Line data Source code
1 : pub(crate) mod analysis;
2 : pub(crate) mod compaction;
3 : pub mod delete;
4 : pub(crate) mod detach_ancestor;
5 : mod eviction_task;
6 : pub(crate) mod handle;
7 : pub(crate) mod import_pgdata;
8 : mod init;
9 : pub mod layer_manager;
10 : pub(crate) mod logical_size;
11 : pub mod offload;
12 : pub mod span;
13 : pub mod uninit;
14 : mod walreceiver;
15 :
16 : use anyhow::{anyhow, bail, ensure, Context, Result};
17 : use arc_swap::{ArcSwap, ArcSwapOption};
18 : use bytes::Bytes;
19 : use camino::Utf8Path;
20 : use chrono::{DateTime, Utc};
21 : use enumset::EnumSet;
22 : use fail::fail_point;
23 : use futures::{stream::FuturesUnordered, StreamExt};
24 : use handle::ShardTimelineId;
25 : use layer_manager::Shutdown;
26 : use offload::OffloadError;
27 : use once_cell::sync::Lazy;
28 : use pageserver_api::models::PageTraceEvent;
29 : use pageserver_api::{
30 : key::{
31 : KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
32 : SPARSE_RANGE,
33 : },
34 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
35 : models::{
36 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
37 : DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy,
38 : InMemoryLayerInfo, LayerMapInfo, LsnLease, TimelineState,
39 : },
40 : reltag::BlockNumber,
41 : shard::{ShardIdentity, ShardNumber, TenantShardId},
42 : };
43 : use rand::Rng;
44 : use remote_storage::DownloadError;
45 : use serde_with::serde_as;
46 : use storage_broker::BrokerClientChannel;
47 : use tokio::sync::mpsc::Sender;
48 : use tokio::{
49 : runtime::Handle,
50 : sync::{oneshot, watch},
51 : };
52 : use tokio_util::sync::CancellationToken;
53 : use tracing::*;
54 : use utils::rate_limit::RateLimit;
55 : use utils::{
56 : fs_ext,
57 : guard_arc_swap::GuardArcSwap,
58 : pausable_failpoint,
59 : postgres_client::PostgresClientProtocol,
60 : sync::gate::{Gate, GateGuard},
61 : };
62 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
63 :
64 : use std::array;
65 : use std::cmp::{max, min};
66 : use std::collections::btree_map::Entry;
67 : use std::collections::{BTreeMap, HashMap, HashSet};
68 : use std::ops::{ControlFlow, Deref, Range};
69 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering as AtomicOrdering};
70 : use std::sync::{Arc, Mutex, OnceLock, RwLock, Weak};
71 : use std::time::{Duration, Instant, SystemTime};
72 :
73 : use crate::l0_flush::{self, L0FlushGlobalState};
74 : use crate::tenant::storage_layer::ImageLayerName;
75 : use crate::{
76 : aux_file::AuxFileSizeEstimator,
77 : page_service::TenantManagerTypes,
78 : tenant::{
79 : config::AttachmentMode,
80 : layer_map::{LayerMap, SearchResult},
81 : metadata::TimelineMetadata,
82 : storage_layer::{
83 : inmemory_layer::IndexEntry, BatchLayerWriter, IoConcurrency, PersistentLayerDesc,
84 : ValueReconstructSituation,
85 : },
86 : },
87 : walingest::WalLagCooldown,
88 : walredo,
89 : };
90 : use crate::{
91 : context::{DownloadBehavior, RequestContext},
92 : disk_usage_eviction_task::DiskUsageEvictionInfo,
93 : pgdatadir_mapping::CollectKeySpaceError,
94 : };
95 : use crate::{
96 : disk_usage_eviction_task::finite_f32,
97 : tenant::storage_layer::{
98 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
99 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
100 : ValuesReconstructState,
101 : },
102 : };
103 : use crate::{
104 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
105 : };
106 : use crate::{
107 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
108 : };
109 : use crate::{
110 : pgdatadir_mapping::DirectoryKind,
111 : virtual_file::{MaybeFatalIo, VirtualFile},
112 : };
113 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
114 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
115 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
116 :
117 : use crate::config::PageServerConf;
118 : use crate::keyspace::{KeyPartitioning, KeySpace};
119 : use crate::metrics::{TimelineMetrics, DELTAS_PER_READ_GLOBAL, LAYERS_PER_READ_GLOBAL};
120 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
121 : use crate::tenant::config::TenantConfOpt;
122 : use pageserver_api::reltag::RelTag;
123 : use pageserver_api::shard::ShardIndex;
124 :
125 : use postgres_connection::PgConnectionConfig;
126 : use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
127 : use utils::{
128 : completion,
129 : generation::Generation,
130 : id::TimelineId,
131 : lsn::{AtomicLsn, Lsn, RecordLsn},
132 : seqwait::SeqWait,
133 : simple_rcu::{Rcu, RcuReadGuard},
134 : };
135 :
136 : use crate::task_mgr;
137 : use crate::task_mgr::TaskKind;
138 : use crate::tenant::gc_result::GcResult;
139 : use crate::ZERO_PAGE;
140 : use pageserver_api::key::Key;
141 :
142 : use self::delete::DeleteTimelineFlow;
143 : pub(super) use self::eviction_task::EvictionTaskTenantState;
144 : use self::eviction_task::EvictionTaskTimelineState;
145 : use self::layer_manager::LayerManager;
146 : use self::logical_size::LogicalSize;
147 : use self::walreceiver::{WalReceiver, WalReceiverConf};
148 :
149 : use super::{
150 : config::TenantConf, storage_layer::LayerVisibilityHint, upload_queue::NotInitialized,
151 : MaybeOffloaded,
152 : };
153 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
154 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
155 : use super::{
156 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
157 : storage_layer::ReadableLayer,
158 : };
159 : use super::{
160 : secondary::heatmap::{HeatMapLayer, HeatMapTimeline},
161 : GcError,
162 : };
163 :
164 : #[cfg(test)]
165 : use pageserver_api::value::Value;
166 :
167 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
168 : pub(crate) enum FlushLoopState {
169 : NotStarted,
170 : Running {
171 : #[cfg(test)]
172 : expect_initdb_optimization: bool,
173 : #[cfg(test)]
174 : initdb_optimization_count: usize,
175 : },
176 : Exited,
177 : }
178 :
179 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
180 : pub enum ImageLayerCreationMode {
181 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
182 : Try,
183 : /// Force creating the image layers if possible. For now, no image layers will be created
184 : /// for metadata keys. Used in compaction code path with force flag enabled.
185 : Force,
186 : /// Initial ingestion of the data, and no data should be dropped in this function. This
187 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
188 : /// code path.
189 : Initial,
190 : }
191 :
192 : impl std::fmt::Display for ImageLayerCreationMode {
193 1136 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
194 1136 : write!(f, "{:?}", self)
195 1136 : }
196 : }
197 :
198 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
199 : /// Can be removed after all refactors are done.
200 56 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
201 56 : drop(rlock)
202 56 : }
203 :
204 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
205 : /// Can be removed after all refactors are done.
206 1192 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
207 1192 : drop(rlock)
208 1192 : }
209 :
210 : /// The outward-facing resources required to build a Timeline
211 : pub struct TimelineResources {
212 : pub remote_client: RemoteTimelineClient,
213 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
214 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
215 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
216 : }
217 :
218 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
219 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
220 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
221 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
222 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
223 : pub(crate) struct RelSizeCache {
224 : pub(crate) complete_as_of: Lsn,
225 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
226 : }
227 :
228 : pub struct Timeline {
229 : pub(crate) conf: &'static PageServerConf,
230 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
231 :
232 : myself: Weak<Self>,
233 :
234 : pub(crate) tenant_shard_id: TenantShardId,
235 : pub timeline_id: TimelineId,
236 :
237 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
238 : /// Never changes for the lifetime of this [`Timeline`] object.
239 : ///
240 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
241 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
242 : pub(crate) generation: Generation,
243 :
244 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
245 : /// to shards, and is constant through the lifetime of this Timeline.
246 : shard_identity: ShardIdentity,
247 :
248 : pub pg_version: u32,
249 :
250 : /// The tuple has two elements.
251 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
252 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
253 : ///
254 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
255 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
256 : ///
257 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
258 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
259 : /// `PersistentLayerDesc`'s.
260 : ///
261 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
262 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
263 : /// runtime, e.g., during page reconstruction.
264 : ///
265 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
266 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
267 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
268 :
269 : last_freeze_at: AtomicLsn,
270 : // Atomic would be more appropriate here.
271 : last_freeze_ts: RwLock<Instant>,
272 :
273 : pub(crate) standby_horizon: AtomicLsn,
274 :
275 : // WAL redo manager. `None` only for broken tenants.
276 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
277 :
278 : /// Remote storage client.
279 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
280 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
281 :
282 : // What page versions do we hold in the repository? If we get a
283 : // request > last_record_lsn, we need to wait until we receive all
284 : // the WAL up to the request. The SeqWait provides functions for
285 : // that. TODO: If we get a request for an old LSN, such that the
286 : // versions have already been garbage collected away, we should
287 : // throw an error, but we don't track that currently.
288 : //
289 : // last_record_lsn.load().last points to the end of last processed WAL record.
290 : //
291 : // We also remember the starting point of the previous record in
292 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
293 : // first WAL record when the node is started up. But here, we just
294 : // keep track of it.
295 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
296 :
297 : // All WAL records have been processed and stored durably on files on
298 : // local disk, up to this LSN. On crash and restart, we need to re-process
299 : // the WAL starting from this point.
300 : //
301 : // Some later WAL records might have been processed and also flushed to disk
302 : // already, so don't be surprised to see some, but there's no guarantee on
303 : // them yet.
304 : disk_consistent_lsn: AtomicLsn,
305 :
306 : // Parent timeline that this timeline was branched from, and the LSN
307 : // of the branch point.
308 : ancestor_timeline: Option<Arc<Timeline>>,
309 : ancestor_lsn: Lsn,
310 :
311 : pub(super) metrics: TimelineMetrics,
312 :
313 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
314 : // in `crate::page_service` writes these metrics.
315 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
316 :
317 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
318 :
319 : /// Ensures layers aren't frozen by checkpointer between
320 : /// [`Timeline::get_layer_for_write`] and layer reads.
321 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
322 : /// Must always be acquired before the layer map/individual layer lock
323 : /// to avoid deadlock.
324 : ///
325 : /// The state is cleared upon freezing.
326 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
327 :
328 : /// Used to avoid multiple `flush_loop` tasks running
329 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
330 :
331 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
332 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
333 : /// The flush cycle counter is sent back on the layer_flush_done channel when
334 : /// the flush finishes. You can use that to wait for the flush to finish.
335 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
336 : /// read by whoever sends an update
337 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
338 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
339 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
340 :
341 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
342 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
343 :
344 : // List of child timelines and their branch points. This is needed to avoid
345 : // garbage collecting data that is still needed by the child timelines.
346 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
347 :
348 : // It may change across major versions so for simplicity
349 : // keep it after running initdb for a timeline.
350 : // It is needed in checks when we want to error on some operations
351 : // when they are requested for pre-initdb lsn.
352 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
353 : // though let's keep them both for better error visibility.
354 : pub initdb_lsn: Lsn,
355 :
356 : /// The repartitioning result. Allows a single writer and multiple readers.
357 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
358 :
359 : /// Configuration: how often should the partitioning be recalculated.
360 : repartition_threshold: u64,
361 :
362 : last_image_layer_creation_check_at: AtomicLsn,
363 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
364 :
365 : /// Current logical size of the "datadir", at the last LSN.
366 : current_logical_size: LogicalSize,
367 :
368 : /// Information about the last processed message by the WAL receiver,
369 : /// or None if WAL receiver has not received anything for this timeline
370 : /// yet.
371 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
372 : pub walreceiver: Mutex<Option<WalReceiver>>,
373 :
374 : /// Relation size cache
375 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
376 :
377 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
378 :
379 : state: watch::Sender<TimelineState>,
380 :
381 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
382 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
383 : pub delete_progress: TimelineDeleteProgress,
384 :
385 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
386 :
387 : /// Load or creation time information about the disk_consistent_lsn and when the loading
388 : /// happened. Used for consumption metrics.
389 : pub(crate) loaded_at: (Lsn, SystemTime),
390 :
391 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
392 : pub(crate) gate: Gate,
393 :
394 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
395 : /// to the timeline should drop out when this token fires.
396 : pub(crate) cancel: CancellationToken,
397 :
398 : /// Make sure we only have one running compaction at a time in tests.
399 : ///
400 : /// Must only be taken in two places:
401 : /// - [`Timeline::compact`] (this file)
402 : /// - [`delete::delete_local_timeline_directory`]
403 : ///
404 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
405 : compaction_lock: tokio::sync::Mutex<()>,
406 :
407 : /// If true, the last compaction failed.
408 : compaction_failed: AtomicBool,
409 :
410 : /// Make sure we only have one running gc at a time.
411 : ///
412 : /// Must only be taken in two places:
413 : /// - [`Timeline::gc`] (this file)
414 : /// - [`delete::delete_local_timeline_directory`]
415 : ///
416 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
417 : gc_lock: tokio::sync::Mutex<()>,
418 :
419 : /// Cloned from [`super::Tenant::pagestream_throttle`] on construction.
420 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
421 :
422 : /// Size estimator for aux file v2
423 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
424 :
425 : /// Some test cases directly place keys into the timeline without actually modifying the directory
426 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
427 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
428 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
429 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
430 : #[cfg(test)]
431 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
432 :
433 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
434 :
435 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
436 :
437 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
438 :
439 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
440 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
441 :
442 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
443 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
444 : }
445 :
446 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
447 :
448 : pub struct WalReceiverInfo {
449 : pub wal_source_connconf: PgConnectionConfig,
450 : pub last_received_msg_lsn: Lsn,
451 : pub last_received_msg_ts: u128,
452 : }
453 :
454 : /// Information about how much history needs to be retained, needed by
455 : /// Garbage Collection.
456 : #[derive(Default)]
457 : pub(crate) struct GcInfo {
458 : /// Specific LSNs that are needed.
459 : ///
460 : /// Currently, this includes all points where child branches have
461 : /// been forked off from. In the future, could also include
462 : /// explicit user-defined snapshot points.
463 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
464 :
465 : /// The cutoff coordinates, which are combined by selecting the minimum.
466 : pub(crate) cutoffs: GcCutoffs,
467 :
468 : /// Leases granted to particular LSNs.
469 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
470 :
471 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
472 : pub(crate) within_ancestor_pitr: bool,
473 : }
474 :
475 : impl GcInfo {
476 564 : pub(crate) fn min_cutoff(&self) -> Lsn {
477 564 : self.cutoffs.select_min()
478 564 : }
479 :
480 464 : pub(super) fn insert_child(
481 464 : &mut self,
482 464 : child_id: TimelineId,
483 464 : child_lsn: Lsn,
484 464 : is_offloaded: MaybeOffloaded,
485 464 : ) {
486 464 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
487 464 : self.retain_lsns.sort_by_key(|i| i.0);
488 464 : }
489 :
490 8 : pub(super) fn remove_child_maybe_offloaded(
491 8 : &mut self,
492 8 : child_id: TimelineId,
493 8 : maybe_offloaded: MaybeOffloaded,
494 8 : ) -> bool {
495 8 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
496 8 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
497 8 : let mut removed = false;
498 12 : self.retain_lsns.retain(|i| {
499 12 : if removed {
500 4 : return true;
501 8 : }
502 8 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
503 8 : removed |= remove;
504 8 : !remove
505 12 : });
506 8 : removed
507 8 : }
508 :
509 8 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
510 8 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
511 8 : }
512 :
513 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
514 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
515 0 : }
516 : }
517 :
518 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
519 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
520 : /// between time-based and space-based retention for observability and consumption metrics purposes.
521 : #[derive(Debug, Clone)]
522 : pub(crate) struct GcCutoffs {
523 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
524 : /// history we must keep to retain a specified number of bytes of WAL.
525 : pub(crate) space: Lsn,
526 :
527 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
528 : /// history we must keep to enable reading back at least the PITR interval duration.
529 : pub(crate) time: Lsn,
530 : }
531 :
532 : impl Default for GcCutoffs {
533 892 : fn default() -> Self {
534 892 : Self {
535 892 : space: Lsn::INVALID,
536 892 : time: Lsn::INVALID,
537 892 : }
538 892 : }
539 : }
540 :
541 : impl GcCutoffs {
542 564 : fn select_min(&self) -> Lsn {
543 564 : std::cmp::min(self.space, self.time)
544 564 : }
545 : }
546 :
547 : pub(crate) struct TimelineVisitOutcome {
548 : completed_keyspace: KeySpace,
549 : image_covered_keyspace: KeySpace,
550 : }
551 :
552 : /// An error happened in a get() operation.
553 : #[derive(thiserror::Error, Debug)]
554 : pub(crate) enum PageReconstructError {
555 : #[error(transparent)]
556 : Other(anyhow::Error),
557 :
558 : #[error("Ancestor LSN wait error: {0}")]
559 : AncestorLsnTimeout(WaitLsnError),
560 :
561 : #[error("timeline shutting down")]
562 : Cancelled,
563 :
564 : /// An error happened replaying WAL records
565 : #[error(transparent)]
566 : WalRedo(anyhow::Error),
567 :
568 : #[error("{0}")]
569 : MissingKey(MissingKeyError),
570 : }
571 :
572 : impl From<anyhow::Error> for PageReconstructError {
573 0 : fn from(value: anyhow::Error) -> Self {
574 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
575 0 : match value.downcast::<PageReconstructError>() {
576 0 : Ok(pre) => pre,
577 0 : Err(other) => PageReconstructError::Other(other),
578 : }
579 0 : }
580 : }
581 :
582 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
583 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
584 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
585 0 : }
586 : }
587 :
588 : impl From<layer_manager::Shutdown> for PageReconstructError {
589 0 : fn from(_: layer_manager::Shutdown) -> Self {
590 0 : PageReconstructError::Cancelled
591 0 : }
592 : }
593 :
594 : impl GetVectoredError {
595 : #[cfg(test)]
596 12 : pub(crate) fn is_missing_key_error(&self) -> bool {
597 12 : matches!(self, Self::MissingKey(_))
598 12 : }
599 : }
600 :
601 : impl From<layer_manager::Shutdown> for GetVectoredError {
602 0 : fn from(_: layer_manager::Shutdown) -> Self {
603 0 : GetVectoredError::Cancelled
604 0 : }
605 : }
606 :
607 : #[derive(thiserror::Error)]
608 : pub struct MissingKeyError {
609 : key: Key,
610 : shard: ShardNumber,
611 : cont_lsn: Lsn,
612 : request_lsn: Lsn,
613 : ancestor_lsn: Option<Lsn>,
614 : backtrace: Option<std::backtrace::Backtrace>,
615 : }
616 :
617 : impl std::fmt::Debug for MissingKeyError {
618 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
619 0 : write!(f, "{}", self)
620 0 : }
621 : }
622 :
623 : impl std::fmt::Display for MissingKeyError {
624 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
625 0 : write!(
626 0 : f,
627 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
628 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
629 0 : )?;
630 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
631 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
632 0 : }
633 :
634 0 : if let Some(ref backtrace) = self.backtrace {
635 0 : write!(f, "\n{}", backtrace)?;
636 0 : }
637 :
638 0 : Ok(())
639 0 : }
640 : }
641 :
642 : impl PageReconstructError {
643 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
644 0 : pub(crate) fn is_stopping(&self) -> bool {
645 : use PageReconstructError::*;
646 0 : match self {
647 0 : Cancelled => true,
648 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
649 : }
650 0 : }
651 : }
652 :
653 : #[derive(thiserror::Error, Debug)]
654 : pub(crate) enum CreateImageLayersError {
655 : #[error("timeline shutting down")]
656 : Cancelled,
657 :
658 : #[error("read failed")]
659 : GetVectoredError(#[source] GetVectoredError),
660 :
661 : #[error("reconstruction failed")]
662 : PageReconstructError(#[source] PageReconstructError),
663 :
664 : #[error(transparent)]
665 : Other(#[from] anyhow::Error),
666 : }
667 :
668 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
669 0 : fn from(_: layer_manager::Shutdown) -> Self {
670 0 : CreateImageLayersError::Cancelled
671 0 : }
672 : }
673 :
674 : #[derive(thiserror::Error, Debug, Clone)]
675 : pub(crate) enum FlushLayerError {
676 : /// Timeline cancellation token was cancelled
677 : #[error("timeline shutting down")]
678 : Cancelled,
679 :
680 : /// We tried to flush a layer while the Timeline is in an unexpected state
681 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
682 : NotRunning(FlushLoopState),
683 :
684 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
685 : // loop via a watch channel, where we can only borrow it.
686 : #[error("create image layers (shared)")]
687 : CreateImageLayersError(Arc<CreateImageLayersError>),
688 :
689 : #[error("other (shared)")]
690 : Other(#[from] Arc<anyhow::Error>),
691 : }
692 :
693 : impl FlushLayerError {
694 : // When crossing from generic anyhow errors to this error type, we explicitly check
695 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
696 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
697 0 : let cancelled = timeline.cancel.is_cancelled()
698 : // The upload queue might have been shut down before the official cancellation of the timeline.
699 0 : || err
700 0 : .downcast_ref::<NotInitialized>()
701 0 : .map(NotInitialized::is_stopping)
702 0 : .unwrap_or_default();
703 0 : if cancelled {
704 0 : Self::Cancelled
705 : } else {
706 0 : Self::Other(Arc::new(err))
707 : }
708 0 : }
709 : }
710 :
711 : impl From<layer_manager::Shutdown> for FlushLayerError {
712 0 : fn from(_: layer_manager::Shutdown) -> Self {
713 0 : FlushLayerError::Cancelled
714 0 : }
715 : }
716 :
717 : #[derive(thiserror::Error, Debug)]
718 : pub(crate) enum GetVectoredError {
719 : #[error("timeline shutting down")]
720 : Cancelled,
721 :
722 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
723 : Oversized(u64),
724 :
725 : #[error("requested at invalid LSN: {0}")]
726 : InvalidLsn(Lsn),
727 :
728 : #[error("requested key not found: {0}")]
729 : MissingKey(MissingKeyError),
730 :
731 : #[error("ancestry walk")]
732 : GetReadyAncestorError(#[source] GetReadyAncestorError),
733 :
734 : #[error(transparent)]
735 : Other(#[from] anyhow::Error),
736 : }
737 :
738 : impl From<GetReadyAncestorError> for GetVectoredError {
739 4 : fn from(value: GetReadyAncestorError) -> Self {
740 : use GetReadyAncestorError::*;
741 4 : match value {
742 0 : Cancelled => GetVectoredError::Cancelled,
743 : AncestorLsnTimeout(_) | BadState { .. } => {
744 4 : GetVectoredError::GetReadyAncestorError(value)
745 : }
746 : }
747 4 : }
748 : }
749 :
750 : #[derive(thiserror::Error, Debug)]
751 : pub(crate) enum GetReadyAncestorError {
752 : #[error("ancestor LSN wait error")]
753 : AncestorLsnTimeout(#[from] WaitLsnError),
754 :
755 : #[error("bad state on timeline {timeline_id}: {state:?}")]
756 : BadState {
757 : timeline_id: TimelineId,
758 : state: TimelineState,
759 : },
760 :
761 : #[error("cancelled")]
762 : Cancelled,
763 : }
764 :
765 : #[derive(Clone, Copy)]
766 : pub enum LogicalSizeCalculationCause {
767 : Initial,
768 : ConsumptionMetricsSyntheticSize,
769 : EvictionTaskImitation,
770 : TenantSizeHandler,
771 : }
772 :
773 : pub enum GetLogicalSizePriority {
774 : User,
775 : Background,
776 : }
777 :
778 0 : #[derive(Debug, enumset::EnumSetType)]
779 : pub(crate) enum CompactFlags {
780 : ForceRepartition,
781 : ForceImageLayerCreation,
782 : ForceL0Compaction,
783 : EnhancedGcBottomMostCompaction,
784 : DryRun,
785 : }
786 :
787 : #[serde_with::serde_as]
788 0 : #[derive(Debug, Clone, serde::Deserialize)]
789 : pub(crate) struct CompactRequest {
790 : pub compact_key_range: Option<CompactKeyRange>,
791 : pub compact_lsn_range: Option<CompactLsnRange>,
792 : /// Whether the compaction job should be scheduled.
793 : #[serde(default)]
794 : pub scheduled: bool,
795 : /// Whether the compaction job should be split across key ranges.
796 : #[serde(default)]
797 : pub sub_compaction: bool,
798 : /// Max job size for each subcompaction job.
799 : pub sub_compaction_max_job_size_mb: Option<u64>,
800 : }
801 :
802 : #[derive(Debug, Clone, Default)]
803 : pub(crate) struct CompactOptions {
804 : pub flags: EnumSet<CompactFlags>,
805 : /// If set, the compaction will only compact the key range specified by this option.
806 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
807 : pub compact_key_range: Option<CompactKeyRange>,
808 : /// If set, the compaction will only compact the LSN within this value.
809 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
810 : pub compact_lsn_range: Option<CompactLsnRange>,
811 : /// Enable sub-compaction (split compaction job across key ranges).
812 : /// This option is only used by GC compaction.
813 : pub sub_compaction: bool,
814 : /// Set job size for the GC compaction.
815 : /// This option is only used by GC compaction.
816 : pub sub_compaction_max_job_size_mb: Option<u64>,
817 : }
818 :
819 : impl std::fmt::Debug for Timeline {
820 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
821 0 : write!(f, "Timeline<{}>", self.timeline_id)
822 0 : }
823 : }
824 :
825 : #[derive(thiserror::Error, Debug)]
826 : pub(crate) enum WaitLsnError {
827 : // Called on a timeline which is shutting down
828 : #[error("Shutdown")]
829 : Shutdown,
830 :
831 : // Called on an timeline not in active state or shutting down
832 : #[error("Bad timeline state: {0:?}")]
833 : BadState(TimelineState),
834 :
835 : // Timeout expired while waiting for LSN to catch up with goal.
836 : #[error("{0}")]
837 : Timeout(String),
838 : }
839 :
840 : // The impls below achieve cancellation mapping for errors.
841 : // Perhaps there's a way of achieving this with less cruft.
842 :
843 : impl From<CreateImageLayersError> for CompactionError {
844 0 : fn from(e: CreateImageLayersError) -> Self {
845 0 : match e {
846 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
847 0 : CreateImageLayersError::Other(e) => {
848 0 : CompactionError::Other(e.context("create image layers"))
849 : }
850 0 : _ => CompactionError::Other(e.into()),
851 : }
852 0 : }
853 : }
854 :
855 : impl From<CreateImageLayersError> for FlushLayerError {
856 0 : fn from(e: CreateImageLayersError) -> Self {
857 0 : match e {
858 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
859 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
860 : }
861 0 : }
862 : }
863 :
864 : impl From<PageReconstructError> for CreateImageLayersError {
865 0 : fn from(e: PageReconstructError) -> Self {
866 0 : match e {
867 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
868 0 : _ => CreateImageLayersError::PageReconstructError(e),
869 : }
870 0 : }
871 : }
872 :
873 : impl From<GetVectoredError> for CreateImageLayersError {
874 0 : fn from(e: GetVectoredError) -> Self {
875 0 : match e {
876 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
877 0 : _ => CreateImageLayersError::GetVectoredError(e),
878 : }
879 0 : }
880 : }
881 :
882 : impl From<GetVectoredError> for PageReconstructError {
883 12 : fn from(e: GetVectoredError) -> Self {
884 12 : match e {
885 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
886 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
887 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
888 8 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
889 4 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
890 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
891 : }
892 12 : }
893 : }
894 :
895 : impl From<GetReadyAncestorError> for PageReconstructError {
896 4 : fn from(e: GetReadyAncestorError) -> Self {
897 : use GetReadyAncestorError::*;
898 4 : match e {
899 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
900 4 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
901 0 : Cancelled => PageReconstructError::Cancelled,
902 : }
903 4 : }
904 : }
905 :
906 : pub(crate) enum WaitLsnTimeout {
907 : Custom(Duration),
908 : // Use the [`PageServerConf::wait_lsn_timeout`] default
909 : Default,
910 : }
911 :
912 : pub(crate) enum WaitLsnWaiter<'a> {
913 : Timeline(&'a Timeline),
914 : Tenant,
915 : PageService,
916 : HttpEndpoint,
917 : }
918 :
919 : /// Argument to [`Timeline::shutdown`].
