Line data Source code
1 : pub(crate) mod analysis;
2 : pub(crate) mod compaction;
3 : pub mod delete;
4 : pub(crate) mod detach_ancestor;
5 : mod eviction_task;
6 : pub(crate) mod handle;
7 : pub(crate) mod import_pgdata;
8 : mod init;
9 : pub mod layer_manager;
10 : pub(crate) mod logical_size;
11 : pub mod offload;
12 : pub mod span;
13 : pub mod uninit;
14 : mod walreceiver;
15 :
16 : use anyhow::{anyhow, bail, ensure, Context, Result};
17 : use arc_swap::ArcSwap;
18 : use bytes::Bytes;
19 : use camino::Utf8Path;
20 : use chrono::{DateTime, Utc};
21 : use enumset::EnumSet;
22 : use fail::fail_point;
23 : use handle::ShardTimelineId;
24 : use offload::OffloadError;
25 : use once_cell::sync::Lazy;
26 : use pageserver_api::{
27 : config::tenant_conf_defaults::DEFAULT_COMPACTION_THRESHOLD,
28 : key::{
29 : KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
30 : NON_INHERITED_SPARSE_RANGE,
31 : },
32 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
33 : models::{
34 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
35 : DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy,
36 : InMemoryLayerInfo, LayerMapInfo, LsnLease, TimelineState,
37 : },
38 : reltag::BlockNumber,
39 : shard::{ShardIdentity, ShardNumber, TenantShardId},
40 : };
41 : use rand::Rng;
42 : use remote_storage::DownloadError;
43 : use serde_with::serde_as;
44 : use storage_broker::BrokerClientChannel;
45 : use tokio::{
46 : runtime::Handle,
47 : sync::{oneshot, watch},
48 : };
49 : use tokio_util::sync::CancellationToken;
50 : use tracing::*;
51 : use utils::{
52 : fs_ext, pausable_failpoint,
53 : postgres_client::PostgresClientProtocol,
54 : sync::gate::{Gate, GateGuard},
55 : };
56 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
57 :
58 : use std::sync::atomic::Ordering as AtomicOrdering;
59 : use std::sync::{Arc, Mutex, RwLock, Weak};
60 : use std::time::{Duration, Instant, SystemTime};
61 : use std::{
62 : array,
63 : collections::{BTreeMap, HashMap, HashSet},
64 : sync::atomic::AtomicU64,
65 : };
66 : use std::{cmp::min, ops::ControlFlow};
67 : use std::{
68 : collections::btree_map::Entry,
69 : ops::{Deref, Range},
70 : };
71 : use std::{pin::pin, sync::OnceLock};
72 :
73 : use crate::{
74 : aux_file::AuxFileSizeEstimator,
75 : tenant::{
76 : config::AttachmentMode,
77 : layer_map::{LayerMap, SearchResult},
78 : metadata::TimelineMetadata,
79 : storage_layer::{inmemory_layer::IndexEntry, PersistentLayerDesc},
80 : },
81 : walingest::WalLagCooldown,
82 : walredo,
83 : };
84 : use crate::{
85 : context::{DownloadBehavior, RequestContext},
86 : disk_usage_eviction_task::DiskUsageEvictionInfo,
87 : pgdatadir_mapping::CollectKeySpaceError,
88 : };
89 : use crate::{
90 : disk_usage_eviction_task::finite_f32,
91 : tenant::storage_layer::{
92 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
93 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
94 : ValuesReconstructState,
95 : },
96 : };
97 : use crate::{
98 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
99 : };
100 : use crate::{
101 : l0_flush::{self, L0FlushGlobalState},
102 : metrics::GetKind,
103 : };
104 : use crate::{
105 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
106 : };
107 : use crate::{
108 : pgdatadir_mapping::DirectoryKind,
109 : virtual_file::{MaybeFatalIo, VirtualFile},
110 : };
111 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
112 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
113 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
114 :
115 : use crate::config::PageServerConf;
116 : use crate::keyspace::{KeyPartitioning, KeySpace};
117 : use crate::metrics::TimelineMetrics;
118 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
119 : use crate::tenant::config::TenantConfOpt;
120 : use pageserver_api::reltag::RelTag;
121 : use pageserver_api::shard::ShardIndex;
122 :
123 : use postgres_connection::PgConnectionConfig;
124 : use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
125 : use utils::{
126 : completion,
127 : generation::Generation,
128 : id::TimelineId,
129 : lsn::{AtomicLsn, Lsn, RecordLsn},
130 : seqwait::SeqWait,
131 : simple_rcu::{Rcu, RcuReadGuard},
132 : };
133 :
134 : use crate::task_mgr;
135 : use crate::task_mgr::TaskKind;
136 : use crate::tenant::gc_result::GcResult;
137 : use crate::ZERO_PAGE;
138 : use pageserver_api::key::Key;
139 :
140 : use self::delete::DeleteTimelineFlow;
141 : pub(super) use self::eviction_task::EvictionTaskTenantState;
142 : use self::eviction_task::EvictionTaskTimelineState;
143 : use self::layer_manager::LayerManager;
144 : use self::logical_size::LogicalSize;
145 : use self::walreceiver::{WalReceiver, WalReceiverConf};
146 :
147 : use super::config::TenantConf;
148 : use super::remote_timeline_client::index::IndexPart;
149 : use super::remote_timeline_client::RemoteTimelineClient;
150 : use super::secondary::heatmap::{HeatMapLayer, HeatMapTimeline};
151 : use super::storage_layer::{LayerFringe, LayerVisibilityHint, ReadableLayer};
152 : use super::upload_queue::NotInitialized;
153 : use super::GcError;
154 : use super::{
155 : debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf, MaybeOffloaded,
156 : };
157 :
158 : #[cfg(test)]
159 : use pageserver_api::value::Value;
