
1828 768 : // out by other background tasks (including image compaction). We request this via
1829 768 : // `BackgroundLoopKind::L0Compaction`.
1830 768 : //
1831 768 : // If this is a regular compaction pass, and L0-only compaction is enabled in the config,
1832 768 : // then we should yield for immediate L0 compaction if necessary while we're waiting for the
1833 768 : // background task semaphore. There's no point yielding otherwise, since we'd just end up
1834 768 : // right back here.
1835 768 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
1836 768 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
1837 0 : true => BackgroundLoopKind::L0Compaction,
1838 768 : false => BackgroundLoopKind::Compaction,
1839 : };
1840 768 : let yield_for_l0 = !is_l0_only
1841 768 : && self.get_compaction_l0_first()
1842 0 : && !options.flags.contains(CompactFlags::NoYield);
1843 :
1844 768 : let acquire = async move {
1845 768 : let guard = self.compaction_lock.lock().await;
1846 768 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
1847 768 : (guard, permit)
1848 768 : };
1849 :
1850 768 : let (_guard, _permit) = tokio::select! {
1851 768 : (guard, permit) = acquire => (guard, permit),
1852 768 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
1853 0 : return Ok(CompactionOutcome::YieldForL0);
1854 : }
1855 768 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
1856 768 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
1857 : };
1858 :
1859 768 : let last_record_lsn = self.get_last_record_lsn();
1860 768 :
1861 768 : // Last record Lsn could be zero in case the timeline was just created
1862 768 : if !last_record_lsn.is_valid() {
1863 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1864 0 : return Ok(CompactionOutcome::Skipped);
1865 768 : }
1866 :
1867 768 : let result = match self.get_compaction_algorithm_settings().kind {
1868 : CompactionAlgorithm::Tiered => {
1869 0 : self.compact_tiered(cancel, ctx).await?;
1870 0 : Ok(CompactionOutcome::Done)
1871 : }
1872 768 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1873 : };
1874 :
1875 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
1876 0 : match result {
1877 768 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
1878 : Err(CompactionError::Other(_)) => {
1879 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
1880 : }
1881 : // Don't change the current value on offload failure or shutdown. We don't want to
1882 : // abruptly stall nor resume L0 flushes in these cases.
1883 0 : Err(CompactionError::Offload(_)) => {}
1884 0 : Err(CompactionError::ShuttingDown) => {}
1885 : };
1886 :
1887 768 : result
1888 768 : }
1889 :
1890 : /// Mutate the timeline with a [`TimelineWriter`].
1891 10266385 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1892 10266385 : TimelineWriter {
1893 10266385 : tl: self,
1894 10266385 : write_guard: self.write_lock.lock().await,
1895 : }
1896 10266385 : }
1897 :
1898 0 : pub(crate) fn activate(
1899 0 : self: &Arc<Self>,
1900 0 : parent: Arc<crate::tenant::Tenant>,
1901 0 : broker_client: BrokerClientChannel,
1902 0 : background_jobs_can_start: Option<&completion::Barrier>,
1903 0 : ctx: &RequestContext,
1904 0 : ) {
1905 0 : if self.tenant_shard_id.is_shard_zero() {
1906 0 : // Logical size is only maintained accurately on shard zero.
1907 0 : self.spawn_initial_logical_size_computation_task(ctx);
1908 0 : }
1909 0 : self.launch_wal_receiver(ctx, broker_client);
1910 0 : self.set_state(TimelineState::Active);
1911 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1912 0 : }
1913 :
1914 : /// After this function returns, there are no timeline-scoped tasks are left running.
1915 : ///
1916 : /// The preferred pattern for is:
1917 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1918 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1919 : /// go the extra mile and keep track of JoinHandles
1920 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1921 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1922 : ///
1923 : /// For legacy reasons, we still have multiple tasks spawned using
1924 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1925 : /// We refer to these as "timeline-scoped task_mgr tasks".
1926 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1927 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1928 : /// or [`task_mgr::shutdown_watcher`].
1929 : /// We want to gradually convert the code base away from these.
1930 : ///
1931 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1932 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1933 : /// ones that aren't mentioned here):
1934 : /// - [`TaskKind::TimelineDeletionWorker`]
1935 : /// - NB: also used for tenant deletion
1936 : /// - [`TaskKind::RemoteUploadTask`]`
1937 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1938 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1939 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1940 : /// - [`TaskKind::Eviction`]
1941 : /// - [`TaskKind::LayerFlushTask`]
1942 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1943 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1944 20 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1945 20 : debug_assert_current_span_has_tenant_and_timeline_id();
1946 20 :
1947 20 : // Regardless of whether we're going to try_freeze_and_flush
1948 20 : // or not, stop ingesting any more data. Walreceiver only provides
1949 20 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1950 20 : // So, only after the self.gate.close() below will we know for sure that
1951 20 : // no walreceiver tasks are left.
1952 20 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1953 20 : // data during the call to `self.freeze_and_flush()` below.
1954 20 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1955 20 : // which is what we'd need to properly model early shutdown of the walreceiver
1956 20 : // task sub-tree before the other Timeline task sub-trees.
1957 20 : let walreceiver = self.walreceiver.lock().unwrap().take();
1958 20 : tracing::debug!(
1959 0 : is_some = walreceiver.is_some(),
1960 0 : "Waiting for WalReceiverManager..."
1961 : );
1962 20 : if let Some(walreceiver) = walreceiver {
1963 0 : walreceiver.cancel();
1964 20 : }
1965 : // ... and inform any waiters for newer LSNs that there won't be any.
1966 20 : self.last_record_lsn.shutdown();
1967 20 :
1968 20 : if let ShutdownMode::FreezeAndFlush = mode {
1969 12 : let do_flush = if let Some((open, frozen)) = self
1970 12 : .layers
1971 12 : .read()
1972 12 : .await
1973 12 : .layer_map()
1974 12 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1975 12 : .ok()
1976 12 : .filter(|(open, frozen)| *open || *frozen > 0)
1977 : {
1978 0 : if self.remote_client.is_archived() == Some(true) {
1979 : // No point flushing on shutdown for an archived timeline: it is not important
1980 : // to have it nice and fresh after our restart, and trying to flush here might
1981 : // race with trying to offload it (which also stops the flush loop)
1982 0 : false
1983 : } else {
1984 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1985 0 : true
1986 : }
1987 : } else {
1988 : // this is double-shutdown, it'll be a no-op
1989 12 : true
1990 : };
1991 :
1992 : // we shut down walreceiver above, so, we won't add anything more
1993 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1994 : // to reach the disk & upload queue, then shut the upload queue and
1995 : // wait for it to drain.
1996 12 : if do_flush {
1997 12 : match self.freeze_and_flush().await {
1998 : Ok(_) => {
1999 : // drain the upload queue
2000 : // if we did not wait for completion here, it might be our shutdown process
2001 : // didn't wait for remote uploads to complete at all, as new tasks can forever
2002 : // be spawned.
2003 : //
2004 : // what is problematic is the shutting down of RemoteTimelineClient, because
2005 : // obviously it does not make sense to stop while we wait for it, but what
2006 : // about corner cases like s3 suddenly hanging up?
2007 12 : self.remote_client.shutdown().await;
2008 : }
2009 : Err(FlushLayerError::Cancelled) => {
2010 : // this is likely the second shutdown, ignore silently.
2011 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2012 0 : debug_assert!(self.cancel.is_cancelled());
2013 : }
2014 0 : Err(e) => {
2015 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
2016 0 : // we have some extra WAL replay to do next time the timeline starts.
2017 0 : warn!("failed to freeze and flush: {e:#}");
2018 : }
2019 : }
2020 :
2021 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2022 : // we also do a final check here to ensure that the queue is empty.
2023 12 : if !self.remote_client.no_pending_work() {
2024 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
2025 12 : }
2026 0 : }
2027 8 : }
2028 :
2029 20 : if let ShutdownMode::Reload = mode {
2030 : // drain the upload queue
2031 4 : self.remote_client.shutdown().await;
2032 4 : if !self.remote_client.no_pending_work() {
2033 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
2034 4 : }
2035 16 : }
2036 :
2037 : // Signal any subscribers to our cancellation token to drop out
2038 20 : tracing::debug!("Cancelling CancellationToken");
2039 20 : self.cancel.cancel();
2040 20 :
2041 20 : // Ensure Prevent new page service requests from starting.
2042 20 : self.handles.shutdown();
2043 20 :
2044 20 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2045 20 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2046 20 : self.remote_client.stop();
2047 20 :
2048 20 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2049 20 : // to shut down the upload queue tasks.
2050 20 : // TODO: fix that, task management should be encapsulated inside remote_client.
2051 20 : task_mgr::shutdown_tasks(
2052 20 : Some(TaskKind::RemoteUploadTask),
2053 20 : Some(self.tenant_shard_id),
2054 20 : Some(self.timeline_id),
2055 20 : )
2056 20 : .await;
2057 :
2058 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2059 20 : tracing::debug!("Waiting for tasks...");
2060 20 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2061 :
2062 : {
2063 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2064 : // open.
2065 20 : let mut write_guard = self.write_lock.lock().await;
2066 20 : self.layers.write().await.shutdown(&mut write_guard);
2067 20 : }
2068 20 :
2069 20 : // Finally wait until any gate-holders are complete.
2070 20 : //
2071 20 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2072 20 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2073 20 : self.gate.close().await;
2074 :
2075 20 : self.metrics.shutdown();
2076 20 : }
2077 :
2078 901 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2079 901 : match (self.current_state(), new_state) {
2080 901 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2081 4 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2082 : }
2083 0 : (st, TimelineState::Loading) => {
2084 0 : error!("ignoring transition from {st:?} into Loading state");
2085 : }
2086 0 : (TimelineState::Broken { .. }, new_state) => {
2087 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2088 : }
2089 : (TimelineState::Stopping, TimelineState::Active) => {
2090 0 : error!("Not activating a Stopping timeline");
2091 : }
2092 897 : (_, new_state) => {
2093 897 : self.state.send_replace(new_state);
2094 897 : }
2095 : }
2096 901 : }
2097 :
2098 4 : pub(crate) fn set_broken(&self, reason: String) {
2099 4 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2100 4 : let broken_state = TimelineState::Broken {
2101 4 : reason,
2102 4 : backtrace: backtrace_str,
2103 4 : };
2104 4 : self.set_state(broken_state);
2105 4 :
2106 4 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2107 4 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2108 4 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2109 4 : self.cancel.cancel();
2110 4 : }
2111 :
2112 453824 : pub(crate) fn current_state(&self) -> TimelineState {
2113 453824 : self.state.borrow().clone()
2114 453824 : }
2115 :
2116 12 : pub(crate) fn is_broken(&self) -> bool {
2117 12 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2118 12 : }
2119 :
2120 475 : pub(crate) fn is_active(&self) -> bool {
2121 475 : self.current_state() == TimelineState::Active
2122 475 : }
2123 :
2124 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
2125 0 : self.remote_client.is_archived()
2126 0 : }
2127 :
2128 776 : pub(crate) fn is_stopping(&self) -> bool {
2129 776 : self.current_state() == TimelineState::Stopping
2130 776 : }
2131 :
2132 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2133 0 : self.state.subscribe()
2134 0 : }
2135 :
2136 451676 : pub(crate) async fn wait_to_become_active(
2137 451676 : &self,
2138 451676 : _ctx: &RequestContext, // Prepare for use by cancellation
2139 451676 : ) -> Result<(), TimelineState> {
2140 451676 : let mut receiver = self.state.subscribe();
2141 : loop {
2142 451676 : let current_state = receiver.borrow().clone();
2143 451676 : match current_state {
2144 : TimelineState::Loading => {
2145 0 : receiver
2146 0 : .changed()
2147 0 : .await
2148 0 : .expect("holding a reference to self");
2149 : }
2150 : TimelineState::Active { .. } => {
2151 451672 : return Ok(());
2152 : }
2153 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2154 : // There's no chance the timeline can transition back into ::Active
2155 4 : return Err(current_state);
2156 : }
2157 : }
2158 : }
2159 451676 : }
2160 :
2161 0 : pub(crate) async fn layer_map_info(
2162 0 : &self,
2163 0 : reset: LayerAccessStatsReset,
2164 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2165 0 : let guard = self.layers.read().await;
2166 0 : let layer_map = guard.layer_map()?;
2167 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2168 0 : if let Some(open_layer) = &layer_map.open_layer {
2169 0 : in_memory_layers.push(open_layer.info());
2170 0 : }
2171 0 : for frozen_layer in &layer_map.frozen_layers {
2172 0 : in_memory_layers.push(frozen_layer.info());
2173 0 : }
2174 :
2175 0 : let historic_layers = layer_map
2176 0 : .iter_historic_layers()
2177 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2178 0 : .collect();
2179 0 :
2180 0 : Ok(LayerMapInfo {
2181 0 : in_memory_layers,
2182 0 : historic_layers,
2183 0 : })
2184 0 : }
2185 :
2186 : #[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))]
2187 : pub(crate) async fn download_layer(
2188 : &self,
2189 : layer_file_name: &LayerName,
2190 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2191 : let Some(layer) = self
2192 : .find_layer(layer_file_name)
2193 : .await
2194 0 : .map_err(|e| match e {
2195 0 : layer_manager::Shutdown => {
2196 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2197 0 : }
2198 0 : })?
2199 : else {
2200 : return Ok(None);
2201 : };
2202 :
2203 : layer.download().await?;
2204 :
2205 : Ok(Some(true))
2206 : }
2207 :
2208 : /// Evict just one layer.
2209 : ///
2210 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2211 0 : pub(crate) async fn evict_layer(
2212 0 : &self,
2213 0 : layer_file_name: &LayerName,
2214 0 : ) -> anyhow::Result<Option<bool>> {
2215 0 : let _gate = self
2216 0 : .gate
2217 0 : .enter()
2218 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2219 :
2220 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2221 0 : return Ok(None);
2222 : };
2223 :
2224 : // curl has this by default
2225 0 : let timeout = std::time::Duration::from_secs(120);
2226 0 :
2227 0 : match local_layer.evict_and_wait(timeout).await {
2228 0 : Ok(()) => Ok(Some(true)),
2229 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2230 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2231 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2232 : }
2233 0 : }
2234 :
2235 9606020 : fn should_roll(
2236 9606020 : &self,
2237 9606020 : layer_size: u64,
2238 9606020 : projected_layer_size: u64,
2239 9606020 : checkpoint_distance: u64,
2240 9606020 : projected_lsn: Lsn,
2241 9606020 : last_freeze_at: Lsn,
2242 9606020 : opened_at: Instant,
2243 9606020 : ) -> bool {
2244 9606020 : let distance = projected_lsn.widening_sub(last_freeze_at);
2245 9606020 :
2246 9606020 : // Rolling the open layer can be triggered by:
2247 9606020 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2248 9606020 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2249 9606020 : // account for how writes are distributed across shards: we expect each node to consume
2250 9606020 : // 1/count of the LSN on average.
2251 9606020 : // 2. The size of the currently open layer.
2252 9606020 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2253 9606020 : // up and suspend activity.
2254 9606020 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2255 0 : info!(
2256 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2257 : projected_lsn, layer_size, distance
2258 : );
2259 :
2260 0 : true
2261 9606020 : } else if projected_layer_size >= checkpoint_distance {
2262 : // NB: this check is relied upon by:
2263 160 : let _ = IndexEntry::validate_checkpoint_distance;
2264 160 : info!(
2265 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2266 : projected_lsn, layer_size, projected_layer_size
2267 : );
2268 :
2269 160 : true
2270 9605860 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2271 0 : info!(
2272 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2273 0 : projected_lsn,
2274 0 : layer_size,
2275 0 : opened_at.elapsed()
2276 : );
2277 :
2278 0 : true
2279 : } else {
2280 9605860 : false
2281 : }
2282 9606020 : }
2283 : }
2284 :
2285 : /// Number of times we will compute partition within a checkpoint distance.
2286 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2287 :
2288 : // Private functions
2289 : impl Timeline {
2290 24 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2291 24 : let tenant_conf = self.tenant_conf.load();
2292 24 : tenant_conf
2293 24 : .tenant_conf
2294 24 : .lsn_lease_length
2295 24 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2296 24 : }
2297 :
2298 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2299 0 : let tenant_conf = self.tenant_conf.load();
2300 0 : tenant_conf
2301 0 : .tenant_conf
2302 0 : .lsn_lease_length_for_ts
2303 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2304 0 : }
2305 :
2306 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2307 0 : let tenant_conf = self.tenant_conf.load();
2308 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2309 0 : }
2310 :
2311 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2312 0 : let tenant_conf = self.tenant_conf.load();
2313 0 : tenant_conf
2314 0 : .tenant_conf
2315 0 : .lazy_slru_download
2316 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2317 0 : }
2318 :
2319 9609118 : fn get_checkpoint_distance(&self) -> u64 {
2320 9609118 : let tenant_conf = self.tenant_conf.load();
2321 9609118 : tenant_conf
2322 9609118 : .tenant_conf
2323 9609118 : .checkpoint_distance
2324 9609118 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2325 9609118 : }
2326 :
2327 9605860 : fn get_checkpoint_timeout(&self) -> Duration {
2328 9605860 : let tenant_conf = self.tenant_conf.load();
2329 9605860 : tenant_conf
2330 9605860 : .tenant_conf
2331 9605860 : .checkpoint_timeout
2332 9605860 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2333 9605860 : }
2334 :
2335 5018 : fn get_compaction_period(&self) -> Duration {
2336 5018 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2337 5018 : tenant_conf
2338 5018 : .compaction_period
2339 5018 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2340 5018 : }
2341 :
2342 1369 : fn get_compaction_target_size(&self) -> u64 {
2343 1369 : let tenant_conf = self.tenant_conf.load();
2344 1369 : tenant_conf
2345 1369 : .tenant_conf
2346 1369 : .compaction_target_size
2347 1369 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2348 1369 : }
2349 :
2350 3209 : fn get_compaction_threshold(&self) -> usize {
2351 3209 : let tenant_conf = self.tenant_conf.load();
2352 3209 : tenant_conf
2353 3209 : .tenant_conf
2354 3209 : .compaction_threshold
2355 3209 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2356 3209 : }
2357 :
2358 92 : fn get_compaction_upper_limit(&self) -> usize {
2359 92 : let tenant_conf = self.tenant_conf.load();
2360 92 : tenant_conf
2361 92 : .tenant_conf
2362 92 : .compaction_upper_limit
2363 92 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2364 92 : }
2365 :
2366 768 : pub fn get_compaction_l0_first(&self) -> bool {
2367 768 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2368 768 : tenant_conf
2369 768 : .compaction_l0_first
2370 768 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2371 768 : }
2372 :
2373 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2374 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2375 0 : tenant_conf
2376 0 : .compaction_l0_semaphore
2377 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2378 0 : }
2379 :
2380 2509 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2381 : // Disable L0 flushes by default. This and compaction needs further tuning.
2382 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 0; // TODO: default to e.g. 3
2383 :
2384 : // If compaction is disabled, don't delay.
2385 2509 : if self.get_compaction_period() == Duration::ZERO {
2386 2509 : return None;
2387 0 : }
2388 0 :
2389 0 : let compaction_threshold = self.get_compaction_threshold();
2390 0 : let tenant_conf = self.tenant_conf.load();
2391 0 : let l0_flush_delay_threshold = tenant_conf
2392 0 : .tenant_conf
2393 0 : .l0_flush_delay_threshold
2394 0 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2395 0 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2396 0 :
2397 0 : // 0 disables backpressure.
2398 0 : if l0_flush_delay_threshold == 0 {
2399 0 : return None;
2400 0 : }
2401 0 :
2402 0 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2403 0 : // backpressure flushes below this.
2404 0 : // TODO: the tenant config should have validation to prevent this instead.
2405 0 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2406 0 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2407 2509 : }
2408 :
2409 2509 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2410 : // Disable L0 stalls by default. In ingest benchmarks, we see image compaction take >10
2411 : // minutes, blocking L0 compaction, and we can't stall L0 flushes for that long.
2412 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2413 :
2414 : // If compaction is disabled, don't stall.
2415 2509 : if self.get_compaction_period() == Duration::ZERO {
2416 2509 : return None;
2417 0 : }
2418 0 :
2419 0 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2420 0 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2421 0 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2422 0 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2423 0 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2424 0 : return None;
2425 0 : }
2426 0 :
2427 0 : let compaction_threshold = self.get_compaction_threshold();
2428 0 : let tenant_conf = self.tenant_conf.load();
2429 0 : let l0_flush_stall_threshold = tenant_conf
2430 0 : .tenant_conf
2431 0 : .l0_flush_stall_threshold
2432 0 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2433 0 :
2434 0 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2435 0 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2436 0 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2437 0 : if cfg!(feature = "testing")
2438 0 : && compaction_threshold == 1
2439 0 : && l0_flush_stall_threshold.is_none()
2440 : {
2441 0 : return None;
2442 0 : }
2443 0 :
2444 0 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2445 0 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2446 0 :
2447 0 : // 0 disables backpressure.
2448 0 : if l0_flush_stall_threshold == 0 {
2449 0 : return None;
2450 0 : }
2451 0 :
2452 0 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2453 0 : // backpressure flushes below this.
2454 0 : // TODO: the tenant config should have validation to prevent this instead.
2455 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2456 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2457 2509 : }
2458 :
2459 2349 : fn get_l0_flush_wait_upload(&self) -> bool {
2460 2349 : let tenant_conf = self.tenant_conf.load();
2461 2349 : tenant_conf
2462 2349 : .tenant_conf
2463 2349 : .l0_flush_wait_upload
2464 2349 : .unwrap_or(self.conf.default_tenant_conf.l0_flush_wait_upload)
2465 2349 : }
2466 :
2467 0 : fn get_image_creation_threshold(&self) -> usize {
2468 0 : let tenant_conf = self.tenant_conf.load();
2469 0 : tenant_conf
2470 0 : .tenant_conf
2471 0 : .image_creation_threshold
2472 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2473 0 : }
2474 :
2475 768 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2476 768 : let tenant_conf = &self.tenant_conf.load();
2477 768 : tenant_conf
2478 768 : .tenant_conf
2479 768 : .compaction_algorithm
2480 768 : .as_ref()
2481 768 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2482 768 : .clone()
2483 768 : }
2484 :
2485 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2486 0 : let tenant_conf = self.tenant_conf.load();
2487 0 : tenant_conf
2488 0 : .tenant_conf
2489 0 : .eviction_policy
2490 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2491 0 : }
2492 :
2493 897 : fn get_evictions_low_residence_duration_metric_threshold(
2494 897 : tenant_conf: &TenantConfOpt,
2495 897 : default_tenant_conf: &TenantConf,
2496 897 : ) -> Duration {
2497 897 : tenant_conf
2498 897 : .evictions_low_residence_duration_metric_threshold
2499 897 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2500 897 : }
2501 :
2502 1181 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2503 1181 : let tenant_conf = self.tenant_conf.load();
2504 1181 : tenant_conf
2505 1181 : .tenant_conf
2506 1181 : .image_layer_creation_check_threshold
2507 1181 : .unwrap_or(
2508 1181 : self.conf
2509 1181 : .default_tenant_conf
2510 1181 : .image_layer_creation_check_threshold,
2511 1181 : )
2512 1181 : }
2513 :
2514 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
2515 0 : let tenant_conf = self.tenant_conf.load();
2516 0 : tenant_conf
2517 0 : .tenant_conf
2518 0 : .image_creation_preempt_threshold
2519 0 : .unwrap_or(
2520 0 : self.conf
2521 0 : .default_tenant_conf
2522 0 : .image_creation_preempt_threshold,
2523 0 : )
2524 0 : }
2525 :
2526 : /// Resolve the effective WAL receiver protocol to use for this tenant.