920 : #[derive(Debug, Clone, Copy)]
921 : pub(crate) enum ShutdownMode {
922 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
923 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
924 : ///
925 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
926 : /// the call to [`Timeline::shutdown`].
927 : FreezeAndFlush,
928 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
929 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
930 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
931 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
932 : Reload,
933 : /// Shut down immediately, without waiting for any open layers to flush.
934 : Hard,
935 : }
936 :
937 : struct ImageLayerCreationOutcome {
938 : unfinished_image_layer: Option<ImageLayerWriter>,
939 : next_start_key: Key,
940 : }
941 :
942 : /// Public interface functions
943 : impl Timeline {
944 : /// Get the LSN where this branch was created
945 8 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
946 8 : self.ancestor_lsn
947 8 : }
948 :
949 : /// Get the ancestor's timeline id
950 24 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
951 24 : self.ancestor_timeline
952 24 : .as_ref()
953 24 : .map(|ancestor| ancestor.timeline_id)
954 24 : }
955 :
956 : /// Get the ancestor timeline
957 4 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
958 4 : self.ancestor_timeline.as_ref()
959 4 : }
960 :
961 : /// Get the bytes written since the PITR cutoff on this branch, and
962 : /// whether this branch's ancestor_lsn is within its parent's PITR.
963 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
964 0 : let gc_info = self.gc_info.read().unwrap();
965 0 : let history = self
966 0 : .get_last_record_lsn()
967 0 : .checked_sub(gc_info.cutoffs.time)
968 0 : .unwrap_or(Lsn(0))
969 0 : .0;
970 0 : (history, gc_info.within_ancestor_pitr)
971 0 : }
972 :
973 : /// Lock and get timeline's GC cutoff
974 1706762 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
975 1706762 : self.latest_gc_cutoff_lsn.read()
976 1706762 : }
977 :
978 : /// Look up given page version.
979 : ///
980 : /// If a remote layer file is needed, it is downloaded as part of this
981 : /// call.
982 : ///
983 : /// This method enforces [`Self::pagestream_throttle`] internally.
984 : ///
985 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
986 : /// abstraction above this needs to store suitable metadata to track what
987 : /// data exists with what keys, in separate metadata entries. If a
988 : /// non-existent key is requested, we may incorrectly return a value from
989 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
990 : /// non-existing key.
991 : ///
992 : /// # Cancel-Safety
993 : ///
994 : /// This method is cancellation-safe.
995 : #[inline(always)]
996 1215172 : pub(crate) async fn get(
997 1215172 : &self,
998 1215172 : key: Key,
999 1215172 : lsn: Lsn,
1000 1215172 : ctx: &RequestContext,
1001 1215172 : ) -> Result<Bytes, PageReconstructError> {
1002 1215172 : if !lsn.is_valid() {
1003 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1004 1215172 : }
1005 1215172 :
1006 1215172 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1007 1215172 : // already checked the key against the shard_identity when looking up the Timeline from
1008 1215172 : // page_service.
1009 1215172 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1010 :
1011 1215172 : let keyspace = KeySpace {
1012 1215172 : ranges: vec![key..key.next()],
1013 1215172 : };
1014 1215172 :
1015 1215172 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1016 :
1017 1215172 : let vectored_res = self
1018 1215172 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
1019 1215172 : .await;
1020 :
1021 1215172 : let key_value = vectored_res?.pop_first();
1022 1215160 : match key_value {
1023 1215136 : Some((got_key, value)) => {
1024 1215136 : if got_key != key {
1025 0 : error!(
1026 0 : "Expected {}, but singular vectored get returned {}",
1027 : key, got_key
1028 : );
1029 0 : Err(PageReconstructError::Other(anyhow!(
1030 0 : "Singular vectored get returned wrong key"
1031 0 : )))
1032 : } else {
1033 1215136 : value
1034 : }
1035 : }
1036 24 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1037 24 : key,
1038 24 : shard: self.shard_identity.get_shard_number(&key),
1039 24 : cont_lsn: Lsn(0),
1040 24 : request_lsn: lsn,
1041 24 : ancestor_lsn: None,
1042 24 : backtrace: None,
1043 24 : })),
1044 : }
1045 1215172 : }
1046 :
1047 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1048 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1049 :
1050 : /// Look up multiple page versions at a given LSN
1051 : ///
1052 : /// This naive implementation will be replaced with a more efficient one
1053 : /// which actually vectorizes the read path.
1054 39356 : pub(crate) async fn get_vectored(
1055 39356 : &self,
1056 39356 : keyspace: KeySpace,
1057 39356 : lsn: Lsn,
1058 39356 : io_concurrency: super::storage_layer::IoConcurrency,
1059 39356 : ctx: &RequestContext,
1060 39356 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1061 39356 : if !lsn.is_valid() {
1062 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1063 39356 : }
1064 39356 :
1065 39356 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1066 39356 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1067 0 : return Err(GetVectoredError::Oversized(key_count));
1068 39356 : }
1069 :
1070 78712 : for range in &keyspace.ranges {
1071 39356 : let mut key = range.start;
1072 79180 : while key != range.end {
1073 39824 : assert!(!self.shard_identity.is_key_disposable(&key));
1074 39824 : key = key.next();
1075 : }
1076 : }
1077 :
1078 39356 : trace!(
1079 0 : "get vectored request for {:?}@{} from task kind {:?}",
1080 0 : keyspace,
1081 0 : lsn,
1082 0 : ctx.task_kind(),
1083 : );
1084 :
1085 39356 : let start = crate::metrics::GET_VECTORED_LATENCY
1086 39356 : .for_task_kind(ctx.task_kind())
1087 39356 : .map(|metric| (metric, Instant::now()));
1088 :
1089 39356 : let res = self
1090 39356 : .get_vectored_impl(
1091 39356 : keyspace.clone(),
1092 39356 : lsn,
1093 39356 : &mut ValuesReconstructState::new(io_concurrency),
1094 39356 : ctx,
1095 39356 : )
1096 39356 : .await;
1097 :
1098 39356 : if let Some((metric, start)) = start {
1099 0 : let elapsed = start.elapsed();
1100 0 : metric.observe(elapsed.as_secs_f64());
1101 39356 : }
1102 :
1103 39356 : res
1104 39356 : }
1105 :
1106 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1107 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1108 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1109 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1110 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1111 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1112 : /// the scan operation will not cause OOM in the future.
1113 24 : pub(crate) async fn scan(
1114 24 : &self,
1115 24 : keyspace: KeySpace,
1116 24 : lsn: Lsn,
1117 24 : ctx: &RequestContext,
1118 24 : io_concurrency: super::storage_layer::IoConcurrency,
1119 24 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1120 24 : if !lsn.is_valid() {
1121 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1122 24 : }
1123 24 :
1124 24 : trace!(
1125 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1126 0 : keyspace,
1127 0 : lsn,
1128 0 : ctx.task_kind()
1129 : );
1130 :
1131 : // We should generalize this into Keyspace::contains in the future.
1132 48 : for range in &keyspace.ranges {
1133 24 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1134 24 : || range.end.field1 > METADATA_KEY_END_PREFIX
1135 : {
1136 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1137 0 : "only metadata keyspace can be scanned"
1138 0 : )));
1139 24 : }
1140 : }
1141 :
1142 24 : let start = crate::metrics::SCAN_LATENCY
1143 24 : .for_task_kind(ctx.task_kind())
1144 24 : .map(ScanLatencyOngoingRecording::start_recording);
1145 :
1146 24 : let vectored_res = self
1147 24 : .get_vectored_impl(
1148 24 : keyspace.clone(),
1149 24 : lsn,
1150 24 : &mut ValuesReconstructState::new(io_concurrency),
1151 24 : ctx,
1152 24 : )
1153 24 : .await;
1154 :
1155 24 : if let Some(recording) = start {
1156 0 : recording.observe();
1157 24 : }
1158 :
1159 24 : vectored_res
1160 24 : }
1161 :
1162 1255260 : pub(super) async fn get_vectored_impl(
1163 1255260 : &self,
1164 1255260 : keyspace: KeySpace,
1165 1255260 : lsn: Lsn,
1166 1255260 : reconstruct_state: &mut ValuesReconstructState,
1167 1255260 : ctx: &RequestContext,
1168 1255260 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1169 1255260 : let traversal_res: Result<(), _> = self
1170 1255260 : .get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1171 1255260 : .await;
1172 1255260 : if let Err(err) = traversal_res {
1173 : // Wait for all the spawned IOs to complete.
1174 : // See comments on `spawn_io` inside `storage_layer` for more details.
1175 32 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1176 32 : .into_values()
1177 32 : .map(|state| state.collect_pending_ios())
1178 32 : .collect::<FuturesUnordered<_>>();
1179 32 : while collect_futs.next().await.is_some() {}
1180 32 : return Err(err);
1181 1255228 : };
1182 1255228 :
1183 1255228 : let layers_visited = reconstruct_state.get_layers_visited();
1184 1255228 :
1185 1255228 : let futs = FuturesUnordered::new();
1186 1335896 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1187 1335896 : futs.push({
1188 1335896 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1189 1335896 : async move {
1190 1335896 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1191 :
1192 1335896 : let converted = match state.collect_pending_ios().await {
1193 1335896 : Ok(ok) => ok,
1194 0 : Err(err) => {
1195 0 : return (key, Err(err));
1196 : }
1197 : };
1198 1335896 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1199 1335896 :
1200 1335896 : // The walredo module expects the records to be descending in terms of Lsn.
1201 1335896 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1202 1335896 : debug_assert!(
1203 1335896 : converted
1204 1335896 : .records
1205 1335896 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1206 0 : "{converted:?}"
1207 : );
1208 :
1209 : (
1210 1335896 : key,
1211 1335896 : walredo_self.reconstruct_value(key, lsn, converted).await,
1212 : )
1213 1335896 : }
1214 1335896 : });
1215 1335896 : }
1216 :
1217 1255228 : let results = futs
1218 1255228 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1219 1255228 : .await;
1220 :
1221 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1222 : // when they're missing. Instead they are omitted from the resulting btree
1223 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1224 : // to avoid infinite results.
1225 1255228 : if !results.is_empty() {
1226 : // Record the total number of layers visited towards each key in the batch. While some
1227 : // layers may not intersect with a given read, and the cost of layer visits are
1228 : // amortized across the batch, each visited layer contributes directly to the observed
1229 : // latency for every read in the batch, which is what we care about.
1230 1254748 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1231 0 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1232 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1233 0 : LOG_PACER.lock().unwrap().call(|| {
1234 0 : let num_keys = keyspace.total_raw_size();
1235 0 : let num_pages = results.len();
1236 0 : tracing::info!(
1237 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1238 0 : lsn = %lsn,
1239 0 : "Vectored read for {keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1240 : );
1241 0 : });
1242 1254748 : }
1243 :
1244 2590644 : for _ in &results {
1245 1335896 : self.metrics.layers_per_read.observe(layers_visited as f64);
1246 1335896 : LAYERS_PER_READ_GLOBAL.observe(layers_visited as f64);
1247 1335896 : }
1248 480 : }
1249 :
1250 1255228 : Ok(results)
1251 1255260 : }
1252 :
1253 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1254 548848 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1255 548848 : self.last_record_lsn.load().last
1256 548848 : }
1257 :
1258 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1259 0 : self.last_record_lsn.load().prev
1260 0 : }
1261 :
1262 : /// Atomically get both last and prev.
1263 456 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1264 456 : self.last_record_lsn.load()
1265 456 : }
1266 :
1267 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1268 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1269 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1270 0 : self.last_record_lsn.status_receiver()
1271 0 : }
1272 :
1273 884 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1274 884 : self.disk_consistent_lsn.load()
1275 884 : }
1276 :
1277 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1278 : /// not validated with control plane yet.
1279 : /// See [`Self::get_remote_consistent_lsn_visible`].
1280 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1281 0 : self.remote_client.remote_consistent_lsn_projected()
1282 0 : }
1283 :
1284 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1285 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1286 : /// generation validation in the deletion queue.
1287 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1288 0 : self.remote_client.remote_consistent_lsn_visible()
1289 0 : }
1290 :
1291 : /// The sum of the file size of all historic layers in the layer map.
1292 : /// This method makes no distinction between local and remote layers.
1293 : /// Hence, the result **does not represent local filesystem usage**.
1294 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1295 0 : let guard = self.layers.read().await;
1296 0 : guard.layer_size_sum()
1297 0 : }
1298 :
1299 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1300 0 : self.metrics.resident_physical_size_get()
1301 0 : }
1302 :
1303 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1304 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1305 0 : }
1306 :
1307 : ///
1308 : /// Wait until WAL has been received and processed up to this LSN.
1309 : ///
1310 : /// You should call this before any of the other get_* or list_* functions. Calling
1311 : /// those functions with an LSN that has been processed yet is an error.
1312 : ///
1313 450898 : pub(crate) async fn wait_lsn(
1314 450898 : &self,
1315 450898 : lsn: Lsn,
1316 450898 : who_is_waiting: WaitLsnWaiter<'_>,
1317 450898 : timeout: WaitLsnTimeout,
1318 450898 : ctx: &RequestContext, /* Prepare for use by cancellation */
1319 450898 : ) -> Result<(), WaitLsnError> {
1320 450898 : let state = self.current_state();
1321 450898 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1322 0 : return Err(WaitLsnError::Shutdown);
1323 450898 : } else if !matches!(state, TimelineState::Active) {
1324 0 : return Err(WaitLsnError::BadState(state));
1325 450898 : }
1326 450898 :
1327 450898 : if cfg!(debug_assertions) {
1328 450898 : match ctx.task_kind() {
1329 : TaskKind::WalReceiverManager
1330 : | TaskKind::WalReceiverConnectionHandler
1331 : | TaskKind::WalReceiverConnectionPoller => {
1332 0 : let is_myself = match who_is_waiting {
1333 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1334 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService | WaitLsnWaiter::HttpEndpoint => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1335 : };
1336 0 : if is_myself {
1337 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1338 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1339 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1340 0 : }
1341 0 : } else {
1342 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1343 0 : // our walreceiver task can make progress independent of theirs
1344 0 : }
1345 : }
1346 450898 : _ => {}
1347 : }
1348 0 : }
1349 :
1350 450898 : let timeout = match timeout {
1351 0 : WaitLsnTimeout::Custom(t) => t,
1352 450898 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1353 : };
1354 :
1355 450898 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1356 450898 :
1357 450898 : match self.last_record_lsn.wait_for_timeout(lsn, timeout).await {
1358 450898 : Ok(()) => Ok(()),
1359 0 : Err(e) => {
1360 : use utils::seqwait::SeqWaitError::*;
1361 0 : match e {
1362 0 : Shutdown => Err(WaitLsnError::Shutdown),
1363 : Timeout => {
1364 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1365 0 : drop(_timer);
1366 0 : let walreceiver_status = self.walreceiver_status();
1367 0 : Err(WaitLsnError::Timeout(format!(
1368 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1369 0 : lsn,
1370 0 : self.get_last_record_lsn(),
1371 0 : self.get_disk_consistent_lsn(),
1372 0 : walreceiver_status,
1373 0 : )))
1374 : }
1375 : }
1376 : }
1377 : }
1378 450898 : }
1379 :
1380 0 : pub(crate) fn walreceiver_status(&self) -> String {
1381 0 : match &*self.walreceiver.lock().unwrap() {
1382 0 : None => "stopping or stopped".to_string(),
1383 0 : Some(walreceiver) => match walreceiver.status() {
1384 0 : Some(status) => status.to_human_readable_string(),
1385 0 : None => "Not active".to_string(),
1386 : },
1387 : }
1388 0 : }
1389 :
1390 : /// Check that it is valid to request operations with that lsn.
1391 464 : pub(crate) fn check_lsn_is_in_scope(
1392 464 : &self,
1393 464 : lsn: Lsn,
1394 464 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1395 464 : ) -> anyhow::Result<()> {
1396 464 : ensure!(
1397 464 : lsn >= **latest_gc_cutoff_lsn,
1398 8 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1399 8 : lsn,
1400 8 : **latest_gc_cutoff_lsn,
1401 : );
1402 456 : Ok(())
1403 464 : }
1404 :
1405 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1406 20 : pub(crate) fn init_lsn_lease(
1407 20 : &self,
1408 20 : lsn: Lsn,
1409 20 : length: Duration,
1410 20 : ctx: &RequestContext,
1411 20 : ) -> anyhow::Result<LsnLease> {
1412 20 : self.make_lsn_lease(lsn, length, true, ctx)
1413 20 : }
1414 :
1415 : /// 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.
1416 8 : pub(crate) fn renew_lsn_lease(
1417 8 : &self,
1418 8 : lsn: Lsn,
1419 8 : length: Duration,
1420 8 : ctx: &RequestContext,
1421 8 : ) -> anyhow::Result<LsnLease> {
1422 8 : self.make_lsn_lease(lsn, length, false, ctx)
1423 8 : }
1424 :
1425 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1426 : ///
1427 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1428 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1429 : ///
1430 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1431 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1432 28 : fn make_lsn_lease(
1433 28 : &self,
1434 28 : lsn: Lsn,
1435 28 : length: Duration,
1436 28 : init: bool,
1437 28 : _ctx: &RequestContext,
1438 28 : ) -> anyhow::Result<LsnLease> {
1439 24 : let lease = {
1440 : // Normalize the requested LSN to be aligned, and move to the first record
1441 : // if it points to the beginning of the page (header).
1442 28 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1443 28 :
1444 28 : let mut gc_info = self.gc_info.write().unwrap();
1445 28 :
1446 28 : let valid_until = SystemTime::now() + length;
1447 28 :
1448 28 : let entry = gc_info.leases.entry(lsn);
1449 28 :
1450 28 : match entry {
1451 12 : Entry::Occupied(mut occupied) => {
1452 12 : let existing_lease = occupied.get_mut();
1453 12 : if valid_until > existing_lease.valid_until {
1454 4 : existing_lease.valid_until = valid_until;
1455 4 : let dt: DateTime<Utc> = valid_until.into();
1456 4 : info!("lease extended to {}", dt);
1457 : } else {
1458 8 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1459 8 : info!("existing lease covers greater length, valid until {}", dt);
1460 : }
1461 :
1462 12 : existing_lease.clone()
1463 : }
1464 16 : Entry::Vacant(vacant) => {
1465 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
1466 : // not blocked by the lsn lease deadline.
1467 16 : let validate = {
1468 16 : let conf = self.tenant_conf.load();
1469 16 : conf.location.attach_mode == AttachmentMode::Single
1470 16 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1471 : };
1472 :
1473 16 : if init || validate {
1474 16 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1475 16 : if lsn < *latest_gc_cutoff_lsn {
1476 4 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1477 12 : }
1478 0 : }
1479 :
1480 12 : let dt: DateTime<Utc> = valid_until.into();
1481 12 : info!("lease created, valid until {}", dt);
1482 12 : vacant.insert(LsnLease { valid_until }).clone()
1483 : }
1484 : }
1485 : };
1486 :
1487 24 : Ok(lease)
1488 28 : }
1489 :
1490 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1491 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1492 : #[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))]
1493 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1494 : self.freeze0().await
1495 : }
1496 :
1497 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1498 : #[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))]
1499 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1500 : self.freeze_and_flush0().await
1501 : }
1502 :
1503 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1504 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1505 2240 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1506 2240 : let mut g = self.write_lock.lock().await;
1507 2240 : let to_lsn = self.get_last_record_lsn();
1508 2240 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1509 2240 : }
1510 :
1511 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1512 : // polluting the span hierarchy.
1513 2240 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1514 2240 : let token = self.freeze0().await?;
1515 2240 : self.wait_flush_completion(token).await
1516 2240 : }
1517 :
1518 : // Check if an open ephemeral layer should be closed: this provides
1519 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1520 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1521 : // ephemeral layer bytes has been breached.
1522 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1523 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1524 : // If the write lock is held, there is an active wal receiver: rolling open layers
1525 : // is their responsibility while they hold this lock.
1526 0 : return;
1527 : };
1528 :
1529 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1530 : // time, and this was missed.
1531 : // if write_guard.is_none() { return; }
1532 :
1533 0 : let Ok(layers_guard) = self.layers.try_read() else {
1534 : // Don't block if the layer lock is busy
1535 0 : return;
1536 : };
1537 :
1538 0 : let Ok(lm) = layers_guard.layer_map() else {
1539 0 : return;
1540 : };
1541 :
1542 0 : let Some(open_layer) = &lm.open_layer else {
1543 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1544 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1545 : // that didn't result in writes to this shard.
1546 :
1547 : // Must not hold the layers lock while waiting for a flush.
1548 0 : drop(layers_guard);
1549 0 :
1550 0 : let last_record_lsn = self.get_last_record_lsn();
1551 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1552 0 : if last_record_lsn > disk_consistent_lsn {
1553 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1554 : // we are a sharded tenant and have skipped some WAL
1555 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1556 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1557 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1558 : // without any data ingested (yet) doesn't write a remote index as soon as it
1559 : // sees its LSN advance: we only do this if we've been layer-less
1560 : // for some time.
1561 0 : tracing::debug!(
1562 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1563 : disk_consistent_lsn,
1564 : last_record_lsn
1565 : );
1566 :
1567 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1568 : // We know there is no open layer, so we can request freezing without actually
1569 : // freezing anything. This is true even if we have dropped the layers_guard, we
1570 : // still hold the write_guard.
1571 0 : let _ = async {
1572 0 : let token = self
1573 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1574 0 : .await?;
1575 0 : self.wait_flush_completion(token).await
1576 0 : }
1577 0 : .await;
1578 0 : }
1579 0 : }
1580 :
1581 0 : return;
1582 : };
1583 :
1584 0 : let Some(current_size) = open_layer.try_len() else {
1585 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1586 : // read lock to get size should always succeed.
1587 0 : tracing::warn!("Lock conflict while reading size of open layer");
1588 0 : return;
1589 : };
1590 :
1591 0 : let current_lsn = self.get_last_record_lsn();
1592 :
1593 0 : let checkpoint_distance_override = open_layer.tick().await;
1594 :
1595 0 : if let Some(size_override) = checkpoint_distance_override {
1596 0 : if current_size > size_override {
1597 : // This is not harmful, but it only happens in relatively rare cases where
1598 : // time-based checkpoints are not happening fast enough to keep the amount of
1599 : // ephemeral data within configured limits. It's a sign of stress on the system.
1600 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1601 0 : }
1602 0 : }
1603 :
1604 0 : let checkpoint_distance =
1605 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1606 0 :
1607 0 : if self.should_roll(
1608 0 : current_size,
1609 0 : current_size,
1610 0 : checkpoint_distance,
1611 0 : self.get_last_record_lsn(),
1612 0 : self.last_freeze_at.load(),
1613 0 : open_layer.get_opened_at(),
1614 0 : ) {
1615 0 : match open_layer.info() {
1616 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1617 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1618 0 : // happens asynchronously in the background.
1619 0 : tracing::debug!(
1620 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1621 : );
1622 : }
1623 : InMemoryLayerInfo::Open { .. } => {
1624 : // Upgrade to a write lock and freeze the layer
1625 0 : drop(layers_guard);
1626 0 : let res = self
1627 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1628 0 : .await;
1629 :
1630 0 : if let Err(e) = res {
1631 0 : tracing::info!(
1632 0 : "failed to flush frozen layer after background freeze: {e:#}"
1633 : );
1634 0 : }
1635 : }
1636 : }
1637 0 : }
1638 0 : }
1639 :
1640 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1641 : ///
1642 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1643 : /// child timelines that are not offloaded yet.
1644 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1645 0 : if self.remote_client.is_archived() != Some(true) {
1646 0 : return (false, "the timeline is not archived");
1647 0 : }
1648 0 : if !self.remote_client.no_pending_work() {
1649 : // if the remote client is still processing some work, we can't offload
1650 0 : return (false, "the upload queue is not drained yet");
1651 0 : }
1652 0 :
1653 0 : (true, "ok")
1654 0 : }
1655 :
1656 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1657 : /// compaction tasks.
1658 728 : pub(crate) async fn compact(
1659 728 : self: &Arc<Self>,
1660 728 : cancel: &CancellationToken,
1661 728 : flags: EnumSet<CompactFlags>,
1662 728 : ctx: &RequestContext,
1663 728 : ) -> Result<bool, CompactionError> {
1664 728 : self.compact_with_options(
1665 728 : cancel,
1666 728 : CompactOptions {
1667 728 : flags,
1668 728 : compact_key_range: None,
1669 728 : compact_lsn_range: None,
1670 728 : sub_compaction: false,
1671 728 : sub_compaction_max_job_size_mb: None,
1672 728 : },
1673 728 : ctx,
1674 728 : )
1675 728 : .await
1676 728 : }
1677 :
1678 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1679 : /// compaction tasks.
1680 728 : pub(crate) async fn compact_with_options(
1681 728 : self: &Arc<Self>,
1682 728 : cancel: &CancellationToken,
1683 728 : options: CompactOptions,
1684 728 : ctx: &RequestContext,
1685 728 : ) -> Result<bool, CompactionError> {
1686 728 : // most likely the cancellation token is from background task, but in tests it could be the
1687 728 : // request task as well.
1688 728 :
1689 728 : let prepare = async move {
1690 728 : let guard = self.compaction_lock.lock().await;
1691 :
1692 728 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1693 728 : BackgroundLoopKind::Compaction,
1694 728 : ctx,
1695 728 : )
1696 728 : .await;
1697 :
1698 728 : (guard, permit)
1699 728 : };
1700 :
1701 : // this wait probably never needs any "long time spent" logging, because we already nag if
1702 : // compaction task goes over it's period (20s) which is quite often in production.
1703 728 : let (_guard, _permit) = tokio::select! {
1704 728 : tuple = prepare => { tuple },
1705 728 : _ = self.cancel.cancelled() => return Ok(false),
1706 728 : _ = cancel.cancelled() => return Ok(false),
1707 : };
1708 :
1709 728 : let last_record_lsn = self.get_last_record_lsn();
1710 728 :
1711 728 : // Last record Lsn could be zero in case the timeline was just created
1712 728 : if !last_record_lsn.is_valid() {
1713 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1714 0 : return Ok(false);
1715 728 : }
1716 :
1717 728 : let result = match self.get_compaction_algorithm_settings().kind {
1718 : CompactionAlgorithm::Tiered => {
1719 0 : self.compact_tiered(cancel, ctx).await?;
1720 0 : Ok(false)
1721 : }
1722 728 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1723 : };
1724 :
1725 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
1726 0 : match result {
1727 728 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
1728 : Err(CompactionError::Other(_)) => {
1729 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
1730 : }
1731 : // Don't change the current value on offload failure or shutdown. We don't want to
1732 : // abruptly stall nor resume L0 flushes in these cases.