160 :
161 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
162 : pub(crate) enum FlushLoopState {
163 : NotStarted,
164 : Running {
165 : #[cfg(test)]
166 : expect_initdb_optimization: bool,
167 : #[cfg(test)]
168 : initdb_optimization_count: usize,
169 : },
170 : Exited,
171 : }
172 :
173 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
174 : pub enum ImageLayerCreationMode {
175 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
176 : Try,
177 : /// Force creating the image layers if possible. For now, no image layers will be created
178 : /// for metadata keys. Used in compaction code path with force flag enabled.
179 : Force,
180 : /// Initial ingestion of the data, and no data should be dropped in this function. This
181 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
182 : /// code path.
183 : Initial,
184 : }
185 :
186 : impl std::fmt::Display for ImageLayerCreationMode {
187 724 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
188 724 : write!(f, "{:?}", self)
189 724 : }
190 : }
191 :
192 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
193 : /// Can be removed after all refactors are done.
194 28 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
195 28 : drop(rlock)
196 28 : }
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.
200 752 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
201 752 : drop(rlock)
202 752 : }
203 :
204 : /// The outward-facing resources required to build a Timeline
205 : pub struct TimelineResources {
206 : pub remote_client: RemoteTimelineClient,
207 : pub pagestream_throttle:
208 : Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::Pagestream>>,
209 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
210 : }
211 :
212 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
213 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
214 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
215 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
216 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
217 : pub(crate) struct RelSizeCache {
218 : pub(crate) complete_as_of: Lsn,
219 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
220 : }
221 :
222 : pub struct Timeline {
223 : pub(crate) conf: &'static PageServerConf,
224 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
225 :
226 : myself: Weak<Self>,
227 :
228 : pub(crate) tenant_shard_id: TenantShardId,
229 : pub timeline_id: TimelineId,
230 :
231 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
232 : /// Never changes for the lifetime of this [`Timeline`] object.
233 : ///
234 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
235 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
236 : pub(crate) generation: Generation,
237 :
238 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
239 : /// to shards, and is constant through the lifetime of this Timeline.
240 : shard_identity: ShardIdentity,
241 :
242 : pub pg_version: u32,
243 :
244 : /// The tuple has two elements.
245 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
246 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
247 : ///
248 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
249 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
250 : ///
251 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
252 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
253 : /// `PersistentLayerDesc`'s.
254 : ///
255 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
256 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
257 : /// runtime, e.g., during page reconstruction.
258 : ///
259 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
260 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
261 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
262 :
263 : last_freeze_at: AtomicLsn,
264 : // Atomic would be more appropriate here.
265 : last_freeze_ts: RwLock<Instant>,
266 :
267 : pub(crate) standby_horizon: AtomicLsn,
268 :
269 : // WAL redo manager. `None` only for broken tenants.
270 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
271 :
272 : /// Remote storage client.
273 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
274 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
275 :
276 : // What page versions do we hold in the repository? If we get a
277 : // request > last_record_lsn, we need to wait until we receive all
278 : // the WAL up to the request. The SeqWait provides functions for
279 : // that. TODO: If we get a request for an old LSN, such that the
280 : // versions have already been garbage collected away, we should
281 : // throw an error, but we don't track that currently.
282 : //
283 : // last_record_lsn.load().last points to the end of last processed WAL record.
284 : //
285 : // We also remember the starting point of the previous record in
286 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
287 : // first WAL record when the node is started up. But here, we just
288 : // keep track of it.
289 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
290 :
291 : // All WAL records have been processed and stored durably on files on
292 : // local disk, up to this LSN. On crash and restart, we need to re-process
293 : // the WAL starting from this point.