2527 : ///
2528 : /// Priority order is:
2529 : /// 1. Tenant config override
2530 : /// 2. Default value for tenant config override
2531 : /// 3. Pageserver config override
2532 : /// 4. Pageserver config default
2533 0 : pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
2534 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2535 0 : tenant_conf
2536 0 : .wal_receiver_protocol_override
2537 0 : .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
2538 0 : .unwrap_or(self.conf.wal_receiver_protocol)
2539 0 : }
2540 :
2541 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2542 0 : // NB: Most tenant conf options are read by background loops, so,
2543 0 : // changes will automatically be picked up.
2544 0 :
2545 0 : // The threshold is embedded in the metric. So, we need to update it.
2546 0 : {
2547 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2548 0 : &new_conf.tenant_conf,
2549 0 : &self.conf.default_tenant_conf,
2550 0 : );
2551 0 :
2552 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2553 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2554 0 :
2555 0 : let timeline_id_str = self.timeline_id.to_string();
2556 0 :
2557 0 : self.remote_client.update_config(&new_conf.location);
2558 0 :
2559 0 : self.metrics
2560 0 : .evictions_with_low_residence_duration
2561 0 : .write()
2562 0 : .unwrap()
2563 0 : .change_threshold(
2564 0 : &tenant_id_str,
2565 0 : &shard_id_str,
2566 0 : &timeline_id_str,
2567 0 : new_threshold,
2568 0 : );
2569 0 : }
2570 0 : }
2571 :
2572 : /// Open a Timeline handle.
2573 : ///
2574 : /// Loads the metadata for the timeline into memory, but not the layer map.
2575 : #[allow(clippy::too_many_arguments)]
2576 897 : pub(super) fn new(
2577 897 : conf: &'static PageServerConf,
2578 897 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2579 897 : metadata: &TimelineMetadata,
2580 897 : previous_heatmap: Option<PreviousHeatmap>,
2581 897 : ancestor: Option<Arc<Timeline>>,
2582 897 : timeline_id: TimelineId,
2583 897 : tenant_shard_id: TenantShardId,
2584 897 : generation: Generation,
2585 897 : shard_identity: ShardIdentity,
2586 897 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2587 897 : resources: TimelineResources,
2588 897 : pg_version: u32,
2589 897 : state: TimelineState,
2590 897 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2591 897 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2592 897 : cancel: CancellationToken,
2593 897 : ) -> Arc<Self> {
2594 897 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2595 897 : let (state, _) = watch::channel(state);
2596 897 :
2597 897 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2598 897 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2599 897 :
2600 897 : let evictions_low_residence_duration_metric_threshold = {
2601 897 : let loaded_tenant_conf = tenant_conf.load();
2602 897 : Self::get_evictions_low_residence_duration_metric_threshold(
2603 897 : &loaded_tenant_conf.tenant_conf,
2604 897 : &conf.default_tenant_conf,
2605 897 : )
2606 : };
2607 :
2608 897 : if let Some(ancestor) = &ancestor {
2609 460 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2610 460 : // If we construct an explicit timeline object, it's obviously not offloaded
2611 460 : let is_offloaded = MaybeOffloaded::No;
2612 460 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2613 460 : }
2614 :
2615 897 : Arc::new_cyclic(|myself| {
2616 897 : let metrics = TimelineMetrics::new(
2617 897 : &tenant_shard_id,
2618 897 : &timeline_id,
2619 897 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2620 897 : "mtime",
2621 897 : evictions_low_residence_duration_metric_threshold,
2622 897 : ),
2623 897 : );
2624 897 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2625 :
2626 897 : let mut result = Timeline {
2627 897 : conf,
2628 897 : tenant_conf,
2629 897 : myself: myself.clone(),
2630 897 : timeline_id,
2631 897 : tenant_shard_id,
2632 897 : generation,
2633 897 : shard_identity,
2634 897 : pg_version,
2635 897 : layers: Default::default(),
2636 897 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
2637 897 :
2638 897 : walredo_mgr,
2639 897 : walreceiver: Mutex::new(None),
2640 897 :
2641 897 : remote_client: Arc::new(resources.remote_client),
2642 897 :
2643 897 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2644 897 : last_record_lsn: SeqWait::new(RecordLsn {
2645 897 : last: disk_consistent_lsn,
2646 897 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2647 897 : }),
2648 897 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2649 897 :
2650 897 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2651 897 : last_freeze_ts: RwLock::new(Instant::now()),
2652 897 :
2653 897 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2654 897 :
2655 897 : ancestor_timeline: ancestor,
2656 897 : ancestor_lsn: metadata.ancestor_lsn(),
2657 897 :
2658 897 : metrics,
2659 897 :
2660 897 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2661 897 : &tenant_shard_id,
2662 897 : &timeline_id,
2663 897 : resources.pagestream_throttle_metrics,
2664 897 : ),
2665 897 :
2666 6279 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2667 897 :
2668 897 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2669 897 :
2670 897 : layer_flush_start_tx,
2671 897 : layer_flush_done_tx,
2672 897 :
2673 897 : write_lock: tokio::sync::Mutex::new(None),
2674 897 :
2675 897 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2676 897 :
2677 897 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
2678 897 : LastImageLayerCreationStatus::default(),
2679 897 : )),
2680 897 :
2681 897 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2682 897 : initdb_lsn: metadata.initdb_lsn(),
2683 897 :
2684 897 : current_logical_size: if disk_consistent_lsn.is_valid() {
2685 : // we're creating timeline data with some layer files existing locally,
2686 : // need to recalculate timeline's logical size based on data in the layers.
2687 468 : LogicalSize::deferred_initial(disk_consistent_lsn)
2688 : } else {
2689 : // we're creating timeline data without any layers existing locally,
2690 : // initial logical size is 0.
2691 429 : LogicalSize::empty_initial()
2692 : },
2693 :
2694 897 : partitioning: GuardArcSwap::new((
2695 897 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2696 897 : Lsn(0),
2697 897 : )),
2698 897 : repartition_threshold: 0,
2699 897 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2700 897 : last_image_layer_creation_check_instant: Mutex::new(None),
2701 897 :
2702 897 : last_received_wal: Mutex::new(None),
2703 897 : rel_size_cache: RwLock::new(RelSizeCache {
2704 897 : complete_as_of: disk_consistent_lsn,
2705 897 : map: HashMap::new(),
2706 897 : }),
2707 897 :
2708 897 : download_all_remote_layers_task_info: RwLock::new(None),
2709 897 :
2710 897 : state,
2711 897 :
2712 897 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2713 897 : EvictionTaskTimelineState::default(),
2714 897 : ),
2715 897 : delete_progress: TimelineDeleteProgress::default(),
2716 897 :
2717 897 : cancel,
2718 897 : gate: Gate::default(),
2719 897 :
2720 897 : compaction_lock: tokio::sync::Mutex::default(),
2721 897 : compaction_failed: AtomicBool::default(),
2722 897 : l0_compaction_trigger: resources.l0_compaction_trigger,
2723 897 : gc_lock: tokio::sync::Mutex::default(),
2724 897 :
2725 897 : standby_horizon: AtomicLsn::new(0),
2726 897 :
2727 897 : pagestream_throttle: resources.pagestream_throttle,
2728 897 :
2729 897 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2730 897 :
2731 897 : #[cfg(test)]
2732 897 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2733 897 :
2734 897 : l0_flush_global_state: resources.l0_flush_global_state,
2735 897 :
2736 897 : handles: Default::default(),
2737 897 :
2738 897 : attach_wal_lag_cooldown,
2739 897 :
2740 897 : create_idempotency,
2741 897 :
2742 897 : page_trace: Default::default(),
2743 897 :
2744 897 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
2745 897 : };
2746 897 :
2747 897 : result.repartition_threshold =
2748 897 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2749 897 :
2750 897 : result
2751 897 : .metrics
2752 897 : .last_record_lsn_gauge
2753 897 : .set(disk_consistent_lsn.0 as i64);
2754 897 : result
2755 897 : })
2756 897 : }
2757 :
2758 1298 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2759 1298 : let Ok(guard) = self.gate.enter() else {
2760 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2761 0 : return;
2762 : };
2763 1298 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2764 1298 : match *flush_loop_state {
2765 885 : FlushLoopState::NotStarted => (),
2766 : FlushLoopState::Running { .. } => {
2767 413 : info!(
2768 0 : "skipping attempt to start flush_loop twice {}/{}",
2769 0 : self.tenant_shard_id, self.timeline_id
2770 : );
2771 413 : return;
2772 : }
2773 : FlushLoopState::Exited => {
2774 0 : info!(
2775 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2776 0 : self.tenant_shard_id, self.timeline_id
2777 : );
2778 0 : return;
2779 : }
2780 : }
2781 :
2782 885 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2783 885 : let self_clone = Arc::clone(self);
2784 885 :
2785 885 : debug!("spawning flush loop");
2786 885 : *flush_loop_state = FlushLoopState::Running {
2787 885 : #[cfg(test)]
2788 885 : expect_initdb_optimization: false,
2789 885 : #[cfg(test)]
2790 885 : initdb_optimization_count: 0,
2791 885 : };
2792 885 : task_mgr::spawn(
2793 885 : task_mgr::BACKGROUND_RUNTIME.handle(),
2794 885 : task_mgr::TaskKind::LayerFlushTask,
2795 885 : self.tenant_shard_id,
2796 885 : Some(self.timeline_id),
2797 885 : "layer flush task",
2798 885 : async move {
2799 885 : let _guard = guard;
2800 885 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2801 885 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2802 20 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2803 20 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2804 20 : *flush_loop_state = FlushLoopState::Exited;
2805 20 : Ok(())
2806 20 : }
2807 885 : .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))
2808 : );
2809 1298 : }
2810 :
2811 : /// Creates and starts the wal receiver.
2812 : ///
2813 : /// This function is expected to be called at most once per Timeline's lifecycle
2814 : /// when the timeline is activated.
2815 0 : fn launch_wal_receiver(
2816 0 : self: &Arc<Self>,
2817 0 : ctx: &RequestContext,
2818 0 : broker_client: BrokerClientChannel,
2819 0 : ) {
2820 0 : info!(
2821 0 : "launching WAL receiver for timeline {} of tenant {}",
2822 0 : self.timeline_id, self.tenant_shard_id
2823 : );
2824 :
2825 0 : let tenant_conf = self.tenant_conf.load();
2826 0 : let wal_connect_timeout = tenant_conf
2827 0 : .tenant_conf
2828 0 : .walreceiver_connect_timeout
2829 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2830 0 : let lagging_wal_timeout = tenant_conf
2831 0 : .tenant_conf
2832 0 : .lagging_wal_timeout
2833 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2834 0 : let max_lsn_wal_lag = tenant_conf
2835 0 : .tenant_conf
2836 0 : .max_lsn_wal_lag
2837 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2838 0 :
2839 0 : let mut guard = self.walreceiver.lock().unwrap();
2840 0 : assert!(
2841 0 : guard.is_none(),
2842 0 : "multiple launches / re-launches of WAL receiver are not supported"
2843 : );
2844 0 : *guard = Some(WalReceiver::start(
2845 0 : Arc::clone(self),
2846 0 : WalReceiverConf {
2847 0 : protocol: self.resolve_wal_receiver_protocol(),
2848 0 : wal_connect_timeout,
2849 0 : lagging_wal_timeout,
2850 0 : max_lsn_wal_lag,
2851 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2852 0 : availability_zone: self.conf.availability_zone.clone(),
2853 0 : ingest_batch_size: self.conf.ingest_batch_size,
2854 0 : },
2855 0 : broker_client,
2856 0 : ctx,
2857 0 : ));
2858 0 : }
2859 :
2860 : /// Initialize with an empty layer map. Used when creating a new timeline.
2861 885 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2862 885 : let mut layers = self.layers.try_write().expect(
2863 885 : "in the context where we call this function, no other task has access to the object",
2864 885 : );
2865 885 : layers
2866 885 : .open_mut()
2867 885 : .expect("in this context the LayerManager must still be open")
2868 885 : .initialize_empty(Lsn(start_lsn.0));
2869 885 : }
2870 :
2871 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2872 : /// files.
2873 12 : pub(super) async fn load_layer_map(
2874 12 : &self,
2875 12 : disk_consistent_lsn: Lsn,
2876 12 : index_part: IndexPart,
2877 12 : ) -> anyhow::Result<()> {
2878 : use init::{Decision::*, Discovered, DismissedLayer};
2879 : use LayerName::*;
2880 :
2881 12 : let mut guard = self.layers.write().await;
2882 :
2883 12 : let timer = self.metrics.load_layer_map_histo.start_timer();
2884 12 :
2885 12 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2886 12 : // structs representing all files on disk
2887 12 : let timeline_path = self
2888 12 : .conf
2889 12 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2890 12 : let conf = self.conf;
2891 12 : let span = tracing::Span::current();
2892 12 :
2893 12 : // Copy to move into the task we're about to spawn
2894 12 : let this = self.myself.upgrade().expect("&self method holds the arc");
2895 :
2896 12 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2897 12 : move || {
2898 12 : let _g = span.entered();
2899 12 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2900 12 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2901 12 : let mut unrecognized_files = Vec::new();
2902 12 :
2903 12 : let mut path = timeline_path;
2904 :
2905 44 : for discovered in discovered {
2906 32 : let (name, kind) = match discovered {
2907 32 : Discovered::Layer(layer_file_name, local_metadata) => {
2908 32 : discovered_layers.push((layer_file_name, local_metadata));
2909 32 : continue;
2910 : }
2911 0 : Discovered::IgnoredBackup(path) => {
2912 0 : std::fs::remove_file(path)
2913 0 : .or_else(fs_ext::ignore_not_found)
2914 0 : .fatal_err("Removing .old file");
2915 0 : continue;
2916 : }
2917 0 : Discovered::Unknown(file_name) => {
2918 0 : // we will later error if there are any
2919 0 : unrecognized_files.push(file_name);
2920 0 : continue;
2921 : }
2922 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2923 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2924 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2925 : };
2926 0 : path.push(Utf8Path::new(&name));
2927 0 : init::cleanup(&path, kind)?;
2928 0 : path.pop();
2929 : }
2930 :
2931 12 : if !unrecognized_files.is_empty() {
2932 : // assume that if there are any there are many many.
2933 0 : let n = unrecognized_files.len();
2934 0 : let first = &unrecognized_files[..n.min(10)];
2935 0 : anyhow::bail!(
2936 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2937 0 : );
2938 12 : }
2939 12 :
2940 12 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2941 12 :
2942 12 : let mut loaded_layers = Vec::new();
2943 12 : let mut needs_cleanup = Vec::new();
2944 12 : let mut total_physical_size = 0;
2945 :
2946 44 : for (name, decision) in decided {
2947 32 : let decision = match decision {
2948 32 : Ok(decision) => decision,
2949 0 : Err(DismissedLayer::Future { local }) => {
2950 0 : if let Some(local) = local {
2951 0 : init::cleanup_future_layer(
2952 0 : &local.local_path,
2953 0 : &name,
2954 0 : disk_consistent_lsn,
2955 0 : )?;
2956 0 : }
2957 0 : needs_cleanup.push(name);
2958 0 : continue;
2959 : }
2960 0 : Err(DismissedLayer::LocalOnly(local)) => {
2961 0 : init::cleanup_local_only_file(&name, &local)?;
2962 : // this file never existed remotely, we will have to do rework
2963 0 : continue;
2964 : }
2965 0 : Err(DismissedLayer::BadMetadata(local)) => {
2966 0 : init::cleanup_local_file_for_remote(&local)?;
2967 : // this file never existed remotely, we will have to do rework
2968 0 : continue;
2969 : }
2970 : };
2971 :
2972 32 : match &name {
2973 24 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2974 8 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2975 : }
2976 :
2977 32 : tracing::debug!(layer=%name, ?decision, "applied");
2978 :
2979 32 : let layer = match decision {
2980 32 : Resident { local, remote } => {
2981 32 : total_physical_size += local.file_size;
2982 32 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2983 32 : .drop_eviction_guard()
2984 : }
2985 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2986 : };
2987 :
2988 32 : loaded_layers.push(layer);
2989 : }
2990 12 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2991 12 : }
2992 12 : })
2993 12 : .await
2994 12 : .map_err(anyhow::Error::new)
2995 12 : .and_then(|x| x)?;
2996 :
2997 12 : let num_layers = loaded_layers.len();
2998 12 :
2999 12 : guard
3000 12 : .open_mut()
3001 12 : .expect("layermanager must be open during init")
3002 12 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3003 12 :
3004 12 : self.remote_client
3005 12 : .schedule_layer_file_deletion(&needs_cleanup)?;
3006 12 : self.remote_client
3007 12 : .schedule_index_upload_for_file_changes()?;
3008 : // This barrier orders above DELETEs before any later operations.
3009 : // This is critical because code executing after the barrier might
3010 : // create again objects with the same key that we just scheduled for deletion.
3011 : // For example, if we just scheduled deletion of an image layer "from the future",
3012 : // later compaction might run again and re-create the same image layer.
3013 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3014 : // "same" here means same key range and LSN.
3015 : //
3016 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3017 : // the upload queue may execute the PUT first, then the DELETE.
3018 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3019 : //
3020 : // 1. a future image layer is created and uploaded
3021 : // 2. ps restart
3022 : // 3. the future layer from (1) is deleted during load layer map
3023 : // 4. image layer is re-created and uploaded
3024 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3025 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3026 : //
3027 : // See https://github.com/neondatabase/neon/issues/5878
3028 : //
3029 : // NB: generation numbers naturally protect against this because they disambiguate
3030 : // (1) and (4)
3031 : // TODO: this is basically a no-op now, should we remove it?
3032 12 : self.remote_client.schedule_barrier()?;
3033 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
3034 : // on retry.
3035 :
3036 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
3037 12 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
3038 12 : self.update_layer_visibility().await?;
3039 :
3040 12 : info!(
3041 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3042 : num_layers, disk_consistent_lsn, total_physical_size
3043 : );
3044 :
3045 12 : timer.stop_and_record();
3046 12 : Ok(())
3047 12 : }
3048 :
3049 : /// Retrieve current logical size of the timeline.
3050 : ///
3051 : /// The size could be lagging behind the actual number, in case
3052 : /// the initial size calculation has not been run (gets triggered on the first size access).
3053 : ///
3054 : /// return size and boolean flag that shows if the size is exact
3055 0 : pub(crate) fn get_current_logical_size(
3056 0 : self: &Arc<Self>,
3057 0 : priority: GetLogicalSizePriority,
3058 0 : ctx: &RequestContext,
3059 0 : ) -> logical_size::CurrentLogicalSize {
3060 0 : if !self.tenant_shard_id.is_shard_zero() {
3061 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3062 : // when HTTP API is serving a GET for timeline zero, return zero
3063 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3064 0 : }
3065 0 :
3066 0 : let current_size = self.current_logical_size.current_size();
3067 0 : debug!("Current size: {current_size:?}");
3068 :
3069 0 : match (current_size.accuracy(), priority) {
3070 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3071 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3072 0 : // background task will eventually deliver an exact value, we're in no rush
3073 0 : }
3074 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3075 : // background task is not ready, but user is asking for it now;
3076 : // => make the background task skip the line
3077 : // (The alternative would be to calculate the size here, but,
3078 : // it can actually take a long time if the user has a lot of rels.
3079 : // And we'll inevitable need it again; So, let the background task do the work.)
3080 0 : match self
3081 0 : .current_logical_size
3082 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3083 0 : .get()
3084 : {
3085 0 : Some(cancel) => cancel.cancel(),
3086 : None => {
3087 0 : match self.current_state() {
3088 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3089 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3090 0 : // Don't make noise.
3091 0 : }
3092 : TimelineState::Loading => {
3093 : // Import does not return an activated timeline.
3094 0 : info!("discarding priority boost for logical size calculation because timeline is not yet active");
3095 : }
3096 : TimelineState::Active => {
3097 : // activation should be setting the once cell
3098 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
3099 0 : debug_assert!(false);
3100 : }
3101 : }
3102 : }
3103 : }
3104 : }
3105 : }
3106 :
3107 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3108 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3109 0 : let first = self
3110 0 : .current_logical_size
3111 0 : .did_return_approximate_to_walreceiver
3112 0 : .compare_exchange(
3113 0 : false,
3114 0 : true,
3115 0 : AtomicOrdering::Relaxed,
3116 0 : AtomicOrdering::Relaxed,
3117 0 : )
3118 0 : .is_ok();
3119 0 : if first {
3120 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3121 0 : }
3122 0 : }
3123 0 : }
3124 :
3125 0 : current_size
3126 0 : }
3127 :
3128 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3129 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3130 : // nothing to do for freshly created timelines;
3131 0 : assert_eq!(
3132 0 : self.current_logical_size.current_size().accuracy(),
3133 0 : logical_size::Accuracy::Exact,
3134 0 : );
3135 0 : self.current_logical_size.initialized.add_permits(1);
3136 0 : return;
3137 : };
3138 :
3139 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3140 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3141 0 : self.current_logical_size
3142 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3143 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3144 0 :
3145 0 : let self_clone = Arc::clone(self);
3146 0 : let background_ctx = ctx.detached_child(
3147 0 : TaskKind::InitialLogicalSizeCalculation,
3148 0 : DownloadBehavior::Download,
3149 0 : );
3150 0 : task_mgr::spawn(
3151 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3152 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3153 0 : self.tenant_shard_id,
3154 0 : Some(self.timeline_id),
3155 0 : "initial size calculation",
3156 : // NB: don't log errors here, task_mgr will do that.