1733 0 : Err(CompactionError::Offload(_)) => {}
1734 0 : Err(CompactionError::ShuttingDown) => {}
1735 : };
1736 :
1737 728 : result
1738 728 : }
1739 :
1740 : /// Mutate the timeline with a [`TimelineWriter`].
1741 10266380 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1742 10266380 : TimelineWriter {
1743 10266380 : tl: self,
1744 10266380 : write_guard: self.write_lock.lock().await,
1745 : }
1746 10266380 : }
1747 :
1748 0 : pub(crate) fn activate(
1749 0 : self: &Arc<Self>,
1750 0 : parent: Arc<crate::tenant::Tenant>,
1751 0 : broker_client: BrokerClientChannel,
1752 0 : background_jobs_can_start: Option<&completion::Barrier>,
1753 0 : ctx: &RequestContext,
1754 0 : ) {
1755 0 : if self.tenant_shard_id.is_shard_zero() {
1756 0 : // Logical size is only maintained accurately on shard zero.
1757 0 : self.spawn_initial_logical_size_computation_task(ctx);
1758 0 : }
1759 0 : self.launch_wal_receiver(ctx, broker_client);
1760 0 : self.set_state(TimelineState::Active);
1761 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1762 0 : }
1763 :
1764 : /// After this function returns, there are no timeline-scoped tasks are left running.
1765 : ///
1766 : /// The preferred pattern for is:
1767 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1768 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1769 : /// go the extra mile and keep track of JoinHandles
1770 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1771 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1772 : ///
1773 : /// For legacy reasons, we still have multiple tasks spawned using
1774 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1775 : /// We refer to these as "timeline-scoped task_mgr tasks".
1776 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1777 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1778 : /// or [`task_mgr::shutdown_watcher`].
1779 : /// We want to gradually convert the code base away from these.
1780 : ///
1781 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1782 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1783 : /// ones that aren't mentioned here):
1784 : /// - [`TaskKind::TimelineDeletionWorker`]
1785 : /// - NB: also used for tenant deletion
1786 : /// - [`TaskKind::RemoteUploadTask`]`
1787 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1788 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1789 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1790 : /// - [`TaskKind::Eviction`]
1791 : /// - [`TaskKind::LayerFlushTask`]
1792 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1793 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1794 20 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1795 20 : debug_assert_current_span_has_tenant_and_timeline_id();
1796 20 :
1797 20 : // Regardless of whether we're going to try_freeze_and_flush
1798 20 : // or not, stop ingesting any more data. Walreceiver only provides
1799 20 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1800 20 : // So, only after the self.gate.close() below will we know for sure that
1801 20 : // no walreceiver tasks are left.
1802 20 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1803 20 : // data during the call to `self.freeze_and_flush()` below.
1804 20 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1805 20 : // which is what we'd need to properly model early shutdown of the walreceiver
1806 20 : // task sub-tree before the other Timeline task sub-trees.
1807 20 : let walreceiver = self.walreceiver.lock().unwrap().take();
1808 20 : tracing::debug!(
1809 0 : is_some = walreceiver.is_some(),
1810 0 : "Waiting for WalReceiverManager..."
1811 : );
1812 20 : if let Some(walreceiver) = walreceiver {
1813 0 : walreceiver.cancel();
1814 20 : }
1815 : // ... and inform any waiters for newer LSNs that there won't be any.
1816 20 : self.last_record_lsn.shutdown();
1817 20 :
1818 20 : if let ShutdownMode::FreezeAndFlush = mode {
1819 12 : if let Some((open, frozen)) = self
1820 12 : .layers
1821 12 : .read()
1822 12 : .await
1823 12 : .layer_map()
1824 12 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1825 12 : .ok()
1826 12 : .filter(|(open, frozen)| *open || *frozen > 0)
1827 : {
1828 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1829 12 : } else {
1830 12 : // this is double-shutdown, ignore it
1831 12 : }
1832 :
1833 : // we shut down walreceiver above, so, we won't add anything more
1834 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1835 : // to reach the disk & upload queue, then shut the upload queue and
1836 : // wait for it to drain.
1837 12 : match self.freeze_and_flush().await {
1838 : Ok(_) => {
1839 : // drain the upload queue
1840 : // if we did not wait for completion here, it might be our shutdown process
1841 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1842 : // be spawned.
1843 : //
1844 : // what is problematic is the shutting down of RemoteTimelineClient, because
1845 : // obviously it does not make sense to stop while we wait for it, but what
1846 : // about corner cases like s3 suddenly hanging up?
1847 12 : self.remote_client.shutdown().await;
1848 : }
1849 : Err(FlushLayerError::Cancelled) => {
1850 : // this is likely the second shutdown, ignore silently.
1851 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1852 0 : debug_assert!(self.cancel.is_cancelled());
1853 : }
1854 0 : Err(e) => {
1855 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1856 0 : // we have some extra WAL replay to do next time the timeline starts.
1857 0 : warn!("failed to freeze and flush: {e:#}");
1858 : }
1859 : }
1860 :
1861 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
1862 : // we also do a final check here to ensure that the queue is empty.
1863 12 : if !self.remote_client.no_pending_work() {
1864 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1865 12 : }
1866 8 : }
1867 :
1868 20 : if let ShutdownMode::Reload = mode {
1869 : // drain the upload queue
1870 4 : self.remote_client.shutdown().await;
1871 4 : if !self.remote_client.no_pending_work() {
1872 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1873 4 : }
1874 16 : }
1875 :
1876 : // Signal any subscribers to our cancellation token to drop out
1877 20 : tracing::debug!("Cancelling CancellationToken");
1878 20 : self.cancel.cancel();
1879 20 :
1880 20 : // Ensure Prevent new page service requests from starting.
1881 20 : self.handles.shutdown();
1882 20 :
1883 20 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1884 20 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1885 20 : self.remote_client.stop();
1886 20 :
1887 20 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1888 20 : // to shut down the upload queue tasks.
1889 20 : // TODO: fix that, task management should be encapsulated inside remote_client.
1890 20 : task_mgr::shutdown_tasks(
1891 20 : Some(TaskKind::RemoteUploadTask),
1892 20 : Some(self.tenant_shard_id),
1893 20 : Some(self.timeline_id),
1894 20 : )
1895 20 : .await;
1896 :
1897 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
1898 20 : tracing::debug!("Waiting for tasks...");
1899 20 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1900 :
1901 : {
1902 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
1903 : // open.
1904 20 : let mut write_guard = self.write_lock.lock().await;
1905 20 : self.layers.write().await.shutdown(&mut write_guard);
1906 20 : }
1907 20 :
1908 20 : // Finally wait until any gate-holders are complete.
1909 20 : //
1910 20 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1911 20 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1912 20 : self.gate.close().await;
1913 :
1914 20 : self.metrics.shutdown();
1915 20 : }
1916 :
1917 896 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1918 896 : match (self.current_state(), new_state) {
1919 896 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1920 4 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1921 : }
1922 0 : (st, TimelineState::Loading) => {
1923 0 : error!("ignoring transition from {st:?} into Loading state");
1924 : }
1925 0 : (TimelineState::Broken { .. }, new_state) => {
1926 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1927 : }
1928 : (TimelineState::Stopping, TimelineState::Active) => {
1929 0 : error!("Not activating a Stopping timeline");
1930 : }
1931 892 : (_, new_state) => {
1932 892 : self.state.send_replace(new_state);
1933 892 : }
1934 : }
1935 896 : }
1936 :
1937 4 : pub(crate) fn set_broken(&self, reason: String) {
1938 4 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1939 4 : let broken_state = TimelineState::Broken {
1940 4 : reason,
1941 4 : backtrace: backtrace_str,
1942 4 : };
1943 4 : self.set_state(broken_state);
1944 4 :
1945 4 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1946 4 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1947 4 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1948 4 : self.cancel.cancel();
1949 4 : }
1950 :
1951 452974 : pub(crate) fn current_state(&self) -> TimelineState {
1952 452974 : self.state.borrow().clone()
1953 452974 : }
1954 :
1955 12 : pub(crate) fn is_broken(&self) -> bool {
1956 12 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1957 12 : }
1958 :
1959 444 : pub(crate) fn is_active(&self) -> bool {
1960 444 : self.current_state() == TimelineState::Active
1961 444 : }
1962 :
1963 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
1964 0 : self.remote_client.is_archived()
1965 0 : }
1966 :
1967 736 : pub(crate) fn is_stopping(&self) -> bool {
1968 736 : self.current_state() == TimelineState::Stopping
1969 736 : }
1970 :
1971 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1972 0 : self.state.subscribe()
1973 0 : }
1974 :
1975 450902 : pub(crate) async fn wait_to_become_active(
1976 450902 : &self,
1977 450902 : _ctx: &RequestContext, // Prepare for use by cancellation
1978 450902 : ) -> Result<(), TimelineState> {
1979 450902 : let mut receiver = self.state.subscribe();
1980 : loop {
1981 450902 : let current_state = receiver.borrow().clone();
1982 450902 : match current_state {
1983 : TimelineState::Loading => {
1984 0 : receiver
1985 0 : .changed()
1986 0 : .await
1987 0 : .expect("holding a reference to self");
1988 : }
1989 : TimelineState::Active { .. } => {
1990 450898 : return Ok(());
1991 : }
1992 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1993 : // There's no chance the timeline can transition back into ::Active
1994 4 : return Err(current_state);
1995 : }
1996 : }
1997 : }
1998 450902 : }
1999 :
2000 0 : pub(crate) async fn layer_map_info(
2001 0 : &self,
2002 0 : reset: LayerAccessStatsReset,
2003 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2004 0 : let guard = self.layers.read().await;
2005 0 : let layer_map = guard.layer_map()?;
2006 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2007 0 : if let Some(open_layer) = &layer_map.open_layer {
2008 0 : in_memory_layers.push(open_layer.info());
2009 0 : }
2010 0 : for frozen_layer in &layer_map.frozen_layers {
2011 0 : in_memory_layers.push(frozen_layer.info());
2012 0 : }
2013 :
2014 0 : let historic_layers = layer_map
2015 0 : .iter_historic_layers()
2016 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2017 0 : .collect();
2018 0 :
2019 0 : Ok(LayerMapInfo {
2020 0 : in_memory_layers,
2021 0 : historic_layers,
2022 0 : })
2023 0 : }
2024 :
2025 : #[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))]
2026 : pub(crate) async fn download_layer(
2027 : &self,
2028 : layer_file_name: &LayerName,
2029 : ) -> anyhow::Result<Option<bool>> {
2030 : let Some(layer) = self.find_layer(layer_file_name).await? else {
2031 : return Ok(None);
2032 : };
2033 :
2034 : layer.download().await?;
2035 :
2036 : Ok(Some(true))
2037 : }
2038 :
2039 : /// Evict just one layer.
2040 : ///
2041 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2042 0 : pub(crate) async fn evict_layer(
2043 0 : &self,
2044 0 : layer_file_name: &LayerName,
2045 0 : ) -> anyhow::Result<Option<bool>> {
2046 0 : let _gate = self
2047 0 : .gate
2048 0 : .enter()
2049 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2050 :
2051 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2052 0 : return Ok(None);
2053 : };
2054 :
2055 : // curl has this by default
2056 0 : let timeout = std::time::Duration::from_secs(120);
2057 0 :
2058 0 : match local_layer.evict_and_wait(timeout).await {
2059 0 : Ok(()) => Ok(Some(true)),
2060 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2061 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2062 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2063 : }
2064 0 : }
2065 :
2066 9606020 : fn should_roll(
2067 9606020 : &self,
2068 9606020 : layer_size: u64,
2069 9606020 : projected_layer_size: u64,
2070 9606020 : checkpoint_distance: u64,
2071 9606020 : projected_lsn: Lsn,
2072 9606020 : last_freeze_at: Lsn,
2073 9606020 : opened_at: Instant,
2074 9606020 : ) -> bool {
2075 9606020 : let distance = projected_lsn.widening_sub(last_freeze_at);
2076 9606020 :
2077 9606020 : // Rolling the open layer can be triggered by:
2078 9606020 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2079 9606020 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2080 9606020 : // account for how writes are distributed across shards: we expect each node to consume
2081 9606020 : // 1/count of the LSN on average.
2082 9606020 : // 2. The size of the currently open layer.
2083 9606020 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2084 9606020 : // up and suspend activity.
2085 9606020 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2086 0 : info!(
2087 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2088 : projected_lsn, layer_size, distance
2089 : );
2090 :
2091 0 : true
2092 9606020 : } else if projected_layer_size >= checkpoint_distance {
2093 : // NB: this check is relied upon by:
2094 160 : let _ = IndexEntry::validate_checkpoint_distance;
2095 160 : info!(
2096 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2097 : projected_lsn, layer_size, projected_layer_size
2098 : );
2099 :
2100 160 : true
2101 9605860 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2102 0 : info!(
2103 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2104 0 : projected_lsn,
2105 0 : layer_size,
2106 0 : opened_at.elapsed()
2107 : );
2108 :
2109 0 : true
2110 : } else {
2111 9605860 : false
2112 : }
2113 9606020 : }
2114 : }
2115 :
2116 : /// Number of times we will compute partition within a checkpoint distance.
2117 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2118 :
2119 : // Private functions
2120 : impl Timeline {
2121 24 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2122 24 : let tenant_conf = self.tenant_conf.load();
2123 24 : tenant_conf
2124 24 : .tenant_conf
2125 24 : .lsn_lease_length
2126 24 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2127 24 : }
2128 :
2129 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2130 0 : let tenant_conf = self.tenant_conf.load();
2131 0 : tenant_conf
2132 0 : .tenant_conf
2133 0 : .lsn_lease_length_for_ts
2134 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2135 0 : }
2136 :
2137 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2138 0 : let tenant_conf = self.tenant_conf.load();
2139 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2140 0 : }
2141 :
2142 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2143 0 : let tenant_conf = self.tenant_conf.load();
2144 0 : tenant_conf
2145 0 : .tenant_conf
2146 0 : .lazy_slru_download
2147 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2148 0 : }
2149 :
2150 9608992 : fn get_checkpoint_distance(&self) -> u64 {
2151 9608992 : let tenant_conf = self.tenant_conf.load();
2152 9608992 : tenant_conf
2153 9608992 : .tenant_conf
2154 9608992 : .checkpoint_distance
2155 9608992 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2156 9608992 : }
2157 :
2158 9605860 : fn get_checkpoint_timeout(&self) -> Duration {
2159 9605860 : let tenant_conf = self.tenant_conf.load();
2160 9605860 : tenant_conf
2161 9605860 : .tenant_conf
2162 9605860 : .checkpoint_timeout
2163 9605860 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2164 9605860 : }
2165 :
2166 5008 : fn get_compaction_period(&self) -> Duration {
2167 5008 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2168 5008 : tenant_conf
2169 5008 : .compaction_period
2170 5008 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2171 5008 : }
2172 :
2173 1324 : fn get_compaction_target_size(&self) -> u64 {
2174 1324 : let tenant_conf = self.tenant_conf.load();
2175 1324 : tenant_conf
2176 1324 : .tenant_conf
2177 1324 : .compaction_target_size
2178 1324 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2179 1324 : }
2180 :
2181 784 : fn get_compaction_threshold(&self) -> usize {
2182 784 : let tenant_conf = self.tenant_conf.load();
2183 784 : tenant_conf
2184 784 : .tenant_conf
2185 784 : .compaction_threshold
2186 784 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2187 784 : }
2188 :
2189 56 : fn get_compaction_upper_limit(&self) -> usize {
2190 56 : let tenant_conf = self.tenant_conf.load();
2191 56 : tenant_conf
2192 56 : .tenant_conf
2193 56 : .compaction_upper_limit
2194 56 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2195 56 : }
2196 :
2197 2504 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2198 : // Disable L0 flushes by default. This and compaction needs further tuning.
2199 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 0; // TODO: default to e.g. 3
2200 :
2201 : // If compaction is disabled, don't delay.
2202 2504 : if self.get_compaction_period() == Duration::ZERO {
2203 2504 : return None;
2204 0 : }
2205 0 :
2206 0 : let compaction_threshold = self.get_compaction_threshold();
2207 0 : let tenant_conf = self.tenant_conf.load();
2208 0 : let l0_flush_delay_threshold = tenant_conf
2209 0 : .tenant_conf
2210 0 : .l0_flush_delay_threshold
2211 0 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2212 0 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2213 0 :
2214 0 : // 0 disables backpressure.
2215 0 : if l0_flush_delay_threshold == 0 {
2216 0 : return None;
2217 0 : }
2218 0 :
2219 0 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2220 0 : // backpressure flushes below this.
2221 0 : // TODO: the tenant config should have validation to prevent this instead.
2222 0 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2223 0 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2224 2504 : }
2225 :
2226 2504 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2227 : // Disable L0 stalls by default. In ingest benchmarks, we see image compaction take >10
2228 : // minutes, blocking L0 compaction, and we can't stall L0 flushes for that long.
2229 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2230 :
2231 : // If compaction is disabled, don't stall.
2232 2504 : if self.get_compaction_period() == Duration::ZERO {
2233 2504 : return None;
2234 0 : }
2235 0 :
2236 0 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2237 0 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2238 0 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2239 0 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2240 0 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2241 0 : return None;
2242 0 : }
2243 0 :
2244 0 : let compaction_threshold = self.get_compaction_threshold();
2245 0 : let tenant_conf = self.tenant_conf.load();
2246 0 : let l0_flush_stall_threshold = tenant_conf
2247 0 : .tenant_conf
2248 0 : .l0_flush_stall_threshold
2249 0 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2250 0 :
2251 0 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2252 0 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2253 0 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2254 0 : if cfg!(feature = "testing")
2255 0 : && compaction_threshold == 1
2256 0 : && l0_flush_stall_threshold.is_none()
2257 : {
2258 0 : return None;
2259 0 : }
2260 0 :
2261 0 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2262 0 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2263 0 :
2264 0 : // 0 disables backpressure.
2265 0 : if l0_flush_stall_threshold == 0 {
2266 0 : return None;
2267 0 : }
2268 0 :
2269 0 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2270 0 : // backpressure flushes below this.
2271 0 : // TODO: the tenant config should have validation to prevent this instead.
2272 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2273 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2274 2504 : }
2275 :
2276 2344 : fn get_l0_flush_wait_upload(&self) -> bool {
2277 2344 : let tenant_conf = self.tenant_conf.load();
2278 2344 : tenant_conf
2279 2344 : .tenant_conf
2280 2344 : .l0_flush_wait_upload
2281 2344 : .unwrap_or(self.conf.default_tenant_conf.l0_flush_wait_upload)
2282 2344 : }
2283 :
2284 28 : fn get_image_creation_threshold(&self) -> usize {
2285 28 : let tenant_conf = self.tenant_conf.load();
2286 28 : tenant_conf
2287 28 : .tenant_conf
2288 28 : .image_creation_threshold
2289 28 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2290 28 : }
2291 :
2292 728 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2293 728 : let tenant_conf = &self.tenant_conf.load();
2294 728 : tenant_conf
2295 728 : .tenant_conf
2296 728 : .compaction_algorithm
2297 728 : .as_ref()
2298 728 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2299 728 : .clone()
2300 728 : }
2301 :
2302 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2303 0 : let tenant_conf = self.tenant_conf.load();
2304 0 : tenant_conf
2305 0 : .tenant_conf
2306 0 : .eviction_policy
2307 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2308 0 : }
2309 :
2310 892 : fn get_evictions_low_residence_duration_metric_threshold(
2311 892 : tenant_conf: &TenantConfOpt,
2312 892 : default_tenant_conf: &TenantConf,
2313 892 : ) -> Duration {
2314 892 : tenant_conf
2315 892 : .evictions_low_residence_duration_metric_threshold
2316 892 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2317 892 : }
2318 :
2319 1136 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2320 1136 : let tenant_conf = self.tenant_conf.load();
2321 1136 : tenant_conf
2322 1136 : .tenant_conf
2323 1136 : .image_layer_creation_check_threshold
2324 1136 : .unwrap_or(
2325 1136 : self.conf
2326 1136 : .default_tenant_conf
2327 1136 : .image_layer_creation_check_threshold,
2328 1136 : )
2329 1136 : }
2330 :
2331 : /// Resolve the effective WAL receiver protocol to use for this tenant.
2332 : ///
2333 : /// Priority order is:
2334 : /// 1. Tenant config override
2335 : /// 2. Default value for tenant config override
2336 : /// 3. Pageserver config override
2337 : /// 4. Pageserver config default
2338 0 : pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
2339 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2340 0 : tenant_conf
2341 0 : .wal_receiver_protocol_override
2342 0 : .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
2343 0 : .unwrap_or(self.conf.wal_receiver_protocol)
2344 0 : }
2345 :
2346 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2347 0 : // NB: Most tenant conf options are read by background loops, so,
2348 0 : // changes will automatically be picked up.
2349 0 :
2350 0 : // The threshold is embedded in the metric. So, we need to update it.
2351 0 : {
2352 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2353 0 : &new_conf.tenant_conf,
2354 0 : &self.conf.default_tenant_conf,
2355 0 : );
2356 0 :
2357 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2358 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2359 0 :
2360 0 : let timeline_id_str = self.timeline_id.to_string();
2361 0 :
2362 0 : self.remote_client.update_config(&new_conf.location);
2363 0 :
2364 0 : self.metrics
2365 0 : .evictions_with_low_residence_duration
2366 0 : .write()
2367 0 : .unwrap()
2368 0 : .change_threshold(
2369 0 : &tenant_id_str,
2370 0 : &shard_id_str,
2371 0 : &timeline_id_str,
2372 0 : new_threshold,
2373 0 : );
2374 0 : }
2375 0 : }
2376 :
2377 : /// Open a Timeline handle.
2378 : ///
2379 : /// Loads the metadata for the timeline into memory, but not the layer map.
2380 : #[allow(clippy::too_many_arguments)]
2381 892 : pub(super) fn new(
2382 892 : conf: &'static PageServerConf,
2383 892 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2384 892 : metadata: &TimelineMetadata,
2385 892 : ancestor: Option<Arc<Timeline>>,
2386 892 : timeline_id: TimelineId,
2387 892 : tenant_shard_id: TenantShardId,
2388 892 : generation: Generation,
2389 892 : shard_identity: ShardIdentity,
2390 892 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2391 892 : resources: TimelineResources,
2392 892 : pg_version: u32,
2393 892 : state: TimelineState,
2394 892 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2395 892 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2396 892 : cancel: CancellationToken,
2397 892 : ) -> Arc<Self> {
2398 892 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2399 892 : let (state, _) = watch::channel(state);
2400 892 :
2401 892 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2402 892 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2403 892 :
2404 892 : let evictions_low_residence_duration_metric_threshold = {
2405 892 : let loaded_tenant_conf = tenant_conf.load();
2406 892 : Self::get_evictions_low_residence_duration_metric_threshold(
2407 892 : &loaded_tenant_conf.tenant_conf,
2408 892 : &conf.default_tenant_conf,
2409 892 : )
2410 : };
2411 :
2412 892 : if let Some(ancestor) = &ancestor {
2413 460 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2414 460 : // If we construct an explicit timeline object, it's obviously not offloaded
2415 460 : let is_offloaded = MaybeOffloaded::No;
2416 460 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2417 460 : }
2418 :
2419 892 : Arc::new_cyclic(|myself| {
2420 892 : let metrics = TimelineMetrics::new(
2421 892 : &tenant_shard_id,
2422 892 : &timeline_id,
2423 892 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2424 892 : "mtime",
2425 892 : evictions_low_residence_duration_metric_threshold,
2426 892 : ),
2427 892 : );
2428 892 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2429 :
2430 892 : let mut result = Timeline {
2431 892 : conf,
2432 892 : tenant_conf,
2433 892 : myself: myself.clone(),
2434 892 : timeline_id,
2435 892 : tenant_shard_id,
2436 892 : generation,
2437 892 : shard_identity,
2438 892 : pg_version,
2439 892 : layers: Default::default(),
2440 892 :
2441 892 : walredo_mgr,
2442 892 : walreceiver: Mutex::new(None),
2443 892 :
2444 892 : remote_client: Arc::new(resources.remote_client),
2445 892 :
2446 892 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2447 892 : last_record_lsn: SeqWait::new(RecordLsn {
2448 892 : last: disk_consistent_lsn,
2449 892 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2450 892 : }),
2451 892 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2452 892 :
2453 892 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2454 892 : last_freeze_ts: RwLock::new(Instant::now()),
2455 892 :
2456 892 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2457 892 :
2458 892 : ancestor_timeline: ancestor,
2459 892 : ancestor_lsn: metadata.ancestor_lsn(),
2460 892 :
2461 892 : metrics,
2462 892 :
2463 892 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2464 892 : &tenant_shard_id,
2465 892 : &timeline_id,
2466 892 : resources.pagestream_throttle_metrics,
2467 892 : ),
2468 892 :
2469 6244 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2470 892 :
2471 892 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2472 892 :
2473 892 : layer_flush_start_tx,
2474 892 : layer_flush_done_tx,
2475 892 :
2476 892 : write_lock: tokio::sync::Mutex::new(None),
2477 892 :
2478 892 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2479 892 :
2480 892 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2481 892 : initdb_lsn: metadata.initdb_lsn(),
2482 892 :
2483 892 : current_logical_size: if disk_consistent_lsn.is_valid() {
2484 : // we're creating timeline data with some layer files existing locally,
2485 : // need to recalculate timeline's logical size based on data in the layers.
2486 468 : LogicalSize::deferred_initial(disk_consistent_lsn)
2487 : } else {
2488 : // we're creating timeline data without any layers existing locally,
2489 : // initial logical size is 0.