294 : //
295 : // Some later WAL records might have been processed and also flushed to disk
296 : // already, so don't be surprised to see some, but there's no guarantee on
297 : // them yet.
298 : disk_consistent_lsn: AtomicLsn,
299 :
300 : // Parent timeline that this timeline was branched from, and the LSN
301 : // of the branch point.
302 : ancestor_timeline: Option<Arc<Timeline>>,
303 : ancestor_lsn: Lsn,
304 :
305 : pub(super) metrics: TimelineMetrics,
306 :
307 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
308 : // in `crate::page_service` writes these metrics.
309 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
310 :
311 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
312 :
313 : /// Ensures layers aren't frozen by checkpointer between
314 : /// [`Timeline::get_layer_for_write`] and layer reads.
315 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
316 : /// Must always be acquired before the layer map/individual layer lock
317 : /// to avoid deadlock.
318 : ///
319 : /// The state is cleared upon freezing.
320 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
321 :
322 : /// Used to avoid multiple `flush_loop` tasks running
323 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
324 :
325 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
326 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
327 : /// The flush cycle counter is sent back on the layer_flush_done channel when
328 : /// the flush finishes. You can use that to wait for the flush to finish.
329 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
330 : /// read by whoever sends an update
331 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
332 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
333 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
334 :
335 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
336 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
337 :
338 : // List of child timelines and their branch points. This is needed to avoid
339 : // garbage collecting data that is still needed by the child timelines.
340 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
341 :
342 : // It may change across major versions so for simplicity
343 : // keep it after running initdb for a timeline.
344 : // It is needed in checks when we want to error on some operations
345 : // when they are requested for pre-initdb lsn.
346 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
347 : // though let's keep them both for better error visibility.
348 : pub initdb_lsn: Lsn,
349 :
350 : /// When did we last calculate the partitioning? Make it pub to test cases.
351 : pub(super) partitioning: tokio::sync::Mutex<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
352 :
353 : /// Configuration: how often should the partitioning be recalculated.
354 : repartition_threshold: u64,
355 :
356 : last_image_layer_creation_check_at: AtomicLsn,
357 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
358 :
359 : /// Current logical size of the "datadir", at the last LSN.
360 : current_logical_size: LogicalSize,
361 :
362 : /// Information about the last processed message by the WAL receiver,
363 : /// or None if WAL receiver has not received anything for this timeline
364 : /// yet.
365 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
366 : pub walreceiver: Mutex<Option<WalReceiver>>,
367 :
368 : /// Relation size cache
369 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
370 :
371 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
372 :
373 : state: watch::Sender<TimelineState>,
374 :
375 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
376 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
377 : pub delete_progress: TimelineDeleteProgress,
378 :
379 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
380 :
381 : /// Load or creation time information about the disk_consistent_lsn and when the loading
382 : /// happened. Used for consumption metrics.
383 : pub(crate) loaded_at: (Lsn, SystemTime),
384 :
385 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
386 : pub(crate) gate: Gate,
387 :
388 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
389 : /// to the timeline should drop out when this token fires.
390 : pub(crate) cancel: CancellationToken,
391 :
392 : /// Make sure we only have one running compaction at a time in tests.
393 : ///
394 : /// Must only be taken in two places:
395 : /// - [`Timeline::compact`] (this file)
396 : /// - [`delete::delete_local_timeline_directory`]
397 : ///
398 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
399 : compaction_lock: tokio::sync::Mutex<()>,
400 :
401 : /// Make sure we only have one running gc at a time.
402 : ///
403 : /// Must only be taken in two places:
404 : /// - [`Timeline::gc`] (this file)
405 : /// - [`delete::delete_local_timeline_directory`]
406 : ///
407 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
408 : gc_lock: tokio::sync::Mutex<()>,
409 :
410 : /// Cloned from [`super::Tenant::pagestream_throttle`] on construction.
411 : pub(crate) pagestream_throttle:
412 : Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::Pagestream>>,
413 :
414 : /// Size estimator for aux file v2
415 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
416 :
417 : /// Some test cases directly place keys into the timeline without actually modifying the directory
418 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
419 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
420 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
421 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
422 : #[cfg(test)]
423 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
424 :
425 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
426 :
427 : pub(crate) handles: handle::PerTimelineState<crate::page_service::TenantManagerTypes>,
428 :
429 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
430 :
431 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
432 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
433 : }
434 :
435 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
436 :
437 : pub struct WalReceiverInfo {
438 : pub wal_source_connconf: PgConnectionConfig,
439 : pub last_received_msg_lsn: Lsn,
440 : pub last_received_msg_ts: u128,
441 : }
442 :
443 : /// Information about how much history needs to be retained, needed by
444 : /// Garbage Collection.
445 : #[derive(Default)]
446 : pub(crate) struct GcInfo {
447 : /// Specific LSNs that are needed.