3157 0 : async move {
3158 0 : self_clone
3159 0 : .initial_logical_size_calculation_task(
3160 0 : initial_part_end,
3161 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3162 0 : background_ctx,
3163 0 : )
3164 0 : .await;
3165 0 : Ok(())
3166 0 : }
3167 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)),
3168 : );
3169 0 : }
3170 :
3171 : /// # Cancellation
3172 : ///
3173 : /// This method is sensitive to `Timeline::cancel`.
3174 : ///
3175 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3176 : ///
3177 : /// # Cancel-Safety
3178 : ///
3179 : /// It does Timeline IO, hence this should be polled to completion because
3180 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3181 : /// to reason about.
3182 0 : async fn initial_logical_size_calculation_task(
3183 0 : self: Arc<Self>,
3184 0 : initial_part_end: Lsn,
3185 0 : skip_concurrency_limiter: CancellationToken,
3186 0 : background_ctx: RequestContext,
3187 0 : ) {
3188 0 : scopeguard::defer! {
3189 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3190 0 : self.current_logical_size.initialized.add_permits(1);
3191 0 : }
3192 0 :
3193 0 : let try_once = |attempt: usize| {
3194 0 : let background_ctx = &background_ctx;
3195 0 : let self_ref = &self;
3196 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3197 0 : async move {
3198 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3199 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3200 0 : background_ctx,
3201 0 : );
3202 :
3203 : use crate::metrics::initial_logical_size::StartCircumstances;
3204 0 : let (_maybe_permit, circumstances) = tokio::select! {
3205 0 : permit = wait_for_permit => {
3206 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3207 : }
3208 0 : _ = self_ref.cancel.cancelled() => {
3209 0 : return Err(CalculateLogicalSizeError::Cancelled);
3210 : }
3211 0 : () = skip_concurrency_limiter.cancelled() => {
3212 : // Some action that is part of a end user interaction requested logical size
3213 : // => break out of the rate limit
3214 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3215 : // but then again what happens if they cancel; also, we should just be using
3216 : // one runtime across the entire process, so, let's leave this for now.
3217 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3218 : }
3219 : };
3220 :
3221 0 : let metrics_guard = if attempt == 1 {
3222 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3223 : } else {
3224 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3225 : };
3226 :
3227 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3228 0 : self_ref.conf,
3229 0 : self_ref
3230 0 : .gate
3231 0 : .enter()
3232 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3233 : );
3234 :
3235 0 : let calculated_size = self_ref
3236 0 : .logical_size_calculation_task(
3237 0 : initial_part_end,
3238 0 : LogicalSizeCalculationCause::Initial,
3239 0 : background_ctx,
3240 0 : )
3241 0 : .await?;
3242 :
3243 0 : self_ref
3244 0 : .trigger_aux_file_size_computation(
3245 0 : initial_part_end,
3246 0 : background_ctx,
3247 0 : io_concurrency,
3248 0 : )
3249 0 : .await?;
3250 :
3251 : // TODO: add aux file size to logical size
3252 :
3253 0 : Ok((calculated_size, metrics_guard))
3254 0 : }
3255 0 : };
3256 :
3257 0 : let retrying = async {
3258 0 : let mut attempt = 0;
3259 : loop {
3260 0 : attempt += 1;
3261 0 :
3262 0 : match try_once(attempt).await {
3263 0 : Ok(res) => return ControlFlow::Continue(res),
3264 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3265 : Err(
3266 0 : e @ (CalculateLogicalSizeError::Decode(_)
3267 0 : | CalculateLogicalSizeError::PageRead(_)),
3268 0 : ) => {
3269 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3270 : // exponential back-off doesn't make sense at these long intervals;
3271 : // use fixed retry interval with generous jitter instead
3272 0 : let sleep_duration = Duration::from_secs(
3273 0 : u64::try_from(
3274 0 : // 1hour base
3275 0 : (60_i64 * 60_i64)
3276 0 : // 10min jitter
3277 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3278 0 : )
3279 0 : .expect("10min < 1hour"),
3280 0 : );
3281 0 : tokio::select! {
3282 0 : _ = tokio::time::sleep(sleep_duration) => {}
3283 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3284 : }
3285 : }
3286 : }
3287 : }
3288 0 : };
3289 :
3290 0 : let (calculated_size, metrics_guard) = match retrying.await {
3291 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3292 0 : ControlFlow::Break(()) => return,
3293 : };
3294 :
3295 : // we cannot query current_logical_size.current_size() to know the current
3296 : // *negative* value, only truncated to u64.
3297 0 : let added = self
3298 0 : .current_logical_size
3299 0 : .size_added_after_initial
3300 0 : .load(AtomicOrdering::Relaxed);
3301 0 :
3302 0 : let sum = calculated_size.saturating_add_signed(added);
3303 0 :
3304 0 : // set the gauge value before it can be set in `update_current_logical_size`.
3305 0 : self.metrics.current_logical_size_gauge.set(sum);
3306 0 :
3307 0 : self.current_logical_size
3308 0 : .initial_logical_size
3309 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3310 0 : .ok()
3311 0 : .expect("only this task sets it");
3312 0 : }
3313 :
3314 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3315 0 : self: &Arc<Self>,
3316 0 : lsn: Lsn,
3317 0 : cause: LogicalSizeCalculationCause,
3318 0 : ctx: RequestContext,
3319 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3320 0 : let (sender, receiver) = oneshot::channel();
3321 0 : let self_clone = Arc::clone(self);
3322 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3323 0 : // we should stop the size calculation work and return an error.
3324 0 : // That would require restructuring this function's API to
3325 0 : // return the result directly, instead of a Receiver for the result.
3326 0 : let ctx = ctx.detached_child(
3327 0 : TaskKind::OndemandLogicalSizeCalculation,
3328 0 : DownloadBehavior::Download,
3329 0 : );
3330 0 : task_mgr::spawn(
3331 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3332 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3333 0 : self.tenant_shard_id,
3334 0 : Some(self.timeline_id),
3335 0 : "ondemand logical size calculation",
3336 0 : async move {
3337 0 : let res = self_clone
3338 0 : .logical_size_calculation_task(lsn, cause, &ctx)
3339 0 : .await;
3340 0 : let _ = sender.send(res).ok();
3341 0 : Ok(()) // Receiver is responsible for handling errors
3342 0 : }
3343 0 : .in_current_span(),
3344 0 : );
3345 0 : receiver
3346 0 : }
3347 :
3348 : #[instrument(skip_all)]
3349 : async fn logical_size_calculation_task(
3350 : self: &Arc<Self>,
3351 : lsn: Lsn,
3352 : cause: LogicalSizeCalculationCause,
3353 : ctx: &RequestContext,
3354 : ) -> Result<u64, CalculateLogicalSizeError> {
3355 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3356 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3357 : // accurate relation sizes, and they do not emit consumption metrics.
3358 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3359 :
3360 : let guard = self
3361 : .gate
3362 : .enter()
3363 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3364 :
3365 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3366 : }
3367 :
3368 : /// Calculate the logical size of the database at the latest LSN.
3369 : ///
3370 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3371 : /// especially if we need to download remote layers.
3372 0 : async fn calculate_logical_size(
3373 0 : &self,
3374 0 : up_to_lsn: Lsn,
3375 0 : cause: LogicalSizeCalculationCause,
3376 0 : _guard: &GateGuard,
3377 0 : ctx: &RequestContext,
3378 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3379 0 : info!(
3380 0 : "Calculating logical size for timeline {} at {}",
3381 : self.timeline_id, up_to_lsn
3382 : );
3383 :
3384 0 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
3385 : {
3386 0 : return Err(CalculateLogicalSizeError::Cancelled);
3387 0 : }
3388 :
3389 : // See if we've already done the work for initial size calculation.
3390 : // This is a short-cut for timelines that are mostly unused.
3391 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3392 0 : return Ok(size);
3393 0 : }
3394 0 : let storage_time_metrics = match cause {
3395 : LogicalSizeCalculationCause::Initial
3396 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3397 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3398 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3399 0 : &self.metrics.imitate_logical_size_histo
3400 : }
3401 : };
3402 0 : let timer = storage_time_metrics.start_timer();
3403 0 : let logical_size = self
3404 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3405 0 : .await?;
3406 0 : debug!("calculated logical size: {logical_size}");
3407 0 : timer.stop_and_record();
3408 0 : Ok(logical_size)
3409 0 : }
3410 :
3411 : /// Update current logical size, adding `delta' to the old value.
3412 541140 : fn update_current_logical_size(&self, delta: i64) {
3413 541140 : let logical_size = &self.current_logical_size;
3414 541140 : logical_size.increment_size(delta);
3415 541140 :
3416 541140 : // Also set the value in the prometheus gauge. Note that
3417 541140 : // there is a race condition here: if this is is called by two
3418 541140 : // threads concurrently, the prometheus gauge might be set to
3419 541140 : // one value while current_logical_size is set to the
3420 541140 : // other.
3421 541140 : match logical_size.current_size() {
3422 541140 : CurrentLogicalSize::Exact(ref new_current_size) => self
3423 541140 : .metrics
3424 541140 : .current_logical_size_gauge
3425 541140 : .set(new_current_size.into()),
3426 0 : CurrentLogicalSize::Approximate(_) => {
3427 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3428 0 : // forth between the initial size calculation task.
3429 0 : }
3430 : }
3431 541140 : }
3432 :
3433 5937 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3434 5937 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3435 5937 : let aux_metric =
3436 5937 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3437 5937 :
3438 5937 : let sum_of_entries = self
3439 5937 : .directory_metrics
3440 5937 : .iter()
3441 41559 : .map(|v| v.load(AtomicOrdering::Relaxed))
3442 5937 : .sum();
3443 : // Set a high general threshold and a lower threshold for the auxiliary files,
3444 : // as we can have large numbers of relations in the db directory.
3445 : const SUM_THRESHOLD: u64 = 5000;
3446 : const AUX_THRESHOLD: u64 = 1000;
3447 5937 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3448 0 : self.metrics
3449 0 : .directory_entries_count_gauge
3450 0 : .set(sum_of_entries);
3451 5937 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3452 0 : metric.set(sum_of_entries);
3453 5937 : }
3454 5937 : }
3455 :
3456 0 : async fn find_layer(
3457 0 : &self,
3458 0 : layer_name: &LayerName,
3459 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3460 0 : let guard = self.layers.read().await;
3461 0 : let layer = guard
3462 0 : .layer_map()?
3463 0 : .iter_historic_layers()
3464 0 : .find(|l| &l.layer_name() == layer_name)
3465 0 : .map(|found| guard.get_from_desc(&found));
3466 0 : Ok(layer)
3467 0 : }
3468 :
3469 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3470 : /// indicating which layers are currently on-disk on the primary.
3471 : ///
3472 : /// None is returned if the Timeline is in a state where uploading a heatmap
3473 : /// doesn't make sense, such as shutting down or initializing. The caller
3474 : /// should treat this as a cue to simply skip doing any heatmap uploading
3475 : /// for this timeline.
3476 35 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3477 35 : if !self.is_active() {
3478 0 : return None;
3479 35 : }
3480 :
3481 35 : let guard = self.layers.read().await;
3482 :
3483 : // Firstly, if there's any heatmap left over from when this location
3484 : // was a secondary, take that into account. Keep layers that are:
3485 : // * present in the layer map
3486 : // * visible
3487 : // * non-resident
3488 : // * not evicted since we read the heatmap
3489 : //
3490 : // Without this, a new cold, attached location would clobber the previous
3491 : // heatamp.
3492 35 : let previous_heatmap = self.previous_heatmap.load();
3493 35 : let visible_non_resident = match previous_heatmap.as_deref() {
3494 26 : Some(PreviousHeatmap::Active { heatmap, read_at }) => {
3495 100 : Some(heatmap.layers.iter().filter_map(|hl| {
3496 100 : let desc: PersistentLayerDesc = hl.name.clone().into();
3497 100 : let layer = guard.try_get_from_key(&desc.key())?;
3498 :
3499 100 : if layer.visibility() == LayerVisibilityHint::Covered {
3500 0 : return None;
3501 100 : }
3502 100 :
3503 100 : if layer.is_likely_resident() {
3504 41 : return None;
3505 59 : }
3506 59 :
3507 59 : if layer.last_evicted_at().happened_after(*read_at) {
3508 13 : return None;
3509 46 : }
3510 46 :
3511 46 : Some((desc, hl.metadata.clone(), hl.access_time))
3512 100 : }))
3513 : }
3514 0 : Some(PreviousHeatmap::Obsolete) => None,
3515 9 : None => None,
3516 : };
3517 :
3518 : // Secondly, all currently visible, resident layers are included.
3519 78 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3520 78 : match layer.visibility() {
3521 : LayerVisibilityHint::Visible => {
3522 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3523 73 : let last_activity_ts = layer.latest_activity();
3524 73 : Some((
3525 73 : layer.layer_desc().clone(),
3526 73 : layer.metadata(),
3527 73 : last_activity_ts,
3528 73 : ))
3529 : }
3530 : LayerVisibilityHint::Covered => {
3531 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3532 5 : None
3533 : }
3534 : }
3535 78 : });
3536 :
3537 35 : let mut layers = match visible_non_resident {
3538 26 : Some(non_resident) => {
3539 26 : let mut non_resident = non_resident.peekable();
3540 26 : if non_resident.peek().is_none() {
3541 8 : self.previous_heatmap
3542 8 : .store(Some(PreviousHeatmap::Obsolete.into()));
3543 18 : }
3544 :
3545 26 : non_resident.chain(resident).collect::<Vec<_>>()
3546 : }
3547 9 : None => resident.collect::<Vec<_>>(),
3548 : };
3549 :
3550 : // Sort layers in order of which to download first. For a large set of layers to download, we
3551 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3552 : // or hours later:
3553 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3554 : // only exist for a few minutes before being compacted into L1s.
3555 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3556 : // the layer is likely to be covered by an image layer during compaction.
3557 262 : layers.sort_by_key(|(desc, _meta, _atime)| {
3558 262 : std::cmp::Reverse((
3559 262 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3560 262 : desc.lsn_range.end,
3561 262 : ))
3562 262 : });
3563 35 :
3564 35 : let layers = layers
3565 35 : .into_iter()
3566 119 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3567 35 : .collect();
3568 35 :
3569 35 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3570 35 : }
3571 :
3572 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3573 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3574 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3575 0 : // branchpoint in the value in IndexPart::lineage
3576 0 : self.ancestor_lsn == lsn
3577 0 : || (self.ancestor_lsn == Lsn::INVALID
3578 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3579 0 : }
3580 : }
3581 :
3582 : impl Timeline {
3583 : #[allow(clippy::doc_lazy_continuation)]
3584 : /// Get the data needed to reconstruct all keys in the provided keyspace
3585 : ///
3586 : /// The algorithm is as follows:
3587 : /// 1. While some keys are still not done and there's a timeline to visit:
3588 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3589 : /// 2.1: Build the fringe for the current keyspace
3590 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3591 : /// intersects
3592 : /// 2.3. Pop the timeline from the fringe
3593 : /// 2.4. If the fringe is empty, go back to 1
3594 1255313 : async fn get_vectored_reconstruct_data(
3595 1255313 : &self,
3596 1255313 : mut keyspace: KeySpace,
3597 1255313 : request_lsn: Lsn,
3598 1255313 : reconstruct_state: &mut ValuesReconstructState,
3599 1255313 : ctx: &RequestContext,
3600 1255313 : ) -> Result<(), GetVectoredError> {
3601 1255313 : let mut timeline_owned: Arc<Timeline>;
3602 1255313 : let mut timeline = self;
3603 1255313 :
3604 1255313 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3605 :
3606 1255309 : let missing_keyspace = loop {
3607 1706985 : if self.cancel.is_cancelled() {
3608 0 : return Err(GetVectoredError::Cancelled);
3609 1706985 : }
3610 :
3611 : let TimelineVisitOutcome {
3612 1706985 : completed_keyspace: completed,
3613 1706985 : image_covered_keyspace,
3614 1706985 : } = Self::get_vectored_reconstruct_data_timeline(
3615 1706985 : timeline,
3616 1706985 : keyspace.clone(),
3617 1706985 : cont_lsn,
3618 1706985 : reconstruct_state,
3619 1706985 : &self.cancel,
3620 1706985 : ctx,
3621 1706985 : )
3622 1706985 : .await?;
3623 :
3624 1706985 : keyspace.remove_overlapping_with(&completed);
3625 1706985 :
3626 1706985 : // Do not descend into the ancestor timeline for aux files.
3627 1706985 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3628 1706985 : // stalling compaction.
3629 1706985 : keyspace.remove_overlapping_with(&KeySpace {
3630 1706985 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
3631 1706985 : });
3632 1706985 :
3633 1706985 : // Keyspace is fully retrieved
3634 1706985 : if keyspace.is_empty() {
3635 1255253 : break None;
3636 451732 : }
3637 :
3638 451732 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3639 : // Not fully retrieved but no ancestor timeline.
3640 56 : break Some(keyspace);
3641 : };
3642 :
3643 : // Now we see if there are keys covered by the image layer but does not exist in the
3644 : // image layer, which means that the key does not exist.
3645 :
3646 : // The block below will stop the vectored search if any of the keys encountered an image layer
3647 : // which did not contain a snapshot for said key. Since we have already removed all completed
3648 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3649 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3650 : // and stop the search as a result of that.
3651 451676 : let mut removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3652 451676 : // Do not fire missing key error for sparse keys.
3653 451676 : removed.remove_overlapping_with(&KeySpace {
3654 451676 : ranges: vec![SPARSE_RANGE],
3655 451676 : });
3656 451676 : if !removed.is_empty() {
3657 0 : break Some(removed);
3658 451676 : }
3659 451676 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3660 451676 : // keyspace.
3661 451676 :
3662 451676 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3663 451676 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3664 451676 : timeline_owned = timeline
3665 451676 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3666 451676 : .await?;
3667 451672 : timeline = &*timeline_owned;
3668 : };
3669 :
3670 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
3671 1255309 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
3672 56 : let mut missing_keyspace = missing_keyspace;
3673 56 : missing_keyspace.remove_overlapping_with(&KeySpace {
3674 56 : ranges: vec![SPARSE_RANGE],
3675 56 : });
3676 56 : if missing_keyspace.is_empty() {
3677 28 : None
3678 : } else {
3679 28 : Some(missing_keyspace)
3680 : }
3681 : } else {
3682 1255253 : None
3683 : };
3684 :
3685 1255309 : if let Some(missing_keyspace) = missing_keyspace {
3686 28 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3687 28 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3688 28 : shard: self
3689 28 : .shard_identity
3690 28 : .get_shard_number(&missing_keyspace.start().unwrap()),
3691 28 : cont_lsn,
3692 28 : request_lsn,
3693 28 : ancestor_lsn: Some(timeline.ancestor_lsn),
3694 28 : backtrace: None,
3695 28 : read_path: std::mem::take(&mut reconstruct_state.read_path),
3696 28 : }));
3697 1255281 : }
3698 1255281 :
3699 1255281 : Ok(())
3700 1255313 : }
3701 :
3702 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3703 : ///
3704 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3705 : /// the current keyspace. The current keyspace of the search at any given timeline
3706 : /// is the original keyspace minus all the keys that have been completed minus
3707 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3708 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3709 : ///
3710 : /// This is basically a depth-first search visitor implementation where a vertex
3711 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3712 : ///
3713 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3714 : /// and get all the required reconstruct data from the layer in one go.
3715 : ///
3716 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3717 : /// decides how to deal with these two keyspaces.
3718 1706985 : async fn get_vectored_reconstruct_data_timeline(
3719 1706985 : timeline: &Timeline,
3720 1706985 : keyspace: KeySpace,
3721 1706985 : mut cont_lsn: Lsn,
3722 1706985 : reconstruct_state: &mut ValuesReconstructState,
3723 1706985 : cancel: &CancellationToken,
3724 1706985 : ctx: &RequestContext,
3725 1706985 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3726 1706985 : let mut unmapped_keyspace = keyspace.clone();
3727 1706985 : let mut fringe = LayerFringe::new();
3728 1706985 :
3729 1706985 : let mut completed_keyspace = KeySpace::default();
3730 1706985 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3731 1706985 :
3732 1706985 : // Prevent GC from progressing while visiting the current timeline.
3733 1706985 : // If we are GC-ing because a new image layer was added while traversing
3734 1706985 : // the timeline, then it will remove layers that are required for fulfilling
3735 1706985 : // the current get request (read-path cannot "look back" and notice the new
3736 1706985 : // image layer).
3737 1706985 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
3738 :
3739 : // See `compaction::compact_with_gc` for why we need this.
3740 1706985 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
3741 :
3742 : loop {
3743 3474918 : if cancel.is_cancelled() {
3744 0 : return Err(GetVectoredError::Cancelled);
3745 3474918 : }
3746 3474918 :
3747 3474918 : let (keys_done_last_step, keys_with_image_coverage) =
3748 3474918 : reconstruct_state.consume_done_keys();
3749 3474918 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3750 3474918 : completed_keyspace.merge(&keys_done_last_step);
3751 3474918 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3752 4996 : unmapped_keyspace
3753 4996 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3754 4996 : image_covered_keyspace.add_range(keys_with_image_coverage);
3755 3469922 : }
3756 :
3757 : // Do not descent any further if the last layer we visited
3758 : // completed all keys in the keyspace it inspected. This is not
3759 : // required for correctness, but avoids visiting extra layers
3760 : // which turns out to be a perf bottleneck in some cases.
3761 3474918 : if !unmapped_keyspace.is_empty() {
3762 2264198 : let guard = timeline.layers.read().await;
3763 2264198 : let layers = guard.layer_map()?;
3764 :
3765 2264198 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3766 1829134 : let start_lsn = l.get_lsn_range().start;
3767 1829134 : cont_lsn > start_lsn
3768 2264198 : });
3769 2264198 :
3770 2264198 : match in_memory_layer {
3771 1213320 : Some(l) => {
3772 1213320 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3773 1213320 : fringe.update(
3774 1213320 : ReadableLayer::InMemoryLayer(l),
3775 1213320 : unmapped_keyspace.clone(),
3776 1213320 : lsn_range,
3777 1213320 : );
3778 1213320 : }
3779 : None => {
3780 1050928 : for range in unmapped_keyspace.ranges.iter() {
3781 1050928 : let results = layers.range_search(range.clone(), cont_lsn);
3782 1050928 :
3783 1050928 : results
3784 1050928 : .found
3785 1050928 : .into_iter()
3786 1050928 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3787 554665 : (
3788 554665 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3789 554665 : keyspace_accum.to_keyspace(),
3790 554665 : lsn_floor..cont_lsn,
3791 554665 : )
3792 1050928 : })
3793 1050928 : .for_each(|(layer, keyspace, lsn_range)| {
3794 554665 : fringe.update(layer, keyspace, lsn_range)
3795 1050928 : });
3796 1050928 : }
3797 : }
3798 : }
3799 :
3800 : // It's safe to drop the layer map lock after planning the next round of reads.