2490 424 : LogicalSize::empty_initial()
2491 : },
2492 :
2493 892 : partitioning: GuardArcSwap::new((
2494 892 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2495 892 : Lsn(0),
2496 892 : )),
2497 892 : repartition_threshold: 0,
2498 892 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2499 892 : last_image_layer_creation_check_instant: Mutex::new(None),
2500 892 :
2501 892 : last_received_wal: Mutex::new(None),
2502 892 : rel_size_cache: RwLock::new(RelSizeCache {
2503 892 : complete_as_of: disk_consistent_lsn,
2504 892 : map: HashMap::new(),
2505 892 : }),
2506 892 :
2507 892 : download_all_remote_layers_task_info: RwLock::new(None),
2508 892 :
2509 892 : state,
2510 892 :
2511 892 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2512 892 : EvictionTaskTimelineState::default(),
2513 892 : ),
2514 892 : delete_progress: TimelineDeleteProgress::default(),
2515 892 :
2516 892 : cancel,
2517 892 : gate: Gate::default(),
2518 892 :
2519 892 : compaction_lock: tokio::sync::Mutex::default(),
2520 892 : compaction_failed: AtomicBool::default(),
2521 892 : gc_lock: tokio::sync::Mutex::default(),
2522 892 :
2523 892 : standby_horizon: AtomicLsn::new(0),
2524 892 :
2525 892 : pagestream_throttle: resources.pagestream_throttle,
2526 892 :
2527 892 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2528 892 :
2529 892 : #[cfg(test)]
2530 892 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2531 892 :
2532 892 : l0_flush_global_state: resources.l0_flush_global_state,
2533 892 :
2534 892 : handles: Default::default(),
2535 892 :
2536 892 : attach_wal_lag_cooldown,
2537 892 :
2538 892 : create_idempotency,
2539 892 :
2540 892 : page_trace: Default::default(),
2541 892 : };
2542 892 :
2543 892 : result.repartition_threshold =
2544 892 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2545 892 :
2546 892 : result
2547 892 : .metrics
2548 892 : .last_record_lsn_gauge
2549 892 : .set(disk_consistent_lsn.0 as i64);
2550 892 : result
2551 892 : })
2552 892 : }
2553 :
2554 1288 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2555 1288 : let Ok(guard) = self.gate.enter() else {
2556 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2557 0 : return;
2558 : };
2559 1288 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2560 1288 : match *flush_loop_state {
2561 880 : FlushLoopState::NotStarted => (),
2562 : FlushLoopState::Running { .. } => {
2563 408 : info!(
2564 0 : "skipping attempt to start flush_loop twice {}/{}",
2565 0 : self.tenant_shard_id, self.timeline_id
2566 : );
2567 408 : return;
2568 : }
2569 : FlushLoopState::Exited => {
2570 0 : warn!(
2571 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2572 0 : self.tenant_shard_id, self.timeline_id
2573 : );
2574 0 : return;
2575 : }
2576 : }
2577 :
2578 880 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2579 880 : let self_clone = Arc::clone(self);
2580 880 :
2581 880 : debug!("spawning flush loop");
2582 880 : *flush_loop_state = FlushLoopState::Running {
2583 880 : #[cfg(test)]
2584 880 : expect_initdb_optimization: false,
2585 880 : #[cfg(test)]
2586 880 : initdb_optimization_count: 0,
2587 880 : };
2588 880 : task_mgr::spawn(
2589 880 : task_mgr::BACKGROUND_RUNTIME.handle(),
2590 880 : task_mgr::TaskKind::LayerFlushTask,
2591 880 : self.tenant_shard_id,
2592 880 : Some(self.timeline_id),
2593 880 : "layer flush task",
2594 880 : async move {
2595 880 : let _guard = guard;
2596 880 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2597 880 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2598 20 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2599 20 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2600 20 : *flush_loop_state = FlushLoopState::Exited;
2601 20 : Ok(())
2602 20 : }
2603 880 : .instrument(info_span!(parent: None, "layer flush task", tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id))
2604 : );
2605 1288 : }
2606 :
2607 : /// Creates and starts the wal receiver.
2608 : ///
2609 : /// This function is expected to be called at most once per Timeline's lifecycle
2610 : /// when the timeline is activated.
2611 0 : fn launch_wal_receiver(
2612 0 : self: &Arc<Self>,
2613 0 : ctx: &RequestContext,
2614 0 : broker_client: BrokerClientChannel,
2615 0 : ) {
2616 0 : info!(
2617 0 : "launching WAL receiver for timeline {} of tenant {}",
2618 0 : self.timeline_id, self.tenant_shard_id
2619 : );
2620 :
2621 0 : let tenant_conf = self.tenant_conf.load();
2622 0 : let wal_connect_timeout = tenant_conf
2623 0 : .tenant_conf
2624 0 : .walreceiver_connect_timeout
2625 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2626 0 : let lagging_wal_timeout = tenant_conf
2627 0 : .tenant_conf
2628 0 : .lagging_wal_timeout
2629 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2630 0 : let max_lsn_wal_lag = tenant_conf
2631 0 : .tenant_conf
2632 0 : .max_lsn_wal_lag
2633 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2634 0 :
2635 0 : let mut guard = self.walreceiver.lock().unwrap();
2636 0 : assert!(
2637 0 : guard.is_none(),
2638 0 : "multiple launches / re-launches of WAL receiver are not supported"
2639 : );
2640 0 : *guard = Some(WalReceiver::start(
2641 0 : Arc::clone(self),
2642 0 : WalReceiverConf {
2643 0 : protocol: self.resolve_wal_receiver_protocol(),
2644 0 : wal_connect_timeout,
2645 0 : lagging_wal_timeout,
2646 0 : max_lsn_wal_lag,
2647 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2648 0 : availability_zone: self.conf.availability_zone.clone(),
2649 0 : ingest_batch_size: self.conf.ingest_batch_size,
2650 0 : },
2651 0 : broker_client,
2652 0 : ctx,
2653 0 : ));
2654 0 : }
2655 :
2656 : /// Initialize with an empty layer map. Used when creating a new timeline.
2657 880 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2658 880 : let mut layers = self.layers.try_write().expect(
2659 880 : "in the context where we call this function, no other task has access to the object",
2660 880 : );
2661 880 : layers
2662 880 : .open_mut()
2663 880 : .expect("in this context the LayerManager must still be open")
2664 880 : .initialize_empty(Lsn(start_lsn.0));
2665 880 : }
2666 :
2667 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2668 : /// files.
2669 12 : pub(super) async fn load_layer_map(
2670 12 : &self,
2671 12 : disk_consistent_lsn: Lsn,
2672 12 : index_part: IndexPart,
2673 12 : ) -> anyhow::Result<()> {
2674 : use init::{Decision::*, Discovered, DismissedLayer};
2675 : use LayerName::*;
2676 :
2677 12 : let mut guard = self.layers.write().await;
2678 :
2679 12 : let timer = self.metrics.load_layer_map_histo.start_timer();
2680 12 :
2681 12 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2682 12 : // structs representing all files on disk
2683 12 : let timeline_path = self
2684 12 : .conf
2685 12 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2686 12 : let conf = self.conf;
2687 12 : let span = tracing::Span::current();
2688 12 :
2689 12 : // Copy to move into the task we're about to spawn
2690 12 : let this = self.myself.upgrade().expect("&self method holds the arc");
2691 :
2692 12 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2693 12 : move || {
2694 12 : let _g = span.entered();
2695 12 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2696 12 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2697 12 : let mut unrecognized_files = Vec::new();
2698 12 :
2699 12 : let mut path = timeline_path;
2700 :
2701 44 : for discovered in discovered {
2702 32 : let (name, kind) = match discovered {
2703 32 : Discovered::Layer(layer_file_name, local_metadata) => {
2704 32 : discovered_layers.push((layer_file_name, local_metadata));
2705 32 : continue;
2706 : }
2707 0 : Discovered::IgnoredBackup(path) => {
2708 0 : std::fs::remove_file(path)
2709 0 : .or_else(fs_ext::ignore_not_found)
2710 0 : .fatal_err("Removing .old file");
2711 0 : continue;
2712 : }
2713 0 : Discovered::Unknown(file_name) => {
2714 0 : // we will later error if there are any
2715 0 : unrecognized_files.push(file_name);
2716 0 : continue;
2717 : }
2718 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2719 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2720 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2721 : };
2722 0 : path.push(Utf8Path::new(&name));
2723 0 : init::cleanup(&path, kind)?;
2724 0 : path.pop();
2725 : }
2726 :
2727 12 : if !unrecognized_files.is_empty() {
2728 : // assume that if there are any there are many many.
2729 0 : let n = unrecognized_files.len();
2730 0 : let first = &unrecognized_files[..n.min(10)];
2731 0 : anyhow::bail!(
2732 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2733 0 : );
2734 12 : }
2735 12 :
2736 12 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2737 12 :
2738 12 : let mut loaded_layers = Vec::new();
2739 12 : let mut needs_cleanup = Vec::new();
2740 12 : let mut total_physical_size = 0;
2741 :
2742 44 : for (name, decision) in decided {
2743 32 : let decision = match decision {
2744 32 : Ok(decision) => decision,
2745 0 : Err(DismissedLayer::Future { local }) => {
2746 0 : if let Some(local) = local {
2747 0 : init::cleanup_future_layer(
2748 0 : &local.local_path,
2749 0 : &name,
2750 0 : disk_consistent_lsn,
2751 0 : )?;
2752 0 : }
2753 0 : needs_cleanup.push(name);
2754 0 : continue;
2755 : }
2756 0 : Err(DismissedLayer::LocalOnly(local)) => {
2757 0 : init::cleanup_local_only_file(&name, &local)?;
2758 : // this file never existed remotely, we will have to do rework
2759 0 : continue;
2760 : }
2761 0 : Err(DismissedLayer::BadMetadata(local)) => {
2762 0 : init::cleanup_local_file_for_remote(&local)?;
2763 : // this file never existed remotely, we will have to do rework
2764 0 : continue;
2765 : }
2766 : };
2767 :
2768 32 : match &name {
2769 24 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2770 8 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2771 : }
2772 :
2773 32 : tracing::debug!(layer=%name, ?decision, "applied");
2774 :
2775 32 : let layer = match decision {
2776 32 : Resident { local, remote } => {
2777 32 : total_physical_size += local.file_size;
2778 32 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2779 32 : .drop_eviction_guard()
2780 : }
2781 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2782 : };
2783 :
2784 32 : loaded_layers.push(layer);
2785 : }
2786 12 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2787 12 : }
2788 12 : })
2789 12 : .await
2790 12 : .map_err(anyhow::Error::new)
2791 12 : .and_then(|x| x)?;
2792 :
2793 12 : let num_layers = loaded_layers.len();
2794 12 :
2795 12 : guard
2796 12 : .open_mut()
2797 12 : .expect("layermanager must be open during init")
2798 12 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2799 12 :
2800 12 : self.remote_client
2801 12 : .schedule_layer_file_deletion(&needs_cleanup)?;
2802 12 : self.remote_client
2803 12 : .schedule_index_upload_for_file_changes()?;
2804 : // This barrier orders above DELETEs before any later operations.
2805 : // This is critical because code executing after the barrier might
2806 : // create again objects with the same key that we just scheduled for deletion.
2807 : // For example, if we just scheduled deletion of an image layer "from the future",
2808 : // later compaction might run again and re-create the same image layer.
2809 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2810 : // "same" here means same key range and LSN.
2811 : //
2812 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2813 : // the upload queue may execute the PUT first, then the DELETE.
2814 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2815 : //
2816 : // 1. a future image layer is created and uploaded
2817 : // 2. ps restart
2818 : // 3. the future layer from (1) is deleted during load layer map
2819 : // 4. image layer is re-created and uploaded
2820 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2821 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2822 : //
2823 : // See https://github.com/neondatabase/neon/issues/5878
2824 : //
2825 : // NB: generation numbers naturally protect against this because they disambiguate
2826 : // (1) and (4)
2827 : // TODO: this is basically a no-op now, should we remove it?
2828 12 : self.remote_client.schedule_barrier()?;
2829 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2830 : // on retry.
2831 :
2832 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2833 12 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2834 12 : self.update_layer_visibility().await?;
2835 :
2836 12 : info!(
2837 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2838 : num_layers, disk_consistent_lsn, total_physical_size
2839 : );
2840 :
2841 12 : timer.stop_and_record();
2842 12 : Ok(())
2843 12 : }
2844 :
2845 : /// Retrieve current logical size of the timeline.
2846 : ///
2847 : /// The size could be lagging behind the actual number, in case
2848 : /// the initial size calculation has not been run (gets triggered on the first size access).
2849 : ///
2850 : /// return size and boolean flag that shows if the size is exact
2851 0 : pub(crate) fn get_current_logical_size(
2852 0 : self: &Arc<Self>,
2853 0 : priority: GetLogicalSizePriority,
2854 0 : ctx: &RequestContext,
2855 0 : ) -> logical_size::CurrentLogicalSize {
2856 0 : if !self.tenant_shard_id.is_shard_zero() {
2857 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2858 : // when HTTP API is serving a GET for timeline zero, return zero
2859 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2860 0 : }
2861 0 :
2862 0 : let current_size = self.current_logical_size.current_size();
2863 0 : debug!("Current size: {current_size:?}");
2864 :
2865 0 : match (current_size.accuracy(), priority) {
2866 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2867 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2868 0 : // background task will eventually deliver an exact value, we're in no rush
2869 0 : }
2870 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2871 : // background task is not ready, but user is asking for it now;
2872 : // => make the background task skip the line
2873 : // (The alternative would be to calculate the size here, but,
2874 : // it can actually take a long time if the user has a lot of rels.
2875 : // And we'll inevitable need it again; So, let the background task do the work.)
2876 0 : match self
2877 0 : .current_logical_size
2878 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2879 0 : .get()
2880 : {
2881 0 : Some(cancel) => cancel.cancel(),
2882 : None => {
2883 0 : match self.current_state() {
2884 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2885 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2886 0 : // Don't make noise.
2887 0 : }
2888 : TimelineState::Loading => {
2889 : // Import does not return an activated timeline.
2890 0 : info!("discarding priority boost for logical size calculation because timeline is not yet active");
2891 : }
2892 : TimelineState::Active => {
2893 : // activation should be setting the once cell
2894 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2895 0 : debug_assert!(false);
2896 : }
2897 : }
2898 : }
2899 : }
2900 : }
2901 : }
2902 :
2903 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2904 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2905 0 : let first = self
2906 0 : .current_logical_size
2907 0 : .did_return_approximate_to_walreceiver
2908 0 : .compare_exchange(
2909 0 : false,
2910 0 : true,
2911 0 : AtomicOrdering::Relaxed,
2912 0 : AtomicOrdering::Relaxed,
2913 0 : )
2914 0 : .is_ok();
2915 0 : if first {
2916 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2917 0 : }
2918 0 : }
2919 0 : }
2920 :
2921 0 : current_size
2922 0 : }
2923 :
2924 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2925 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2926 : // nothing to do for freshly created timelines;
2927 0 : assert_eq!(
2928 0 : self.current_logical_size.current_size().accuracy(),
2929 0 : logical_size::Accuracy::Exact,
2930 0 : );
2931 0 : self.current_logical_size.initialized.add_permits(1);
2932 0 : return;
2933 : };
2934 :
2935 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2936 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2937 0 : self.current_logical_size
2938 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2939 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2940 0 :
2941 0 : let self_clone = Arc::clone(self);
2942 0 : let background_ctx = ctx.detached_child(
2943 0 : TaskKind::InitialLogicalSizeCalculation,
2944 0 : DownloadBehavior::Download,
2945 0 : );
2946 0 : task_mgr::spawn(
2947 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2948 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2949 0 : self.tenant_shard_id,
2950 0 : Some(self.timeline_id),
2951 0 : "initial size calculation",
2952 : // NB: don't log errors here, task_mgr will do that.
2953 0 : async move {
2954 0 : self_clone
2955 0 : .initial_logical_size_calculation_task(
2956 0 : initial_part_end,
2957 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2958 0 : background_ctx,
2959 0 : )
2960 0 : .await;
2961 0 : Ok(())
2962 0 : }
2963 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)),
2964 : );
2965 0 : }
2966 :
2967 : /// # Cancellation
2968 : ///
2969 : /// This method is sensitive to `Timeline::cancel`.
2970 : ///
2971 : /// It is _not_ sensitive to task_mgr::shutdown_token().
2972 : ///
2973 : /// # Cancel-Safety
2974 : ///
2975 : /// It does Timeline IO, hence this should be polled to completion because
2976 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
2977 : /// to reason about.
2978 0 : async fn initial_logical_size_calculation_task(
2979 0 : self: Arc<Self>,
2980 0 : initial_part_end: Lsn,
2981 0 : skip_concurrency_limiter: CancellationToken,
2982 0 : background_ctx: RequestContext,
2983 0 : ) {
2984 0 : scopeguard::defer! {
2985 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2986 0 : self.current_logical_size.initialized.add_permits(1);
2987 0 : }
2988 0 :
2989 0 : let try_once = |attempt: usize| {
2990 0 : let background_ctx = &background_ctx;
2991 0 : let self_ref = &self;
2992 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2993 0 : async move {
2994 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2995 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2996 0 : background_ctx,
2997 0 : );
2998 :
2999 : use crate::metrics::initial_logical_size::StartCircumstances;
3000 0 : let (_maybe_permit, circumstances) = tokio::select! {
3001 0 : permit = wait_for_permit => {
3002 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3003 : }
3004 0 : _ = self_ref.cancel.cancelled() => {
3005 0 : return Err(CalculateLogicalSizeError::Cancelled);
3006 : }
3007 0 : () = skip_concurrency_limiter.cancelled() => {
3008 : // Some action that is part of a end user interaction requested logical size
3009 : // => break out of the rate limit
3010 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3011 : // but then again what happens if they cancel; also, we should just be using
3012 : // one runtime across the entire process, so, let's leave this for now.
3013 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3014 : }
3015 : };
3016 :
3017 0 : let metrics_guard = if attempt == 1 {
3018 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3019 : } else {
3020 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3021 : };
3022 :
3023 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3024 0 : self_ref.conf,
3025 0 : self_ref
3026 0 : .gate
3027 0 : .enter()
3028 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3029 : );
3030 :
3031 0 : let calculated_size = self_ref
3032 0 : .logical_size_calculation_task(
3033 0 : initial_part_end,
3034 0 : LogicalSizeCalculationCause::Initial,
3035 0 : background_ctx,
3036 0 : )
3037 0 : .await?;
3038 :
3039 0 : self_ref
3040 0 : .trigger_aux_file_size_computation(
3041 0 : initial_part_end,
3042 0 : background_ctx,
3043 0 : io_concurrency,
3044 0 : )
3045 0 : .await?;
3046 :
3047 : // TODO: add aux file size to logical size
3048 :
3049 0 : Ok((calculated_size, metrics_guard))
3050 0 : }
3051 0 : };
3052 :
3053 0 : let retrying = async {
3054 0 : let mut attempt = 0;
3055 : loop {
3056 0 : attempt += 1;
3057 0 :
3058 0 : match try_once(attempt).await {
3059 0 : Ok(res) => return ControlFlow::Continue(res),
3060 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3061 : Err(
3062 0 : e @ (CalculateLogicalSizeError::Decode(_)
3063 0 : | CalculateLogicalSizeError::PageRead(_)),
3064 0 : ) => {
3065 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3066 : // exponential back-off doesn't make sense at these long intervals;
3067 : // use fixed retry interval with generous jitter instead
3068 0 : let sleep_duration = Duration::from_secs(
3069 0 : u64::try_from(
3070 0 : // 1hour base
3071 0 : (60_i64 * 60_i64)
3072 0 : // 10min jitter
3073 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3074 0 : )
3075 0 : .expect("10min < 1hour"),
3076 0 : );
3077 0 : tokio::select! {
3078 0 : _ = tokio::time::sleep(sleep_duration) => {}
3079 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3080 : }
3081 : }
3082 : }
3083 : }
3084 0 : };
3085 :
3086 0 : let (calculated_size, metrics_guard) = match retrying.await {
3087 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3088 0 : ControlFlow::Break(()) => return,
3089 : };
3090 :
3091 : // we cannot query current_logical_size.current_size() to know the current
3092 : // *negative* value, only truncated to u64.
3093 0 : let added = self
3094 0 : .current_logical_size
3095 0 : .size_added_after_initial
3096 0 : .load(AtomicOrdering::Relaxed);
3097 0 :
3098 0 : let sum = calculated_size.saturating_add_signed(added);
3099 0 :
3100 0 : // set the gauge value before it can be set in `update_current_logical_size`.
3101 0 : self.metrics.current_logical_size_gauge.set(sum);
3102 0 :
3103 0 : self.current_logical_size
3104 0 : .initial_logical_size
3105 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3106 0 : .ok()
3107 0 : .expect("only this task sets it");
3108 0 : }
3109 :
3110 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3111 0 : self: &Arc<Self>,
3112 0 : lsn: Lsn,
3113 0 : cause: LogicalSizeCalculationCause,
3114 0 : ctx: RequestContext,
3115 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3116 0 : let (sender, receiver) = oneshot::channel();
3117 0 : let self_clone = Arc::clone(self);
3118 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3119 0 : // we should stop the size calculation work and return an error.
3120 0 : // That would require restructuring this function's API to
3121 0 : // return the result directly, instead of a Receiver for the result.
3122 0 : let ctx = ctx.detached_child(
3123 0 : TaskKind::OndemandLogicalSizeCalculation,
3124 0 : DownloadBehavior::Download,
3125 0 : );
3126 0 : task_mgr::spawn(
3127 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3128 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3129 0 : self.tenant_shard_id,
3130 0 : Some(self.timeline_id),
3131 0 : "ondemand logical size calculation",
3132 0 : async move {
3133 0 : let res = self_clone
3134 0 : .logical_size_calculation_task(lsn, cause, &ctx)
3135 0 : .await;
3136 0 : let _ = sender.send(res).ok();
3137 0 : Ok(()) // Receiver is responsible for handling errors
3138 0 : }
3139 0 : .in_current_span(),
3140 0 : );
3141 0 : receiver
3142 0 : }
3143 :
3144 : #[instrument(skip_all)]
3145 : async fn logical_size_calculation_task(
3146 : self: &Arc<Self>,
3147 : lsn: Lsn,
3148 : cause: LogicalSizeCalculationCause,
3149 : ctx: &RequestContext,
3150 : ) -> Result<u64, CalculateLogicalSizeError> {
3151 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3152 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3153 : // accurate relation sizes, and they do not emit consumption metrics.
3154 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3155 :
3156 : let guard = self
3157 : .gate
3158 : .enter()
3159 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3160 :
3161 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3162 : }
3163 :
3164 : /// Calculate the logical size of the database at the latest LSN.
3165 : ///
3166 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3167 : /// especially if we need to download remote layers.
3168 0 : async fn calculate_logical_size(
3169 0 : &self,
3170 0 : up_to_lsn: Lsn,
3171 0 : cause: LogicalSizeCalculationCause,
3172 0 : _guard: &GateGuard,
3173 0 : ctx: &RequestContext,
3174 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3175 0 : info!(
3176 0 : "Calculating logical size for timeline {} at {}",
3177 : self.timeline_id, up_to_lsn
3178 : );
3179 :
3180 0 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
3181 : {
3182 0 : return Err(CalculateLogicalSizeError::Cancelled);
3183 0 : }
3184 :
3185 : // See if we've already done the work for initial size calculation.
3186 : // This is a short-cut for timelines that are mostly unused.
3187 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3188 0 : return Ok(size);
3189 0 : }
3190 0 : let storage_time_metrics = match cause {
3191 : LogicalSizeCalculationCause::Initial
3192 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3193 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3194 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3195 0 : &self.metrics.imitate_logical_size_histo
3196 : }
3197 : };
3198 0 : let timer = storage_time_metrics.start_timer();
3199 0 : let logical_size = self
3200 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3201 0 : .await?;
3202 0 : debug!("calculated logical size: {logical_size}");
3203 0 : timer.stop_and_record();
3204 0 : Ok(logical_size)
3205 0 : }
3206 :
3207 : /// Update current logical size, adding `delta' to the old value.
3208 541140 : fn update_current_logical_size(&self, delta: i64) {
3209 541140 : let logical_size = &self.current_logical_size;
3210 541140 : logical_size.increment_size(delta);
3211 541140 :
3212 541140 : // Also set the value in the prometheus gauge. Note that
3213 541140 : // there is a race condition here: if this is is called by two
3214 541140 : // threads concurrently, the prometheus gauge might be set to
3215 541140 : // one value while current_logical_size is set to the
3216 541140 : // other.
3217 541140 : match logical_size.current_size() {
3218 541140 : CurrentLogicalSize::Exact(ref new_current_size) => self
3219 541140 : .metrics
3220 541140 : .current_logical_size_gauge
3221 541140 : .set(new_current_size.into()),
3222 0 : CurrentLogicalSize::Approximate(_) => {
3223 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3224 0 : // forth between the initial size calculation task.
3225 0 : }
3226 : }
3227 541140 : }
3228 :
3229 5912 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3230 5912 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3231 5912 : let aux_metric =
3232 5912 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3233 5912 :
3234 5912 : let sum_of_entries = self
3235 5912 : .directory_metrics
3236 5912 : .iter()
3237 41384 : .map(|v| v.load(AtomicOrdering::Relaxed))
3238 5912 : .sum();
3239 : // Set a high general threshold and a lower threshold for the auxiliary files,
3240 : // as we can have large numbers of relations in the db directory.
3241 : const SUM_THRESHOLD: u64 = 5000;
3242 : const AUX_THRESHOLD: u64 = 1000;
3243 5912 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3244 0 : self.metrics
3245 0 : .directory_entries_count_gauge
3246 0 : .set(sum_of_entries);
3247 5912 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3248 0 : metric.set(sum_of_entries);
3249 5912 : }
3250 5912 : }
3251 :
3252 0 : async fn find_layer(
3253 0 : &self,
3254 0 : layer_name: &LayerName,
3255 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3256 0 : let guard = self.layers.read().await;
3257 0 : let layer = guard
3258 0 : .layer_map()?
3259 0 : .iter_historic_layers()
3260 0 : .find(|l| &l.layer_name() == layer_name)
3261 0 : .map(|found| guard.get_from_desc(&found));
3262 0 : Ok(layer)
3263 0 : }
3264 :
3265 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3266 : /// indicating which layers are currently on-disk on the primary.
3267 : ///
3268 : /// None is returned if the Timeline is in a state where uploading a heatmap
3269 : /// doesn't make sense, such as shutting down or initializing. The caller
3270 : /// should treat this as a cue to simply skip doing any heatmap uploading
3271 : /// for this timeline.
3272 4 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3273 4 : if !self.is_active() {
3274 0 : return None;
3275 4 : }
3276 :
3277 4 : let guard = self.layers.read().await;
3278 :
3279 20 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3280 20 : match layer.visibility() {
3281 : LayerVisibilityHint::Visible => {
3282 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3283 16 : let last_activity_ts = layer.latest_activity();
3284 16 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
3285 : }
3286 : LayerVisibilityHint::Covered => {
3287 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3288 4 : None
3289 : }
3290 : }
3291 20 : });
3292 4 :
3293 4 : let mut layers = resident.collect::<Vec<_>>();
3294 4 :
3295 4 : // Sort layers in order of which to download first. For a large set of layers to download, we
3296 4 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3297 4 : // or hours later:
3298 4 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3299 4 : // only exist for a few minutes before being compacted into L1s.
3300 4 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3301 4 : // the layer is likely to be covered by an image layer during compaction.