448 : ///
449 : /// Currently, this includes all points where child branches have
450 : /// been forked off from. In the future, could also include
451 : /// explicit user-defined snapshot points.
452 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
453 :
454 : /// The cutoff coordinates, which are combined by selecting the minimum.
455 : pub(crate) cutoffs: GcCutoffs,
456 :
457 : /// Leases granted to particular LSNs.
458 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
459 :
460 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
461 : pub(crate) within_ancestor_pitr: bool,
462 : }
463 :
464 : impl GcInfo {
465 228 : pub(crate) fn min_cutoff(&self) -> Lsn {
466 228 : self.cutoffs.select_min()
467 228 : }
468 :
469 232 : pub(super) fn insert_child(
470 232 : &mut self,
471 232 : child_id: TimelineId,
472 232 : child_lsn: Lsn,
473 232 : is_offloaded: MaybeOffloaded,
474 232 : ) {
475 232 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
476 232 : self.retain_lsns.sort_by_key(|i| i.0);
477 232 : }
478 :
479 4 : pub(super) fn remove_child_maybe_offloaded(
480 4 : &mut self,
481 4 : child_id: TimelineId,
482 4 : maybe_offloaded: MaybeOffloaded,
483 4 : ) -> bool {
484 4 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
485 4 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
486 4 : let mut removed = false;
487 6 : self.retain_lsns.retain(|i| {
488 6 : if removed {
489 2 : return true;
490 4 : }
491 4 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
492 4 : removed |= remove;
493 4 : !remove
494 6 : });
495 4 : removed
496 4 : }
497 :
498 4 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
499 4 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
500 4 : }
501 :
502 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
503 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
504 0 : }
505 : }
506 :
507 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
508 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
509 : /// between time-based and space-based retention for observability and consumption metrics purposes.
510 : #[derive(Debug, Clone)]
511 : pub(crate) struct GcCutoffs {
512 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
513 : /// history we must keep to retain a specified number of bytes of WAL.
514 : pub(crate) space: Lsn,
515 :
516 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
517 : /// history we must keep to enable reading back at least the PITR interval duration.
518 : pub(crate) time: Lsn,
519 : }
520 :
521 : impl Default for GcCutoffs {
522 422 : fn default() -> Self {
523 422 : Self {
524 422 : space: Lsn::INVALID,
525 422 : time: Lsn::INVALID,
526 422 : }
527 422 : }
528 : }
529 :
530 : impl GcCutoffs {
531 228 : fn select_min(&self) -> Lsn {
532 228 : std::cmp::min(self.space, self.time)
533 228 : }
534 : }
535 :
536 : pub(crate) struct TimelineVisitOutcome {
537 : completed_keyspace: KeySpace,
538 : image_covered_keyspace: KeySpace,
539 : }
540 :
541 : /// An error happened in a get() operation.
542 : #[derive(thiserror::Error, Debug)]
543 : pub(crate) enum PageReconstructError {
544 : #[error(transparent)]
545 : Other(anyhow::Error),
546 :
547 : #[error("Ancestor LSN wait error: {0}")]
548 : AncestorLsnTimeout(WaitLsnError),
549 :
550 : #[error("timeline shutting down")]
551 : Cancelled,
552 :
553 : /// An error happened replaying WAL records
554 : #[error(transparent)]
555 : WalRedo(anyhow::Error),
556 :
557 : #[error("{0}")]
558 : MissingKey(MissingKeyError),
559 : }
560 :
561 : impl From<anyhow::Error> for PageReconstructError {
562 0 : fn from(value: anyhow::Error) -> Self {
563 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
564 0 : match value.downcast::<PageReconstructError>() {
565 0 : Ok(pre) => pre,
566 0 : Err(other) => PageReconstructError::Other(other),
567 : }
568 0 : }
569 : }
570 :
571 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
572 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
573 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
574 0 : }
575 : }
576 :
577 : impl From<layer_manager::Shutdown> for PageReconstructError {
578 0 : fn from(_: layer_manager::Shutdown) -> Self {
579 0 : PageReconstructError::Cancelled
580 0 : }
581 : }
582 :
583 : impl GetVectoredError {
584 : #[cfg(test)]
585 6 : pub(crate) fn is_missing_key_error(&self) -> bool {
586 6 : matches!(self, Self::MissingKey(_))
587 6 : }
588 : }
589 :
590 : impl From<layer_manager::Shutdown> for GetVectoredError {
591 0 : fn from(_: layer_manager::Shutdown) -> Self {
592 0 : GetVectoredError::Cancelled
593 0 : }
594 : }
595 :
596 : #[derive(thiserror::Error)]
597 : pub struct MissingKeyError {
598 : key: Key,
599 : shard: ShardNumber,
600 : cont_lsn: Lsn,
601 : request_lsn: Lsn,
602 : ancestor_lsn: Option<Lsn>,
603 : backtrace: Option<std::backtrace::Backtrace>,
604 : }
605 :
606 : impl std::fmt::Debug for MissingKeyError {
607 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
608 0 : write!(f, "{}", self)
609 0 : }
610 : }
611 :
612 : impl std::fmt::Display for MissingKeyError {