3801 : // The fringe keeps readable handles for the layers which are safe to read even
3802 : // if layers were compacted or flushed.
3803 : //
3804 : // The more interesting consideration is: "Why is the read algorithm still correct
3805 : // if the layer map changes while it is operating?". Doing a vectored read on a
3806 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3807 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3808 : // range at *a particular point in time*. It is fine for the answer to be different
3809 : // at two different time points.
3810 2264198 : drop(guard);
3811 1210720 : }
3812 :
3813 3474918 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3814 1767933 : if let Some(ref mut read_path) = reconstruct_state.read_path {
3815 1767933 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
3816 1767933 : }
3817 1767933 : let next_cont_lsn = lsn_range.start;
3818 1767933 : layer_to_read
3819 1767933 : .get_values_reconstruct_data(
3820 1767933 : keyspace_to_read.clone(),
3821 1767933 : lsn_range,
3822 1767933 : reconstruct_state,
3823 1767933 : ctx,
3824 1767933 : )
3825 1767933 : .await?;
3826 :
3827 1767933 : unmapped_keyspace = keyspace_to_read;
3828 1767933 : cont_lsn = next_cont_lsn;
3829 1767933 :
3830 1767933 : reconstruct_state.on_layer_visited(&layer_to_read);
3831 : } else {
3832 1706985 : break;
3833 1706985 : }
3834 1706985 : }
3835 1706985 :
3836 1706985 : Ok(TimelineVisitOutcome {
3837 1706985 : completed_keyspace,
3838 1706985 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3839 1706985 : })
3840 1706985 : }
3841 :
3842 451676 : async fn get_ready_ancestor_timeline(
3843 451676 : &self,
3844 451676 : ancestor: &Arc<Timeline>,
3845 451676 : ctx: &RequestContext,
3846 451676 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3847 451676 : // It's possible that the ancestor timeline isn't active yet, or
3848 451676 : // is active but hasn't yet caught up to the branch point. Wait
3849 451676 : // for it.
3850 451676 : //
3851 451676 : // This cannot happen while the pageserver is running normally,
3852 451676 : // because you cannot create a branch from a point that isn't
3853 451676 : // present in the pageserver yet. However, we don't wait for the
3854 451676 : // branch point to be uploaded to cloud storage before creating
3855 451676 : // a branch. I.e., the branch LSN need not be remote consistent
3856 451676 : // for the branching operation to succeed.
3857 451676 : //
3858 451676 : // Hence, if we try to load a tenant in such a state where
3859 451676 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3860 451676 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3861 451676 : // then we will need to wait for the ancestor timeline to
3862 451676 : // re-stream WAL up to branch_lsn before we access it.
3863 451676 : //
3864 451676 : // How can a tenant get in such a state?
3865 451676 : // - ungraceful pageserver process exit
3866 451676 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3867 451676 : //
3868 451676 : // NB: this could be avoided by requiring
3869 451676 : // branch_lsn >= remote_consistent_lsn
3870 451676 : // during branch creation.
3871 451676 : match ancestor.wait_to_become_active(ctx).await {
3872 451672 : Ok(()) => {}
3873 : Err(TimelineState::Stopping) => {
3874 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3875 0 : return Err(GetReadyAncestorError::Cancelled);
3876 : }
3877 4 : Err(state) => {
3878 4 : return Err(GetReadyAncestorError::BadState {
3879 4 : timeline_id: ancestor.timeline_id,
3880 4 : state,
3881 4 : });
3882 : }
3883 : }
3884 451672 : ancestor
3885 451672 : .wait_lsn(
3886 451672 : self.ancestor_lsn,
3887 451672 : WaitLsnWaiter::Timeline(self),
3888 451672 : WaitLsnTimeout::Default,
3889 451672 : ctx,
3890 451672 : )
3891 451672 : .await
3892 451672 : .map_err(|e| match e {
3893 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3894 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3895 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3896 0 : timeline_id: ancestor.timeline_id,
3897 0 : state,
3898 0 : },
3899 451672 : })?;
3900 :
3901 451672 : Ok(ancestor.clone())
3902 451676 : }
3903 :
3904 594312 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3905 594312 : &self.shard_identity
3906 594312 : }
3907 :
3908 : #[inline(always)]
3909 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3910 0 : ShardTimelineId {
3911 0 : shard_index: ShardIndex {
3912 0 : shard_number: self.shard_identity.number,
3913 0 : shard_count: self.shard_identity.count,
3914 0 : },
3915 0 : timeline_id: self.timeline_id,
3916 0 : }
3917 0 : }
3918 :
3919 : /// Returns a non-frozen open in-memory layer for ingestion.
3920 : ///
3921 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3922 : /// this function without holding the mutex.
3923 2597 : async fn get_layer_for_write(
3924 2597 : &self,
3925 2597 : lsn: Lsn,
3926 2597 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3927 2597 : ctx: &RequestContext,
3928 2597 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3929 2597 : let mut guard = self.layers.write().await;
3930 :
3931 2597 : let last_record_lsn = self.get_last_record_lsn();
3932 2597 : ensure!(
3933 2597 : lsn > last_record_lsn,
3934 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3935 : lsn,
3936 : last_record_lsn,
3937 : );
3938 :
3939 2597 : let layer = guard
3940 2597 : .open_mut()?
3941 2597 : .get_layer_for_write(
3942 2597 : lsn,
3943 2597 : self.conf,
3944 2597 : self.timeline_id,
3945 2597 : self.tenant_shard_id,
3946 2597 : &self.gate,
3947 2597 : ctx,
3948 2597 : )
3949 2597 : .await?;
3950 2597 : Ok(layer)
3951 2597 : }
3952 :
3953 10558189 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3954 10558189 : assert!(new_lsn.is_aligned());
3955 :
3956 10558189 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
3957 10558189 : self.last_record_lsn.advance(new_lsn);
3958 10558189 : }
3959 :
3960 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3961 : ///
3962 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3963 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3964 2405 : async fn freeze_inmem_layer_at(
3965 2405 : &self,
3966 2405 : at: Lsn,
3967 2405 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3968 2405 : ) -> Result<u64, FlushLayerError> {
3969 2405 : let frozen = {
3970 2405 : let mut guard = self.layers.write().await;
3971 2405 : guard
3972 2405 : .open_mut()?
3973 2405 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
3974 2405 : .await
3975 : };
3976 :
3977 2405 : if frozen {
3978 2349 : let now = Instant::now();
3979 2349 : *(self.last_freeze_ts.write().unwrap()) = now;
3980 2349 : }
3981 :
3982 : // Increment the flush cycle counter and wake up the flush task.
3983 : // Remember the new value, so that when we listen for the flush
3984 : // to finish, we know when the flush that we initiated has
3985 : // finished, instead of some other flush that was started earlier.
3986 2405 : let mut my_flush_request = 0;
3987 2405 :
3988 2405 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3989 2405 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3990 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3991 2405 : }
3992 2405 :
3993 2405 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3994 2405 : my_flush_request = *counter + 1;
3995 2405 : *counter = my_flush_request;
3996 2405 : *lsn = std::cmp::max(at, *lsn);
3997 2405 : });
3998 2405 :
3999 2405 : assert_ne!(my_flush_request, 0);
4000 :
4001 2405 : Ok(my_flush_request)
4002 2405 : }
4003 :
4004 : /// Layer flusher task's main loop.
4005 885 : async fn flush_loop(
4006 885 : self: &Arc<Self>,
4007 885 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4008 885 : ctx: &RequestContext,
4009 885 : ) {
4010 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4011 885 : let mut watch_l0 = match self.layers.read().await.layer_map() {
4012 885 : Ok(lm) => lm.watch_level0_deltas(),
4013 0 : Err(Shutdown) => return,
4014 : };
4015 :
4016 885 : info!("started flush loop");
4017 : loop {
4018 3215 : tokio::select! {
4019 3215 : _ = self.cancel.cancelled() => {
4020 20 : info!("shutting down layer flush task due to Timeline::cancel");
4021 20 : break;
4022 : },
4023 3215 : _ = layer_flush_start_rx.changed() => {}
4024 2330 : }
4025 2330 : trace!("waking up");
4026 2330 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4027 2330 :
4028 2330 : // The highest LSN to which we flushed in the loop over frozen layers
4029 2330 : let mut flushed_to_lsn = Lsn(0);
4030 :
4031 2330 : let result = loop {
4032 4679 : if self.cancel.is_cancelled() {
4033 0 : info!("dropping out of flush loop for timeline shutdown");
4034 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
4035 : // anyone waiting on that will respect self.cancel as well: they will stop
4036 : // waiting at the same time we as drop out of this loop.
4037 0 : return;
4038 4679 : }
4039 4679 :
4040 4679 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4041 4679 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4042 4679 : if flushed_to_lsn >= frozen_to_lsn {
4043 2274 : break Ok(());
4044 2405 : }
4045 :
4046 : // Fetch the next layer to flush, if any.
4047 2405 : let (layer, l0_count, frozen_count, frozen_size) = {
4048 2405 : let layers = self.layers.read().await;
4049 2405 : let Ok(lm) = layers.layer_map() else {
4050 0 : info!("dropping out of flush loop for timeline shutdown");
4051 0 : return;
4052 : };
4053 2405 : let l0_count = lm.level0_deltas().len();
4054 2405 : let frozen_count = lm.frozen_layers.len();
4055 2405 : let frozen_size: u64 = lm
4056 2405 : .frozen_layers
4057 2405 : .iter()
4058 2425 : .map(|l| l.estimated_in_mem_size())
4059 2405 : .sum();
4060 2405 : let layer = lm.frozen_layers.front().cloned();
4061 2405 : (layer, l0_count, frozen_count, frozen_size)
4062 2405 : // drop 'layers' lock
4063 2405 : };
4064 2405 : let Some(layer) = layer else {
4065 56 : break Ok(());
4066 : };
4067 :
4068 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
4069 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
4070 : //
4071 : // NB: the compaction loop only checks `compaction_threshold` every 20 seconds, so
4072 : // we can end up stalling before compaction even starts. Consider making it more
4073 : // responsive (e.g. via `watch_level0_deltas`).
4074 2349 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
4075 0 : if l0_count >= stall_threshold {
4076 0 : warn!(
4077 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
4078 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4079 : );
4080 0 : let stall_timer = self
4081 0 : .metrics
4082 0 : .flush_delay_histo
4083 0 : .start_timer()
4084 0 : .record_on_drop();
4085 0 : tokio::select! {
4086 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
4087 0 : if let Ok(l0) = result.as_deref() {
4088 0 : let delay = stall_timer.elapsed().as_secs_f64();
4089 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
4090 0 : }
4091 : },
4092 0 : _ = self.cancel.cancelled() => {},
4093 : }
4094 0 : continue; // check again
4095 0 : }
4096 2349 : }
4097 :
4098 : // Flush the layer.
4099 2349 : let flush_timer = self.metrics.flush_time_histo.start_timer();
4100 2349 : match self.flush_frozen_layer(layer, ctx).await {
4101 2349 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
4102 : Err(FlushLayerError::Cancelled) => {
4103 0 : info!("dropping out of flush loop for timeline shutdown");
4104 0 : return;
4105 : }
4106 0 : err @ Err(
4107 0 : FlushLayerError::NotRunning(_)
4108 0 : | FlushLayerError::Other(_)
4109 0 : | FlushLayerError::CreateImageLayersError(_),
4110 0 : ) => {
4111 0 : error!("could not flush frozen layer: {err:?}");
4112 0 : break err.map(|_| ());
4113 : }
4114 : }
4115 2349 : let flush_duration = flush_timer.stop_and_record();
4116 2349 :
4117 2349 : // Notify the tenant compaction loop if L0 compaction is needed.
4118 2349 : let l0_count = *watch_l0.borrow();
4119 2349 : if l0_count >= self.get_compaction_threshold() {
4120 952 : self.l0_compaction_trigger.notify_one();
4121 1397 : }
4122 :
4123 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
4124 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
4125 : // ingestion by having ephemeral layer rolls wait for flushes.
4126 2349 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
4127 0 : if l0_count >= delay_threshold {
4128 0 : let delay = flush_duration.as_secs_f64();
4129 0 : info!(
4130 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
4131 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4132 : );
4133 0 : let _delay_timer = self
4134 0 : .metrics
4135 0 : .flush_delay_histo
4136 0 : .start_timer()
4137 0 : .record_on_drop();
4138 0 : tokio::select! {
4139 0 : _ = tokio::time::sleep(flush_duration) => {},
4140 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
4141 0 : _ = self.cancel.cancelled() => {},
4142 : }
4143 0 : }
4144 2349 : }
4145 : };
4146 :
4147 : // Unsharded tenants should never advance their LSN beyond the end of the
4148 : // highest layer they write: such gaps between layer data and the frozen LSN
4149 : // are only legal on sharded tenants.
4150 2330 : debug_assert!(
4151 2330 : self.shard_identity.count.count() > 1
4152 2330 : || flushed_to_lsn >= frozen_to_lsn
4153 56 : || !flushed_to_lsn.is_valid()
4154 : );
4155 :
4156 2330 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
4157 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
4158 : // to us via layer_flush_start_rx, then advance it here.
4159 : //
4160 : // This path is only taken for tenants with multiple shards: single sharded tenants should
4161 : // never encounter a gap in the wal.
4162 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
4163 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
4164 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
4165 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
4166 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
4167 0 : }
4168 0 : }
4169 2330 : }
4170 :
4171 : // Notify any listeners that we're done
4172 2330 : let _ = self
4173 2330 : .layer_flush_done_tx
4174 2330 : .send_replace((flush_counter, result));
4175 : }
4176 20 : }
4177 :
4178 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
4179 2245 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
4180 2245 : let mut rx = self.layer_flush_done_tx.subscribe();
4181 : loop {
4182 : {
4183 4567 : let (last_result_counter, last_result) = &*rx.borrow();
4184 4567 : if *last_result_counter >= request {
4185 2245 : if let Err(err) = last_result {
4186 : // We already logged the original error in
4187 : // flush_loop. We cannot propagate it to the caller
4188 : // here, because it might not be Cloneable
4189 0 : return Err(err.clone());
4190 : } else {
4191 2245 : return Ok(());
4192 : }
4193 2322 : }
4194 2322 : }
4195 2322 : trace!("waiting for flush to complete");
4196 2322 : tokio::select! {
4197 2322 : rx_e = rx.changed() => {
4198 2322 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
4199 : },
4200 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
4201 : // the notification from [`flush_loop`] that it completed.
4202 2322 : _ = self.cancel.cancelled() => {
4203 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
4204 0 : return Ok(())
4205 : }
4206 : };
4207 2322 : trace!("done")
4208 : }
4209 2245 : }
4210 :
4211 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
4212 : ///
4213 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
4214 : #[instrument(skip_all, fields(layer=%frozen_layer))]
4215 : async fn flush_frozen_layer(
4216 : self: &Arc<Self>,
4217 : frozen_layer: Arc<InMemoryLayer>,
4218 : ctx: &RequestContext,
4219 : ) -> Result<Lsn, FlushLayerError> {
4220 : debug_assert_current_span_has_tenant_and_timeline_id();
4221 :
4222 : // As a special case, when we have just imported an image into the repository,
4223 : // instead of writing out a L0 delta layer, we directly write out image layer
4224 : // files instead. This is possible as long as *all* the data imported into the
4225 : // repository have the same LSN.
4226 : let lsn_range = frozen_layer.get_lsn_range();
4227 :
4228 : // Whether to directly create image layers for this flush, or flush them as delta layers
4229 : let create_image_layer =
4230 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
4231 :
4232 : #[cfg(test)]
4233 : {
4234 : match &mut *self.flush_loop_state.lock().unwrap() {
4235 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
4236 : panic!("flush loop not running")
4237 : }
4238 : FlushLoopState::Running {
4239 : expect_initdb_optimization,
4240 : initdb_optimization_count,
4241 : ..
4242 : } => {
4243 : if create_image_layer {
4244 : *initdb_optimization_count += 1;
4245 : } else {
4246 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
4247 : }
4248 : }
4249 : }
4250 : }
4251 :
4252 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
4253 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
4254 : // require downloading anything during initial import.
4255 : let ((rel_partition, metadata_partition), _lsn) = self
4256 : .repartition(
4257 : self.initdb_lsn,
4258 : self.get_compaction_target_size(),
4259 : EnumSet::empty(),
4260 : ctx,
4261 : )
4262 : .await
4263 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
4264 :
4265 : if self.cancel.is_cancelled() {
4266 : return Err(FlushLayerError::Cancelled);
4267 : }
4268 :
4269 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
4270 : // So that the key ranges don't overlap.
4271 : let mut partitions = KeyPartitioning::default();
4272 : partitions.parts.extend(rel_partition.parts);
4273 : if !metadata_partition.parts.is_empty() {
4274 : assert_eq!(
4275 : metadata_partition.parts.len(),
4276 : 1,
4277 : "currently sparse keyspace should only contain a single metadata keyspace"
4278 : );
4279 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
4280 : // every single key within the keyspace, and therefore, it's safe to force converting it
4281 : // into a dense keyspace before calling this function.
4282 : partitions
4283 : .parts
4284 : .extend(metadata_partition.into_dense().parts);
4285 : }
4286 :
4287 : let mut layers_to_upload = Vec::new();
4288 : let (generated_image_layers, is_complete) = self
4289 : .create_image_layers(
4290 : &partitions,
4291 : self.initdb_lsn,
4292 : ImageLayerCreationMode::Initial,
4293 : ctx,
4294 : LastImageLayerCreationStatus::Initial,
4295 : false, // don't yield for L0, we're flushing L0
4296 : )
4297 : .await?;
4298 : debug_assert!(
4299 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
4300 : "init image generation mode must fully cover the keyspace"
4301 : );
4302 : layers_to_upload.extend(generated_image_layers);
4303 :
4304 : (layers_to_upload, None)
4305 : } else {
4306 : // Normal case, write out a L0 delta layer file.
4307 : // `create_delta_layer` will not modify the layer map.
4308 : // We will remove frozen layer and add delta layer in one atomic operation later.
4309 : let Some(layer) = self
4310 : .create_delta_layer(&frozen_layer, None, ctx)
4311 : .await
4312 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
4313 : else {
4314 : panic!("delta layer cannot be empty if no filter is applied");
4315 : };
4316 : (
4317 : // FIXME: even though we have a single image and single delta layer assumption
4318 : // we push them to vec
4319 : vec![layer.clone()],
4320 : Some(layer),
4321 : )
4322 : };
4323 :
4324 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
4325 :
4326 : if self.cancel.is_cancelled() {
4327 : return Err(FlushLayerError::Cancelled);
4328 : }
4329 :
4330 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
4331 :
4332 : // The new on-disk layers are now in the layer map. We can remove the
4333 : // in-memory layer from the map now. The flushed layer is stored in
4334 : // the mapping in `create_delta_layer`.
4335 : {
4336 : let mut guard = self.layers.write().await;
4337 :
4338 : guard.open_mut()?.finish_flush_l0_layer(
4339 : delta_layer_to_add.as_ref(),
4340 : &frozen_layer,
4341 : &self.metrics,
4342 : );
4343 :
4344 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
4345 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
4346 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
4347 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
4348 : }
4349 : // release lock on 'layers'
4350 : };
4351 :
4352 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
4353 : // This makes us refuse ingest until the new layers have been persisted to the remote
4354 : // TODO: remove this, and rely on l0_flush_{delay,stall}_threshold instead.
4355 : if self.get_l0_flush_wait_upload() {
4356 : let start = Instant::now();
4357 : self.remote_client
4358 : .wait_completion()
4359 : .await
4360 0 : .map_err(|e| match e {
4361 : WaitCompletionError::UploadQueueShutDownOrStopped
4362 : | WaitCompletionError::NotInitialized(
4363 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
4364 0 : ) => FlushLayerError::Cancelled,
4365 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
4366 0 : FlushLayerError::Other(anyhow!(e).into())
4367 : }
4368 0 : })?;
4369 : let duration = start.elapsed().as_secs_f64();
4370 : self.metrics.flush_wait_upload_time_gauge_add(duration);
4371 : }
4372 :
4373 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
4374 : // a compaction can delete the file and then it won't be available for uploads any more.
4375 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
4376 : // race situation.
4377 : // See https://github.com/neondatabase/neon/issues/4526
4378 : pausable_failpoint!("flush-frozen-pausable");
4379 :
4380 : // This failpoint is used by another test case `test_pageserver_recovery`.
4381 : fail_point!("flush-frozen-exit");
4382 :
4383 : Ok(Lsn(lsn_range.end.0 - 1))
4384 : }
4385 :
4386 : /// Return true if the value changed
4387 : ///
4388 : /// This function must only be used from the layer flush task.
4389 2349 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
4390 2349 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
4391 2349 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
4392 :
4393 2349 : self.metrics
4394 2349 : .disk_consistent_lsn_gauge
4395 2349 : .set(new_value.0 as i64);
4396 2349 : new_value != old_value
4397 2349 : }
4398 :
4399 : /// Update metadata file
4400 2449 : fn schedule_uploads(
4401 2449 : &self,
4402 2449 : disk_consistent_lsn: Lsn,
4403 2449 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
4404 2449 : ) -> anyhow::Result<()> {
4405 2449 : // We can only save a valid 'prev_record_lsn' value on disk if we
4406 2449 : // flushed *all* in-memory changes to disk. We only track
4407 2449 : // 'prev_record_lsn' in memory for the latest processed record, so we
4408 2449 : // don't remember what the correct value that corresponds to some old
4409 2449 : // LSN is. But if we flush everything, then the value corresponding
4410 2449 : // current 'last_record_lsn' is correct and we can store it on disk.