3302 40 : layers.sort_by_key(|(desc, _meta, _atime)| {
3303 40 : std::cmp::Reverse((
3304 40 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3305 40 : desc.lsn_range.end,
3306 40 : ))
3307 40 : });
3308 4 :
3309 4 : let layers = layers
3310 4 : .into_iter()
3311 16 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3312 4 : .collect();
3313 4 :
3314 4 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3315 4 : }
3316 :
3317 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3318 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3319 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3320 0 : // branchpoint in the value in IndexPart::lineage
3321 0 : self.ancestor_lsn == lsn
3322 0 : || (self.ancestor_lsn == Lsn::INVALID
3323 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3324 0 : }
3325 : }
3326 :
3327 : impl Timeline {
3328 : #[allow(clippy::doc_lazy_continuation)]
3329 : /// Get the data needed to reconstruct all keys in the provided keyspace
3330 : ///
3331 : /// The algorithm is as follows:
3332 : /// 1. While some keys are still not done and there's a timeline to visit:
3333 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3334 : /// 2.1: Build the fringe for the current keyspace
3335 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3336 : /// intersects
3337 : /// 2.3. Pop the timeline from the fringe
3338 : /// 2.4. If the fringe is empty, go back to 1
3339 1255260 : async fn get_vectored_reconstruct_data(
3340 1255260 : &self,
3341 1255260 : mut keyspace: KeySpace,
3342 1255260 : request_lsn: Lsn,
3343 1255260 : reconstruct_state: &mut ValuesReconstructState,
3344 1255260 : ctx: &RequestContext,
3345 1255260 : ) -> Result<(), GetVectoredError> {
3346 1255260 : let mut timeline_owned: Arc<Timeline>;
3347 1255260 : let mut timeline = self;
3348 1255260 :
3349 1255260 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3350 :
3351 1255256 : let missing_keyspace = loop {
3352 1706158 : if self.cancel.is_cancelled() {
3353 0 : return Err(GetVectoredError::Cancelled);
3354 1706158 : }
3355 :
3356 : let TimelineVisitOutcome {
3357 1706158 : completed_keyspace: completed,
3358 1706158 : image_covered_keyspace,
3359 1706158 : } = Self::get_vectored_reconstruct_data_timeline(
3360 1706158 : timeline,
3361 1706158 : keyspace.clone(),
3362 1706158 : cont_lsn,
3363 1706158 : reconstruct_state,
3364 1706158 : &self.cancel,
3365 1706158 : ctx,
3366 1706158 : )
3367 1706158 : .await?;
3368 :
3369 1706158 : keyspace.remove_overlapping_with(&completed);
3370 1706158 :
3371 1706158 : // Do not descend into the ancestor timeline for aux files.
3372 1706158 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3373 1706158 : // stalling compaction.
3374 1706158 : keyspace.remove_overlapping_with(&KeySpace {
3375 1706158 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
3376 1706158 : });
3377 1706158 :
3378 1706158 : // Keyspace is fully retrieved
3379 1706158 : if keyspace.is_empty() {
3380 1255200 : break None;
3381 450958 : }
3382 :
3383 450958 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3384 : // Not fully retrieved but no ancestor timeline.
3385 56 : break Some(keyspace);
3386 : };
3387 :
3388 : // Now we see if there are keys covered by the image layer but does not exist in the
3389 : // image layer, which means that the key does not exist.
3390 :
3391 : // The block below will stop the vectored search if any of the keys encountered an image layer
3392 : // which did not contain a snapshot for said key. Since we have already removed all completed
3393 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3394 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3395 : // and stop the search as a result of that.
3396 450902 : let mut removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3397 450902 : // Do not fire missing key error for sparse keys.
3398 450902 : removed.remove_overlapping_with(&KeySpace {
3399 450902 : ranges: vec![SPARSE_RANGE],
3400 450902 : });
3401 450902 : if !removed.is_empty() {
3402 0 : break Some(removed);
3403 450902 : }
3404 450902 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3405 450902 : // keyspace.
3406 450902 :
3407 450902 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3408 450902 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3409 450902 : timeline_owned = timeline
3410 450902 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3411 450902 : .await?;
3412 450898 : timeline = &*timeline_owned;
3413 : };
3414 :
3415 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
3416 1255256 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
3417 56 : let mut missing_keyspace = missing_keyspace;
3418 56 : missing_keyspace.remove_overlapping_with(&KeySpace {
3419 56 : ranges: vec![SPARSE_RANGE],
3420 56 : });
3421 56 : if missing_keyspace.is_empty() {
3422 28 : None
3423 : } else {
3424 28 : Some(missing_keyspace)
3425 : }
3426 : } else {
3427 1255200 : None
3428 : };
3429 :
3430 1255256 : if let Some(missing_keyspace) = missing_keyspace {
3431 28 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3432 28 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3433 28 : shard: self
3434 28 : .shard_identity
3435 28 : .get_shard_number(&missing_keyspace.start().unwrap()),
3436 28 : cont_lsn,
3437 28 : request_lsn,
3438 28 : ancestor_lsn: Some(timeline.ancestor_lsn),
3439 28 : backtrace: None,
3440 28 : }));
3441 1255228 : }
3442 1255228 :
3443 1255228 : Ok(())
3444 1255260 : }
3445 :
3446 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3447 : ///
3448 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3449 : /// the current keyspace. The current keyspace of the search at any given timeline
3450 : /// is the original keyspace minus all the keys that have been completed minus
3451 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3452 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3453 : ///
3454 : /// This is basically a depth-first search visitor implementation where a vertex
3455 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3456 : ///
3457 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3458 : /// and get all the required reconstruct data from the layer in one go.
3459 : ///
3460 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3461 : /// decides how to deal with these two keyspaces.
3462 1706158 : async fn get_vectored_reconstruct_data_timeline(
3463 1706158 : timeline: &Timeline,
3464 1706158 : keyspace: KeySpace,
3465 1706158 : mut cont_lsn: Lsn,
3466 1706158 : reconstruct_state: &mut ValuesReconstructState,
3467 1706158 : cancel: &CancellationToken,
3468 1706158 : ctx: &RequestContext,
3469 1706158 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3470 1706158 : let mut unmapped_keyspace = keyspace.clone();
3471 1706158 : let mut fringe = LayerFringe::new();
3472 1706158 :
3473 1706158 : let mut completed_keyspace = KeySpace::default();
3474 1706158 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3475 1706158 :
3476 1706158 : // Prevent GC from progressing while visiting the current timeline.
3477 1706158 : // If we are GC-ing because a new image layer was added while traversing
3478 1706158 : // the timeline, then it will remove layers that are required for fulfilling
3479 1706158 : // the current get request (read-path cannot "look back" and notice the new
3480 1706158 : // image layer).
3481 1706158 : let _gc_cutoff_holder = timeline.get_latest_gc_cutoff_lsn();
3482 :
3483 : loop {
3484 3399002 : if cancel.is_cancelled() {
3485 0 : return Err(GetVectoredError::Cancelled);
3486 3399002 : }
3487 3399002 :
3488 3399002 : let (keys_done_last_step, keys_with_image_coverage) =
3489 3399002 : reconstruct_state.consume_done_keys();
3490 3399002 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3491 3399002 : completed_keyspace.merge(&keys_done_last_step);
3492 3399002 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3493 45092 : unmapped_keyspace
3494 45092 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3495 45092 : image_covered_keyspace.add_range(keys_with_image_coverage);
3496 3353910 : }
3497 :
3498 : // Do not descent any further if the last layer we visited
3499 : // completed all keys in the keyspace it inspected. This is not
3500 : // required for correctness, but avoids visiting extra layers
3501 : // which turns out to be a perf bottleneck in some cases.
3502 3399002 : if !unmapped_keyspace.is_empty() {
3503 2148286 : let guard = timeline.layers.read().await;
3504 2148286 : let layers = guard.layer_map()?;
3505 :
3506 2148286 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3507 1828311 : let start_lsn = l.get_lsn_range().start;
3508 1828311 : cont_lsn > start_lsn
3509 2148286 : });
3510 2148286 :
3511 2148286 : match in_memory_layer {
3512 1213336 : Some(l) => {
3513 1213336 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3514 1213336 : fringe.update(
3515 1213336 : ReadableLayer::InMemoryLayer(l),
3516 1213336 : unmapped_keyspace.clone(),
3517 1213336 : lsn_range,
3518 1213336 : );
3519 1213336 : }
3520 : None => {
3521 934998 : for range in unmapped_keyspace.ranges.iter() {
3522 934998 : let results = layers.range_search(range.clone(), cont_lsn);
3523 934998 :
3524 934998 : results
3525 934998 : .found
3526 934998 : .into_iter()
3527 934998 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3528 479536 : (
3529 479536 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3530 479536 : keyspace_accum.to_keyspace(),
3531 479536 : lsn_floor..cont_lsn,
3532 479536 : )
3533 934998 : })
3534 934998 : .for_each(|(layer, keyspace, lsn_range)| {
3535 479536 : fringe.update(layer, keyspace, lsn_range)
3536 934998 : });
3537 934998 : }
3538 : }
3539 : }
3540 :
3541 : // It's safe to drop the layer map lock after planning the next round of reads.
3542 : // The fringe keeps readable handles for the layers which are safe to read even
3543 : // if layers were compacted or flushed.
3544 : //
3545 : // The more interesting consideration is: "Why is the read algorithm still correct
3546 : // if the layer map changes while it is operating?". Doing a vectored read on a
3547 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3548 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3549 : // range at *a particular point in time*. It is fine for the answer to be different
3550 : // at two different time points.
3551 2148286 : drop(guard);
3552 1250716 : }
3553 :
3554 3399002 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3555 1692844 : let next_cont_lsn = lsn_range.start;
3556 1692844 : layer_to_read
3557 1692844 : .get_values_reconstruct_data(
3558 1692844 : keyspace_to_read.clone(),
3559 1692844 : lsn_range,
3560 1692844 : reconstruct_state,
3561 1692844 : ctx,
3562 1692844 : )
3563 1692844 : .await?;
3564 :
3565 1692844 : unmapped_keyspace = keyspace_to_read;
3566 1692844 : cont_lsn = next_cont_lsn;
3567 1692844 :
3568 1692844 : reconstruct_state.on_layer_visited(&layer_to_read);
3569 : } else {
3570 1706158 : break;
3571 1706158 : }
3572 1706158 : }
3573 1706158 :
3574 1706158 : Ok(TimelineVisitOutcome {
3575 1706158 : completed_keyspace,
3576 1706158 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3577 1706158 : })
3578 1706158 : }
3579 :
3580 450902 : async fn get_ready_ancestor_timeline(
3581 450902 : &self,
3582 450902 : ancestor: &Arc<Timeline>,
3583 450902 : ctx: &RequestContext,
3584 450902 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3585 450902 : // It's possible that the ancestor timeline isn't active yet, or
3586 450902 : // is active but hasn't yet caught up to the branch point. Wait
3587 450902 : // for it.
3588 450902 : //
3589 450902 : // This cannot happen while the pageserver is running normally,
3590 450902 : // because you cannot create a branch from a point that isn't
3591 450902 : // present in the pageserver yet. However, we don't wait for the
3592 450902 : // branch point to be uploaded to cloud storage before creating
3593 450902 : // a branch. I.e., the branch LSN need not be remote consistent
3594 450902 : // for the branching operation to succeed.
3595 450902 : //
3596 450902 : // Hence, if we try to load a tenant in such a state where
3597 450902 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3598 450902 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3599 450902 : // then we will need to wait for the ancestor timeline to
3600 450902 : // re-stream WAL up to branch_lsn before we access it.
3601 450902 : //
3602 450902 : // How can a tenant get in such a state?
3603 450902 : // - ungraceful pageserver process exit
3604 450902 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3605 450902 : //
3606 450902 : // NB: this could be avoided by requiring
3607 450902 : // branch_lsn >= remote_consistent_lsn
3608 450902 : // during branch creation.
3609 450902 : match ancestor.wait_to_become_active(ctx).await {
3610 450898 : Ok(()) => {}
3611 : Err(TimelineState::Stopping) => {
3612 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3613 0 : return Err(GetReadyAncestorError::Cancelled);
3614 : }
3615 4 : Err(state) => {
3616 4 : return Err(GetReadyAncestorError::BadState {
3617 4 : timeline_id: ancestor.timeline_id,
3618 4 : state,
3619 4 : });
3620 : }
3621 : }
3622 450898 : ancestor
3623 450898 : .wait_lsn(
3624 450898 : self.ancestor_lsn,
3625 450898 : WaitLsnWaiter::Timeline(self),
3626 450898 : WaitLsnTimeout::Default,
3627 450898 : ctx,
3628 450898 : )
3629 450898 : .await
3630 450898 : .map_err(|e| match e {
3631 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3632 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3633 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3634 0 : timeline_id: ancestor.timeline_id,
3635 0 : state,
3636 0 : },
3637 450898 : })?;
3638 :
3639 450898 : Ok(ancestor.clone())
3640 450902 : }
3641 :
3642 594312 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3643 594312 : &self.shard_identity
3644 594312 : }
3645 :
3646 : #[inline(always)]
3647 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3648 0 : ShardTimelineId {
3649 0 : shard_index: ShardIndex {
3650 0 : shard_number: self.shard_identity.number,
3651 0 : shard_count: self.shard_identity.count,
3652 0 : },
3653 0 : timeline_id: self.timeline_id,
3654 0 : }
3655 0 : }
3656 :
3657 : /// Returns a non-frozen open in-memory layer for ingestion.
3658 : ///
3659 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3660 : /// this function without holding the mutex.
3661 2592 : async fn get_layer_for_write(
3662 2592 : &self,
3663 2592 : lsn: Lsn,
3664 2592 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3665 2592 : ctx: &RequestContext,
3666 2592 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3667 2592 : let mut guard = self.layers.write().await;
3668 :
3669 2592 : let last_record_lsn = self.get_last_record_lsn();
3670 2592 : ensure!(
3671 2592 : lsn > last_record_lsn,
3672 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3673 : lsn,
3674 : last_record_lsn,
3675 : );
3676 :
3677 2592 : let layer = guard
3678 2592 : .open_mut()?
3679 2592 : .get_layer_for_write(
3680 2592 : lsn,
3681 2592 : self.conf,
3682 2592 : self.timeline_id,
3683 2592 : self.tenant_shard_id,
3684 2592 : &self.gate,
3685 2592 : ctx,
3686 2592 : )
3687 2592 : .await?;
3688 2592 : Ok(layer)
3689 2592 : }
3690 :
3691 10558184 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3692 10558184 : assert!(new_lsn.is_aligned());
3693 :
3694 10558184 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
3695 10558184 : self.last_record_lsn.advance(new_lsn);
3696 10558184 : }
3697 :
3698 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3699 : ///
3700 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3701 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3702 2400 : async fn freeze_inmem_layer_at(
3703 2400 : &self,
3704 2400 : at: Lsn,
3705 2400 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3706 2400 : ) -> Result<u64, FlushLayerError> {
3707 2400 : let frozen = {
3708 2400 : let mut guard = self.layers.write().await;
3709 2400 : guard
3710 2400 : .open_mut()?
3711 2400 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
3712 2400 : .await
3713 : };
3714 :
3715 2400 : if frozen {
3716 2344 : let now = Instant::now();
3717 2344 : *(self.last_freeze_ts.write().unwrap()) = now;
3718 2344 : }
3719 :
3720 : // Increment the flush cycle counter and wake up the flush task.
3721 : // Remember the new value, so that when we listen for the flush
3722 : // to finish, we know when the flush that we initiated has
3723 : // finished, instead of some other flush that was started earlier.
3724 2400 : let mut my_flush_request = 0;
3725 2400 :
3726 2400 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3727 2400 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3728 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3729 2400 : }
3730 2400 :
3731 2400 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3732 2400 : my_flush_request = *counter + 1;
3733 2400 : *counter = my_flush_request;
3734 2400 : *lsn = std::cmp::max(at, *lsn);
3735 2400 : });
3736 2400 :
3737 2400 : assert_ne!(my_flush_request, 0);
3738 :
3739 2400 : Ok(my_flush_request)
3740 2400 : }
3741 :
3742 : /// Layer flusher task's main loop.
3743 880 : async fn flush_loop(
3744 880 : self: &Arc<Self>,
3745 880 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3746 880 : ctx: &RequestContext,
3747 880 : ) {
3748 : // Subscribe to L0 delta layer updates, for compaction backpressure.
3749 880 : let mut watch_l0 = match self.layers.read().await.layer_map() {
3750 880 : Ok(lm) => lm.watch_level0_deltas(),
3751 0 : Err(Shutdown) => return,
3752 : };
3753 :
3754 880 : info!("started flush loop");
3755 : loop {
3756 3203 : tokio::select! {
3757 3203 : _ = self.cancel.cancelled() => {
3758 20 : info!("shutting down layer flush task due to Timeline::cancel");
3759 20 : break;
3760 : },
3761 3203 : _ = layer_flush_start_rx.changed() => {}
3762 2323 : }
3763 2323 : trace!("waking up");
3764 2323 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3765 2323 :
3766 2323 : // The highest LSN to which we flushed in the loop over frozen layers
3767 2323 : let mut flushed_to_lsn = Lsn(0);
3768 :
3769 2323 : let result = loop {
3770 4667 : if self.cancel.is_cancelled() {
3771 0 : info!("dropping out of flush loop for timeline shutdown");
3772 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3773 : // anyone waiting on that will respect self.cancel as well: they will stop
3774 : // waiting at the same time we as drop out of this loop.
3775 0 : return;
3776 4667 : }
3777 4667 :
3778 4667 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
3779 4667 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
3780 4667 : if flushed_to_lsn >= frozen_to_lsn {
3781 2267 : break Ok(());
3782 2400 : }
3783 :
3784 : // Fetch the next layer to flush, if any.
3785 2400 : let (layer, l0_count, frozen_count, frozen_size) = {
3786 2400 : let layers = self.layers.read().await;
3787 2400 : let Ok(lm) = layers.layer_map() else {
3788 0 : info!("dropping out of flush loop for timeline shutdown");
3789 0 : return;
3790 : };
3791 2400 : let l0_count = lm.level0_deltas().len();
3792 2400 : let frozen_count = lm.frozen_layers.len();
3793 2400 : let frozen_size: u64 = lm
3794 2400 : .frozen_layers
3795 2400 : .iter()
3796 2425 : .map(|l| l.estimated_in_mem_size())
3797 2400 : .sum();
3798 2400 : let layer = lm.frozen_layers.front().cloned();
3799 2400 : (layer, l0_count, frozen_count, frozen_size)
3800 2400 : // drop 'layers' lock
3801 2400 : };
3802 2400 : let Some(layer) = layer else {
3803 56 : break Ok(());
3804 : };
3805 :
3806 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
3807 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
3808 : //
3809 : // NB: the compaction loop only checks `compaction_threshold` every 20 seconds, so
3810 : // we can end up stalling before compaction even starts. Consider making it more
3811 : // responsive (e.g. via `watch_level0_deltas`).
3812 2344 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
3813 0 : if l0_count >= stall_threshold {
3814 0 : warn!(
3815 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
3816 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
3817 : );
3818 0 : let stall_timer = self
3819 0 : .metrics
3820 0 : .flush_delay_histo
3821 0 : .start_timer()
3822 0 : .record_on_drop();
3823 0 : tokio::select! {
3824 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
3825 0 : if let Ok(l0) = result.as_deref() {
3826 0 : let delay = stall_timer.elapsed().as_secs_f64();
3827 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
3828 0 : }
3829 : },
3830 0 : _ = self.cancel.cancelled() => {},
3831 : }
3832 0 : continue; // check again
3833 0 : }
3834 2344 : }
3835 :
3836 : // Flush the layer.
3837 2344 : let flush_timer = self.metrics.flush_time_histo.start_timer();
3838 2344 : match self.flush_frozen_layer(layer, ctx).await {
3839 2344 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
3840 : Err(FlushLayerError::Cancelled) => {
3841 0 : info!("dropping out of flush loop for timeline shutdown");
3842 0 : return;
3843 : }
3844 0 : err @ Err(
3845 0 : FlushLayerError::NotRunning(_)
3846 0 : | FlushLayerError::Other(_)
3847 0 : | FlushLayerError::CreateImageLayersError(_),
3848 0 : ) => {
3849 0 : error!("could not flush frozen layer: {err:?}");
3850 0 : break err.map(|_| ());
3851 : }
3852 : }
3853 2344 : let flush_duration = flush_timer.stop_and_record();
3854 :
3855 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
3856 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
3857 : // ingestion by having ephemeral layer rolls wait for flushes.
3858 2344 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
3859 0 : if l0_count >= delay_threshold {
3860 0 : let delay = flush_duration.as_secs_f64();
3861 0 : info!(
3862 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
3863 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
3864 : );
3865 0 : let _delay_timer = self
3866 0 : .metrics
3867 0 : .flush_delay_histo
3868 0 : .start_timer()
3869 0 : .record_on_drop();
3870 0 : tokio::select! {
3871 0 : _ = tokio::time::sleep(flush_duration) => {},
3872 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
3873 0 : _ = self.cancel.cancelled() => {},
3874 : }
3875 0 : }
3876 2344 : }
3877 : };
3878 :
3879 : // Unsharded tenants should never advance their LSN beyond the end of the
3880 : // highest layer they write: such gaps between layer data and the frozen LSN
3881 : // are only legal on sharded tenants.
3882 2323 : debug_assert!(
3883 2323 : self.shard_identity.count.count() > 1
3884 2323 : || flushed_to_lsn >= frozen_to_lsn
3885 56 : || !flushed_to_lsn.is_valid()
3886 : );
3887 :
3888 2323 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3889 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3890 : // to us via layer_flush_start_rx, then advance it here.
3891 : //
3892 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3893 : // never encounter a gap in the wal.
3894 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3895 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3896 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3897 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3898 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3899 0 : }
3900 0 : }
3901 2323 : }
3902 :
3903 : // Notify any listeners that we're done
3904 2323 : let _ = self
3905 2323 : .layer_flush_done_tx
3906 2323 : .send_replace((flush_counter, result));
3907 : }
3908 20 : }
3909 :
3910 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
3911 2240 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3912 2240 : let mut rx = self.layer_flush_done_tx.subscribe();
3913 : loop {
3914 : {
3915 4555 : let (last_result_counter, last_result) = &*rx.borrow();
3916 4555 : if *last_result_counter >= request {
3917 2240 : if let Err(err) = last_result {
3918 : // We already logged the original error in
3919 : // flush_loop. We cannot propagate it to the caller
3920 : // here, because it might not be Cloneable
3921 0 : return Err(err.clone());
3922 : } else {
3923 2240 : return Ok(());
3924 : }
3925 2315 : }
3926 2315 : }
3927 2315 : trace!("waiting for flush to complete");
3928 2315 : tokio::select! {
3929 2315 : rx_e = rx.changed() => {
3930 2315 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3931 : },
3932 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3933 : // the notification from [`flush_loop`] that it completed.
3934 2315 : _ = self.cancel.cancelled() => {
3935 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3936 0 : return Ok(())
3937 : }
3938 : };
3939 2315 : trace!("done")
3940 : }
3941 2240 : }
3942 :
3943 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3944 : ///
3945 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3946 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3947 : async fn flush_frozen_layer(
3948 : self: &Arc<Self>,
3949 : frozen_layer: Arc<InMemoryLayer>,
3950 : ctx: &RequestContext,
3951 : ) -> Result<Lsn, FlushLayerError> {
3952 : debug_assert_current_span_has_tenant_and_timeline_id();
3953 :
3954 : // As a special case, when we have just imported an image into the repository,
3955 : // instead of writing out a L0 delta layer, we directly write out image layer
3956 : // files instead. This is possible as long as *all* the data imported into the
3957 : // repository have the same LSN.
3958 : let lsn_range = frozen_layer.get_lsn_range();
3959 :
3960 : // Whether to directly create image layers for this flush, or flush them as delta layers
3961 : let create_image_layer =
3962 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3963 :
3964 : #[cfg(test)]
3965 : {
3966 : match &mut *self.flush_loop_state.lock().unwrap() {
3967 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3968 : panic!("flush loop not running")
3969 : }
3970 : FlushLoopState::Running {
3971 : expect_initdb_optimization,
3972 : initdb_optimization_count,
3973 : ..
3974 : } => {
3975 : if create_image_layer {
3976 : *initdb_optimization_count += 1;
3977 : } else {
3978 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3979 : }
3980 : }
3981 : }
3982 : }
3983 :
3984 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3985 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3986 : // require downloading anything during initial import.
3987 : let ((rel_partition, metadata_partition), _lsn) = self
3988 : .repartition(
3989 : self.initdb_lsn,
3990 : self.get_compaction_target_size(),
3991 : EnumSet::empty(),
3992 : ctx,
3993 : )
3994 : .await
3995 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
3996 :
3997 : if self.cancel.is_cancelled() {
3998 : return Err(FlushLayerError::Cancelled);
3999 : }
4000 :
4001 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
4002 : // So that the key ranges don't overlap.
4003 : let mut partitions = KeyPartitioning::default();
4004 : partitions.parts.extend(rel_partition.parts);
4005 : if !metadata_partition.parts.is_empty() {
4006 : assert_eq!(
4007 : metadata_partition.parts.len(),
4008 : 1,
4009 : "currently sparse keyspace should only contain a single metadata keyspace"
4010 : );
4011 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
4012 : // every single key within the keyspace, and therefore, it's safe to force converting it
4013 : // into a dense keyspace before calling this function.
4014 : partitions
4015 : .parts
4016 : .extend(metadata_partition.into_dense().parts);
4017 : }
4018 :
4019 : let mut layers_to_upload = Vec::new();
4020 : layers_to_upload.extend(
4021 : self.create_image_layers(
4022 : &partitions,
4023 : self.initdb_lsn,
4024 : ImageLayerCreationMode::Initial,
4025 : ctx,
4026 : )
4027 : .await?,
4028 : );
4029 :
4030 : (layers_to_upload, None)
4031 : } else {
4032 : // Normal case, write out a L0 delta layer file.
4033 : // `create_delta_layer` will not modify the layer map.
4034 : // We will remove frozen layer and add delta layer in one atomic operation later.
4035 : let Some(layer) = self
4036 : .create_delta_layer(&frozen_layer, None, ctx)
4037 : .await
4038 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
4039 : else {
4040 : panic!("delta layer cannot be empty if no filter is applied");
4041 : };
4042 : (
4043 : // FIXME: even though we have a single image and single delta layer assumption
4044 : // we push them to vec
4045 : vec![layer.clone()],
4046 : Some(layer),
4047 : )
4048 : };
4049 :
4050 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
4051 :
4052 : if self.cancel.is_cancelled() {
4053 : return Err(FlushLayerError::Cancelled);
4054 : }
4055 :
4056 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
4057 :
4058 : // The new on-disk layers are now in the layer map. We can remove the
4059 : // in-memory layer from the map now. The flushed layer is stored in
4060 : // the mapping in `create_delta_layer`.
4061 : {
4062 : let mut guard = self.layers.write().await;
4063 :
4064 : guard.open_mut()?.finish_flush_l0_layer(
4065 : delta_layer_to_add.as_ref(),
4066 : &frozen_layer,
4067 : &self.metrics,
4068 : );
4069 :
4070 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
4071 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
4072 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
4073 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
4074 : }
4075 : // release lock on 'layers'
4076 : };
4077 :
4078 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
4079 : // This makes us refuse ingest until the new layers have been persisted to the remote
4080 : // TODO: remove this, and rely on l0_flush_{delay,stall}_threshold instead.
4081 : if self.get_l0_flush_wait_upload() {
4082 : let start = Instant::now();
4083 : self.remote_client
4084 : .wait_completion()
4085 : .await
4086 0 : .map_err(|e| match e {
4087 : WaitCompletionError::UploadQueueShutDownOrStopped
4088 : | WaitCompletionError::NotInitialized(
4089 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
4090 0 : ) => FlushLayerError::Cancelled,
4091 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
4092 0 : FlushLayerError::Other(anyhow!(e).into())
4093 : }
4094 0 : })?;
4095 : let duration = start.elapsed().as_secs_f64();
4096 : self.metrics.flush_wait_upload_time_gauge_add(duration);
4097 : }
4098 :
4099 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
4100 : // a compaction can delete the file and then it won't be available for uploads any more.
4101 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
4102 : // race situation.
4103 : // See https://github.com/neondatabase/neon/issues/4526
4104 : pausable_failpoint!("flush-frozen-pausable");
4105 :
4106 : // This failpoint is used by another test case `test_pageserver_recovery`.