613 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
614 0 : write!(
615 0 : f,
616 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
617 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
618 0 : )?;
619 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
620 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
621 0 : }
622 :
623 0 : if let Some(ref backtrace) = self.backtrace {
624 0 : write!(f, "\n{}", backtrace)?;
625 0 : }
626 :
627 0 : Ok(())
628 0 : }
629 : }
630 :
631 : impl PageReconstructError {
632 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
633 0 : pub(crate) fn is_stopping(&self) -> bool {
634 : use PageReconstructError::*;
635 0 : match self {
636 0 : Cancelled => true,
637 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
638 : }
639 0 : }
640 : }
641 :
642 : #[derive(thiserror::Error, Debug)]
643 : pub(crate) enum CreateImageLayersError {
644 : #[error("timeline shutting down")]
645 : Cancelled,
646 :
647 : #[error("read failed")]
648 : GetVectoredError(#[source] GetVectoredError),
649 :
650 : #[error("reconstruction failed")]
651 : PageReconstructError(#[source] PageReconstructError),
652 :
653 : #[error(transparent)]
654 : Other(#[from] anyhow::Error),
655 : }
656 :
657 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
658 0 : fn from(_: layer_manager::Shutdown) -> Self {
659 0 : CreateImageLayersError::Cancelled
660 0 : }
661 : }
662 :
663 : #[derive(thiserror::Error, Debug, Clone)]
664 : pub(crate) enum FlushLayerError {
665 : /// Timeline cancellation token was cancelled
666 : #[error("timeline shutting down")]
667 : Cancelled,
668 :
669 : /// We tried to flush a layer while the Timeline is in an unexpected state
670 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
671 : NotRunning(FlushLoopState),
672 :
673 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
674 : // loop via a watch channel, where we can only borrow it.
675 : #[error("create image layers (shared)")]
676 : CreateImageLayersError(Arc<CreateImageLayersError>),
677 :
678 : #[error("other (shared)")]
679 : Other(#[from] Arc<anyhow::Error>),
680 : }
681 :
682 : impl FlushLayerError {
683 : // When crossing from generic anyhow errors to this error type, we explicitly check
684 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
685 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
686 0 : let cancelled = timeline.cancel.is_cancelled()
687 : // The upload queue might have been shut down before the official cancellation of the timeline.
688 0 : || err
689 0 : .downcast_ref::<NotInitialized>()
690 0 : .map(NotInitialized::is_stopping)
691 0 : .unwrap_or_default();
692 0 : if cancelled {
693 0 : Self::Cancelled
694 : } else {
695 0 : Self::Other(Arc::new(err))
696 : }
697 0 : }
698 : }
699 :
700 : impl From<layer_manager::Shutdown> for FlushLayerError {
701 0 : fn from(_: layer_manager::Shutdown) -> Self {
702 0 : FlushLayerError::Cancelled
703 0 : }
704 : }
705 :
706 : #[derive(thiserror::Error, Debug)]
707 : pub(crate) enum GetVectoredError {
708 : #[error("timeline shutting down")]
709 : Cancelled,
710 :
711 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
712 : Oversized(u64),
713 :
714 : #[error("requested at invalid LSN: {0}")]
715 : InvalidLsn(Lsn),
716 :
717 : #[error("requested key not found: {0}")]
718 : MissingKey(MissingKeyError),
719 :
720 : #[error("ancestry walk")]
721 : GetReadyAncestorError(#[source] GetReadyAncestorError),
722 :
723 : #[error(transparent)]
724 : Other(#[from] anyhow::Error),
725 : }
726 :
727 : impl From<GetReadyAncestorError> for GetVectoredError {
728 2 : fn from(value: GetReadyAncestorError) -> Self {
729 : use GetReadyAncestorError::*;
730 2 : match value {
731 0 : Cancelled => GetVectoredError::Cancelled,
732 : AncestorLsnTimeout(_) | BadState { .. } => {
733 2 : GetVectoredError::GetReadyAncestorError(value)
734 : }
735 : }
736 2 : }
737 : }
738 :
739 : #[derive(thiserror::Error, Debug)]
740 : pub(crate) enum GetReadyAncestorError {
741 : #[error("ancestor LSN wait error")]
742 : AncestorLsnTimeout(#[from] WaitLsnError),
743 :
744 : #[error("bad state on timeline {timeline_id}: {state:?}")]
745 : BadState {
746 : timeline_id: TimelineId,
747 : state: TimelineState,
748 : },
749 :
750 : #[error("cancelled")]
751 : Cancelled,
752 : }
753 :
754 : #[derive(Clone, Copy)]
755 : pub enum LogicalSizeCalculationCause {
756 : Initial,
757 : ConsumptionMetricsSyntheticSize,
758 : EvictionTaskImitation,
759 : TenantSizeHandler,
760 : }
761 :
762 : pub enum GetLogicalSizePriority {
763 : User,
764 : Background,
765 : }
766 :
767 0 : #[derive(Debug, enumset::EnumSetType)]
768 : pub(crate) enum CompactFlags {
769 : ForceRepartition,
770 : ForceImageLayerCreation,
771 : ForceL0Compaction,
772 : EnhancedGcBottomMostCompaction,
773 : DryRun,
774 : }
775 :
776 : #[serde_with::serde_as]
777 0 : #[derive(Debug, Clone, serde::Deserialize)]
778 : pub(crate) struct CompactRequest {
779 : pub compact_key_range: Option<CompactKeyRange>,
780 : pub compact_lsn_range: Option<CompactLsnRange>,
781 : /// Whether the compaction job should be scheduled.