4411 2449 : let RecordLsn {
4412 2449 : last: last_record_lsn,
4413 2449 : prev: prev_record_lsn,
4414 2449 : } = self.last_record_lsn.load();
4415 2449 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
4416 2194 : Some(prev_record_lsn)
4417 : } else {
4418 255 : None
4419 : };
4420 :
4421 2449 : let update = crate::tenant::metadata::MetadataUpdate::new(
4422 2449 : disk_consistent_lsn,
4423 2449 : ondisk_prev_record_lsn,
4424 2449 : *self.applied_gc_cutoff_lsn.read(),
4425 2449 : );
4426 2449 :
4427 2449 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
4428 0 : "{}",
4429 0 : x.unwrap()
4430 2449 : ));
4431 :
4432 4822 : for layer in layers_to_upload {
4433 2373 : self.remote_client.schedule_layer_file_upload(layer)?;
4434 : }
4435 2449 : self.remote_client
4436 2449 : .schedule_index_upload_for_metadata_update(&update)?;
4437 :
4438 2449 : Ok(())
4439 2449 : }
4440 :
4441 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
4442 0 : self.remote_client
4443 0 : .preserve_initdb_archive(
4444 0 : &self.tenant_shard_id.tenant_id,
4445 0 : &self.timeline_id,
4446 0 : &self.cancel,
4447 0 : )
4448 0 : .await
4449 0 : }
4450 :
4451 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
4452 : // in layer map immediately. The caller is responsible to put it into the layer map.
4453 1936 : async fn create_delta_layer(
4454 1936 : self: &Arc<Self>,
4455 1936 : frozen_layer: &Arc<InMemoryLayer>,
4456 1936 : key_range: Option<Range<Key>>,
4457 1936 : ctx: &RequestContext,
4458 1936 : ) -> anyhow::Result<Option<ResidentLayer>> {
4459 1936 : let self_clone = Arc::clone(self);
4460 1936 : let frozen_layer = Arc::clone(frozen_layer);
4461 1936 : let ctx = ctx.attached_child();
4462 1936 : let work = async move {
4463 1936 : let Some((desc, path)) = frozen_layer
4464 1936 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
4465 1936 : .await?
4466 : else {
4467 0 : return Ok(None);
4468 : };
4469 1936 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
4470 :
4471 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
4472 : // We just need to fsync the directory in which these inodes are linked,
4473 : // which we know to be the timeline directory.
4474 : //
4475 : // We use fatal_err() below because the after write_to_disk returns with success,
4476 : // the in-memory state of the filesystem already has the layer file in its final place,
4477 : // and subsequent pageserver code could think it's durable while it really isn't.
4478 1936 : let timeline_dir = VirtualFile::open(
4479 1936 : &self_clone
4480 1936 : .conf
4481 1936 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
4482 1936 : &ctx,
4483 1936 : )
4484 1936 : .await
4485 1936 : .fatal_err("VirtualFile::open for timeline dir fsync");
4486 1936 : timeline_dir
4487 1936 : .sync_all()
4488 1936 : .await
4489 1936 : .fatal_err("VirtualFile::sync_all timeline dir");
4490 1936 : anyhow::Ok(Some(new_delta))
4491 1936 : };
4492 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
4493 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
4494 : use crate::virtual_file::io_engine::IoEngine;
4495 1936 : match crate::virtual_file::io_engine::get() {
4496 0 : IoEngine::NotSet => panic!("io engine not set"),
4497 : IoEngine::StdFs => {
4498 968 : let span = tracing::info_span!("blocking");
4499 968 : tokio::task::spawn_blocking({
4500 968 : move || Handle::current().block_on(work.instrument(span))
4501 968 : })
4502 968 : .await
4503 968 : .context("spawn_blocking")
4504 968 : .and_then(|x| x)
4505 : }
4506 : #[cfg(target_os = "linux")]
4507 968 : IoEngine::TokioEpollUring => work.await,
4508 : }
4509 1936 : }
4510 :
4511 1181 : async fn repartition(
4512 1181 : &self,
4513 1181 : lsn: Lsn,
4514 1181 : partition_size: u64,
4515 1181 : flags: EnumSet<CompactFlags>,
4516 1181 : ctx: &RequestContext,
4517 1181 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
4518 1181 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
4519 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
4520 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
4521 : // and hence before the compaction task starts.
4522 0 : return Err(CompactionError::Other(anyhow!(
4523 0 : "repartition() called concurrently"
4524 0 : )));
4525 : };
4526 1181 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
4527 1181 : if lsn < *partition_lsn {
4528 0 : return Err(CompactionError::Other(anyhow!(
4529 0 : "repartition() called with LSN going backwards, this should not happen"
4530 0 : )));
4531 1181 : }
4532 1181 :
4533 1181 : let distance = lsn.0 - partition_lsn.0;
4534 1181 : if *partition_lsn != Lsn(0)
4535 564 : && distance <= self.repartition_threshold
4536 564 : && !flags.contains(CompactFlags::ForceRepartition)
4537 : {
4538 536 : debug!(
4539 : distance,
4540 : threshold = self.repartition_threshold,
4541 0 : "no repartitioning needed"
4542 : );
4543 536 : return Ok((
4544 536 : (dense_partition.clone(), sparse_partition.clone()),
4545 536 : *partition_lsn,
4546 536 : ));
4547 645 : }
4548 :
4549 645 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4550 645 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4551 645 : let sparse_partitioning = SparseKeyPartitioning {
4552 645 : parts: vec![sparse_ks],
4553 645 : }; // no partitioning for metadata keys for now
4554 645 : let result = ((dense_partitioning, sparse_partitioning), lsn);
4555 645 : guard.write(result.clone());
4556 645 : Ok(result)
4557 1181 : }
4558 :
4559 : // Is it time to create a new image layer for the given partition? True if we want to generate.
4560 0 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4561 0 : let threshold = self.get_image_creation_threshold();
4562 :
4563 0 : let guard = self.layers.read().await;
4564 0 : let Ok(layers) = guard.layer_map() else {
4565 0 : return false;
4566 : };
4567 :
4568 0 : let mut max_deltas = 0;
4569 0 : for part_range in &partition.ranges {
4570 0 : let image_coverage = layers.image_coverage(part_range, lsn);
4571 0 : for (img_range, last_img) in image_coverage {
4572 0 : let img_lsn = if let Some(last_img) = last_img {
4573 0 : last_img.get_lsn_range().end
4574 : } else {
4575 0 : Lsn(0)
4576 : };
4577 : // Let's consider an example:
4578 : //
4579 : // delta layer with LSN range 71-81
4580 : // delta layer with LSN range 81-91
4581 : // delta layer with LSN range 91-101
4582 : // image layer at LSN 100
4583 : //
4584 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4585 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4586 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4587 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4588 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4589 0 : if img_lsn < lsn {
4590 0 : let num_deltas =
4591 0 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4592 0 :
4593 0 : max_deltas = max_deltas.max(num_deltas);
4594 0 : if num_deltas >= threshold {
4595 0 : debug!(
4596 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4597 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4598 : );
4599 0 : return true;
4600 0 : }
4601 0 : }
4602 : }
4603 : }
4604 :
4605 0 : debug!(
4606 : max_deltas,
4607 0 : "none of the partitioned ranges had >= {threshold} deltas"
4608 : );
4609 0 : false
4610 0 : }
4611 :
4612 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4613 : /// so that at most one image layer will be produced from this function.
4614 : #[allow(clippy::too_many_arguments)]
4615 457 : async fn create_image_layer_for_rel_blocks(
4616 457 : self: &Arc<Self>,
4617 457 : partition: &KeySpace,
4618 457 : mut image_layer_writer: ImageLayerWriter,
4619 457 : lsn: Lsn,
4620 457 : ctx: &RequestContext,
4621 457 : img_range: Range<Key>,
4622 457 : io_concurrency: IoConcurrency,
4623 457 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4624 457 : let mut wrote_keys = false;
4625 457 :
4626 457 : let mut key_request_accum = KeySpaceAccum::new();
4627 3023 : for range in &partition.ranges {
4628 2566 : let mut key = range.start;
4629 5561 : while key < range.end {
4630 : // Decide whether to retain this key: usually we do, but sharded tenants may
4631 : // need to drop keys that don't belong to them. If we retain the key, add it
4632 : // to `key_request_accum` for later issuing a vectored get
4633 2995 : if self.shard_identity.is_key_disposable(&key) {
4634 0 : debug!(
4635 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4636 0 : key,
4637 0 : self.shard_identity.get_shard_number(&key)
4638 : );
4639 2995 : } else {
4640 2995 : key_request_accum.add_key(key);
4641 2995 : }
4642 :
4643 2995 : let last_key_in_range = key.next() == range.end;
4644 2995 : key = key.next();
4645 2995 :
4646 2995 : // Maybe flush `key_rest_accum`
4647 2995 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4648 2995 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4649 : {
4650 2566 : let results = self
4651 2566 : .get_vectored(
4652 2566 : key_request_accum.consume_keyspace(),
4653 2566 : lsn,
4654 2566 : io_concurrency.clone(),
4655 2566 : ctx,
4656 2566 : )
4657 2566 : .await?;
4658 :
4659 2566 : if self.cancel.is_cancelled() {
4660 0 : return Err(CreateImageLayersError::Cancelled);
4661 2566 : }
4662 :
4663 5561 : for (img_key, img) in results {
4664 2995 : let img = match img {
4665 2995 : Ok(img) => img,
4666 0 : Err(err) => {
4667 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4668 0 : // page without losing any actual user data. That seems better
4669 0 : // than failing repeatedly and getting stuck.
4670 0 : //
4671 0 : // We had a bug at one point, where we truncated the FSM and VM
4672 0 : // in the pageserver, but the Postgres didn't know about that
4673 0 : // and continued to generate incremental WAL records for pages
4674 0 : // that didn't exist in the pageserver. Trying to replay those
4675 0 : // WAL records failed to find the previous image of the page.
4676 0 : // This special case allows us to recover from that situation.
4677 0 : // See https://github.com/neondatabase/neon/issues/2601.
4678 0 : //
4679 0 : // Unfortunately we cannot do this for the main fork, or for
4680 0 : // any metadata keys, keys, as that would lead to actual data
4681 0 : // loss.
4682 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4683 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4684 0 : ZERO_PAGE.clone()
4685 : } else {
4686 0 : return Err(CreateImageLayersError::from(err));
4687 : }
4688 : }
4689 : };
4690 :
4691 : // Write all the keys we just read into our new image layer.
4692 2995 : image_layer_writer.put_image(img_key, img, ctx).await?;
4693 2995 : wrote_keys = true;
4694 : }
4695 429 : }
4696 : }
4697 : }
4698 :
4699 457 : if wrote_keys {
4700 : // Normal path: we have written some data into the new image layer for this
4701 : // partition, so flush it to disk.
4702 457 : info!(
4703 0 : "produced image layer for rel {}",
4704 0 : ImageLayerName {
4705 0 : key_range: img_range.clone(),
4706 0 : lsn
4707 0 : },
4708 : );
4709 457 : Ok(ImageLayerCreationOutcome::Generated {
4710 457 : unfinished_image_layer: image_layer_writer,
4711 457 : })
4712 : } else {
4713 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4714 0 : Ok(ImageLayerCreationOutcome::Empty)
4715 : }
4716 457 : }
4717 :
4718 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4719 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4720 : /// would not be too large to fit in a single image layer.
4721 : ///
4722 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
4723 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
4724 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
4725 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
4726 : #[allow(clippy::too_many_arguments)]
4727 441 : async fn create_image_layer_for_metadata_keys(
4728 441 : self: &Arc<Self>,
4729 441 : partition: &KeySpace,
4730 441 : mut image_layer_writer: ImageLayerWriter,
4731 441 : lsn: Lsn,
4732 441 : ctx: &RequestContext,
4733 441 : img_range: Range<Key>,
4734 441 : mode: ImageLayerCreationMode,
4735 441 : io_concurrency: IoConcurrency,
4736 441 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4737 441 : // Metadata keys image layer creation.
4738 441 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
4739 441 : let begin = Instant::now();
4740 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
4741 : // not contain too many keys, otherwise this takes a lot of memory.
4742 441 : let data = self
4743 441 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4744 441 : .await?;
4745 441 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4746 441 : let mut new_data = BTreeMap::new();
4747 441 : let mut total_kb_retrieved = 0;
4748 441 : let mut total_keys_retrieved = 0;
4749 12465 : for (k, v) in data {
4750 12024 : let v = v?;
4751 12024 : total_kb_retrieved += KEY_SIZE + v.len();
4752 12024 : total_keys_retrieved += 1;
4753 12024 : new_data.insert(k, v);
4754 : }
4755 441 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4756 441 : };
4757 441 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4758 441 : let elapsed = begin.elapsed();
4759 441 :
4760 441 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4761 441 : info!(
4762 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()
4763 : );
4764 :
4765 441 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4766 0 : return Ok(ImageLayerCreationOutcome::Skip);
4767 441 : }
4768 441 : if self.cancel.is_cancelled() {
4769 0 : return Err(CreateImageLayersError::Cancelled);
4770 441 : }
4771 441 : let mut wrote_any_image = false;
4772 12465 : for (k, v) in data {
4773 12024 : if v.is_empty() {
4774 : // the key has been deleted, it does not need an image
4775 : // in metadata keyspace, an empty image == tombstone
4776 8 : continue;
4777 12016 : }
4778 12016 : wrote_any_image = true;
4779 12016 :
4780 12016 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4781 12016 :
4782 12016 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4783 12016 : // on the normal data path either.
4784 12016 : image_layer_writer.put_image(k, v, ctx).await?;
4785 : }
4786 :
4787 441 : if wrote_any_image {
4788 : // Normal path: we have written some data into the new image layer for this
4789 : // partition, so flush it to disk.
4790 20 : info!(
4791 0 : "created image layer for metadata {}",
4792 0 : ImageLayerName {
4793 0 : key_range: img_range.clone(),
4794 0 : lsn
4795 0 : }
4796 : );
4797 20 : Ok(ImageLayerCreationOutcome::Generated {
4798 20 : unfinished_image_layer: image_layer_writer,
4799 20 : })
4800 : } else {
4801 421 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4802 421 : Ok(ImageLayerCreationOutcome::Empty)
4803 : }
4804 441 : }
4805 :
4806 : /// Predicate function which indicates whether we should check if new image layers
4807 : /// are required. Since checking if new image layers are required is expensive in
4808 : /// terms of CPU, we only do it in the following cases:
4809 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4810 : /// 2. If enough time has passed since the last check:
4811 : /// 1. For large tenants, we wish to perform the check more often since they
4812 : /// suffer from the lack of image layers
4813 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4814 1181 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4815 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4816 :
4817 1181 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4818 1181 : let distance = lsn
4819 1181 : .checked_sub(last_checks_at)
4820 1181 : .expect("Attempt to compact with LSN going backwards");
4821 1181 : let min_distance =
4822 1181 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4823 1181 :
4824 1181 : let distance_based_decision = distance.0 >= min_distance;
4825 1181 :
4826 1181 : let mut time_based_decision = false;
4827 1181 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4828 1181 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4829 977 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4830 : {
4831 0 : self.get_checkpoint_timeout()
4832 : } else {
4833 977 : Duration::from_secs(3600 * 48)
4834 : };
4835 :
4836 977 : time_based_decision = match *last_check_instant {
4837 564 : Some(last_check) => {
4838 564 : let elapsed = last_check.elapsed();
4839 564 : elapsed >= check_required_after
4840 : }
4841 413 : None => true,
4842 : };
4843 204 : }
4844 :
4845 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4846 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4847 : // check.
4848 1181 : let decision = distance_based_decision || time_based_decision;
4849 :
4850 1181 : if decision {
4851 413 : self.last_image_layer_creation_check_at.store(lsn);
4852 413 : *last_check_instant = Some(Instant::now());
4853 768 : }
4854 :
4855 1181 : decision
4856 1181 : }
4857 :
4858 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
4859 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
4860 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4861 : async fn create_image_layers(
4862 : self: &Arc<Timeline>,
4863 : partitioning: &KeyPartitioning,
4864 : lsn: Lsn,
4865 : mode: ImageLayerCreationMode,
4866 : ctx: &RequestContext,
4867 : last_status: LastImageLayerCreationStatus,
4868 : yield_for_l0: bool,
4869 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
4870 : let timer = self.metrics.create_images_time_histo.start_timer();
4871 :
4872 : if partitioning.parts.is_empty() {
4873 : warn!("no partitions to create image layers for");
4874 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
4875 : }
4876 :
4877 : // We need to avoid holes between generated image layers.
4878 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4879 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4880 : //
4881 : // How such hole between partitions can appear?
4882 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4883 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4884 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4885 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4886 : let mut start = Key::MIN;
4887 :
4888 : let check_for_image_layers =
4889 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
4890 : info!(
4891 : "resuming image layer creation: last_status=incomplete, continue from {}",
4892 : last_key
4893 : );
4894 : true
4895 : } else {
4896 : self.should_check_if_image_layers_required(lsn)
4897 : };
4898 :
4899 : let mut batch_image_writer = BatchLayerWriter::new(self.conf).await?;
4900 :
4901 : let mut all_generated = true;
4902 :
4903 : let mut partition_processed = 0;
4904 : let mut total_partitions = partitioning.parts.len();
4905 : let mut last_partition_processed = None;
4906 : let mut partition_parts = partitioning.parts.clone();
4907 :
4908 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
4909 : // We need to skip the partitions that have already been processed.
4910 : let mut found = false;
4911 : for (i, partition) in partition_parts.iter().enumerate() {
4912 : if last_key <= partition.end().unwrap() {
4913 : // ```plain
4914 : // |------|--------|----------|------|
4915 : // ^last_key
4916 : // ^start from this partition
4917 : // ```
4918 : // Why `i+1` instead of `i`?
4919 : // It is possible that the user did some writes after the previous image layer creation attempt so that
4920 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
4921 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
4922 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
4923 : if i + 1 >= total_partitions {
4924 : // In general, this case should not happen -- if last_key is on the last partition, the previous
4925 : // iteration of image layer creation should return a complete status.
4926 : break; // with found=false
4927 : }
4928 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
4929 : total_partitions = partition_parts.len();
4930 : // Update the start key to the partition start.
4931 : start = partition_parts[0].start().unwrap();
4932 : found = true;
4933 : break;
4934 : }
4935 : }
4936 : if !found {
4937 : // Last key is within the last partition, or larger than all partitions.
4938 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
4939 : }
4940 : }
4941 :
4942 : for partition in partition_parts.iter() {
4943 : if self.cancel.is_cancelled() {
4944 : return Err(CreateImageLayersError::Cancelled);
4945 : }
4946 : partition_processed += 1;
4947 : let img_range = start..partition.ranges.last().unwrap().end;
4948 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4949 : if compact_metadata {
4950 : for range in &partition.ranges {
4951 : assert!(
4952 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4953 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4954 : "metadata keys must be partitioned separately"
4955 : );
4956 : }
4957 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4958 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4959 : // might mess up with evictions.
4960 : start = img_range.end;
4961 : continue;
4962 : }
4963 : // For initial and force modes, we always generate image layers for metadata keys.
4964 : } else if let ImageLayerCreationMode::Try = mode {
4965 : // check_for_image_layers = false -> skip
4966 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4967 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4968 : start = img_range.end;
4969 : continue;
4970 : }
4971 : }
4972 : if let ImageLayerCreationMode::Force = mode {
4973 : // When forced to create image layers, we might try and create them where they already
4974 : // exist. This mode is only used in tests/debug.
4975 : let layers = self.layers.read().await;
4976 : if layers.contains_key(&PersistentLayerKey {
4977 : key_range: img_range.clone(),
4978 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4979 : is_delta: false,
4980 : }) {
4981 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
4982 : // in the future.
4983 : tracing::info!(
4984 : "Skipping image layer at {lsn} {}..{}, already exists",
4985 : img_range.start,
4986 : img_range.end
4987 : );
4988 : start = img_range.end;
4989 : continue;
4990 : }
4991 : }
4992 :
4993 : let image_layer_writer = ImageLayerWriter::new(
4994 : self.conf,
4995 : self.timeline_id,
4996 : self.tenant_shard_id,
4997 : &img_range,
4998 : lsn,
4999 : ctx,
5000 : )
5001 : .await?;
5002 :
5003 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
5004 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
5005 0 : "failpoint image-layer-writer-fail-before-finish"
5006 0 : )))
5007 0 : });
5008 :
5009 : let io_concurrency = IoConcurrency::spawn_from_conf(
5010 : self.conf,
5011 : self.gate
5012 : .enter()
5013 0 : .map_err(|_| CreateImageLayersError::Cancelled)?,
5014 : );
5015 :
5016 : let outcome = if !compact_metadata {
5017 : self.create_image_layer_for_rel_blocks(
5018 : partition,
5019 : image_layer_writer,
5020 : lsn,
5021 : ctx,
5022 : img_range.clone(),
5023 : io_concurrency,
5024 : )
5025 : .await?
5026 : } else {
5027 : self.create_image_layer_for_metadata_keys(
5028 : partition,
5029 : image_layer_writer,
5030 : lsn,
5031 : ctx,
5032 : img_range.clone(),
5033 : mode,
5034 : io_concurrency,
5035 : )
5036 : .await?
5037 : };
5038 : match outcome {
5039 : ImageLayerCreationOutcome::Empty => {
5040 : // No data in this partition, so we don't need to create an image layer (for now).
5041 : // The next image layer should cover this key range, so we don't advance the `start`
5042 : // key.
5043 : }
5044 : ImageLayerCreationOutcome::Generated {
5045 : unfinished_image_layer,
5046 : } => {
5047 : batch_image_writer.add_unfinished_image_writer(
5048 : unfinished_image_layer,
5049 : img_range.clone(),
5050 : lsn,
5051 : );
5052 : // The next image layer should be generated right after this one.
5053 : start = img_range.end;
5054 : }
5055 : ImageLayerCreationOutcome::Skip => {
5056 : // We don't need to create an image layer for this partition.
5057 : // The next image layer should NOT cover this range, otherwise
5058 : // the keyspace becomes empty (reads don't go past image layers).
5059 : start = img_range.end;
5060 : }
5061 : }
5062 :
5063 : if let ImageLayerCreationMode::Try = mode {
5064 : // We have at least made some progress
5065 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
5066 : // The `Try` mode is currently only used on the compaction path. We want to avoid
5067 : // image layer generation taking too long time and blocking L0 compaction. So in this
5068 : // mode, we also inspect the current number of L0 layers and skip image layer generation
5069 : // if there are too many of them.
5070 : let num_of_l0_layers = {
5071 : let layers = self.layers.read().await;
5072 : layers.layer_map()?.level0_deltas().len()
5073 : };
5074 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
5075 : * self.get_compaction_threshold();
5076 : if image_preempt_threshold != 0 && num_of_l0_layers >= image_preempt_threshold {
5077 : tracing::info!(
5078 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers {}",
5079 : partition.start().unwrap(), partition.end().unwrap(), num_of_l0_layers
5080 : );
5081 : last_partition_processed = Some(partition.clone());
5082 : all_generated = false;
5083 : break;
5084 : }
5085 : }
5086 : }
5087 : }
5088 :
5089 : let image_layers = batch_image_writer.finish(self, ctx).await?;
5090 :
5091 : let mut guard = self.layers.write().await;
5092 :
5093 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
5094 : // now they are being scheduled outside of write lock; current way is inconsistent with
5095 : // compaction lock order.