4107 : fail_point!("flush-frozen-exit");
4108 :
4109 : Ok(Lsn(lsn_range.end.0 - 1))
4110 : }
4111 :
4112 : /// Return true if the value changed
4113 : ///
4114 : /// This function must only be used from the layer flush task.
4115 2344 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
4116 2344 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
4117 2344 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
4118 :
4119 2344 : self.metrics
4120 2344 : .disk_consistent_lsn_gauge
4121 2344 : .set(new_value.0 as i64);
4122 2344 : new_value != old_value
4123 2344 : }
4124 :
4125 : /// Update metadata file
4126 2444 : fn schedule_uploads(
4127 2444 : &self,
4128 2444 : disk_consistent_lsn: Lsn,
4129 2444 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
4130 2444 : ) -> anyhow::Result<()> {
4131 2444 : // We can only save a valid 'prev_record_lsn' value on disk if we
4132 2444 : // flushed *all* in-memory changes to disk. We only track
4133 2444 : // 'prev_record_lsn' in memory for the latest processed record, so we
4134 2444 : // don't remember what the correct value that corresponds to some old
4135 2444 : // LSN is. But if we flush everything, then the value corresponding
4136 2444 : // current 'last_record_lsn' is correct and we can store it on disk.
4137 2444 : let RecordLsn {
4138 2444 : last: last_record_lsn,
4139 2444 : prev: prev_record_lsn,
4140 2444 : } = self.last_record_lsn.load();
4141 2444 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
4142 2190 : Some(prev_record_lsn)
4143 : } else {
4144 254 : None
4145 : };
4146 :
4147 2444 : let update = crate::tenant::metadata::MetadataUpdate::new(
4148 2444 : disk_consistent_lsn,
4149 2444 : ondisk_prev_record_lsn,
4150 2444 : *self.latest_gc_cutoff_lsn.read(),
4151 2444 : );
4152 2444 :
4153 2444 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
4154 0 : "{}",
4155 0 : x.unwrap()
4156 2444 : ));
4157 :
4158 4812 : for layer in layers_to_upload {
4159 2368 : self.remote_client.schedule_layer_file_upload(layer)?;
4160 : }
4161 2444 : self.remote_client
4162 2444 : .schedule_index_upload_for_metadata_update(&update)?;
4163 :
4164 2444 : Ok(())
4165 2444 : }
4166 :
4167 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
4168 0 : self.remote_client
4169 0 : .preserve_initdb_archive(
4170 0 : &self.tenant_shard_id.tenant_id,
4171 0 : &self.timeline_id,
4172 0 : &self.cancel,
4173 0 : )
4174 0 : .await
4175 0 : }
4176 :
4177 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
4178 : // in layer map immediately. The caller is responsible to put it into the layer map.
4179 1936 : async fn create_delta_layer(
4180 1936 : self: &Arc<Self>,
4181 1936 : frozen_layer: &Arc<InMemoryLayer>,
4182 1936 : key_range: Option<Range<Key>>,
4183 1936 : ctx: &RequestContext,
4184 1936 : ) -> anyhow::Result<Option<ResidentLayer>> {
4185 1936 : let self_clone = Arc::clone(self);
4186 1936 : let frozen_layer = Arc::clone(frozen_layer);
4187 1936 : let ctx = ctx.attached_child();
4188 1936 : let work = async move {
4189 1936 : let Some((desc, path)) = frozen_layer
4190 1936 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
4191 1936 : .await?
4192 : else {
4193 0 : return Ok(None);
4194 : };
4195 1936 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
4196 :
4197 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
4198 : // We just need to fsync the directory in which these inodes are linked,
4199 : // which we know to be the timeline directory.
4200 : //
4201 : // We use fatal_err() below because the after write_to_disk returns with success,
4202 : // the in-memory state of the filesystem already has the layer file in its final place,
4203 : // and subsequent pageserver code could think it's durable while it really isn't.
4204 1936 : let timeline_dir = VirtualFile::open(
4205 1936 : &self_clone
4206 1936 : .conf
4207 1936 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
4208 1936 : &ctx,
4209 1936 : )
4210 1936 : .await
4211 1936 : .fatal_err("VirtualFile::open for timeline dir fsync");
4212 1936 : timeline_dir
4213 1936 : .sync_all()
4214 1936 : .await
4215 1936 : .fatal_err("VirtualFile::sync_all timeline dir");
4216 1936 : anyhow::Ok(Some(new_delta))
4217 1936 : };
4218 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
4219 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
4220 : use crate::virtual_file::io_engine::IoEngine;
4221 1936 : match crate::virtual_file::io_engine::get() {
4222 0 : IoEngine::NotSet => panic!("io engine not set"),
4223 : IoEngine::StdFs => {
4224 968 : let span = tracing::info_span!("blocking");
4225 968 : tokio::task::spawn_blocking({
4226 968 : move || Handle::current().block_on(work.instrument(span))
4227 968 : })
4228 968 : .await
4229 968 : .context("spawn_blocking")
4230 968 : .and_then(|x| x)
4231 : }
4232 : #[cfg(target_os = "linux")]
4233 968 : IoEngine::TokioEpollUring => work.await,
4234 : }
4235 1936 : }
4236 :
4237 1136 : async fn repartition(
4238 1136 : &self,
4239 1136 : lsn: Lsn,
4240 1136 : partition_size: u64,
4241 1136 : flags: EnumSet<CompactFlags>,
4242 1136 : ctx: &RequestContext,
4243 1136 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
4244 1136 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
4245 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
4246 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
4247 : // and hence before the compaction task starts.
4248 0 : return Err(CompactionError::Other(anyhow!(
4249 0 : "repartition() called concurrently"
4250 0 : )));
4251 : };
4252 1136 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
4253 1136 : if lsn < *partition_lsn {
4254 0 : return Err(CompactionError::Other(anyhow!(
4255 0 : "repartition() called with LSN going backwards, this should not happen"
4256 0 : )));
4257 1136 : }
4258 1136 :
4259 1136 : let distance = lsn.0 - partition_lsn.0;
4260 1136 : if *partition_lsn != Lsn(0)
4261 524 : && distance <= self.repartition_threshold
4262 524 : && !flags.contains(CompactFlags::ForceRepartition)
4263 : {
4264 496 : debug!(
4265 : distance,
4266 : threshold = self.repartition_threshold,
4267 0 : "no repartitioning needed"
4268 : );
4269 496 : return Ok((
4270 496 : (dense_partition.clone(), sparse_partition.clone()),
4271 496 : *partition_lsn,
4272 496 : ));
4273 640 : }
4274 :
4275 640 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4276 640 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4277 640 : let sparse_partitioning = SparseKeyPartitioning {
4278 640 : parts: vec![sparse_ks],
4279 640 : }; // no partitioning for metadata keys for now
4280 640 : let result = ((dense_partitioning, sparse_partitioning), lsn);
4281 640 : guard.write(result.clone());
4282 640 : Ok(result)
4283 1136 : }
4284 :
4285 : // Is it time to create a new image layer for the given partition?
4286 28 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4287 28 : let threshold = self.get_image_creation_threshold();
4288 :
4289 28 : let guard = self.layers.read().await;
4290 28 : let Ok(layers) = guard.layer_map() else {
4291 0 : return false;
4292 : };
4293 :
4294 28 : let mut max_deltas = 0;
4295 56 : for part_range in &partition.ranges {
4296 28 : let image_coverage = layers.image_coverage(part_range, lsn);
4297 56 : for (img_range, last_img) in image_coverage {
4298 28 : let img_lsn = if let Some(last_img) = last_img {
4299 0 : last_img.get_lsn_range().end
4300 : } else {
4301 28 : Lsn(0)
4302 : };
4303 : // Let's consider an example:
4304 : //
4305 : // delta layer with LSN range 71-81
4306 : // delta layer with LSN range 81-91
4307 : // delta layer with LSN range 91-101
4308 : // image layer at LSN 100
4309 : //
4310 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4311 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4312 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4313 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4314 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4315 28 : if img_lsn < lsn {
4316 28 : let num_deltas =
4317 28 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4318 28 :
4319 28 : max_deltas = max_deltas.max(num_deltas);
4320 28 : if num_deltas >= threshold {
4321 0 : debug!(
4322 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4323 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4324 : );
4325 0 : return true;
4326 28 : }
4327 0 : }
4328 : }
4329 : }
4330 :
4331 28 : debug!(
4332 : max_deltas,
4333 0 : "none of the partitioned ranges had >= {threshold} deltas"
4334 : );
4335 28 : false
4336 28 : }
4337 :
4338 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4339 : /// so that at most one image layer will be produced from this function.
4340 : #[allow(clippy::too_many_arguments)]
4341 460 : async fn create_image_layer_for_rel_blocks(
4342 460 : self: &Arc<Self>,
4343 460 : partition: &KeySpace,
4344 460 : mut image_layer_writer: ImageLayerWriter,
4345 460 : lsn: Lsn,
4346 460 : ctx: &RequestContext,
4347 460 : img_range: Range<Key>,
4348 460 : start: Key,
4349 460 : io_concurrency: IoConcurrency,
4350 460 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4351 460 : let mut wrote_keys = false;
4352 460 :
4353 460 : let mut key_request_accum = KeySpaceAccum::new();
4354 3044 : for range in &partition.ranges {
4355 2584 : let mut key = range.start;
4356 5600 : while key < range.end {
4357 : // Decide whether to retain this key: usually we do, but sharded tenants may
4358 : // need to drop keys that don't belong to them. If we retain the key, add it
4359 : // to `key_request_accum` for later issuing a vectored get
4360 3016 : if self.shard_identity.is_key_disposable(&key) {
4361 0 : debug!(
4362 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4363 0 : key,
4364 0 : self.shard_identity.get_shard_number(&key)
4365 : );
4366 3016 : } else {
4367 3016 : key_request_accum.add_key(key);
4368 3016 : }
4369 :
4370 3016 : let last_key_in_range = key.next() == range.end;
4371 3016 : key = key.next();
4372 3016 :
4373 3016 : // Maybe flush `key_rest_accum`
4374 3016 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4375 3016 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4376 : {
4377 2584 : let results = self
4378 2584 : .get_vectored(
4379 2584 : key_request_accum.consume_keyspace(),
4380 2584 : lsn,
4381 2584 : io_concurrency.clone(),
4382 2584 : ctx,
4383 2584 : )
4384 2584 : .await?;
4385 :
4386 2584 : if self.cancel.is_cancelled() {
4387 0 : return Err(CreateImageLayersError::Cancelled);
4388 2584 : }
4389 :
4390 5600 : for (img_key, img) in results {
4391 3016 : let img = match img {
4392 3016 : Ok(img) => img,
4393 0 : Err(err) => {
4394 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4395 0 : // page without losing any actual user data. That seems better
4396 0 : // than failing repeatedly and getting stuck.
4397 0 : //
4398 0 : // We had a bug at one point, where we truncated the FSM and VM
4399 0 : // in the pageserver, but the Postgres didn't know about that
4400 0 : // and continued to generate incremental WAL records for pages
4401 0 : // that didn't exist in the pageserver. Trying to replay those
4402 0 : // WAL records failed to find the previous image of the page.
4403 0 : // This special case allows us to recover from that situation.
4404 0 : // See https://github.com/neondatabase/neon/issues/2601.
4405 0 : //
4406 0 : // Unfortunately we cannot do this for the main fork, or for
4407 0 : // any metadata keys, keys, as that would lead to actual data
4408 0 : // loss.
4409 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4410 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4411 0 : ZERO_PAGE.clone()
4412 : } else {
4413 0 : return Err(CreateImageLayersError::from(err));
4414 : }
4415 : }
4416 : };
4417 :
4418 : // Write all the keys we just read into our new image layer.
4419 3016 : image_layer_writer.put_image(img_key, img, ctx).await?;
4420 3016 : wrote_keys = true;
4421 : }
4422 432 : }
4423 : }
4424 : }
4425 :
4426 460 : if wrote_keys {
4427 : // Normal path: we have written some data into the new image layer for this
4428 : // partition, so flush it to disk.
4429 460 : info!(
4430 0 : "produced image layer for rel {}",
4431 0 : ImageLayerName {
4432 0 : key_range: img_range.clone(),
4433 0 : lsn
4434 0 : },
4435 : );
4436 460 : Ok(ImageLayerCreationOutcome {
4437 460 : unfinished_image_layer: Some(image_layer_writer),
4438 460 : next_start_key: img_range.end,
4439 460 : })
4440 : } else {
4441 : // Special case: the image layer may be empty if this is a sharded tenant and the
4442 : // partition does not cover any keys owned by this shard. In this case, to ensure
4443 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4444 : // layer we write will cover the key range that we just scanned.
4445 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4446 0 : Ok(ImageLayerCreationOutcome {
4447 0 : unfinished_image_layer: None,
4448 0 : next_start_key: start,
4449 0 : })
4450 : }
4451 460 : }
4452 :
4453 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4454 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4455 : /// would not be too large to fit in a single image layer.
4456 : #[allow(clippy::too_many_arguments)]
4457 440 : async fn create_image_layer_for_metadata_keys(
4458 440 : self: &Arc<Self>,
4459 440 : partition: &KeySpace,
4460 440 : mut image_layer_writer: ImageLayerWriter,
4461 440 : lsn: Lsn,
4462 440 : ctx: &RequestContext,
4463 440 : img_range: Range<Key>,
4464 440 : mode: ImageLayerCreationMode,
4465 440 : start: Key,
4466 440 : io_concurrency: IoConcurrency,
4467 440 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4468 440 : // Metadata keys image layer creation.
4469 440 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
4470 440 : let begin = Instant::now();
4471 440 : let data = self
4472 440 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4473 440 : .await?;
4474 440 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4475 440 : let mut new_data = BTreeMap::new();
4476 440 : let mut total_kb_retrieved = 0;
4477 440 : let mut total_keys_retrieved = 0;
4478 20464 : for (k, v) in data {
4479 20024 : let v = v?;
4480 20024 : total_kb_retrieved += KEY_SIZE + v.len();
4481 20024 : total_keys_retrieved += 1;
4482 20024 : new_data.insert(k, v);
4483 : }
4484 440 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4485 440 : };
4486 440 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4487 440 : let elapsed = begin.elapsed();
4488 440 :
4489 440 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4490 440 : info!(
4491 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()
4492 : );
4493 :
4494 440 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4495 4 : return Ok(ImageLayerCreationOutcome {
4496 4 : unfinished_image_layer: None,
4497 4 : next_start_key: img_range.end,
4498 4 : });
4499 436 : }
4500 436 : if self.cancel.is_cancelled() {
4501 0 : return Err(CreateImageLayersError::Cancelled);
4502 436 : }
4503 436 : let mut wrote_any_image = false;
4504 20460 : for (k, v) in data {
4505 20024 : if v.is_empty() {
4506 : // the key has been deleted, it does not need an image
4507 : // in metadata keyspace, an empty image == tombstone
4508 16 : continue;
4509 20008 : }
4510 20008 : wrote_any_image = true;
4511 20008 :
4512 20008 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4513 20008 :
4514 20008 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4515 20008 : // on the normal data path either.
4516 20008 : image_layer_writer.put_image(k, v, ctx).await?;
4517 : }
4518 :
4519 436 : if wrote_any_image {
4520 : // Normal path: we have written some data into the new image layer for this
4521 : // partition, so flush it to disk.
4522 24 : info!(
4523 0 : "created image layer for metadata {}",
4524 0 : ImageLayerName {
4525 0 : key_range: img_range.clone(),
4526 0 : lsn
4527 0 : }
4528 : );
4529 24 : Ok(ImageLayerCreationOutcome {
4530 24 : unfinished_image_layer: Some(image_layer_writer),
4531 24 : next_start_key: img_range.end,
4532 24 : })
4533 : } else {
4534 : // Special case: the image layer may be empty if this is a sharded tenant and the
4535 : // partition does not cover any keys owned by this shard. In this case, to ensure
4536 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4537 : // layer we write will cover the key range that we just scanned.
4538 412 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4539 412 : Ok(ImageLayerCreationOutcome {
4540 412 : unfinished_image_layer: None,
4541 412 : next_start_key: start,
4542 412 : })
4543 : }
4544 440 : }
4545 :
4546 : /// Predicate function which indicates whether we should check if new image layers
4547 : /// are required. Since checking if new image layers are required is expensive in
4548 : /// terms of CPU, we only do it in the following cases:
4549 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4550 : /// 2. If enough time has passed since the last check:
4551 : /// 1. For large tenants, we wish to perform the check more often since they
4552 : /// suffer from the lack of image layers
4553 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4554 1136 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4555 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4556 :
4557 1136 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4558 1136 : let distance = lsn
4559 1136 : .checked_sub(last_checks_at)
4560 1136 : .expect("Attempt to compact with LSN going backwards");
4561 1136 : let min_distance =
4562 1136 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4563 1136 :
4564 1136 : let distance_based_decision = distance.0 >= min_distance;
4565 1136 :
4566 1136 : let mut time_based_decision = false;
4567 1136 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4568 1136 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4569 932 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4570 : {
4571 0 : self.get_checkpoint_timeout()
4572 : } else {
4573 932 : Duration::from_secs(3600 * 48)
4574 : };
4575 :
4576 932 : time_based_decision = match *last_check_instant {
4577 524 : Some(last_check) => {
4578 524 : let elapsed = last_check.elapsed();
4579 524 : elapsed >= check_required_after
4580 : }
4581 408 : None => true,
4582 : };
4583 204 : }
4584 :
4585 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4586 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4587 : // check.
4588 1136 : let decision = distance_based_decision || time_based_decision;
4589 :
4590 1136 : if decision {
4591 412 : self.last_image_layer_creation_check_at.store(lsn);
4592 412 : *last_check_instant = Some(Instant::now());
4593 724 : }
4594 :
4595 1136 : decision
4596 1136 : }
4597 :
4598 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4599 : async fn create_image_layers(
4600 : self: &Arc<Timeline>,
4601 : partitioning: &KeyPartitioning,
4602 : lsn: Lsn,
4603 : mode: ImageLayerCreationMode,
4604 : ctx: &RequestContext,
4605 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4606 : let timer = self.metrics.create_images_time_histo.start_timer();
4607 :
4608 : // We need to avoid holes between generated image layers.
4609 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4610 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4611 : //
4612 : // How such hole between partitions can appear?
4613 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4614 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4615 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4616 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4617 : let mut start = Key::MIN;
4618 :
4619 : let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
4620 :
4621 : let mut batch_image_writer = BatchLayerWriter::new(self.conf).await?;
4622 :
4623 : for partition in partitioning.parts.iter() {
4624 : if self.cancel.is_cancelled() {
4625 : return Err(CreateImageLayersError::Cancelled);
4626 : }
4627 :
4628 : let img_range = start..partition.ranges.last().unwrap().end;
4629 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4630 : if compact_metadata {
4631 : for range in &partition.ranges {
4632 : assert!(
4633 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4634 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4635 : "metadata keys must be partitioned separately"
4636 : );
4637 : }
4638 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4639 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4640 : // might mess up with evictions.
4641 : start = img_range.end;
4642 : continue;
4643 : }
4644 : // For initial and force modes, we always generate image layers for metadata keys.
4645 : } else if let ImageLayerCreationMode::Try = mode {
4646 : // check_for_image_layers = false -> skip
4647 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4648 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4649 : start = img_range.end;
4650 : continue;
4651 : }
4652 : }
4653 : if let ImageLayerCreationMode::Force = mode {
4654 : // When forced to create image layers, we might try and create them where they already
4655 : // exist. This mode is only used in tests/debug.
4656 : let layers = self.layers.read().await;
4657 : if layers.contains_key(&PersistentLayerKey {
4658 : key_range: img_range.clone(),
4659 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4660 : is_delta: false,
4661 : }) {
4662 : tracing::info!(
4663 : "Skipping image layer at {lsn} {}..{}, already exists",
4664 : img_range.start,
4665 : img_range.end
4666 : );
4667 : start = img_range.end;
4668 : continue;
4669 : }
4670 : }
4671 :
4672 : let image_layer_writer = ImageLayerWriter::new(
4673 : self.conf,
4674 : self.timeline_id,
4675 : self.tenant_shard_id,
4676 : &img_range,
4677 : lsn,
4678 : ctx,
4679 : )
4680 : .await?;
4681 :
4682 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4683 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4684 0 : "failpoint image-layer-writer-fail-before-finish"
4685 0 : )))
4686 0 : });
4687 :
4688 : let io_concurrency = IoConcurrency::spawn_from_conf(
4689 : self.conf,
4690 : self.gate
4691 : .enter()
4692 0 : .map_err(|_| CreateImageLayersError::Cancelled)?,
4693 : );
4694 :
4695 : let ImageLayerCreationOutcome {
4696 : unfinished_image_layer,
4697 : next_start_key,
4698 : } = if !compact_metadata {
4699 : self.create_image_layer_for_rel_blocks(
4700 : partition,
4701 : image_layer_writer,
4702 : lsn,
4703 : ctx,
4704 : img_range.clone(),
4705 : start,
4706 : io_concurrency,
4707 : )
4708 : .await?
4709 : } else {
4710 : self.create_image_layer_for_metadata_keys(
4711 : partition,
4712 : image_layer_writer,
4713 : lsn,
4714 : ctx,
4715 : img_range.clone(),
4716 : mode,
4717 : start,
4718 : io_concurrency,
4719 : )
4720 : .await?
4721 : };
4722 : start = next_start_key;
4723 : if let Some(unfinished_image_layer) = unfinished_image_layer {
4724 : batch_image_writer.add_unfinished_image_writer(
4725 : unfinished_image_layer,
4726 : img_range,
4727 : lsn,
4728 : );
4729 : }
4730 : }
4731 :
4732 : let image_layers = batch_image_writer.finish(self, ctx).await?;
4733 :
4734 : let mut guard = self.layers.write().await;
4735 :
4736 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4737 : // now they are being scheduled outside of write lock; current way is inconsistent with
4738 : // compaction lock order.
4739 : guard
4740 : .open_mut()?
4741 : .track_new_image_layers(&image_layers, &self.metrics);
4742 : drop_wlock(guard);
4743 : timer.stop_and_record();
4744 :
4745 : // Creating image layers may have caused some previously visible layers to be covered
4746 : if !image_layers.is_empty() {
4747 : self.update_layer_visibility().await?;
4748 : }
4749 :
4750 : Ok(image_layers)
4751 : }
4752 :
4753 : /// Wait until the background initial logical size calculation is complete, or
4754 : /// this Timeline is shut down. Calling this function will cause the initial
4755 : /// logical size calculation to skip waiting for the background jobs barrier.
4756 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4757 0 : if !self.shard_identity.is_shard_zero() {
4758 : // We don't populate logical size on shard >0: skip waiting for it.
4759 0 : return;
4760 0 : }
4761 0 :
4762 0 : if self.remote_client.is_deleting() {
4763 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4764 0 : return;
4765 0 : }
4766 0 :
4767 0 : if self.current_logical_size.current_size().is_exact() {
4768 : // root timelines are initialized with exact count, but never start the background
4769 : // calculation
4770 0 : return;
4771 0 : }
4772 :
4773 0 : if let Some(await_bg_cancel) = self
4774 0 : .current_logical_size
4775 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4776 0 : .get()
4777 0 : {
4778 0 : await_bg_cancel.cancel();
4779 0 : } else {
4780 : // We should not wait if we were not able to explicitly instruct
4781 : // the logical size cancellation to skip the concurrency limit semaphore.
4782 : // TODO: this is an unexpected case. We should restructure so that it
4783 : // can't happen.
4784 0 : tracing::warn!(
4785 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4786 : );
4787 0 : debug_assert!(false);
4788 : }
4789 :
4790 0 : tokio::select!(
4791 0 : _ = self.current_logical_size.initialized.acquire() => {},
4792 0 : _ = self.cancel.cancelled() => {}
4793 : )
4794 0 : }
4795 :
4796 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4797 : /// Timelines layers up to the ancestor_lsn.
4798 : ///
4799 : /// Requires a timeline that:
4800 : /// - has an ancestor to detach from
4801 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4802 : /// a technical requirement
4803 : ///
4804 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4805 : /// polled again until completion.
4806 : ///
4807 : /// During the operation all timelines sharing the data with this timeline will be reparented
4808 : /// from our ancestor to be branches of this timeline.
4809 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4810 0 : self: &Arc<Timeline>,
4811 0 : tenant: &crate::tenant::Tenant,
4812 0 : options: detach_ancestor::Options,
4813 0 : ctx: &RequestContext,
4814 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
4815 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4816 0 : }
4817 :
4818 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
4819 : /// reparents any reparentable children of previous ancestor.
4820 : ///
4821 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
4822 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
4823 : /// successfully, tenant must be reloaded.
4824 : ///
4825 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
4826 : /// resetting the tenant.
4827 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
4828 0 : self: &Arc<Timeline>,
4829 0 : tenant: &crate::tenant::Tenant,
4830 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4831 0 : ctx: &RequestContext,
4832 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
4833 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
4834 0 : }
4835 :
4836 : /// Final step which unblocks the GC.
4837 : ///
4838 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
4839 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4840 0 : self: &Arc<Timeline>,
4841 0 : tenant: &crate::tenant::Tenant,
4842 0 : attempt: detach_ancestor::Attempt,
4843 0 : ctx: &RequestContext,
4844 0 : ) -> Result<(), detach_ancestor::Error> {
4845 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
4846 0 : }
4847 : }
4848 :
4849 : impl Drop for Timeline {
4850 20 : fn drop(&mut self) {
4851 20 : if let Some(ancestor) = &self.ancestor_timeline {
4852 : // This lock should never be poisoned, but in case it is we do a .map() instead of
4853 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
4854 8 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
4855 8 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
4856 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
4857 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
4858 8 : }
4859 0 : }
4860 12 : }
4861 20 : info!(
4862 0 : "Timeline {} for tenant {} is being dropped",
4863 : self.timeline_id, self.tenant_shard_id.tenant_id
4864 : );
4865 20 : }
4866 : }
4867 :
4868 : /// Top-level failure to compact.
4869 : #[derive(Debug, thiserror::Error)]
4870 : pub(crate) enum CompactionError {
4871 : #[error("The timeline or pageserver is shutting down")]
4872 : ShuttingDown,
4873 : /// Compaction tried to offload a timeline and failed
4874 : #[error("Failed to offload timeline: {0}")]
4875 : Offload(OffloadError),
4876 : /// Compaction cannot be done right now; page reconstruction and so on.