782 : #[serde(default)]
783 : pub scheduled: bool,
784 : /// Whether the compaction job should be split across key ranges.
785 : #[serde(default)]
786 : pub sub_compaction: bool,
787 : /// Max job size for each subcompaction job.
788 : pub sub_compaction_max_job_size_mb: Option<u64>,
789 : }
790 :
791 : #[derive(Debug, Clone, Default)]
792 : pub(crate) struct CompactOptions {
793 : pub flags: EnumSet<CompactFlags>,
794 : /// If set, the compaction will only compact the key range specified by this option.
795 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
796 : pub compact_key_range: Option<CompactKeyRange>,
797 : /// If set, the compaction will only compact the LSN within this value.
798 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
799 : pub compact_lsn_range: Option<CompactLsnRange>,
800 : /// Enable sub-compaction (split compaction job across key ranges).
801 : /// This option is only used by GC compaction.
802 : pub sub_compaction: bool,
803 : /// Set job size for the GC compaction.
804 : /// This option is only used by GC compaction.
805 : pub sub_compaction_max_job_size_mb: Option<u64>,
806 : }
807 :
808 : impl std::fmt::Debug for Timeline {
809 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
810 0 : write!(f, "Timeline<{}>", self.timeline_id)
811 0 : }
812 : }
813 :
814 : #[derive(thiserror::Error, Debug)]
815 : pub(crate) enum WaitLsnError {
816 : // Called on a timeline which is shutting down
817 : #[error("Shutdown")]
818 : Shutdown,
819 :
820 : // Called on an timeline not in active state or shutting down
821 : #[error("Bad timeline state: {0:?}")]
822 : BadState(TimelineState),
823 :
824 : // Timeout expired while waiting for LSN to catch up with goal.
825 : #[error("{0}")]
826 : Timeout(String),
827 : }
828 :
829 : // The impls below achieve cancellation mapping for errors.
830 : // Perhaps there's a way of achieving this with less cruft.
831 :
832 : impl From<CreateImageLayersError> for CompactionError {
833 0 : fn from(e: CreateImageLayersError) -> Self {
834 0 : match e {
835 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
836 0 : CreateImageLayersError::Other(e) => {
837 0 : CompactionError::Other(e.context("create image layers"))
838 : }
839 0 : _ => CompactionError::Other(e.into()),
840 : }
841 0 : }
842 : }
843 :
844 : impl From<CreateImageLayersError> for FlushLayerError {
845 0 : fn from(e: CreateImageLayersError) -> Self {
846 0 : match e {
847 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
848 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
849 : }
850 0 : }
851 : }
852 :
853 : impl From<PageReconstructError> for CreateImageLayersError {
854 0 : fn from(e: PageReconstructError) -> Self {
855 0 : match e {
856 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
857 0 : _ => CreateImageLayersError::PageReconstructError(e),
858 : }
859 0 : }
860 : }
861 :
862 : impl From<GetVectoredError> for CreateImageLayersError {
863 0 : fn from(e: GetVectoredError) -> Self {
864 0 : match e {
865 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
866 0 : _ => CreateImageLayersError::GetVectoredError(e),
867 : }
868 0 : }
869 : }
870 :
871 : impl From<GetVectoredError> for PageReconstructError {
872 6 : fn from(e: GetVectoredError) -> Self {
873 6 : match e {
874 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
875 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
876 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
877 4 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
878 2 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
879 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
880 : }
881 6 : }
882 : }
883 :
884 : impl From<GetReadyAncestorError> for PageReconstructError {
885 2 : fn from(e: GetReadyAncestorError) -> Self {
886 : use GetReadyAncestorError::*;
887 2 : match e {
888 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
889 2 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
890 0 : Cancelled => PageReconstructError::Cancelled,
891 : }
892 2 : }
893 : }
894 :
895 : pub(crate) enum WaitLsnWaiter<'a> {
896 : Timeline(&'a Timeline),
897 : Tenant,
898 : PageService,
899 : }
900 :
901 : /// Argument to [`Timeline::shutdown`].