5096 : guard
5097 : .open_mut()?
5098 : .track_new_image_layers(&image_layers, &self.metrics);
5099 : drop_wlock(guard);
5100 : let duration = timer.stop_and_record();
5101 :
5102 : // Creating image layers may have caused some previously visible layers to be covered
5103 : if !image_layers.is_empty() {
5104 : self.update_layer_visibility().await?;
5105 : }
5106 :
5107 : let total_layer_size = image_layers
5108 : .iter()
5109 477 : .map(|l| l.metadata().file_size)
5110 : .sum::<u64>();
5111 :
5112 : info!(
5113 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
5114 : image_layers.len(),
5115 : total_layer_size,
5116 : duration.as_secs_f64(),
5117 : partition_processed,
5118 : total_partitions
5119 : );
5120 :
5121 : Ok((
5122 : image_layers,
5123 : if all_generated {
5124 : LastImageLayerCreationStatus::Complete
5125 : } else {
5126 : LastImageLayerCreationStatus::Incomplete {
5127 : last_key: if let Some(last_partition_processed) = last_partition_processed {
5128 : last_partition_processed.end().unwrap_or(Key::MIN)
5129 : } else {
5130 : // This branch should be unreachable, but in case it happens, we can just return the start key.
5131 : Key::MIN
5132 : },
5133 : }
5134 : },
5135 : ))
5136 : }
5137 :
5138 : /// Wait until the background initial logical size calculation is complete, or
5139 : /// this Timeline is shut down. Calling this function will cause the initial
5140 : /// logical size calculation to skip waiting for the background jobs barrier.
5141 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
5142 0 : if !self.shard_identity.is_shard_zero() {
5143 : // We don't populate logical size on shard >0: skip waiting for it.
5144 0 : return;
5145 0 : }
5146 0 :
5147 0 : if self.remote_client.is_deleting() {
5148 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
5149 0 : return;
5150 0 : }
5151 0 :
5152 0 : if self.current_logical_size.current_size().is_exact() {
5153 : // root timelines are initialized with exact count, but never start the background
5154 : // calculation
5155 0 : return;
5156 0 : }
5157 :
5158 0 : if let Some(await_bg_cancel) = self
5159 0 : .current_logical_size
5160 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
5161 0 : .get()
5162 0 : {
5163 0 : await_bg_cancel.cancel();
5164 0 : } else {
5165 : // We should not wait if we were not able to explicitly instruct
5166 : // the logical size cancellation to skip the concurrency limit semaphore.
5167 : // TODO: this is an unexpected case. We should restructure so that it
5168 : // can't happen.
5169 0 : tracing::warn!(
5170 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
5171 : );
5172 0 : debug_assert!(false);
5173 : }
5174 :
5175 0 : tokio::select!(
5176 0 : _ = self.current_logical_size.initialized.acquire() => {},
5177 0 : _ = self.cancel.cancelled() => {}
5178 : )
5179 0 : }
5180 :
5181 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
5182 : /// Timelines layers up to the ancestor_lsn.
5183 : ///
5184 : /// Requires a timeline that:
5185 : /// - has an ancestor to detach from
5186 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
5187 : /// a technical requirement
5188 : ///
5189 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
5190 : /// polled again until completion.
5191 : ///
5192 : /// During the operation all timelines sharing the data with this timeline will be reparented
5193 : /// from our ancestor to be branches of this timeline.
5194 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
5195 0 : self: &Arc<Timeline>,
5196 0 : tenant: &crate::tenant::Tenant,
5197 0 : options: detach_ancestor::Options,
5198 0 : ctx: &RequestContext,
5199 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
5200 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
5201 0 : }
5202 :
5203 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
5204 : /// reparents any reparentable children of previous ancestor.
5205 : ///
5206 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
5207 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
5208 : /// successfully, tenant must be reloaded.
5209 : ///
5210 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
5211 : /// resetting the tenant.
5212 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
5213 0 : self: &Arc<Timeline>,
5214 0 : tenant: &crate::tenant::Tenant,
5215 0 : prepared: detach_ancestor::PreparedTimelineDetach,
5216 0 : ctx: &RequestContext,
5217 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
5218 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
5219 0 : }
5220 :
5221 : /// Final step which unblocks the GC.
5222 : ///
5223 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
5224 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
5225 0 : self: &Arc<Timeline>,
5226 0 : tenant: &crate::tenant::Tenant,
5227 0 : attempt: detach_ancestor::Attempt,
5228 0 : ctx: &RequestContext,
5229 0 : ) -> Result<(), detach_ancestor::Error> {
5230 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
5231 0 : }
5232 : }
5233 :
5234 : impl Drop for Timeline {
5235 20 : fn drop(&mut self) {
5236 20 : if let Some(ancestor) = &self.ancestor_timeline {
5237 : // This lock should never be poisoned, but in case it is we do a .map() instead of
5238 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
5239 8 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
5240 8 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
5241 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
5242 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
5243 8 : }
5244 0 : }
5245 12 : }
5246 20 : info!(
5247 0 : "Timeline {} for tenant {} is being dropped",
5248 : self.timeline_id, self.tenant_shard_id.tenant_id
5249 : );
5250 20 : }
5251 : }
5252 :
5253 : /// Top-level failure to compact.
5254 : #[derive(Debug, thiserror::Error)]
5255 : pub(crate) enum CompactionError {
5256 : #[error("The timeline or pageserver is shutting down")]
5257 : ShuttingDown,
5258 : /// Compaction tried to offload a timeline and failed
5259 : #[error("Failed to offload timeline: {0}")]
5260 : Offload(OffloadError),
5261 : /// Compaction cannot be done right now; page reconstruction and so on.
5262 : #[error(transparent)]
5263 : Other(anyhow::Error),
5264 : }
5265 :
5266 : impl From<OffloadError> for CompactionError {
5267 0 : fn from(e: OffloadError) -> Self {
5268 0 : match e {
5269 0 : OffloadError::Cancelled => Self::ShuttingDown,
5270 0 : _ => Self::Offload(e),
5271 : }
5272 0 : }
5273 : }
5274 :
5275 : impl CompactionError {
5276 0 : pub fn is_cancelled(&self) -> bool {
5277 0 : matches!(self, CompactionError::ShuttingDown)
5278 0 : }
5279 : }
5280 :
5281 : impl From<CollectKeySpaceError> for CompactionError {
5282 0 : fn from(err: CollectKeySpaceError) -> Self {
5283 0 : match err {
5284 : CollectKeySpaceError::Cancelled
5285 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
5286 0 : CompactionError::ShuttingDown
5287 : }
5288 0 : e => CompactionError::Other(e.into()),
5289 : }
5290 0 : }
5291 : }
5292 :
5293 : impl From<super::upload_queue::NotInitialized> for CompactionError {
5294 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
5295 0 : match value {
5296 : super::upload_queue::NotInitialized::Uninitialized => {
5297 0 : CompactionError::Other(anyhow::anyhow!(value))
5298 : }
5299 : super::upload_queue::NotInitialized::ShuttingDown
5300 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
5301 : }
5302 0 : }
5303 : }
5304 :
5305 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
5306 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
5307 0 : match e {
5308 : super::storage_layer::layer::DownloadError::TimelineShutdown
5309 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
5310 0 : CompactionError::ShuttingDown
5311 : }
5312 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
5313 : | super::storage_layer::layer::DownloadError::DownloadRequired
5314 : | super::storage_layer::layer::DownloadError::NotFile(_)
5315 : | super::storage_layer::layer::DownloadError::DownloadFailed
5316 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
5317 0 : CompactionError::Other(anyhow::anyhow!(e))
5318 : }
5319 : #[cfg(test)]
5320 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
5321 0 : CompactionError::Other(anyhow::anyhow!(e))
5322 : }
5323 : }
5324 0 : }
5325 : }
5326 :
5327 : impl From<layer_manager::Shutdown> for CompactionError {
5328 0 : fn from(_: layer_manager::Shutdown) -> Self {
5329 0 : CompactionError::ShuttingDown
5330 0 : }
5331 : }
5332 :
5333 : #[serde_as]
5334 644 : #[derive(serde::Serialize)]
5335 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
5336 :
5337 : #[derive(Default)]
5338 : enum DurationRecorder {
5339 : #[default]
5340 : NotStarted,
5341 : Recorded(RecordedDuration, tokio::time::Instant),
5342 : }
5343 :
5344 : impl DurationRecorder {
5345 1228 : fn till_now(&self) -> DurationRecorder {
5346 1228 : match self {
5347 : DurationRecorder::NotStarted => {
5348 0 : panic!("must only call on recorded measurements")
5349 : }
5350 1228 : DurationRecorder::Recorded(_, ended) => {
5351 1228 : let now = tokio::time::Instant::now();
5352 1228 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
5353 1228 : }
5354 1228 : }
5355 1228 : }
5356 644 : fn into_recorded(self) -> Option<RecordedDuration> {
5357 644 : match self {
5358 0 : DurationRecorder::NotStarted => None,
5359 644 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
5360 : }
5361 644 : }
5362 : }
5363 :
5364 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
5365 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
5366 : /// the layer descriptor requires the user to provide the ranges, which should cover all
5367 : /// keys specified in the `data` field.
5368 : #[cfg(test)]
5369 : #[derive(Clone)]
5370 : pub struct DeltaLayerTestDesc {
5371 : pub lsn_range: Range<Lsn>,
5372 : pub key_range: Range<Key>,
5373 : pub data: Vec<(Key, Lsn, Value)>,
5374 : }
5375 :
5376 : #[cfg(test)]
5377 : impl DeltaLayerTestDesc {
5378 9 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
5379 9 : Self {
5380 9 : lsn_range,
5381 9 : key_range,
5382 9 : data,
5383 9 : }
5384 9 : }
5385 :
5386 178 : pub fn new_with_inferred_key_range(
5387 178 : lsn_range: Range<Lsn>,
5388 178 : data: Vec<(Key, Lsn, Value)>,
5389 178 : ) -> Self {
5390 442 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
5391 442 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
5392 178 : Self {
5393 178 : key_range: (*key_min)..(key_max.next()),
5394 178 : lsn_range,
5395 178 : data,
5396 178 : }
5397 178 : }
5398 :
5399 25 : pub(crate) fn layer_name(&self) -> LayerName {
5400 25 : LayerName::Delta(super::storage_layer::DeltaLayerName {
5401 25 : key_range: self.key_range.clone(),
5402 25 : lsn_range: self.lsn_range.clone(),
5403 25 : })
5404 25 : }
5405 : }
5406 :
5407 : impl Timeline {
5408 92 : async fn finish_compact_batch(
5409 92 : self: &Arc<Self>,
5410 92 : new_deltas: &[ResidentLayer],
5411 92 : new_images: &[ResidentLayer],
5412 92 : layers_to_remove: &[Layer],
5413 92 : ) -> Result<(), CompactionError> {
5414 92 : let mut guard = tokio::select! {
5415 92 : guard = self.layers.write() => guard,
5416 92 : _ = self.cancel.cancelled() => {
5417 0 : return Err(CompactionError::ShuttingDown);
5418 : }
5419 : };
5420 :
5421 92 : let mut duplicated_layers = HashSet::new();
5422 92 :
5423 92 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
5424 :
5425 744 : for l in new_deltas {
5426 652 : if guard.contains(l.as_ref()) {
5427 : // expected in tests
5428 0 : tracing::error!(layer=%l, "duplicated L1 layer");
5429 :
5430 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
5431 : // `guard` on self.layers. as of writing this, there are no error returns except
5432 : // for compact_level0_phase1 creating an L0, which does not happen in practice
5433 : // because we have not implemented L0 => L0 compaction.
5434 0 : duplicated_layers.insert(l.layer_desc().key());
5435 652 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
5436 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
5437 652 : } else {
5438 652 : insert_layers.push(l.clone());
5439 652 : }
5440 : }
5441 :
5442 : // only remove those inputs which were not outputs
5443 92 : let remove_layers: Vec<Layer> = layers_to_remove
5444 92 : .iter()
5445 804 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
5446 92 : .cloned()
5447 92 : .collect();
5448 92 :
5449 92 : if !new_images.is_empty() {
5450 0 : guard
5451 0 : .open_mut()?
5452 0 : .track_new_image_layers(new_images, &self.metrics);
5453 92 : }
5454 :
5455 92 : guard
5456 92 : .open_mut()?
5457 92 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
5458 92 :
5459 92 : self.remote_client
5460 92 : .schedule_compaction_update(&remove_layers, new_deltas)?;
5461 :
5462 92 : drop_wlock(guard);
5463 92 :
5464 92 : Ok(())
5465 92 : }
5466 :
5467 0 : async fn rewrite_layers(
5468 0 : self: &Arc<Self>,
5469 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
5470 0 : mut drop_layers: Vec<Layer>,
5471 0 : ) -> Result<(), CompactionError> {
5472 0 : let mut guard = self.layers.write().await;
5473 :
5474 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
5475 : // to avoid double-removing, and avoid rewriting something that was removed.
5476 0 : replace_layers.retain(|(l, _)| guard.contains(l));
5477 0 : drop_layers.retain(|l| guard.contains(l));
5478 0 :
5479 0 : guard
5480 0 : .open_mut()?
5481 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
5482 0 :
5483 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
5484 0 :
5485 0 : self.remote_client
5486 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
5487 :
5488 0 : Ok(())
5489 0 : }
5490 :
5491 : /// Schedules the uploads of the given image layers
5492 768 : fn upload_new_image_layers(
5493 768 : self: &Arc<Self>,
5494 768 : new_images: impl IntoIterator<Item = ResidentLayer>,
5495 768 : ) -> Result<(), super::upload_queue::NotInitialized> {
5496 808 : for layer in new_images {
5497 40 : self.remote_client.schedule_layer_file_upload(layer)?;
5498 : }
5499 : // should any new image layer been created, not uploading index_part will
5500 : // result in a mismatch between remote_physical_size and layermap calculated
5501 : // size, which will fail some tests, but should not be an issue otherwise.
5502 768 : self.remote_client
5503 768 : .schedule_index_upload_for_file_changes()?;
5504 768 : Ok(())
5505 768 : }
5506 :
5507 0 : async fn find_gc_time_cutoff(
5508 0 : &self,
5509 0 : now: SystemTime,
5510 0 : pitr: Duration,
5511 0 : cancel: &CancellationToken,
5512 0 : ctx: &RequestContext,
5513 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
5514 0 : debug_assert_current_span_has_tenant_and_timeline_id();
5515 0 : if self.shard_identity.is_shard_zero() {
5516 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
5517 0 : let time_range = if pitr == Duration::ZERO {
5518 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
5519 : } else {
5520 0 : pitr
5521 : };
5522 :
5523 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
5524 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
5525 0 : let timestamp = to_pg_timestamp(time_cutoff);
5526 :
5527 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
5528 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
5529 0 : LsnForTimestamp::Future(lsn) => {
5530 0 : // The timestamp is in the future. That sounds impossible,
5531 0 : // but what it really means is that there hasn't been
5532 0 : // any commits since the cutoff timestamp.
5533 0 : //
5534 0 : // In this case we should use the LSN of the most recent commit,
5535 0 : // which is implicitly the last LSN in the log.
5536 0 : debug!("future({})", lsn);
5537 0 : Some(self.get_last_record_lsn())
5538 : }
5539 0 : LsnForTimestamp::Past(lsn) => {
5540 0 : debug!("past({})", lsn);
5541 0 : None
5542 : }
5543 0 : LsnForTimestamp::NoData(lsn) => {
5544 0 : debug!("nodata({})", lsn);
5545 0 : None
5546 : }
5547 : };
5548 0 : Ok(time_cutoff)
5549 : } else {
5550 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
5551 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
5552 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
5553 : // space cutoff that we would also have respected ourselves.
5554 0 : match self
5555 0 : .remote_client
5556 0 : .download_foreign_index(ShardNumber(0), cancel)
5557 0 : .await
5558 : {
5559 0 : Ok((index_part, index_generation, _index_mtime)) => {
5560 0 : tracing::info!("GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
5561 0 : index_part.metadata.latest_gc_cutoff_lsn());
5562 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
5563 : }
5564 : Err(DownloadError::NotFound) => {
5565 : // This is unexpected, because during timeline creations shard zero persists to remote
5566 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
5567 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
5568 : // state, then shard zero should _eventually_ write an index when it GCs.
5569 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
5570 0 : Ok(None)
5571 : }
5572 0 : Err(e) => {
5573 0 : // TODO: this function should return a different error type than page reconstruct error
5574 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
5575 : }
5576 : }
5577 :
5578 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
5579 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
5580 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
5581 : // new location. This is legal in principle, but problematic in practice because it might result
5582 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
5583 : // because they have GC'd past the branch point.
5584 : }
5585 0 : }
5586 :
5587 : /// Find the Lsns above which layer files need to be retained on
5588 : /// garbage collection.
5589 : ///
5590 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
5591 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
5592 : /// the space-based retention.
5593 : ///
5594 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
5595 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
5596 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
5597 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
5598 : /// in the response as the GC cutoff point for the timeline.
5599 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
5600 : pub(super) async fn find_gc_cutoffs(
5601 : &self,
5602 : now: SystemTime,
5603 : space_cutoff: Lsn,
5604 : pitr: Duration,
5605 : cancel: &CancellationToken,
5606 : ctx: &RequestContext,
5607 : ) -> Result<GcCutoffs, PageReconstructError> {
5608 : let _timer = self
5609 : .metrics
5610 : .find_gc_cutoffs_histo
5611 : .start_timer()
5612 : .record_on_drop();
5613 :
5614 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
5615 :
5616 : if cfg!(test) {
5617 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
5618 : if pitr == Duration::ZERO {
5619 : return Ok(GcCutoffs {
5620 : time: self.get_last_record_lsn(),
5621 : space: space_cutoff,
5622 : });
5623 : }
5624 : }
5625 :
5626 : // Calculate a time-based limit on how much to retain:
5627 : // - if PITR interval is set, then this is our cutoff.
5628 : // - if PITR interval is not set, then we do a lookup
5629 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
5630 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
5631 :
5632 : Ok(match (pitr, time_cutoff) {
5633 : (Duration::ZERO, Some(time_cutoff)) => {
5634 : // PITR is not set. Retain the size-based limit, or the default time retention,
5635 : // whichever requires less data.
5636 : GcCutoffs {
5637 : time: self.get_last_record_lsn(),
5638 : space: std::cmp::max(time_cutoff, space_cutoff),
5639 : }
5640 : }
5641 : (Duration::ZERO, None) => {
5642 : // PITR is not set, and time lookup failed
5643 : GcCutoffs {
5644 : time: self.get_last_record_lsn(),
5645 : space: space_cutoff,
5646 : }
5647 : }
5648 : (_, None) => {
5649 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
5650 : // cannot advance beyond what was already GC'd, and respect space-based retention
5651 : GcCutoffs {
5652 : time: *self.get_applied_gc_cutoff_lsn(),
5653 : space: space_cutoff,
5654 : }
5655 : }
5656 : (_, Some(time_cutoff)) => {
5657 : // PITR interval is set and we looked up timestamp successfully. Ignore
5658 : // size based retention and make time cutoff authoritative
5659 : GcCutoffs {
5660 : time: time_cutoff,
5661 : space: time_cutoff,
5662 : }
5663 : }
5664 : })
5665 : }
5666 :
5667 : /// Garbage collect layer files on a timeline that are no longer needed.
5668 : ///
5669 : /// Currently, we don't make any attempt at removing unneeded page versions
5670 : /// within a layer file. We can only remove the whole file if it's fully
5671 : /// obsolete.
5672 8 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
5673 : // this is most likely the background tasks, but it might be the spawned task from
5674 : // immediate_gc
5675 8 : let _g = tokio::select! {
5676 8 : guard = self.gc_lock.lock() => guard,
5677 8 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
5678 : };
5679 8 : let timer = self.metrics.garbage_collect_histo.start_timer();
5680 8 :
5681 8 : fail_point!("before-timeline-gc");
5682 8 :
5683 8 : // Is the timeline being deleted?
5684 8 : if self.is_stopping() {
5685 0 : return Err(GcError::TimelineCancelled);
5686 8 : }
5687 8 :
5688 8 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
5689 8 : let gc_info = self.gc_info.read().unwrap();
5690 8 :
5691 8 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
5692 8 : let time_cutoff = gc_info.cutoffs.time;
5693 8 : let retain_lsns = gc_info
5694 8 : .retain_lsns
5695 8 : .iter()
5696 8 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
5697 8 : .collect();
5698 8 :
5699 8 : // Gets the maximum LSN that holds the valid lease.
5700 8 : //
5701 8 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
5702 8 : // Here, we do not check for stale leases again.
5703 8 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
5704 8 :
5705 8 : (
5706 8 : space_cutoff,
5707 8 : time_cutoff,
5708 8 : retain_lsns,
5709 8 : max_lsn_with_valid_lease,
5710 8 : )
5711 8 : };
5712 8 :
5713 8 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
5714 8 : let standby_horizon = self.standby_horizon.load();
5715 8 : // Hold GC for the standby, but as a safety guard do it only within some
5716 8 : // reasonable lag.
5717 8 : if standby_horizon != Lsn::INVALID {
5718 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
5719 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
5720 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
5721 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
5722 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
5723 : } else {
5724 0 : warn!(
5725 0 : "standby is lagging for more than {}MB, not holding gc for it",
5726 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
5727 : )
5728 : }
5729 0 : }
5730 8 : }
5731 :
5732 : // Reset standby horizon to ignore it if it is not updated till next GC.
5733 : // It is an easy way to unset it when standby disappears without adding
5734 : // more conf options.
5735 8 : self.standby_horizon.store(Lsn::INVALID);
5736 8 : self.metrics
5737 8 : .standby_horizon_gauge
5738 8 : .set(Lsn::INVALID.0 as i64);
5739 :
5740 8 : let res = self
5741 8 : .gc_timeline(
5742 8 : space_cutoff,
5743 8 : time_cutoff,
5744 8 : retain_lsns,
5745 8 : max_lsn_with_valid_lease,
5746 8 : new_gc_cutoff,
5747 8 : )
5748 8 : .instrument(
5749 8 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5750 : )
5751 8 : .await?;
5752 :
5753 : // only record successes
5754 8 : timer.stop_and_record();
5755 8 :
5756 8 : Ok(res)
5757 8 : }
5758 :
5759 8 : async fn gc_timeline(
5760 8 : &self,
5761 8 : space_cutoff: Lsn,
5762 8 : time_cutoff: Lsn,
5763 8 : retain_lsns: Vec<Lsn>,
5764 8 : max_lsn_with_valid_lease: Option<Lsn>,
5765 8 : new_gc_cutoff: Lsn,
5766 8 : ) -> Result<GcResult, GcError> {
5767 8 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5768 8 :
5769 8 : let now = SystemTime::now();
5770 8 : let mut result: GcResult = GcResult::default();
5771 8 :
5772 8 : // Nothing to GC. Return early.