4877 : #[error(transparent)]
4878 : Other(anyhow::Error),
4879 : }
4880 :
4881 : impl From<OffloadError> for CompactionError {
4882 0 : fn from(e: OffloadError) -> Self {
4883 0 : match e {
4884 0 : OffloadError::Cancelled => Self::ShuttingDown,
4885 0 : _ => Self::Offload(e),
4886 : }
4887 0 : }
4888 : }
4889 :
4890 : impl CompactionError {
4891 0 : pub fn is_cancelled(&self) -> bool {
4892 0 : matches!(self, CompactionError::ShuttingDown)
4893 0 : }
4894 : }
4895 :
4896 : impl From<CollectKeySpaceError> for CompactionError {
4897 0 : fn from(err: CollectKeySpaceError) -> Self {
4898 0 : match err {
4899 : CollectKeySpaceError::Cancelled
4900 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4901 0 : CompactionError::ShuttingDown
4902 : }
4903 0 : e => CompactionError::Other(e.into()),
4904 : }
4905 0 : }
4906 : }
4907 :
4908 : impl From<super::upload_queue::NotInitialized> for CompactionError {
4909 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
4910 0 : match value {
4911 : super::upload_queue::NotInitialized::Uninitialized => {
4912 0 : CompactionError::Other(anyhow::anyhow!(value))
4913 : }
4914 : super::upload_queue::NotInitialized::ShuttingDown
4915 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
4916 : }
4917 0 : }
4918 : }
4919 :
4920 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
4921 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
4922 0 : match e {
4923 : super::storage_layer::layer::DownloadError::TimelineShutdown
4924 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
4925 0 : CompactionError::ShuttingDown
4926 : }
4927 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
4928 : | super::storage_layer::layer::DownloadError::DownloadRequired
4929 : | super::storage_layer::layer::DownloadError::NotFile(_)
4930 : | super::storage_layer::layer::DownloadError::DownloadFailed
4931 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
4932 0 : CompactionError::Other(anyhow::anyhow!(e))
4933 : }
4934 : #[cfg(test)]
4935 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
4936 0 : CompactionError::Other(anyhow::anyhow!(e))
4937 : }
4938 : }
4939 0 : }
4940 : }
4941 :
4942 : impl From<layer_manager::Shutdown> for CompactionError {
4943 0 : fn from(_: layer_manager::Shutdown) -> Self {
4944 0 : CompactionError::ShuttingDown
4945 0 : }
4946 : }
4947 :
4948 : #[serde_as]
4949 392 : #[derive(serde::Serialize)]
4950 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4951 :
4952 : #[derive(Default)]
4953 : enum DurationRecorder {
4954 : #[default]
4955 : NotStarted,
4956 : Recorded(RecordedDuration, tokio::time::Instant),
4957 : }
4958 :
4959 : impl DurationRecorder {
4960 1008 : fn till_now(&self) -> DurationRecorder {
4961 1008 : match self {
4962 : DurationRecorder::NotStarted => {
4963 0 : panic!("must only call on recorded measurements")
4964 : }
4965 1008 : DurationRecorder::Recorded(_, ended) => {
4966 1008 : let now = tokio::time::Instant::now();
4967 1008 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4968 1008 : }
4969 1008 : }
4970 1008 : }
4971 392 : fn into_recorded(self) -> Option<RecordedDuration> {
4972 392 : match self {
4973 0 : DurationRecorder::NotStarted => None,
4974 392 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4975 : }
4976 392 : }
4977 : }
4978 :
4979 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
4980 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
4981 : /// the layer descriptor requires the user to provide the ranges, which should cover all
4982 : /// keys specified in the `data` field.
4983 : #[cfg(test)]
4984 : #[derive(Clone)]
4985 : pub struct DeltaLayerTestDesc {
4986 : pub lsn_range: Range<Lsn>,
4987 : pub key_range: Range<Key>,
4988 : pub data: Vec<(Key, Lsn, Value)>,
4989 : }
4990 :
4991 : #[cfg(test)]
4992 : impl DeltaLayerTestDesc {
4993 4 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
4994 4 : Self {
4995 4 : lsn_range,
4996 4 : key_range,
4997 4 : data,
4998 4 : }
4999 4 : }
5000 :
5001 176 : pub fn new_with_inferred_key_range(
5002 176 : lsn_range: Range<Lsn>,
5003 176 : data: Vec<(Key, Lsn, Value)>,
5004 176 : ) -> Self {
5005 440 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
5006 440 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
5007 176 : Self {
5008 176 : key_range: (*key_min)..(key_max.next()),
5009 176 : lsn_range,
5010 176 : data,
5011 176 : }
5012 176 : }
5013 :
5014 20 : pub(crate) fn layer_name(&self) -> LayerName {
5015 20 : LayerName::Delta(super::storage_layer::DeltaLayerName {
5016 20 : key_range: self.key_range.clone(),
5017 20 : lsn_range: self.lsn_range.clone(),
5018 20 : })
5019 20 : }
5020 : }
5021 :
5022 : impl Timeline {
5023 56 : async fn finish_compact_batch(
5024 56 : self: &Arc<Self>,
5025 56 : new_deltas: &[ResidentLayer],
5026 56 : new_images: &[ResidentLayer],
5027 56 : layers_to_remove: &[Layer],
5028 56 : ) -> Result<(), CompactionError> {
5029 56 : let mut guard = tokio::select! {
5030 56 : guard = self.layers.write() => guard,
5031 56 : _ = self.cancel.cancelled() => {
5032 0 : return Err(CompactionError::ShuttingDown);
5033 : }
5034 : };
5035 :
5036 56 : let mut duplicated_layers = HashSet::new();
5037 56 :
5038 56 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
5039 :
5040 672 : for l in new_deltas {
5041 616 : if guard.contains(l.as_ref()) {
5042 : // expected in tests
5043 0 : tracing::error!(layer=%l, "duplicated L1 layer");
5044 :
5045 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
5046 : // `guard` on self.layers. as of writing this, there are no error returns except
5047 : // for compact_level0_phase1 creating an L0, which does not happen in practice
5048 : // because we have not implemented L0 => L0 compaction.
5049 0 : duplicated_layers.insert(l.layer_desc().key());
5050 616 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
5051 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
5052 616 : } else {
5053 616 : insert_layers.push(l.clone());
5054 616 : }
5055 : }
5056 :
5057 : // only remove those inputs which were not outputs
5058 56 : let remove_layers: Vec<Layer> = layers_to_remove
5059 56 : .iter()
5060 804 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
5061 56 : .cloned()
5062 56 : .collect();
5063 56 :
5064 56 : if !new_images.is_empty() {
5065 0 : guard
5066 0 : .open_mut()?
5067 0 : .track_new_image_layers(new_images, &self.metrics);
5068 56 : }
5069 :
5070 56 : guard
5071 56 : .open_mut()?
5072 56 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
5073 56 :
5074 56 : self.remote_client
5075 56 : .schedule_compaction_update(&remove_layers, new_deltas)?;
5076 :
5077 56 : drop_wlock(guard);
5078 56 :
5079 56 : Ok(())
5080 56 : }
5081 :
5082 0 : async fn rewrite_layers(
5083 0 : self: &Arc<Self>,
5084 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
5085 0 : mut drop_layers: Vec<Layer>,
5086 0 : ) -> Result<(), CompactionError> {
5087 0 : let mut guard = self.layers.write().await;
5088 :
5089 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
5090 : // to avoid double-removing, and avoid rewriting something that was removed.
5091 0 : replace_layers.retain(|(l, _)| guard.contains(l));
5092 0 : drop_layers.retain(|l| guard.contains(l));
5093 0 :
5094 0 : guard
5095 0 : .open_mut()?
5096 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
5097 0 :
5098 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
5099 0 :
5100 0 : self.remote_client
5101 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
5102 :
5103 0 : Ok(())
5104 0 : }
5105 :
5106 : /// Schedules the uploads of the given image layers
5107 728 : fn upload_new_image_layers(
5108 728 : self: &Arc<Self>,
5109 728 : new_images: impl IntoIterator<Item = ResidentLayer>,
5110 728 : ) -> Result<(), super::upload_queue::NotInitialized> {
5111 780 : for layer in new_images {
5112 52 : self.remote_client.schedule_layer_file_upload(layer)?;
5113 : }
5114 : // should any new image layer been created, not uploading index_part will
5115 : // result in a mismatch between remote_physical_size and layermap calculated
5116 : // size, which will fail some tests, but should not be an issue otherwise.
5117 728 : self.remote_client
5118 728 : .schedule_index_upload_for_file_changes()?;
5119 728 : Ok(())
5120 728 : }
5121 :
5122 0 : async fn find_gc_time_cutoff(
5123 0 : &self,
5124 0 : now: SystemTime,
5125 0 : pitr: Duration,
5126 0 : cancel: &CancellationToken,
5127 0 : ctx: &RequestContext,
5128 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
5129 0 : debug_assert_current_span_has_tenant_and_timeline_id();
5130 0 : if self.shard_identity.is_shard_zero() {
5131 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
5132 0 : let time_range = if pitr == Duration::ZERO {
5133 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
5134 : } else {
5135 0 : pitr
5136 : };
5137 :
5138 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
5139 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
5140 0 : let timestamp = to_pg_timestamp(time_cutoff);
5141 :
5142 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
5143 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
5144 0 : LsnForTimestamp::Future(lsn) => {
5145 0 : // The timestamp is in the future. That sounds impossible,
5146 0 : // but what it really means is that there hasn't been
5147 0 : // any commits since the cutoff timestamp.
5148 0 : //
5149 0 : // In this case we should use the LSN of the most recent commit,
5150 0 : // which is implicitly the last LSN in the log.
5151 0 : debug!("future({})", lsn);
5152 0 : Some(self.get_last_record_lsn())
5153 : }
5154 0 : LsnForTimestamp::Past(lsn) => {
5155 0 : debug!("past({})", lsn);
5156 0 : None
5157 : }
5158 0 : LsnForTimestamp::NoData(lsn) => {
5159 0 : debug!("nodata({})", lsn);
5160 0 : None
5161 : }
5162 : };
5163 0 : Ok(time_cutoff)
5164 : } else {
5165 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
5166 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
5167 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
5168 : // space cutoff that we would also have respected ourselves.
5169 0 : match self
5170 0 : .remote_client
5171 0 : .download_foreign_index(ShardNumber(0), cancel)
5172 0 : .await
5173 : {
5174 0 : Ok((index_part, index_generation, _index_mtime)) => {
5175 0 : tracing::info!("GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
5176 0 : index_part.metadata.latest_gc_cutoff_lsn());
5177 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
5178 : }
5179 : Err(DownloadError::NotFound) => {
5180 : // This is unexpected, because during timeline creations shard zero persists to remote
5181 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
5182 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
5183 : // state, then shard zero should _eventually_ write an index when it GCs.
5184 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
5185 0 : Ok(None)
5186 : }
5187 0 : Err(e) => {
5188 0 : // TODO: this function should return a different error type than page reconstruct error
5189 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
5190 : }
5191 : }
5192 :
5193 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
5194 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
5195 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
5196 : // new location. This is legal in principle, but problematic in practice because it might result
5197 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
5198 : // because they have GC'd past the branch point.
5199 : }
5200 0 : }
5201 :
5202 : /// Find the Lsns above which layer files need to be retained on
5203 : /// garbage collection.
5204 : ///
5205 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
5206 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
5207 : /// the space-based retention.
5208 : ///
5209 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
5210 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
5211 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
5212 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
5213 : /// in the response as the GC cutoff point for the timeline.
5214 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
5215 : pub(super) async fn find_gc_cutoffs(
5216 : &self,
5217 : now: SystemTime,
5218 : space_cutoff: Lsn,
5219 : pitr: Duration,
5220 : cancel: &CancellationToken,
5221 : ctx: &RequestContext,
5222 : ) -> Result<GcCutoffs, PageReconstructError> {
5223 : let _timer = self
5224 : .metrics
5225 : .find_gc_cutoffs_histo
5226 : .start_timer()
5227 : .record_on_drop();
5228 :
5229 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
5230 :
5231 : if cfg!(test) {
5232 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
5233 : if pitr == Duration::ZERO {
5234 : return Ok(GcCutoffs {
5235 : time: self.get_last_record_lsn(),
5236 : space: space_cutoff,
5237 : });
5238 : }
5239 : }
5240 :
5241 : // Calculate a time-based limit on how much to retain:
5242 : // - if PITR interval is set, then this is our cutoff.
5243 : // - if PITR interval is not set, then we do a lookup
5244 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
5245 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
5246 :
5247 : Ok(match (pitr, time_cutoff) {
5248 : (Duration::ZERO, Some(time_cutoff)) => {
5249 : // PITR is not set. Retain the size-based limit, or the default time retention,
5250 : // whichever requires less data.
5251 : GcCutoffs {
5252 : time: self.get_last_record_lsn(),
5253 : space: std::cmp::max(time_cutoff, space_cutoff),
5254 : }
5255 : }
5256 : (Duration::ZERO, None) => {
5257 : // PITR is not set, and time lookup failed
5258 : GcCutoffs {
5259 : time: self.get_last_record_lsn(),
5260 : space: space_cutoff,
5261 : }
5262 : }
5263 : (_, None) => {
5264 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
5265 : // cannot advance beyond what was already GC'd, and respect space-based retention
5266 : GcCutoffs {
5267 : time: *self.get_latest_gc_cutoff_lsn(),
5268 : space: space_cutoff,
5269 : }
5270 : }
5271 : (_, Some(time_cutoff)) => {
5272 : // PITR interval is set and we looked up timestamp successfully. Ignore
5273 : // size based retention and make time cutoff authoritative
5274 : GcCutoffs {
5275 : time: time_cutoff,
5276 : space: time_cutoff,
5277 : }
5278 : }
5279 : })
5280 : }
5281 :
5282 : /// Garbage collect layer files on a timeline that are no longer needed.
5283 : ///
5284 : /// Currently, we don't make any attempt at removing unneeded page versions
5285 : /// within a layer file. We can only remove the whole file if it's fully
5286 : /// obsolete.
5287 8 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
5288 : // this is most likely the background tasks, but it might be the spawned task from
5289 : // immediate_gc
5290 8 : let _g = tokio::select! {
5291 8 : guard = self.gc_lock.lock() => guard,
5292 8 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
5293 : };
5294 8 : let timer = self.metrics.garbage_collect_histo.start_timer();
5295 8 :
5296 8 : fail_point!("before-timeline-gc");
5297 8 :
5298 8 : // Is the timeline being deleted?
5299 8 : if self.is_stopping() {
5300 0 : return Err(GcError::TimelineCancelled);
5301 8 : }
5302 8 :
5303 8 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
5304 8 : let gc_info = self.gc_info.read().unwrap();
5305 8 :
5306 8 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
5307 8 : let time_cutoff = gc_info.cutoffs.time;
5308 8 : let retain_lsns = gc_info
5309 8 : .retain_lsns
5310 8 : .iter()
5311 8 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
5312 8 : .collect();
5313 8 :
5314 8 : // Gets the maximum LSN that holds the valid lease.
5315 8 : //
5316 8 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
5317 8 : // Here, we do not check for stale leases again.
5318 8 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
5319 8 :
5320 8 : (
5321 8 : space_cutoff,
5322 8 : time_cutoff,
5323 8 : retain_lsns,
5324 8 : max_lsn_with_valid_lease,
5325 8 : )
5326 8 : };
5327 8 :
5328 8 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
5329 8 : let standby_horizon = self.standby_horizon.load();
5330 8 : // Hold GC for the standby, but as a safety guard do it only within some
5331 8 : // reasonable lag.
5332 8 : if standby_horizon != Lsn::INVALID {
5333 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
5334 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
5335 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
5336 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
5337 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
5338 : } else {
5339 0 : warn!(
5340 0 : "standby is lagging for more than {}MB, not holding gc for it",
5341 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
5342 : )
5343 : }
5344 0 : }
5345 8 : }
5346 :
5347 : // Reset standby horizon to ignore it if it is not updated till next GC.
5348 : // It is an easy way to unset it when standby disappears without adding
5349 : // more conf options.
5350 8 : self.standby_horizon.store(Lsn::INVALID);
5351 8 : self.metrics
5352 8 : .standby_horizon_gauge
5353 8 : .set(Lsn::INVALID.0 as i64);
5354 :
5355 8 : let res = self
5356 8 : .gc_timeline(
5357 8 : space_cutoff,
5358 8 : time_cutoff,
5359 8 : retain_lsns,
5360 8 : max_lsn_with_valid_lease,
5361 8 : new_gc_cutoff,
5362 8 : )
5363 8 : .instrument(
5364 8 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5365 : )
5366 8 : .await?;
5367 :
5368 : // only record successes
5369 8 : timer.stop_and_record();
5370 8 :
5371 8 : Ok(res)
5372 8 : }
5373 :
5374 8 : async fn gc_timeline(
5375 8 : &self,
5376 8 : space_cutoff: Lsn,
5377 8 : time_cutoff: Lsn,
5378 8 : retain_lsns: Vec<Lsn>,
5379 8 : max_lsn_with_valid_lease: Option<Lsn>,
5380 8 : new_gc_cutoff: Lsn,
5381 8 : ) -> Result<GcResult, GcError> {
5382 8 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5383 8 :
5384 8 : let now = SystemTime::now();
5385 8 : let mut result: GcResult = GcResult::default();
5386 8 :
5387 8 : // Nothing to GC. Return early.
5388 8 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
5389 8 : if latest_gc_cutoff >= new_gc_cutoff {
5390 0 : info!(
5391 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5392 : );
5393 0 : return Ok(result);
5394 8 : }
5395 :
5396 : // We need to ensure that no one tries to read page versions or create
5397 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5398 : // for details. This will block until the old value is no longer in use.
5399 : //
5400 : // The GC cutoff should only ever move forwards.
5401 8 : let waitlist = {
5402 8 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
5403 8 : if *write_guard > new_gc_cutoff {
5404 0 : return Err(GcError::BadLsn {
5405 0 : why: format!(
5406 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5407 0 : *write_guard, new_gc_cutoff
5408 0 : ),
5409 0 : });
5410 8 : }
5411 8 :
5412 8 : write_guard.store_and_unlock(new_gc_cutoff)
5413 8 : };
5414 8 : waitlist.wait().await;
5415 :
5416 8 : info!("GC starting");
5417 :
5418 8 : debug!("retain_lsns: {:?}", retain_lsns);
5419 :
5420 8 : let mut layers_to_remove = Vec::new();
5421 :
5422 : // Scan all layers in the timeline (remote or on-disk).
5423 : //
5424 : // Garbage collect the layer if all conditions are satisfied:
5425 : // 1. it is older than cutoff LSN;
5426 : // 2. it is older than PITR interval;
5427 : // 3. it doesn't need to be retained for 'retain_lsns';
5428 : // 4. it does not need to be kept for LSNs holding valid leases.
5429 : // 5. newer on-disk image layers cover the layer's whole key range
5430 : //
5431 : // TODO holding a write lock is too agressive and avoidable
5432 8 : let mut guard = self.layers.write().await;
5433 8 : let layers = guard.layer_map()?;
5434 48 : 'outer: for l in layers.iter_historic_layers() {
5435 48 : result.layers_total += 1;
5436 48 :
5437 48 : // 1. Is it newer than GC horizon cutoff point?
5438 48 : if l.get_lsn_range().end > space_cutoff {
5439 4 : info!(
5440 0 : "keeping {} because it's newer than space_cutoff {}",
5441 0 : l.layer_name(),
5442 : space_cutoff,
5443 : );
5444 4 : result.layers_needed_by_cutoff += 1;
5445 4 : continue 'outer;
5446 44 : }
5447 44 :
5448 44 : // 2. It is newer than PiTR cutoff point?
5449 44 : if l.get_lsn_range().end > time_cutoff {
5450 0 : info!(
5451 0 : "keeping {} because it's newer than time_cutoff {}",
5452 0 : l.layer_name(),
5453 : time_cutoff,
5454 : );
5455 0 : result.layers_needed_by_pitr += 1;
5456 0 : continue 'outer;
5457 44 : }
5458 :
5459 : // 3. Is it needed by a child branch?
5460 : // NOTE With that we would keep data that
5461 : // might be referenced by child branches forever.
5462 : // We can track this in child timeline GC and delete parent layers when
5463 : // they are no longer needed. This might be complicated with long inheritance chains.
5464 : //
5465 : // TODO Vec is not a great choice for `retain_lsns`
5466 44 : for retain_lsn in &retain_lsns {
5467 : // start_lsn is inclusive
5468 0 : if &l.get_lsn_range().start <= retain_lsn {
5469 0 : info!(
5470 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5471 0 : l.layer_name(),
5472 0 : retain_lsn,
5473 0 : l.is_incremental(),
5474 : );
5475 0 : result.layers_needed_by_branches += 1;
5476 0 : continue 'outer;
5477 0 : }
5478 : }
5479 :
5480 : // 4. Is there a valid lease that requires us to keep this layer?
5481 44 : if let Some(lsn) = &max_lsn_with_valid_lease {
5482 : // keep if layer start <= any of the lease
5483 36 : if &l.get_lsn_range().start <= lsn {
5484 28 : info!(
5485 0 : "keeping {} because there is a valid lease preventing GC at {}",
5486 0 : l.layer_name(),
5487 : lsn,
5488 : );
5489 28 : result.layers_needed_by_leases += 1;
5490 28 : continue 'outer;
5491 8 : }
5492 8 : }
5493 :
5494 : // 5. Is there a later on-disk layer for this relation?
5495 : //
5496 : // The end-LSN is exclusive, while disk_consistent_lsn is
5497 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5498 : // OK for a delta layer to have end LSN 101, but if the end LSN
5499 : // is 102, then it might not have been fully flushed to disk
5500 : // before crash.
5501 : //
5502 : // For example, imagine that the following layers exist:
5503 : //
5504 : // 1000 - image (A)
5505 : // 1000-2000 - delta (B)
5506 : // 2000 - image (C)
5507 : // 2000-3000 - delta (D)
5508 : // 3000 - image (E)
5509 : //
5510 : // If GC horizon is at 2500, we can remove layers A and B, but
5511 : // we cannot remove C, even though it's older than 2500, because
5512 : // the delta layer 2000-3000 depends on it.
5513 16 : if !layers
5514 16 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5515 : {
5516 12 : info!("keeping {} because it is the latest layer", l.layer_name());
5517 12 : result.layers_not_updated += 1;
5518 12 : continue 'outer;
5519 4 : }
5520 4 :
5521 4 : // We didn't find any reason to keep this file, so remove it.
5522 4 : info!(
5523 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5524 0 : l.layer_name(),
5525 0 : l.is_incremental(),
5526 : );
5527 4 : layers_to_remove.push(l);
5528 : }
5529 :
5530 8 : if !layers_to_remove.is_empty() {
5531 : // Persist the new GC cutoff value before we actually remove anything.
5532 : // This unconditionally schedules also an index_part.json update, even though, we will
5533 : // be doing one a bit later with the unlinked gc'd layers.
5534 4 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5535 4 : self.schedule_uploads(disk_consistent_lsn, None)
5536 4 : .map_err(|e| {
5537 0 : if self.cancel.is_cancelled() {
5538 0 : GcError::TimelineCancelled
5539 : } else {
5540 0 : GcError::Remote(e)
5541 : }
5542 4 : })?;
5543 :
5544 4 : let gc_layers = layers_to_remove
5545 4 : .iter()
5546 4 : .map(|x| guard.get_from_desc(x))
5547 4 : .collect::<Vec<Layer>>();
5548 4 :
5549 4 : result.layers_removed = gc_layers.len() as u64;
5550 4 :
5551 4 : self.remote_client.schedule_gc_update(&gc_layers)?;
5552 :
5553 4 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5554 4 :
5555 4 : #[cfg(feature = "testing")]
5556 4 : {
5557 4 : result.doomed_layers = gc_layers;
5558 4 : }
5559 4 : }
5560 :
5561 8 : info!(
5562 0 : "GC completed removing {} layers, cutoff {}",
5563 : result.layers_removed, new_gc_cutoff
5564 : );
5565 :
5566 8 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5567 8 : Ok(result)
5568 8 : }
5569 :
5570 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5571 1339976 : async fn reconstruct_value(
5572 1339976 : &self,
5573 1339976 : key: Key,
5574 1339976 : request_lsn: Lsn,
5575 1339976 : mut data: ValueReconstructState,
5576 1339976 : ) -> Result<Bytes, PageReconstructError> {
5577 1339976 : // Perform WAL redo if needed
5578 1339976 : data.records.reverse();
5579 1339976 :
5580 1339976 : // If we have a page image, and no WAL, we're all set
5581 1339976 : if data.records.is_empty() {
5582 1338340 : if let Some((img_lsn, img)) = &data.img {
5583 1338340 : trace!(
5584 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5585 : key,
5586 : img_lsn,
5587 : request_lsn,
5588 : );
5589 1338340 : Ok(img.clone())
5590 : } else {
5591 0 : Err(PageReconstructError::from(anyhow!(
5592 0 : "base image for {key} at {request_lsn} not found"
5593 0 : )))
5594 : }
5595 : } else {
5596 : // We need to do WAL redo.
5597 : //
5598 : // If we don't have a base image, then the oldest WAL record better initialize
5599 : // the page
5600 1636 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5601 0 : Err(PageReconstructError::from(anyhow!(
5602 0 : "Base image for {} at {} not found, but got {} WAL records",
5603 0 : key,
5604 0 : request_lsn,
5605 0 : data.records.len()
5606 0 : )))
5607 : } else {
5608 1636 : if data.img.is_some() {
5609 1504 : trace!(
5610 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5611 0 : data.records.len(),
5612 : key,
5613 : request_lsn
5614 : );
5615 : } else {
5616 132 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5617 : };
5618 1636 : let res = self
5619 1636 : .walredo_mgr
5620 1636 : .as_ref()
5621 1636 : .context("timeline has no walredo manager")
5622 1636 : .map_err(PageReconstructError::WalRedo)?
5623 1636 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5624 1636 : .await;
5625 1636 : let img = match res {
5626 1636 : Ok(img) => img,
5627 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5628 0 : Err(walredo::Error::Other(e)) => {
5629 0 : return Err(PageReconstructError::WalRedo(
5630 0 : e.context("reconstruct a page image"),
5631 0 : ))
5632 : }
5633 : };
5634 1636 : Ok(img)
5635 : }
5636 : }
5637 1339976 : }
5638 :
5639 0 : pub(crate) async fn spawn_download_all_remote_layers(
5640 0 : self: Arc<Self>,
5641 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5642 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5643 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5644 :
5645 : // this is not really needed anymore; it has tests which really check the return value from
5646 : // http api. it would be better not to maintain this anymore.
5647 :
5648 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5649 0 : if let Some(st) = &*status_guard {
5650 0 : match &st.state {
5651 : DownloadRemoteLayersTaskState::Running => {
5652 0 : return Err(st.clone());
5653 : }
5654 : DownloadRemoteLayersTaskState::ShutDown
5655 0 : | DownloadRemoteLayersTaskState::Completed => {
5656 0 : *status_guard = None;
5657 0 : }
5658 : }
5659 0 : }
5660 :
5661 0 : let self_clone = Arc::clone(&self);
5662 0 : let task_id = task_mgr::spawn(
5663 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5664 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5665 0 : self.tenant_shard_id,
5666 0 : Some(self.timeline_id),
5667 0 : "download all remote layers task",
5668 0 : async move {
5669 0 : self_clone.download_all_remote_layers(request).await;
5670 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5671 0 : match &mut *status_guard {
5672 : None => {
5673 0 : warn!("tasks status is supposed to be Some(), since we are running");
5674 : }
5675 0 : Some(st) => {
5676 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5677 0 : if st.task_id != exp_task_id {
5678 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5679 0 : } else {
5680 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5681 0 : }
5682 : }
5683 : };
5684 0 : Ok(())
5685 0 : }
5686 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))
5687 : );
5688 :
5689 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5690 0 : task_id: format!("{task_id}"),
5691 0 : state: DownloadRemoteLayersTaskState::Running,
5692 0 : total_layer_count: 0,
5693 0 : successful_download_count: 0,
5694 0 : failed_download_count: 0,
5695 0 : };
5696 0 : *status_guard = Some(initial_info.clone());
5697 0 :
5698 0 : Ok(initial_info)
5699 0 : }
5700 :
5701 0 : async fn download_all_remote_layers(
5702 0 : self: &Arc<Self>,
5703 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5704 0 : ) {
5705 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5706 :
5707 0 : let remaining = {
5708 0 : let guard = self.layers.read().await;
5709 0 : let Ok(lm) = guard.layer_map() else {
5710 : // technically here we could look into iterating accessible layers, but downloading
5711 : // all layers of a shutdown timeline makes no sense regardless.