902 : #[derive(Debug, Clone, Copy)]
903 : pub(crate) enum ShutdownMode {
904 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
905 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
906 : ///
907 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
908 : /// the call to [`Timeline::shutdown`].
909 : FreezeAndFlush,
910 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
911 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
912 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
913 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
914 : Reload,
915 : /// Shut down immediately, without waiting for any open layers to flush.
916 : Hard,
917 : }
918 :
919 : struct ImageLayerCreationOutcome {
920 : image: Option<ResidentLayer>,
921 : next_start_key: Key,
922 : }
923 :
924 : /// Public interface functions
925 : impl Timeline {
926 : /// Get the LSN where this branch was created
927 4 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
928 4 : self.ancestor_lsn
929 4 : }
930 :
931 : /// Get the ancestor's timeline id
932 12 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
933 12 : self.ancestor_timeline
934 12 : .as_ref()
935 12 : .map(|ancestor| ancestor.timeline_id)
936 12 : }
937 :
938 : /// Get the ancestor timeline
939 2 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
940 2 : self.ancestor_timeline.as_ref()
941 2 : }
942 :
943 : /// Get the bytes written since the PITR cutoff on this branch, and
944 : /// whether this branch's ancestor_lsn is within its parent's PITR.
945 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
946 0 : let gc_info = self.gc_info.read().unwrap();
947 0 : let history = self
948 0 : .get_last_record_lsn()
949 0 : .checked_sub(gc_info.cutoffs.time)
950 0 : .unwrap_or(Lsn(0))
951 0 : .0;
952 0 : (history, gc_info.within_ancestor_pitr)
953 0 : }
954 :
955 : /// Lock and get timeline's GC cutoff
956 302 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
957 302 : self.latest_gc_cutoff_lsn.read()
958 302 : }
959 :
960 : /// Look up given page version.
961 : ///
962 : /// If a remote layer file is needed, it is downloaded as part of this
963 : /// call.
964 : ///
965 : /// This method enforces [`Self::pagestream_throttle`] internally.
966 : ///
967 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
968 : /// abstraction above this needs to store suitable metadata to track what
969 : /// data exists with what keys, in separate metadata entries. If a
970 : /// non-existent key is requested, we may incorrectly return a value from
971 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
972 : /// non-existing key.
973 : ///
974 : /// # Cancel-Safety
975 : ///
976 : /// This method is cancellation-safe.
977 : #[inline(always)]
978 607472 : pub(crate) async fn get(
979 607472 : &self,
980 607472 : key: Key,
981 607472 : lsn: Lsn,
982 607472 : ctx: &RequestContext,
983 607472 : ) -> Result<Bytes, PageReconstructError> {
984 607472 : if !lsn.is_valid() {
985 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
986 607472 : }
987 607472 :
988 607472 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
989 607472 : // already checked the key against the shard_identity when looking up the Timeline from
990 607472 : // page_service.
991 607472 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
992 :
993 607472 : let keyspace = KeySpace {
994 607472 : ranges: vec![key..key.next()],
995 607472 : };
996 607472 :
997 607472 : // Initialise the reconstruct state for the key with the cache
998 607472 : // entry returned above.
999 607472 : let mut reconstruct_state = ValuesReconstructState::new();
1000 :
1001 607472 : let vectored_res = self
1002 607472 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
1003 607472 : .await;
1004 :
1005 607472 : let key_value = vectored_res?.pop_first();
1006 607466 : match key_value {
1007 607454 : Some((got_key, value)) => {
1008 607454 : if got_key != key {
1009 0 : error!(
1010 0 : "Expected {}, but singular vectored get returned {}",
1011 : key, got_key
1012 : );
1013 0 : Err(PageReconstructError::Other(anyhow!(
1014 0 : "Singular vectored get returned wrong key"
1015 0 : )))
1016 : } else {
1017 607454 : value
1018 : }
1019 : }
1020 12 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1021 12 : key,
1022 12 : shard: self.shard_identity.get_shard_number(&key),
1023 12 : cont_lsn: Lsn(0),
1024 12 : request_lsn: lsn,
1025 12 : ancestor_lsn: None,
1026 12 : backtrace: None,
1027 12 : })),
1028 : }
1029 607472 : }
1030 :
1031 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1032 : pub(crate) const VEC_GET_LAYERS_VISITED_WARN_THRESH: f64 = 512.0;
1033 :
1034 : /// Look up multiple page versions at a given LSN
1035 : ///
1036 : /// This naive implementation will be replaced with a more efficient one
1037 : /// which actually vectorizes the read path.