5773 8 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
5774 8 : if latest_gc_cutoff >= new_gc_cutoff {
5775 0 : info!(
5776 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5777 : );
5778 0 : return Ok(result);
5779 8 : }
5780 :
5781 : // We need to ensure that no one tries to read page versions or create
5782 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5783 : // for details. This will block until the old value is no longer in use.
5784 : //
5785 : // The GC cutoff should only ever move forwards.
5786 8 : let waitlist = {
5787 8 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
5788 8 : if *write_guard > new_gc_cutoff {
5789 0 : return Err(GcError::BadLsn {
5790 0 : why: format!(
5791 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5792 0 : *write_guard, new_gc_cutoff
5793 0 : ),
5794 0 : });
5795 8 : }
5796 8 :
5797 8 : write_guard.store_and_unlock(new_gc_cutoff)
5798 8 : };
5799 8 : waitlist.wait().await;
5800 :
5801 8 : info!("GC starting");
5802 :
5803 8 : debug!("retain_lsns: {:?}", retain_lsns);
5804 :
5805 8 : let mut layers_to_remove = Vec::new();
5806 :
5807 : // Scan all layers in the timeline (remote or on-disk).
5808 : //
5809 : // Garbage collect the layer if all conditions are satisfied:
5810 : // 1. it is older than cutoff LSN;
5811 : // 2. it is older than PITR interval;
5812 : // 3. it doesn't need to be retained for 'retain_lsns';
5813 : // 4. it does not need to be kept for LSNs holding valid leases.
5814 : // 5. newer on-disk image layers cover the layer's whole key range
5815 : //
5816 : // TODO holding a write lock is too agressive and avoidable
5817 8 : let mut guard = self.layers.write().await;
5818 8 : let layers = guard.layer_map()?;
5819 48 : 'outer: for l in layers.iter_historic_layers() {
5820 48 : result.layers_total += 1;
5821 48 :
5822 48 : // 1. Is it newer than GC horizon cutoff point?
5823 48 : if l.get_lsn_range().end > space_cutoff {
5824 4 : info!(
5825 0 : "keeping {} because it's newer than space_cutoff {}",
5826 0 : l.layer_name(),
5827 : space_cutoff,
5828 : );
5829 4 : result.layers_needed_by_cutoff += 1;
5830 4 : continue 'outer;
5831 44 : }
5832 44 :
5833 44 : // 2. It is newer than PiTR cutoff point?
5834 44 : if l.get_lsn_range().end > time_cutoff {
5835 0 : info!(
5836 0 : "keeping {} because it's newer than time_cutoff {}",
5837 0 : l.layer_name(),
5838 : time_cutoff,
5839 : );
5840 0 : result.layers_needed_by_pitr += 1;
5841 0 : continue 'outer;
5842 44 : }
5843 :
5844 : // 3. Is it needed by a child branch?
5845 : // NOTE With that we would keep data that
5846 : // might be referenced by child branches forever.
5847 : // We can track this in child timeline GC and delete parent layers when
5848 : // they are no longer needed. This might be complicated with long inheritance chains.
5849 : //
5850 : // TODO Vec is not a great choice for `retain_lsns`
5851 44 : for retain_lsn in &retain_lsns {
5852 : // start_lsn is inclusive
5853 0 : if &l.get_lsn_range().start <= retain_lsn {
5854 0 : info!(
5855 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5856 0 : l.layer_name(),
5857 0 : retain_lsn,
5858 0 : l.is_incremental(),
5859 : );
5860 0 : result.layers_needed_by_branches += 1;
5861 0 : continue 'outer;
5862 0 : }
5863 : }
5864 :
5865 : // 4. Is there a valid lease that requires us to keep this layer?
5866 44 : if let Some(lsn) = &max_lsn_with_valid_lease {
5867 : // keep if layer start <= any of the lease
5868 36 : if &l.get_lsn_range().start <= lsn {
5869 28 : info!(
5870 0 : "keeping {} because there is a valid lease preventing GC at {}",
5871 0 : l.layer_name(),
5872 : lsn,
5873 : );
5874 28 : result.layers_needed_by_leases += 1;
5875 28 : continue 'outer;
5876 8 : }
5877 8 : }
5878 :
5879 : // 5. Is there a later on-disk layer for this relation?
5880 : //
5881 : // The end-LSN is exclusive, while disk_consistent_lsn is
5882 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5883 : // OK for a delta layer to have end LSN 101, but if the end LSN
5884 : // is 102, then it might not have been fully flushed to disk
5885 : // before crash.
5886 : //
5887 : // For example, imagine that the following layers exist:
5888 : //
5889 : // 1000 - image (A)
5890 : // 1000-2000 - delta (B)
5891 : // 2000 - image (C)
5892 : // 2000-3000 - delta (D)
5893 : // 3000 - image (E)
5894 : //
5895 : // If GC horizon is at 2500, we can remove layers A and B, but
5896 : // we cannot remove C, even though it's older than 2500, because
5897 : // the delta layer 2000-3000 depends on it.
5898 16 : if !layers
5899 16 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5900 : {
5901 12 : info!("keeping {} because it is the latest layer", l.layer_name());
5902 12 : result.layers_not_updated += 1;
5903 12 : continue 'outer;
5904 4 : }
5905 4 :
5906 4 : // We didn't find any reason to keep this file, so remove it.
5907 4 : info!(
5908 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5909 0 : l.layer_name(),
5910 0 : l.is_incremental(),
5911 : );
5912 4 : layers_to_remove.push(l);
5913 : }
5914 :
5915 8 : if !layers_to_remove.is_empty() {
5916 : // Persist the new GC cutoff value before we actually remove anything.
5917 : // This unconditionally schedules also an index_part.json update, even though, we will
5918 : // be doing one a bit later with the unlinked gc'd layers.
5919 4 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5920 4 : self.schedule_uploads(disk_consistent_lsn, None)
5921 4 : .map_err(|e| {
5922 0 : if self.cancel.is_cancelled() {
5923 0 : GcError::TimelineCancelled
5924 : } else {
5925 0 : GcError::Remote(e)
5926 : }
5927 4 : })?;
5928 :
5929 4 : let gc_layers = layers_to_remove
5930 4 : .iter()
5931 4 : .map(|x| guard.get_from_desc(x))
5932 4 : .collect::<Vec<Layer>>();
5933 4 :
5934 4 : result.layers_removed = gc_layers.len() as u64;
5935 4 :
5936 4 : self.remote_client.schedule_gc_update(&gc_layers)?;
5937 :
5938 4 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5939 4 :
5940 4 : #[cfg(feature = "testing")]
5941 4 : {
5942 4 : result.doomed_layers = gc_layers;
5943 4 : }
5944 4 : }
5945 :
5946 8 : info!(
5947 0 : "GC completed removing {} layers, cutoff {}",
5948 : result.layers_removed, new_gc_cutoff
5949 : );
5950 :
5951 8 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5952 8 : Ok(result)
5953 8 : }
5954 :
5955 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5956 1332025 : async fn reconstruct_value(
5957 1332025 : &self,
5958 1332025 : key: Key,
5959 1332025 : request_lsn: Lsn,
5960 1332025 : mut data: ValueReconstructState,
5961 1332025 : ) -> Result<Bytes, PageReconstructError> {
5962 1332025 : // Perform WAL redo if needed
5963 1332025 : data.records.reverse();
5964 1332025 :
5965 1332025 : // If we have a page image, and no WAL, we're all set
5966 1332025 : if data.records.is_empty() {
5967 1330389 : if let Some((img_lsn, img)) = &data.img {
5968 1330389 : trace!(
5969 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5970 : key,
5971 : img_lsn,
5972 : request_lsn,
5973 : );
5974 1330389 : Ok(img.clone())
5975 : } else {
5976 0 : Err(PageReconstructError::from(anyhow!(
5977 0 : "base image for {key} at {request_lsn} not found"
5978 0 : )))
5979 : }
5980 : } else {
5981 : // We need to do WAL redo.
5982 : //
5983 : // If we don't have a base image, then the oldest WAL record better initialize
5984 : // the page
5985 1636 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5986 0 : Err(PageReconstructError::from(anyhow!(
5987 0 : "Base image for {} at {} not found, but got {} WAL records",
5988 0 : key,
5989 0 : request_lsn,
5990 0 : data.records.len()
5991 0 : )))
5992 : } else {
5993 1636 : if data.img.is_some() {
5994 1504 : trace!(
5995 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5996 0 : data.records.len(),
5997 : key,
5998 : request_lsn
5999 : );
6000 : } else {
6001 132 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
6002 : };
6003 1636 : let res = self
6004 1636 : .walredo_mgr
6005 1636 : .as_ref()
6006 1636 : .context("timeline has no walredo manager")
6007 1636 : .map_err(PageReconstructError::WalRedo)?
6008 1636 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
6009 1636 : .await;
6010 1636 : let img = match res {
6011 1636 : Ok(img) => img,
6012 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
6013 0 : Err(walredo::Error::Other(err)) => {
6014 0 : critical!("walredo failure during page reconstruction: {err:?}");
6015 0 : return Err(PageReconstructError::WalRedo(
6016 0 : err.context("reconstruct a page image"),
6017 0 : ));
6018 : }
6019 : };
6020 1636 : Ok(img)
6021 : }
6022 : }
6023 1332025 : }
6024 :
6025 0 : pub(crate) async fn spawn_download_all_remote_layers(
6026 0 : self: Arc<Self>,
6027 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6028 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
6029 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6030 :
6031 : // this is not really needed anymore; it has tests which really check the return value from
6032 : // http api. it would be better not to maintain this anymore.
6033 :
6034 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
6035 0 : if let Some(st) = &*status_guard {
6036 0 : match &st.state {
6037 : DownloadRemoteLayersTaskState::Running => {
6038 0 : return Err(st.clone());
6039 : }
6040 : DownloadRemoteLayersTaskState::ShutDown
6041 0 : | DownloadRemoteLayersTaskState::Completed => {
6042 0 : *status_guard = None;
6043 0 : }
6044 : }
6045 0 : }
6046 :
6047 0 : let self_clone = Arc::clone(&self);
6048 0 : let task_id = task_mgr::spawn(
6049 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
6050 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
6051 0 : self.tenant_shard_id,
6052 0 : Some(self.timeline_id),
6053 0 : "download all remote layers task",
6054 0 : async move {
6055 0 : self_clone.download_all_remote_layers(request).await;
6056 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
6057 0 : match &mut *status_guard {
6058 : None => {
6059 0 : warn!("tasks status is supposed to be Some(), since we are running");
6060 : }
6061 0 : Some(st) => {
6062 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
6063 0 : if st.task_id != exp_task_id {
6064 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
6065 0 : } else {
6066 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6067 0 : }
6068 : }
6069 : };
6070 0 : Ok(())
6071 0 : }
6072 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))
6073 : );
6074 :
6075 0 : let initial_info = DownloadRemoteLayersTaskInfo {
6076 0 : task_id: format!("{task_id}"),
6077 0 : state: DownloadRemoteLayersTaskState::Running,
6078 0 : total_layer_count: 0,
6079 0 : successful_download_count: 0,
6080 0 : failed_download_count: 0,
6081 0 : };
6082 0 : *status_guard = Some(initial_info.clone());
6083 0 :
6084 0 : Ok(initial_info)
6085 0 : }
6086 :
6087 0 : async fn download_all_remote_layers(
6088 0 : self: &Arc<Self>,
6089 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6090 0 : ) {
6091 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6092 :
6093 0 : let remaining = {
6094 0 : let guard = self.layers.read().await;
6095 0 : let Ok(lm) = guard.layer_map() else {
6096 : // technically here we could look into iterating accessible layers, but downloading
6097 : // all layers of a shutdown timeline makes no sense regardless.
6098 0 : tracing::info!("attempted to download all layers of shutdown timeline");
6099 0 : return;
6100 : };
6101 0 : lm.iter_historic_layers()
6102 0 : .map(|desc| guard.get_from_desc(&desc))
6103 0 : .collect::<Vec<_>>()
6104 0 : };
6105 0 : let total_layer_count = remaining.len();
6106 :
6107 : macro_rules! lock_status {
6108 : ($st:ident) => {
6109 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
6110 : let st = st
6111 : .as_mut()
6112 : .expect("this function is only called after the task has been spawned");
6113 : assert_eq!(
6114 : st.task_id,
6115 : format!(
6116 : "{}",
6117 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
6118 : )
6119 : );
6120 : let $st = st;
6121 : };
6122 : }
6123 :
6124 : {
6125 0 : lock_status!(st);
6126 0 : st.total_layer_count = total_layer_count as u64;
6127 0 : }
6128 0 :
6129 0 : let mut remaining = remaining.into_iter();
6130 0 : let mut have_remaining = true;
6131 0 : let mut js = tokio::task::JoinSet::new();
6132 0 :
6133 0 : let cancel = task_mgr::shutdown_token();
6134 0 :
6135 0 : let limit = request.max_concurrent_downloads;
6136 :
6137 : loop {
6138 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
6139 0 : let Some(next) = remaining.next() else {
6140 0 : have_remaining = false;
6141 0 : break;
6142 : };
6143 :
6144 0 : let span = tracing::info_span!("download", layer = %next);
6145 :
6146 0 : js.spawn(
6147 0 : async move {
6148 0 : let res = next.download().await;
6149 0 : (next, res)
6150 0 : }
6151 0 : .instrument(span),
6152 0 : );
6153 0 : }
6154 :
6155 0 : while let Some(res) = js.join_next().await {
6156 0 : match res {
6157 : Ok((_, Ok(_))) => {
6158 0 : lock_status!(st);
6159 0 : st.successful_download_count += 1;
6160 : }
6161 0 : Ok((layer, Err(e))) => {
6162 0 : tracing::error!(%layer, "download failed: {e:#}");
6163 0 : lock_status!(st);
6164 0 : st.failed_download_count += 1;
6165 : }
6166 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
6167 0 : Err(je) if je.is_panic() => {
6168 0 : lock_status!(st);
6169 0 : st.failed_download_count += 1;
6170 : }
6171 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
6172 : }
6173 : }
6174 :
6175 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
6176 0 : break;
6177 0 : }
6178 : }
6179 :
6180 : {
6181 0 : lock_status!(st);
6182 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6183 : }
6184 0 : }
6185 :
6186 0 : pub(crate) fn get_download_all_remote_layers_task_info(
6187 0 : &self,
6188 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
6189 0 : self.download_all_remote_layers_task_info
6190 0 : .read()
6191 0 : .unwrap()
6192 0 : .clone()
6193 0 : }
6194 : }
6195 :
6196 : impl Timeline {
6197 : /// Returns non-remote layers for eviction.
6198 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
6199 0 : let guard = self.layers.read().await;
6200 0 : let mut max_layer_size: Option<u64> = None;
6201 0 :
6202 0 : let resident_layers = guard
6203 0 : .likely_resident_layers()
6204 0 : .map(|layer| {
6205 0 : let file_size = layer.layer_desc().file_size;
6206 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
6207 0 :
6208 0 : let last_activity_ts = layer.latest_activity();
6209 0 :
6210 0 : EvictionCandidate {
6211 0 : layer: layer.to_owned().into(),
6212 0 : last_activity_ts,
6213 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
6214 0 : visibility: layer.visibility(),
6215 0 : }
6216 0 : })
6217 0 : .collect();
6218 0 :
6219 0 : DiskUsageEvictionInfo {
6220 0 : max_layer_size,
6221 0 : resident_layers,
6222 0 : }
6223 0 : }
6224 :
6225 3761 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
6226 3761 : ShardIndex {
6227 3761 : shard_number: self.tenant_shard_id.shard_number,
6228 3761 : shard_count: self.tenant_shard_id.shard_count,
6229 3761 : }
6230 3761 : }
6231 :
6232 : /// Persistently blocks gc for `Manual` reason.
6233 : ///
6234 : /// Returns true if no such block existed before, false otherwise.
6235 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
6236 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6237 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6238 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
6239 0 : }
6240 :
6241 : /// Persistently unblocks gc for `Manual` reason.
6242 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
6243 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6244 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6245 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
6246 0 : }
6247 :
6248 : #[cfg(test)]
6249 97 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
6250 97 : self.last_record_lsn.advance(new_lsn);
6251 97 : }
6252 :
6253 : #[cfg(test)]
6254 8 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
6255 8 : self.disk_consistent_lsn.store(new_value);
6256 8 : }
6257 :
6258 : /// Force create an image layer and place it into the layer map.
6259 : ///
6260 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
6261 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
6262 : /// placed into the layer map in one run AND be validated.
6263 : #[cfg(test)]
6264 121 : pub(super) async fn force_create_image_layer(
6265 121 : self: &Arc<Timeline>,
6266 121 : lsn: Lsn,
6267 121 : mut images: Vec<(Key, Bytes)>,
6268 121 : check_start_lsn: Option<Lsn>,
6269 121 : ctx: &RequestContext,
6270 121 : ) -> anyhow::Result<()> {
6271 121 : let last_record_lsn = self.get_last_record_lsn();
6272 121 : assert!(
6273 121 : lsn <= last_record_lsn,
6274 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
6275 : );
6276 121 : if let Some(check_start_lsn) = check_start_lsn {
6277 121 : assert!(lsn >= check_start_lsn);
6278 0 : }
6279 348 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
6280 121 : let min_key = *images.first().map(|(k, _)| k).unwrap();
6281 121 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
6282 121 : let mut image_layer_writer = ImageLayerWriter::new(
6283 121 : self.conf,
6284 121 : self.timeline_id,
6285 121 : self.tenant_shard_id,
6286 121 : &(min_key..end_key),
6287 121 : lsn,
6288 121 : ctx,
6289 121 : )
6290 121 : .await?;
6291 590 : for (key, img) in images {
6292 469 : image_layer_writer.put_image(key, img, ctx).await?;
6293 : }
6294 121 : let (desc, path) = image_layer_writer.finish(ctx).await?;
6295 121 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6296 121 : info!("force created image layer {}", image_layer.local_path());
6297 : {
6298 121 : let mut guard = self.layers.write().await;
6299 121 : guard
6300 121 : .open_mut()
6301 121 : .unwrap()
6302 121 : .force_insert_layer(image_layer.clone());
6303 121 : }
6304 121 :
6305 121 : // Update remote_timeline_client state to reflect existence of this layer
6306 121 : self.remote_client
6307 121 : .schedule_layer_file_upload(image_layer)
6308 121 : .unwrap();
6309 121 :
6310 121 : Ok(())
6311 121 : }
6312 :
6313 : /// Force create a delta layer and place it into the layer map.
6314 : ///
6315 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
6316 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
6317 : /// placed into the layer map in one run AND be validated.
6318 : #[cfg(test)]
6319 187 : pub(super) async fn force_create_delta_layer(
6320 187 : self: &Arc<Timeline>,
6321 187 : mut deltas: DeltaLayerTestDesc,
6322 187 : check_start_lsn: Option<Lsn>,
6323 187 : ctx: &RequestContext,
6324 187 : ) -> anyhow::Result<()> {
6325 187 : let last_record_lsn = self.get_last_record_lsn();
6326 187 : deltas
6327 187 : .data
6328 264 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
6329 187 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
6330 187 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
6331 638 : for (_, lsn, _) in &deltas.data {
6332 451 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
6333 : }
6334 187 : assert!(
6335 187 : deltas.lsn_range.end <= last_record_lsn,
6336 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
6337 : deltas.lsn_range.end,
6338 : last_record_lsn
6339 : );
6340 187 : if let Some(check_start_lsn) = check_start_lsn {
6341 187 : assert!(deltas.lsn_range.start >= check_start_lsn);
6342 0 : }
6343 187 : let mut delta_layer_writer = DeltaLayerWriter::new(
6344 187 : self.conf,
6345 187 : self.timeline_id,
6346 187 : self.tenant_shard_id,
6347 187 : deltas.key_range.start,
6348 187 : deltas.lsn_range,
6349 187 : ctx,
6350 187 : )
6351 187 : .await?;
6352 638 : for (key, lsn, val) in deltas.data {
6353 451 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
6354 : }
6355 187 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
6356 187 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6357 187 : info!("force created delta layer {}", delta_layer.local_path());
6358 : {
6359 187 : let mut guard = self.layers.write().await;
6360 187 : guard
6361 187 : .open_mut()
6362 187 : .unwrap()
6363 187 : .force_insert_layer(delta_layer.clone());
6364 187 : }
6365 187 :
6366 187 : // Update remote_timeline_client state to reflect existence of this layer
6367 187 : self.remote_client
6368 187 : .schedule_layer_file_upload(delta_layer)
6369 187 : .unwrap();
6370 187 :
6371 187 : Ok(())
6372 187 : }
6373 :
6374 : /// Return all keys at the LSN in the image layers
6375 : #[cfg(test)]
6376 12 : pub(crate) async fn inspect_image_layers(
6377 12 : self: &Arc<Timeline>,
6378 12 : lsn: Lsn,
6379 12 : ctx: &RequestContext,
6380 12 : io_concurrency: IoConcurrency,
6381 12 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
6382 12 : let mut all_data = Vec::new();
6383 12 : let guard = self.layers.read().await;
6384 68 : for layer in guard.layer_map()?.iter_historic_layers() {
6385 68 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
6386 12 : let layer = guard.get_from_desc(&layer);
6387 12 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
6388 12 : layer
6389 12 : .get_values_reconstruct_data(
6390 12 : KeySpace::single(Key::MIN..Key::MAX),
6391 12 : lsn..Lsn(lsn.0 + 1),
6392 12 : &mut reconstruct_data,
6393 12 : ctx,
6394 12 : )
6395 12 : .await?;
6396 104 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
6397 104 : let v = v.collect_pending_ios().await?;
6398 104 : all_data.push((k, v.img.unwrap().1));
6399 : }
6400 56 : }
6401 : }
6402 12 : all_data.sort();
6403 12 : Ok(all_data)
6404 12 : }
6405 :
6406 : /// Get all historic layer descriptors in the layer map
6407 : #[cfg(test)]
6408 48 : pub(crate) async fn inspect_historic_layers(
6409 48 : self: &Arc<Timeline>,
6410 48 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
6411 48 : let mut layers = Vec::new();
6412 48 : let guard = self.layers.read().await;
6413 228 : for layer in guard.layer_map()?.iter_historic_layers() {
6414 228 : layers.push(layer.key());
6415 228 : }
6416 48 : Ok(layers)
6417 48 : }
6418 :
6419 : #[cfg(test)]
6420 20 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
6421 20 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
6422 20 : keyspace.merge(&ks);
6423 20 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
6424 20 : }
6425 : }
6426 :
6427 : /// Tracking writes ingestion does to a particular in-memory layer.