5712 0 : tracing::info!("attempted to download all layers of shutdown timeline");
5713 0 : return;
5714 : };
5715 0 : lm.iter_historic_layers()
5716 0 : .map(|desc| guard.get_from_desc(&desc))
5717 0 : .collect::<Vec<_>>()
5718 0 : };
5719 0 : let total_layer_count = remaining.len();
5720 :
5721 : macro_rules! lock_status {
5722 : ($st:ident) => {
5723 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5724 : let st = st
5725 : .as_mut()
5726 : .expect("this function is only called after the task has been spawned");
5727 : assert_eq!(
5728 : st.task_id,
5729 : format!(
5730 : "{}",
5731 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5732 : )
5733 : );
5734 : let $st = st;
5735 : };
5736 : }
5737 :
5738 : {
5739 0 : lock_status!(st);
5740 0 : st.total_layer_count = total_layer_count as u64;
5741 0 : }
5742 0 :
5743 0 : let mut remaining = remaining.into_iter();
5744 0 : let mut have_remaining = true;
5745 0 : let mut js = tokio::task::JoinSet::new();
5746 0 :
5747 0 : let cancel = task_mgr::shutdown_token();
5748 0 :
5749 0 : let limit = request.max_concurrent_downloads;
5750 :
5751 : loop {
5752 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5753 0 : let Some(next) = remaining.next() else {
5754 0 : have_remaining = false;
5755 0 : break;
5756 : };
5757 :
5758 0 : let span = tracing::info_span!("download", layer = %next);
5759 :
5760 0 : js.spawn(
5761 0 : async move {
5762 0 : let res = next.download().await;
5763 0 : (next, res)
5764 0 : }
5765 0 : .instrument(span),
5766 0 : );
5767 0 : }
5768 :
5769 0 : while let Some(res) = js.join_next().await {
5770 0 : match res {
5771 : Ok((_, Ok(_))) => {
5772 0 : lock_status!(st);
5773 0 : st.successful_download_count += 1;
5774 : }
5775 0 : Ok((layer, Err(e))) => {
5776 0 : tracing::error!(%layer, "download failed: {e:#}");
5777 0 : lock_status!(st);
5778 0 : st.failed_download_count += 1;
5779 : }
5780 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5781 0 : Err(je) if je.is_panic() => {
5782 0 : lock_status!(st);
5783 0 : st.failed_download_count += 1;
5784 : }
5785 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5786 : }
5787 : }
5788 :
5789 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5790 0 : break;
5791 0 : }
5792 : }
5793 :
5794 : {
5795 0 : lock_status!(st);
5796 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5797 : }
5798 0 : }
5799 :
5800 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5801 0 : &self,
5802 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5803 0 : self.download_all_remote_layers_task_info
5804 0 : .read()
5805 0 : .unwrap()
5806 0 : .clone()
5807 0 : }
5808 : }
5809 :
5810 : impl Timeline {
5811 : /// Returns non-remote layers for eviction.
5812 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5813 0 : let guard = self.layers.read().await;
5814 0 : let mut max_layer_size: Option<u64> = None;
5815 0 :
5816 0 : let resident_layers = guard
5817 0 : .likely_resident_layers()
5818 0 : .map(|layer| {
5819 0 : let file_size = layer.layer_desc().file_size;
5820 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5821 0 :
5822 0 : let last_activity_ts = layer.latest_activity();
5823 0 :
5824 0 : EvictionCandidate {
5825 0 : layer: layer.to_owned().into(),
5826 0 : last_activity_ts,
5827 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5828 0 : visibility: layer.visibility(),
5829 0 : }
5830 0 : })
5831 0 : .collect();
5832 0 :
5833 0 : DiskUsageEvictionInfo {
5834 0 : max_layer_size,
5835 0 : resident_layers,
5836 0 : }
5837 0 : }
5838 :
5839 3720 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5840 3720 : ShardIndex {
5841 3720 : shard_number: self.tenant_shard_id.shard_number,
5842 3720 : shard_count: self.tenant_shard_id.shard_count,
5843 3720 : }
5844 3720 : }
5845 :
5846 : /// Persistently blocks gc for `Manual` reason.
5847 : ///
5848 : /// Returns true if no such block existed before, false otherwise.
5849 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
5850 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5851 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5852 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
5853 0 : }
5854 :
5855 : /// Persistently unblocks gc for `Manual` reason.
5856 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
5857 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5858 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5859 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
5860 0 : }
5861 :
5862 : #[cfg(test)]
5863 92 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5864 92 : self.last_record_lsn.advance(new_lsn);
5865 92 : }
5866 :
5867 : #[cfg(test)]
5868 4 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
5869 4 : self.disk_consistent_lsn.store(new_value);
5870 4 : }
5871 :
5872 : /// Force create an image layer and place it into the layer map.
5873 : ///
5874 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5875 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5876 : /// placed into the layer map in one run AND be validated.
5877 : #[cfg(test)]
5878 116 : pub(super) async fn force_create_image_layer(
5879 116 : self: &Arc<Timeline>,
5880 116 : lsn: Lsn,
5881 116 : mut images: Vec<(Key, Bytes)>,
5882 116 : check_start_lsn: Option<Lsn>,
5883 116 : ctx: &RequestContext,
5884 116 : ) -> anyhow::Result<()> {
5885 116 : let last_record_lsn = self.get_last_record_lsn();
5886 116 : assert!(
5887 116 : lsn <= last_record_lsn,
5888 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5889 : );
5890 116 : if let Some(check_start_lsn) = check_start_lsn {
5891 116 : assert!(lsn >= check_start_lsn);
5892 0 : }
5893 348 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5894 116 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5895 116 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5896 116 : let mut image_layer_writer = ImageLayerWriter::new(
5897 116 : self.conf,
5898 116 : self.timeline_id,
5899 116 : self.tenant_shard_id,
5900 116 : &(min_key..end_key),
5901 116 : lsn,
5902 116 : ctx,
5903 116 : )
5904 116 : .await?;
5905 580 : for (key, img) in images {
5906 464 : image_layer_writer.put_image(key, img, ctx).await?;
5907 : }
5908 116 : let (desc, path) = image_layer_writer.finish(ctx).await?;
5909 116 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5910 116 : info!("force created image layer {}", image_layer.local_path());
5911 : {
5912 116 : let mut guard = self.layers.write().await;
5913 116 : guard
5914 116 : .open_mut()
5915 116 : .unwrap()
5916 116 : .force_insert_layer(image_layer.clone());
5917 116 : }
5918 116 :
5919 116 : // Update remote_timeline_client state to reflect existence of this layer
5920 116 : self.remote_client
5921 116 : .schedule_layer_file_upload(image_layer)
5922 116 : .unwrap();
5923 116 :
5924 116 : Ok(())
5925 116 : }
5926 :
5927 : /// Force create a delta layer and place it into the layer map.
5928 : ///
5929 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5930 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5931 : /// placed into the layer map in one run AND be validated.
5932 : #[cfg(test)]
5933 180 : pub(super) async fn force_create_delta_layer(
5934 180 : self: &Arc<Timeline>,
5935 180 : mut deltas: DeltaLayerTestDesc,
5936 180 : check_start_lsn: Option<Lsn>,
5937 180 : ctx: &RequestContext,
5938 180 : ) -> anyhow::Result<()> {
5939 180 : let last_record_lsn = self.get_last_record_lsn();
5940 180 : deltas
5941 180 : .data
5942 264 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5943 180 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5944 180 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5945 624 : for (_, lsn, _) in &deltas.data {
5946 444 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5947 : }
5948 180 : assert!(
5949 180 : deltas.lsn_range.end <= last_record_lsn,
5950 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
5951 : deltas.lsn_range.end,
5952 : last_record_lsn
5953 : );
5954 180 : if let Some(check_start_lsn) = check_start_lsn {
5955 180 : assert!(deltas.lsn_range.start >= check_start_lsn);
5956 0 : }
5957 180 : let mut delta_layer_writer = DeltaLayerWriter::new(
5958 180 : self.conf,
5959 180 : self.timeline_id,
5960 180 : self.tenant_shard_id,
5961 180 : deltas.key_range.start,
5962 180 : deltas.lsn_range,
5963 180 : ctx,
5964 180 : )
5965 180 : .await?;
5966 624 : for (key, lsn, val) in deltas.data {
5967 444 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5968 : }
5969 180 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
5970 180 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5971 180 : info!("force created delta layer {}", delta_layer.local_path());
5972 : {
5973 180 : let mut guard = self.layers.write().await;
5974 180 : guard
5975 180 : .open_mut()
5976 180 : .unwrap()
5977 180 : .force_insert_layer(delta_layer.clone());
5978 180 : }
5979 180 :
5980 180 : // Update remote_timeline_client state to reflect existence of this layer
5981 180 : self.remote_client
5982 180 : .schedule_layer_file_upload(delta_layer)
5983 180 : .unwrap();
5984 180 :
5985 180 : Ok(())
5986 180 : }
5987 :
5988 : /// Return all keys at the LSN in the image layers
5989 : #[cfg(test)]
5990 12 : pub(crate) async fn inspect_image_layers(
5991 12 : self: &Arc<Timeline>,
5992 12 : lsn: Lsn,
5993 12 : ctx: &RequestContext,
5994 12 : io_concurrency: IoConcurrency,
5995 12 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
5996 12 : let mut all_data = Vec::new();
5997 12 : let guard = self.layers.read().await;
5998 68 : for layer in guard.layer_map()?.iter_historic_layers() {
5999 68 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
6000 16 : let layer = guard.get_from_desc(&layer);
6001 16 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
6002 16 : layer
6003 16 : .get_values_reconstruct_data(
6004 16 : KeySpace::single(Key::MIN..Key::MAX),
6005 16 : lsn..Lsn(lsn.0 + 1),
6006 16 : &mut reconstruct_data,
6007 16 : ctx,
6008 16 : )
6009 16 : .await?;
6010 132 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
6011 132 : let v = v.collect_pending_ios().await?;
6012 132 : all_data.push((k, v.img.unwrap().1));
6013 : }
6014 52 : }
6015 : }
6016 12 : all_data.sort();
6017 12 : Ok(all_data)
6018 12 : }
6019 :
6020 : /// Get all historic layer descriptors in the layer map
6021 : #[cfg(test)]
6022 48 : pub(crate) async fn inspect_historic_layers(
6023 48 : self: &Arc<Timeline>,
6024 48 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
6025 48 : let mut layers = Vec::new();
6026 48 : let guard = self.layers.read().await;
6027 228 : for layer in guard.layer_map()?.iter_historic_layers() {
6028 228 : layers.push(layer.key());
6029 228 : }
6030 48 : Ok(layers)
6031 48 : }
6032 :
6033 : #[cfg(test)]
6034 20 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
6035 20 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
6036 20 : keyspace.merge(&ks);
6037 20 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
6038 20 : }
6039 : }
6040 :
6041 : /// Tracking writes ingestion does to a particular in-memory layer.
6042 : ///
6043 : /// Cleared upon freezing a layer.
6044 : pub(crate) struct TimelineWriterState {
6045 : open_layer: Arc<InMemoryLayer>,
6046 : current_size: u64,
6047 : // Previous Lsn which passed through
6048 : prev_lsn: Option<Lsn>,
6049 : // Largest Lsn which passed through the current writer
6050 : max_lsn: Option<Lsn>,
6051 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
6052 : cached_last_freeze_at: Lsn,
6053 : }
6054 :
6055 : impl TimelineWriterState {
6056 2592 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
6057 2592 : Self {
6058 2592 : open_layer,
6059 2592 : current_size,
6060 2592 : prev_lsn: None,
6061 2592 : max_lsn: None,
6062 2592 : cached_last_freeze_at: last_freeze_at,
6063 2592 : }
6064 2592 : }
6065 : }
6066 :
6067 : /// Various functions to mutate the timeline.
6068 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
6069 : // This is probably considered a bad practice in Rust and should be fixed eventually,
6070 : // but will cause large code changes.
6071 : pub(crate) struct TimelineWriter<'a> {
6072 : tl: &'a Timeline,
6073 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
6074 : }
6075 :
6076 : impl Deref for TimelineWriter<'_> {
6077 : type Target = Timeline;
6078 :
6079 19796336 : fn deref(&self) -> &Self::Target {
6080 19796336 : self.tl
6081 19796336 : }
6082 : }
6083 :
6084 : #[derive(PartialEq)]
6085 : enum OpenLayerAction {
6086 : Roll,
6087 : Open,
6088 : None,
6089 : }
6090 :
6091 : impl TimelineWriter<'_> {
6092 9608464 : async fn handle_open_layer_action(
6093 9608464 : &mut self,
6094 9608464 : at: Lsn,
6095 9608464 : action: OpenLayerAction,
6096 9608464 : ctx: &RequestContext,
6097 9608464 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
6098 9608464 : match action {
6099 : OpenLayerAction::Roll => {
6100 160 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
6101 160 : self.roll_layer(freeze_at).await?;
6102 160 : self.open_layer(at, ctx).await?;
6103 : }
6104 2432 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
6105 : OpenLayerAction::None => {
6106 9605872 : assert!(self.write_guard.is_some());
6107 : }
6108 : }
6109 :
6110 9608464 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
6111 9608464 : }
6112 :
6113 2592 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
6114 2592 : let layer = self
6115 2592 : .tl
6116 2592 : .get_layer_for_write(at, &self.write_guard, ctx)
6117 2592 : .await?;
6118 2592 : let initial_size = layer.size().await?;
6119 :
6120 2592 : let last_freeze_at = self.last_freeze_at.load();
6121 2592 : self.write_guard.replace(TimelineWriterState::new(
6122 2592 : layer,
6123 2592 : initial_size,
6124 2592 : last_freeze_at,
6125 2592 : ));
6126 2592 :
6127 2592 : Ok(())
6128 2592 : }
6129 :
6130 160 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
6131 160 : let current_size = self.write_guard.as_ref().unwrap().current_size;
6132 :
6133 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
6134 : // to propagate the backpressure up into WAL ingestion.
6135 160 : let l0_count = self
6136 160 : .tl
6137 160 : .layers
6138 160 : .read()
6139 160 : .await
6140 160 : .layer_map()?
6141 160 : .level0_deltas()
6142 160 : .len();
6143 160 : let wait_thresholds = [
6144 160 : self.get_l0_flush_delay_threshold(),
6145 160 : self.get_l0_flush_stall_threshold(),
6146 160 : ];
6147 160 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
6148 :
6149 : // self.write_guard will be taken by the freezing
6150 160 : let flush_id = self
6151 160 : .tl
6152 160 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
6153 160 : .await?;
6154 :
6155 160 : assert!(self.write_guard.is_none());
6156 :
6157 160 : if let Some(wait_threshold) = wait_threshold {
6158 0 : if l0_count >= wait_threshold {
6159 0 : info!("layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers");
6160 0 : self.tl.wait_flush_completion(flush_id).await?;
6161 0 : }
6162 160 : }
6163 :
6164 160 : if current_size >= self.get_checkpoint_distance() * 2 {
6165 0 : warn!("Flushed oversized open layer with size {}", current_size)
6166 160 : }
6167 :
6168 160 : Ok(())
6169 160 : }
6170 :
6171 9608464 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
6172 9608464 : let state = &*self.write_guard;
6173 9608464 : let Some(state) = &state else {
6174 2432 : return OpenLayerAction::Open;
6175 : };
6176 :
6177 : #[cfg(feature = "testing")]
6178 9606032 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
6179 : // this check and assertion are not really needed because
6180 : // LayerManager::try_freeze_in_memory_layer will always clear out the
6181 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
6182 : // is no TimelineWriterState.
6183 0 : assert!(
6184 0 : state.open_layer.end_lsn.get().is_some(),
6185 0 : "our open_layer must be outdated"
6186 : );
6187 :
6188 : // this would be a memory leak waiting to happen because the in-memory layer always has
6189 : // an index
6190 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
6191 9606032 : }
6192 9606032 :
6193 9606032 : if state.prev_lsn == Some(lsn) {
6194 : // Rolling mid LSN is not supported by [downstream code].
6195 : // Hence, only roll at LSN boundaries.
6196 : //
6197 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
6198 12 : return OpenLayerAction::None;
6199 9606020 : }
6200 9606020 :
6201 9606020 : if state.current_size == 0 {
6202 : // Don't roll empty layers
6203 0 : return OpenLayerAction::None;
6204 9606020 : }
6205 9606020 :
6206 9606020 : if self.tl.should_roll(
6207 9606020 : state.current_size,
6208 9606020 : state.current_size + new_value_size,
6209 9606020 : self.get_checkpoint_distance(),
6210 9606020 : lsn,
6211 9606020 : state.cached_last_freeze_at,
6212 9606020 : state.open_layer.get_opened_at(),
6213 9606020 : ) {
6214 160 : OpenLayerAction::Roll
6215 : } else {
6216 9605860 : OpenLayerAction::None
6217 : }
6218 9608464 : }
6219 :
6220 : /// Put a batch of keys at the specified Lsns.
6221 9608460 : pub(crate) async fn put_batch(
6222 9608460 : &mut self,
6223 9608460 : batch: SerializedValueBatch,
6224 9608460 : ctx: &RequestContext,
6225 9608460 : ) -> anyhow::Result<()> {
6226 9608460 : if !batch.has_data() {
6227 0 : return Ok(());
6228 9608460 : }
6229 9608460 :
6230 9608460 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
6231 9608460 : // We don't assert this in release builds, since key ownership policies may change over
6232 9608460 : // time. Stray keys will be removed during compaction.
6233 9608460 : if cfg!(debug_assertions) {
6234 19789792 : for metadata in &batch.metadata {
6235 10181332 : if let ValueMeta::Serialized(metadata) = metadata {
6236 10181332 : let key = Key::from_compact(metadata.key);
6237 10181332 : assert!(
6238 10181332 : self.shard_identity.is_key_local(&key)
6239 0 : || self.shard_identity.is_key_global(&key),
6240 0 : "key {key} does not belong on shard {}",
6241 0 : self.shard_identity.shard_index()
6242 : );
6243 0 : }
6244 : }
6245 0 : }
6246 :
6247 9608460 : let batch_max_lsn = batch.max_lsn;
6248 9608460 : let buf_size: u64 = batch.buffer_size() as u64;
6249 9608460 :
6250 9608460 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
6251 9608460 : let layer = self
6252 9608460 : .handle_open_layer_action(batch_max_lsn, action, ctx)
6253 9608460 : .await?;
6254 :
6255 9608460 : let res = layer.put_batch(batch, ctx).await;
6256 :
6257 9608460 : if res.is_ok() {
6258 9608460 : // Update the current size only when the entire write was ok.
6259 9608460 : // In case of failures, we may have had partial writes which
6260 9608460 : // render the size tracking out of sync. That's ok because
6261 9608460 : // the checkpoint distance should be significantly smaller
6262 9608460 : // than the S3 single shot upload limit of 5GiB.
6263 9608460 : let state = self.write_guard.as_mut().unwrap();
6264 9608460 :
6265 9608460 : state.current_size += buf_size;
6266 9608460 : state.prev_lsn = Some(batch_max_lsn);
6267 9608460 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
6268 9608460 : }
6269 :
6270 9608460 : res
6271 9608460 : }
6272 :
6273 : #[cfg(test)]
6274 : /// Test helper, for tests that would like to poke individual values without composing a batch
6275 8780308 : pub(crate) async fn put(
6276 8780308 : &mut self,
6277 8780308 : key: Key,
6278 8780308 : lsn: Lsn,
6279 8780308 : value: &Value,
6280 8780308 : ctx: &RequestContext,
6281 8780308 : ) -> anyhow::Result<()> {
6282 : use utils::bin_ser::BeSer;
6283 8780308 : if !key.is_valid_key_on_write_path() {
6284 0 : bail!(
6285 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
6286 0 : key
6287 0 : );
6288 8780308 : }
6289 8780308 : let val_ser_size = value.serialized_size().unwrap() as usize;
6290 8780308 : let batch = SerializedValueBatch::from_values(vec![(
6291 8780308 : key.to_compact(),
6292 8780308 : lsn,
6293 8780308 : val_ser_size,
6294 8780308 : value.clone(),
6295 8780308 : )]);
6296 8780308 :
6297 8780308 : self.put_batch(batch, ctx).await
6298 8780308 : }
6299 :
6300 4 : pub(crate) async fn delete_batch(
6301 4 : &mut self,
6302 4 : batch: &[(Range<Key>, Lsn)],
6303 4 : ctx: &RequestContext,
6304 4 : ) -> anyhow::Result<()> {
6305 4 : if let Some((_, lsn)) = batch.first() {
6306 4 : let action = self.get_open_layer_action(*lsn, 0);
6307 4 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
6308 4 : layer.put_tombstones(batch).await?;
6309 0 : }
6310 :
6311 4 : Ok(())
6312 4 : }
6313 :
6314 : /// Track the end of the latest digested WAL record.
6315 : /// Remember the (end of) last valid WAL record remembered in the timeline.
6316 : ///
6317 : /// Call this after you have finished writing all the WAL up to 'lsn'.
6318 : ///
6319 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
6320 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
6321 : /// the latest and can be read.
6322 10558184 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
6323 10558184 : self.tl.finish_write(new_lsn);
6324 10558184 : }
6325 :
6326 541140 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
6327 541140 : self.tl.update_current_logical_size(delta)
6328 541140 : }
6329 : }
6330 :
6331 : // We need TimelineWriter to be send in upcoming conversion of
6332 : // Timeline::layers to tokio::sync::RwLock.
6333 : #[test]
6334 4 : fn is_send() {
6335 4 : fn _assert_send<T: Send>() {}
6336 4 : _assert_send::<TimelineWriter<'_>>();
6337 4 : }
6338 :
6339 : #[cfg(test)]
6340 : mod tests {
6341 : use pageserver_api::key::Key;
6342 : use pageserver_api::value::Value;
6343 : use utils::{id::TimelineId, lsn::Lsn};
6344 :
6345 : use crate::tenant::{
6346 : harness::{test_img, TenantHarness},
6347 : layer_map::LayerMap,
6348 : storage_layer::{Layer, LayerName},
6349 : timeline::{DeltaLayerTestDesc, EvictionError},
6350 : Timeline,
6351 : };
6352 :
6353 : #[tokio::test]
6354 4 : async fn test_heatmap_generation() {
6355 4 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
6356 4 :
6357 4 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6358 4 : Lsn(0x10)..Lsn(0x20),
6359 4 : vec![(
6360 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6361 4 : Lsn(0x11),
6362 4 : Value::Image(test_img("foo")),
6363 4 : )],
6364 4 : );
6365 4 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6366 4 : Lsn(0x10)..Lsn(0x20),
6367 4 : vec![(
6368 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6369 4 : Lsn(0x11),
6370 4 : Value::Image(test_img("foo")),
6371 4 : )],
6372 4 : );
6373 4 : let l0_delta = DeltaLayerTestDesc::new(
6374 4 : Lsn(0x20)..Lsn(0x30),
6375 4 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6376 4 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6377 4 : vec![(
6378 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6379 4 : Lsn(0x25),
6380 4 : Value::Image(test_img("foo")),
6381 4 : )],
6382 4 : );
6383 4 : let delta_layers = vec![
6384 4 : covered_delta.clone(),
6385 4 : visible_delta.clone(),
6386 4 : l0_delta.clone(),
6387 4 : ];
6388 4 :
6389 4 : let image_layer = (
6390 4 : Lsn(0x40),
6391 4 : vec![(
6392 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6393 4 : test_img("bar"),
6394 4 : )],
6395 4 : );
6396 4 : let image_layers = vec![image_layer];
6397 4 :
6398 4 : let (tenant, ctx) = harness.load().await;
6399 4 : let timeline = tenant
6400 4 : .create_test_timeline_with_layers(
6401 4 : TimelineId::generate(),
6402 4 : Lsn(0x10),
6403 4 : 14,
6404 4 : &ctx,
6405 4 : delta_layers,
6406 4 : image_layers,
6407 4 : Lsn(0x100),
6408 4 : )
6409 4 : .await
6410 4 : .unwrap();
6411 4 :
6412 4 : // Layer visibility is an input to heatmap generation, so refresh it first
6413 4 : timeline.update_layer_visibility().await.unwrap();
6414 4 :
6415 4 : let heatmap = timeline
6416 4 : .generate_heatmap()
6417 4 : .await
6418 4 : .expect("Infallible while timeline is not shut down");
6419 4 :
6420 4 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
6421 4 :
6422 4 : // L0 should come last
6423 4 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
6424 4 :
6425 4 : let mut last_lsn = Lsn::MAX;
6426 20 : for layer in heatmap.layers {
6427 4 : // Covered layer should be omitted
6428 16 : assert!(layer.name != covered_delta.layer_name());
6429 4 :
6430 16 : let layer_lsn = match &layer.name {
6431 8 : LayerName::Delta(d) => d.lsn_range.end,
6432 8 : LayerName::Image(i) => i.lsn,
6433 4 : };
6434 4 :
6435 4 : // Apart from L0s, newest Layers should come first
6436 16 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
6437 12 : assert!(layer_lsn <= last_lsn);
6438 12 : last_lsn = layer_lsn;
6439 4 : }
6440 4 : }
6441 4 : }
6442 :
6443 : #[tokio::test]
6444 4 : async fn two_layer_eviction_attempts_at_the_same_time() {
6445 4 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6446 4 : .await
6447 4 : .unwrap();
6448 4 :
6449 4 : let (tenant, ctx) = harness.load().await;
6450 4 : let timeline = tenant
6451 4 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6452 4 : .await
6453 4 : .unwrap();
6454 4 :
6455 4 : let layer = find_some_layer(&timeline).await;
6456 4 : let layer = layer
6457 4 : .keep_resident()
6458 4 : .await
6459 4 : .expect("no download => no downloading errors")
6460 4 : .drop_eviction_guard();
6461 4 :
6462 4 : let forever = std::time::Duration::from_secs(120);
6463 4 :
6464 4 : let first = layer.evict_and_wait(forever);
6465 4 : let second = layer.evict_and_wait(forever);
6466 4 :
6467 4 : let (first, second) = tokio::join!(first, second);
6468 4 :
6469 4 : let res = layer.keep_resident().await;
6470 4 : assert!(res.is_none(), "{res:?}");
6471 4 :
6472 4 : match (first, second) {
6473 4 : (Ok(()), Ok(())) => {
6474 4 : // because there are no more timeline locks being taken on eviction path, we can
6475 4 : // witness all three outcomes here.
6476 4 : }
6477 4 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
6478 0 : // if one completes before the other, this is fine just as well.
6479 0 : }
6480 4 : other => unreachable!("unexpected {:?}", other),
6481 4 : }
6482 4 : }
6483 :
6484 4 : async fn find_some_layer(timeline: &Timeline) -> Layer {
6485 4 : let layers = timeline.layers.read().await;
6486 4 : let desc = layers
6487 4 : .layer_map()
6488 4 : .unwrap()
6489 4 : .iter_historic_layers()
6490 4 : .next()
6491 4 : .expect("must find one layer to evict");
6492 4 :
6493 4 : layers.get_from_desc(&desc)
6494 4 : }
6495 : }
|