6428 : ///
6429 : /// Cleared upon freezing a layer.
6430 : pub(crate) struct TimelineWriterState {
6431 : open_layer: Arc<InMemoryLayer>,
6432 : current_size: u64,
6433 : // Previous Lsn which passed through
6434 : prev_lsn: Option<Lsn>,
6435 : // Largest Lsn which passed through the current writer
6436 : max_lsn: Option<Lsn>,
6437 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
6438 : cached_last_freeze_at: Lsn,
6439 : }
6440 :
6441 : impl TimelineWriterState {
6442 2597 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
6443 2597 : Self {
6444 2597 : open_layer,
6445 2597 : current_size,
6446 2597 : prev_lsn: None,
6447 2597 : max_lsn: None,
6448 2597 : cached_last_freeze_at: last_freeze_at,
6449 2597 : }
6450 2597 : }
6451 : }
6452 :
6453 : /// Various functions to mutate the timeline.
6454 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
6455 : // This is probably considered a bad practice in Rust and should be fixed eventually,
6456 : // but will cause large code changes.
6457 : pub(crate) struct TimelineWriter<'a> {
6458 : tl: &'a Timeline,
6459 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
6460 : }
6461 :
6462 : impl Deref for TimelineWriter<'_> {
6463 : type Target = Timeline;
6464 :
6465 19796401 : fn deref(&self) -> &Self::Target {
6466 19796401 : self.tl
6467 19796401 : }
6468 : }
6469 :
6470 : #[derive(PartialEq)]
6471 : enum OpenLayerAction {
6472 : Roll,
6473 : Open,
6474 : None,
6475 : }
6476 :
6477 : impl TimelineWriter<'_> {
6478 9608469 : async fn handle_open_layer_action(
6479 9608469 : &mut self,
6480 9608469 : at: Lsn,
6481 9608469 : action: OpenLayerAction,
6482 9608469 : ctx: &RequestContext,
6483 9608469 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
6484 9608469 : match action {
6485 : OpenLayerAction::Roll => {
6486 160 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
6487 160 : self.roll_layer(freeze_at).await?;
6488 160 : self.open_layer(at, ctx).await?;
6489 : }
6490 2437 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
6491 : OpenLayerAction::None => {
6492 9605872 : assert!(self.write_guard.is_some());
6493 : }
6494 : }
6495 :
6496 9608469 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
6497 9608469 : }
6498 :
6499 2597 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
6500 2597 : let layer = self
6501 2597 : .tl
6502 2597 : .get_layer_for_write(at, &self.write_guard, ctx)
6503 2597 : .await?;
6504 2597 : let initial_size = layer.size().await?;
6505 :
6506 2597 : let last_freeze_at = self.last_freeze_at.load();
6507 2597 : self.write_guard.replace(TimelineWriterState::new(
6508 2597 : layer,
6509 2597 : initial_size,
6510 2597 : last_freeze_at,
6511 2597 : ));
6512 2597 :
6513 2597 : Ok(())
6514 2597 : }
6515 :
6516 160 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
6517 160 : let current_size = self.write_guard.as_ref().unwrap().current_size;
6518 :
6519 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
6520 : // to propagate the backpressure up into WAL ingestion.
6521 160 : let l0_count = self
6522 160 : .tl
6523 160 : .layers
6524 160 : .read()
6525 160 : .await
6526 160 : .layer_map()?
6527 160 : .level0_deltas()
6528 160 : .len();
6529 160 : let wait_thresholds = [
6530 160 : self.get_l0_flush_delay_threshold(),
6531 160 : self.get_l0_flush_stall_threshold(),
6532 160 : ];
6533 160 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
6534 :
6535 : // self.write_guard will be taken by the freezing
6536 160 : let flush_id = self
6537 160 : .tl
6538 160 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
6539 160 : .await?;
6540 :
6541 160 : assert!(self.write_guard.is_none());
6542 :
6543 160 : if let Some(wait_threshold) = wait_threshold {
6544 0 : if l0_count >= wait_threshold {
6545 0 : info!("layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers");
6546 0 : self.tl.wait_flush_completion(flush_id).await?;
6547 0 : }
6548 160 : }
6549 :
6550 160 : if current_size >= self.get_checkpoint_distance() * 2 {
6551 0 : warn!("Flushed oversized open layer with size {}", current_size)
6552 160 : }
6553 :
6554 160 : Ok(())
6555 160 : }
6556 :
6557 9608469 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
6558 9608469 : let state = &*self.write_guard;
6559 9608469 : let Some(state) = &state else {
6560 2437 : return OpenLayerAction::Open;
6561 : };
6562 :
6563 : #[cfg(feature = "testing")]
6564 9606032 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
6565 : // this check and assertion are not really needed because
6566 : // LayerManager::try_freeze_in_memory_layer will always clear out the
6567 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
6568 : // is no TimelineWriterState.
6569 0 : assert!(
6570 0 : state.open_layer.end_lsn.get().is_some(),
6571 0 : "our open_layer must be outdated"
6572 : );
6573 :
6574 : // this would be a memory leak waiting to happen because the in-memory layer always has
6575 : // an index
6576 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
6577 9606032 : }
6578 9606032 :
6579 9606032 : if state.prev_lsn == Some(lsn) {
6580 : // Rolling mid LSN is not supported by [downstream code].
6581 : // Hence, only roll at LSN boundaries.
6582 : //
6583 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
6584 12 : return OpenLayerAction::None;
6585 9606020 : }
6586 9606020 :
6587 9606020 : if state.current_size == 0 {
6588 : // Don't roll empty layers
6589 0 : return OpenLayerAction::None;
6590 9606020 : }
6591 9606020 :
6592 9606020 : if self.tl.should_roll(
6593 9606020 : state.current_size,
6594 9606020 : state.current_size + new_value_size,
6595 9606020 : self.get_checkpoint_distance(),
6596 9606020 : lsn,
6597 9606020 : state.cached_last_freeze_at,
6598 9606020 : state.open_layer.get_opened_at(),
6599 9606020 : ) {
6600 160 : OpenLayerAction::Roll
6601 : } else {
6602 9605860 : OpenLayerAction::None
6603 : }
6604 9608469 : }
6605 :
6606 : /// Put a batch of keys at the specified Lsns.
6607 9608465 : pub(crate) async fn put_batch(
6608 9608465 : &mut self,
6609 9608465 : batch: SerializedValueBatch,
6610 9608465 : ctx: &RequestContext,
6611 9608465 : ) -> anyhow::Result<()> {
6612 9608465 : if !batch.has_data() {
6613 0 : return Ok(());
6614 9608465 : }
6615 9608465 :
6616 9608465 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
6617 9608465 : // We don't assert this in release builds, since key ownership policies may change over
6618 9608465 : // time. Stray keys will be removed during compaction.
6619 9608465 : if cfg!(debug_assertions) {
6620 19789832 : for metadata in &batch.metadata {
6621 10181367 : if let ValueMeta::Serialized(metadata) = metadata {
6622 10181367 : let key = Key::from_compact(metadata.key);
6623 10181367 : assert!(
6624 10181367 : self.shard_identity.is_key_local(&key)
6625 0 : || self.shard_identity.is_key_global(&key),
6626 0 : "key {key} does not belong on shard {}",
6627 0 : self.shard_identity.shard_index()
6628 : );
6629 0 : }
6630 : }
6631 0 : }
6632 :
6633 9608465 : let batch_max_lsn = batch.max_lsn;
6634 9608465 : let buf_size: u64 = batch.buffer_size() as u64;
6635 9608465 :
6636 9608465 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
6637 9608465 : let layer = self
6638 9608465 : .handle_open_layer_action(batch_max_lsn, action, ctx)
6639 9608465 : .await?;
6640 :
6641 9608465 : let res = layer.put_batch(batch, ctx).await;
6642 :
6643 9608465 : if res.is_ok() {
6644 9608465 : // Update the current size only when the entire write was ok.
6645 9608465 : // In case of failures, we may have had partial writes which
6646 9608465 : // render the size tracking out of sync. That's ok because
6647 9608465 : // the checkpoint distance should be significantly smaller
6648 9608465 : // than the S3 single shot upload limit of 5GiB.
6649 9608465 : let state = self.write_guard.as_mut().unwrap();
6650 9608465 :
6651 9608465 : state.current_size += buf_size;
6652 9608465 : state.prev_lsn = Some(batch_max_lsn);
6653 9608465 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
6654 9608465 : }
6655 :
6656 9608465 : res
6657 9608465 : }
6658 :
6659 : #[cfg(test)]
6660 : /// Test helper, for tests that would like to poke individual values without composing a batch
6661 8780308 : pub(crate) async fn put(
6662 8780308 : &mut self,
6663 8780308 : key: Key,
6664 8780308 : lsn: Lsn,
6665 8780308 : value: &Value,
6666 8780308 : ctx: &RequestContext,
6667 8780308 : ) -> anyhow::Result<()> {
6668 : use utils::bin_ser::BeSer;
6669 8780308 : if !key.is_valid_key_on_write_path() {
6670 0 : bail!(
6671 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
6672 0 : key
6673 0 : );
6674 8780308 : }
6675 8780308 : let val_ser_size = value.serialized_size().unwrap() as usize;
6676 8780308 : let batch = SerializedValueBatch::from_values(vec![(
6677 8780308 : key.to_compact(),
6678 8780308 : lsn,
6679 8780308 : val_ser_size,
6680 8780308 : value.clone(),
6681 8780308 : )]);
6682 8780308 :
6683 8780308 : self.put_batch(batch, ctx).await
6684 8780308 : }
6685 :
6686 4 : pub(crate) async fn delete_batch(
6687 4 : &mut self,
6688 4 : batch: &[(Range<Key>, Lsn)],
6689 4 : ctx: &RequestContext,
6690 4 : ) -> anyhow::Result<()> {
6691 4 : if let Some((_, lsn)) = batch.first() {
6692 4 : let action = self.get_open_layer_action(*lsn, 0);
6693 4 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
6694 4 : layer.put_tombstones(batch).await?;
6695 0 : }
6696 :
6697 4 : Ok(())
6698 4 : }
6699 :
6700 : /// Track the end of the latest digested WAL record.
6701 : /// Remember the (end of) last valid WAL record remembered in the timeline.
6702 : ///
6703 : /// Call this after you have finished writing all the WAL up to 'lsn'.
6704 : ///
6705 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
6706 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
6707 : /// the latest and can be read.
6708 10558189 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
6709 10558189 : self.tl.finish_write(new_lsn);
6710 10558189 : }
6711 :
6712 541140 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
6713 541140 : self.tl.update_current_logical_size(delta)
6714 541140 : }
6715 : }
6716 :
6717 : // We need TimelineWriter to be send in upcoming conversion of
6718 : // Timeline::layers to tokio::sync::RwLock.
6719 : #[test]
6720 4 : fn is_send() {
6721 4 : fn _assert_send<T: Send>() {}
6722 4 : _assert_send::<TimelineWriter<'_>>();
6723 4 : }
6724 :
6725 : #[cfg(test)]
6726 : mod tests {
6727 : use std::sync::Arc;
6728 :
6729 : use pageserver_api::key::Key;
6730 : use pageserver_api::value::Value;
6731 : use tracing::Instrument;
6732 : use utils::{id::TimelineId, lsn::Lsn};
6733 :
6734 : use crate::tenant::{
6735 : harness::{test_img, TenantHarness},
6736 : layer_map::LayerMap,
6737 : storage_layer::{Layer, LayerName, LayerVisibilityHint},
6738 : timeline::{DeltaLayerTestDesc, EvictionError},
6739 : PreviousHeatmap, Timeline,
6740 : };
6741 :
6742 : use super::HeatMapTimeline;
6743 :
6744 22 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
6745 22 : assert_eq!(lhs.layers.len(), rhs.layers.len());
6746 21 : let lhs_rhs = lhs.layers.iter().zip(rhs.layers.iter());
6747 105 : for (l, r) in lhs_rhs {
6748 84 : assert_eq!(l.name, r.name);
6749 84 : assert_eq!(l.metadata, r.metadata);
6750 : }
6751 21 : }
6752 :
6753 : #[tokio::test]
6754 5 : async fn test_heatmap_generation() {
6755 5 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
6756 5 :
6757 5 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6758 5 : Lsn(0x10)..Lsn(0x20),
6759 5 : vec![(
6760 5 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6761 5 : Lsn(0x11),
6762 5 : Value::Image(test_img("foo")),
6763 5 : )],
6764 5 : );
6765 5 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6766 5 : Lsn(0x10)..Lsn(0x20),
6767 5 : vec![(
6768 5 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6769 5 : Lsn(0x11),
6770 5 : Value::Image(test_img("foo")),
6771 5 : )],
6772 5 : );
6773 5 : let l0_delta = DeltaLayerTestDesc::new(
6774 5 : Lsn(0x20)..Lsn(0x30),
6775 5 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6776 5 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6777 5 : vec![(
6778 5 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6779 5 : Lsn(0x25),
6780 5 : Value::Image(test_img("foo")),
6781 5 : )],
6782 5 : );
6783 5 : let delta_layers = vec![
6784 5 : covered_delta.clone(),
6785 5 : visible_delta.clone(),
6786 5 : l0_delta.clone(),
6787 5 : ];
6788 5 :
6789 5 : let image_layer = (
6790 5 : Lsn(0x40),
6791 5 : vec![(
6792 5 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6793 5 : test_img("bar"),
6794 5 : )],
6795 5 : );
6796 5 : let image_layers = vec![image_layer];
6797 5 :
6798 5 : let (tenant, ctx) = harness.load().await;
6799 5 : let timeline = tenant
6800 5 : .create_test_timeline_with_layers(
6801 5 : TimelineId::generate(),
6802 5 : Lsn(0x10),
6803 5 : 14,
6804 5 : &ctx,
6805 5 : delta_layers,
6806 5 : image_layers,
6807 5 : Lsn(0x100),
6808 5 : )
6809 5 : .await
6810 5 : .unwrap();
6811 5 :
6812 5 : // Layer visibility is an input to heatmap generation, so refresh it first
6813 5 : timeline.update_layer_visibility().await.unwrap();
6814 5 :
6815 5 : let heatmap = timeline
6816 5 : .generate_heatmap()
6817 5 : .await
6818 5 : .expect("Infallible while timeline is not shut down");
6819 5 :
6820 5 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
6821 5 :
6822 5 : // L0 should come last
6823 5 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
6824 5 :
6825 5 : let mut last_lsn = Lsn::MAX;
6826 25 : for layer in &heatmap.layers {
6827 5 : // Covered layer should be omitted
6828 20 : assert!(layer.name != covered_delta.layer_name());
6829 5 :
6830 20 : let layer_lsn = match &layer.name {
6831 10 : LayerName::Delta(d) => d.lsn_range.end,
6832 10 : LayerName::Image(i) => i.lsn,
6833 5 : };
6834 5 :
6835 5 : // Apart from L0s, newest Layers should come first
6836 20 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
6837 15 : assert!(layer_lsn <= last_lsn);
6838 15 : last_lsn = layer_lsn;
6839 5 : }
6840 5 : }
6841 5 :
6842 5 : // Evict all the layers and stash the old heatmap in the timeline.
6843 5 : // This simulates a migration to a cold secondary location.
6844 5 :
6845 5 : let guard = timeline.layers.read().await;
6846 5 : let mut all_layers = Vec::new();
6847 5 : let forever = std::time::Duration::from_secs(120);
6848 25 : for layer in guard.likely_resident_layers() {
6849 25 : all_layers.push(layer.clone());
6850 25 : layer.evict_and_wait(forever).await.unwrap();
6851 5 : }
6852 5 : drop(guard);
6853 5 :
6854 5 : timeline
6855 5 : .previous_heatmap
6856 5 : .store(Some(Arc::new(PreviousHeatmap::Active {
6857 5 : heatmap: heatmap.clone(),
6858 5 : read_at: std::time::Instant::now(),
6859 5 : })));
6860 5 :
6861 5 : // Generate a new heatmap and assert that it contains the same layers as the old one.
6862 5 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
6863 5 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
6864 5 :
6865 5 : // Download each layer one by one. Generate the heatmap at each step and check
6866 5 : // that it's stable.
6867 26 : for layer in all_layers {
6868 22 : if layer.visibility() == LayerVisibilityHint::Covered {
6869 5 : continue;
6870 18 : }
6871 18 :
6872 18 : eprintln!("Downloading {layer} and re-generating heatmap");
6873 5 :
6874 18 : let _resident = layer
6875 18 : .download_and_keep_resident()
6876 18 : .instrument(tracing::info_span!(
6877 17 : parent: None,
6878 5 : "download_layer",
6879 5 : tenant_id = %timeline.tenant_shard_id.tenant_id,
6880 0 : shard_id = %timeline.tenant_shard_id.shard_slug(),
6881 0 : timeline_id = %timeline.timeline_id
6882 5 : ))
6883 17 : .await
6884 17 : .unwrap();
6885 5 :
6886 17 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
6887 17 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
6888 5 : }
6889 5 :
6890 5 : // Everything from the post-migration heatmap is now resident.
6891 5 : // Check that we drop it from memory.
6892 5 : assert!(matches!(
6893 5 : timeline.previous_heatmap.load().as_deref(),
6894 5 : Some(PreviousHeatmap::Obsolete)
6895 5 : ));
6896 5 : }
6897 :
6898 : #[tokio::test]
6899 4 : async fn test_previous_heatmap_obsoletion() {
6900 4 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
6901 4 : .await
6902 4 : .unwrap();
6903 4 :
6904 4 : let l0_delta = DeltaLayerTestDesc::new(
6905 4 : Lsn(0x20)..Lsn(0x30),
6906 4 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6907 4 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6908 4 : vec![(
6909 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6910 4 : Lsn(0x25),
6911 4 : Value::Image(test_img("foo")),
6912 4 : )],
6913 4 : );
6914 4 :
6915 4 : let image_layer = (
6916 4 : Lsn(0x40),
6917 4 : vec![(
6918 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6919 4 : test_img("bar"),
6920 4 : )],
6921 4 : );
6922 4 :
6923 4 : let delta_layers = vec![l0_delta];
6924 4 : let image_layers = vec![image_layer];
6925 4 :
6926 4 : let (tenant, ctx) = harness.load().await;
6927 4 : let timeline = tenant
6928 4 : .create_test_timeline_with_layers(
6929 4 : TimelineId::generate(),
6930 4 : Lsn(0x10),
6931 4 : 14,
6932 4 : &ctx,
6933 4 : delta_layers,
6934 4 : image_layers,
6935 4 : Lsn(0x100),
6936 4 : )
6937 4 : .await
6938 4 : .unwrap();
6939 4 :
6940 4 : // Layer visibility is an input to heatmap generation, so refresh it first
6941 4 : timeline.update_layer_visibility().await.unwrap();
6942 4 :
6943 4 : let heatmap = timeline
6944 4 : .generate_heatmap()
6945 4 : .await
6946 4 : .expect("Infallible while timeline is not shut down");
6947 4 :
6948 4 : // Both layers should be in the heatmap
6949 4 : assert!(!heatmap.layers.is_empty());
6950 4 :
6951 4 : // Now simulate a migration.
6952 4 : timeline
6953 4 : .previous_heatmap
6954 4 : .store(Some(Arc::new(PreviousHeatmap::Active {
6955 4 : heatmap: heatmap.clone(),
6956 4 : read_at: std::time::Instant::now(),
6957 4 : })));
6958 4 :
6959 4 : // Evict all the layers in the previous heatmap
6960 4 : let guard = timeline.layers.read().await;
6961 4 : let forever = std::time::Duration::from_secs(120);
6962 12 : for layer in guard.likely_resident_layers() {
6963 12 : layer.evict_and_wait(forever).await.unwrap();
6964 4 : }
6965 4 : drop(guard);
6966 4 :
6967 4 : // Generate a new heatmap and check that the previous heatmap
6968 4 : // has been marked obsolete.
6969 4 : let post_eviction_heatmap = timeline
6970 4 : .generate_heatmap()
6971 4 : .await
6972 4 : .expect("Infallible while timeline is not shut down");
6973 4 :
6974 4 : assert!(post_eviction_heatmap.layers.is_empty());
6975 4 : assert!(matches!(
6976 4 : timeline.previous_heatmap.load().as_deref(),
6977 4 : Some(PreviousHeatmap::Obsolete)
6978 4 : ));
6979 4 : }
6980 :
6981 : #[tokio::test]
6982 4 : async fn two_layer_eviction_attempts_at_the_same_time() {
6983 4 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6984 4 : .await
6985 4 : .unwrap();
6986 4 :
6987 4 : let (tenant, ctx) = harness.load().await;
6988 4 : let timeline = tenant
6989 4 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6990 4 : .await
6991 4 : .unwrap();
6992 4 :
6993 4 : let layer = find_some_layer(&timeline).await;
6994 4 : let layer = layer
6995 4 : .keep_resident()
6996 4 : .await
6997 4 : .expect("no download => no downloading errors")
6998 4 : .drop_eviction_guard();
6999 4 :
7000 4 : let forever = std::time::Duration::from_secs(120);
7001 4 :
7002 4 : let first = layer.evict_and_wait(forever);
7003 4 : let second = layer.evict_and_wait(forever);
7004 4 :
7005 4 : let (first, second) = tokio::join!(first, second);
7006 4 :
7007 4 : let res = layer.keep_resident().await;
7008 4 : assert!(res.is_none(), "{res:?}");
7009 4 :
7010 4 : match (first, second) {
7011 4 : (Ok(()), Ok(())) => {
7012 4 : // because there are no more timeline locks being taken on eviction path, we can
7013 4 : // witness all three outcomes here.
7014 4 : }
7015 4 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
7016 0 : // if one completes before the other, this is fine just as well.
7017 0 : }
7018 4 : other => unreachable!("unexpected {:?}", other),
7019 4 : }
7020 4 : }
7021 :
7022 4 : async fn find_some_layer(timeline: &Timeline) -> Layer {
7023 4 : let layers = timeline.layers.read().await;
7024 4 : let desc = layers
7025 4 : .layer_map()
7026 4 : .unwrap()
7027 4 : .iter_historic_layers()
7028 4 : .next()
7029 4 : .expect("must find one layer to evict");
7030 4 :
7031 4 : layers.get_from_desc(&desc)
7032 4 : }
7033 : }
|