1038 19534 : pub(crate) async fn get_vectored(
1039 19534 : &self,
1040 19534 : keyspace: KeySpace,
1041 19534 : lsn: Lsn,
1042 19534 : ctx: &RequestContext,
1043 19534 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1044 19534 : if !lsn.is_valid() {
1045 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1046 19534 : }
1047 19534 :
1048 19534 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1049 19534 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1050 0 : return Err(GetVectoredError::Oversized(key_count));
1051 19534 : }
1052 :
1053 39068 : for range in &keyspace.ranges {
1054 19534 : let mut key = range.start;
1055 39278 : while key != range.end {
1056 19744 : assert!(!self.shard_identity.is_key_disposable(&key));
1057 19744 : key = key.next();
1058 : }
1059 : }
1060 :
1061 19534 : trace!(
1062 0 : "get vectored request for {:?}@{} from task kind {:?}",
1063 0 : keyspace,
1064 0 : lsn,
1065 0 : ctx.task_kind(),
1066 : );
1067 :
1068 19534 : let start = crate::metrics::GET_VECTORED_LATENCY
1069 19534 : .for_task_kind(ctx.task_kind())
1070 19534 : .map(|metric| (metric, Instant::now()));
1071 :
1072 19534 : let res = self
1073 19534 : .get_vectored_impl(
1074 19534 : keyspace.clone(),
1075 19534 : lsn,
1076 19534 : &mut ValuesReconstructState::new(),
1077 19534 : ctx,
1078 19534 : )
1079 19534 : .await;
1080 :
1081 19534 : if let Some((metric, start)) = start {
1082 0 : let elapsed = start.elapsed();
1083 0 : metric.observe(elapsed.as_secs_f64());
1084 19534 : }
1085 :
1086 19534 : res
1087 19534 : }
1088 :
1089 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1090 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1091 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1092 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1093 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1094 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1095 : /// the scan operation will not cause OOM in the future.
1096 12 : pub(crate) async fn scan(
1097 12 : &self,
1098 12 : keyspace: KeySpace,
1099 12 : lsn: Lsn,
1100 12 : ctx: &RequestContext,
1101 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1102 12 : if !lsn.is_valid() {
1103 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1104 12 : }
1105 12 :
1106 12 : trace!(
1107 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1108 0 : keyspace,
1109 0 : lsn,
1110 0 : ctx.task_kind()
1111 : );
1112 :
1113 : // We should generalize this into Keyspace::contains in the future.
1114 24 : for range in &keyspace.ranges {
1115 12 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1116 12 : || range.end.field1 > METADATA_KEY_END_PREFIX
1117 : {
1118 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1119 0 : "only metadata keyspace can be scanned"
1120 0 : )));
1121 12 : }
1122 : }
1123 :
1124 12 : let start = crate::metrics::SCAN_LATENCY
1125 12 : .for_task_kind(ctx.task_kind())
1126 12 : .map(ScanLatencyOngoingRecording::start_recording);
1127 :
1128 12 : let vectored_res = self
1129 12 : .get_vectored_impl(
1130 12 : keyspace.clone(),
1131 12 : lsn,
1132 12 : &mut ValuesReconstructState::default(),
1133 12 : ctx,
1134 12 : )
1135 12 : .await;
1136 :
1137 12 : if let Some(recording) = start {
1138 0 : recording.observe();
1139 12 : }
1140 :
1141 12 : vectored_res
1142 12 : }
1143 :
1144 627324 : pub(super) async fn get_vectored_impl(
1145 627324 : &self,
1146 627324 : keyspace: KeySpace,
1147 627324 : lsn: Lsn,
1148 627324 : reconstruct_state: &mut ValuesReconstructState,
1149 627324 : ctx: &RequestContext,
1150 627324 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1151 627324 : let get_kind = if keyspace.total_raw_size() == 1 {
1152 626880 : GetKind::Singular
1153 : } else {
1154 444 : GetKind::Vectored
1155 : };
1156 :
1157 627324 : let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
1158 627324 : .for_get_kind(get_kind)
1159 627324 : .start_timer();
1160 627324 : self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1161 627324 : .await?;
1162 627308 : get_data_timer.stop_and_record();
1163 627308 :
1164 627308 : let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
1165 627308 : .for_get_kind(get_kind)
1166 627308 : .start_timer();
1167 627308 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
1168 627308 : let layers_visited = reconstruct_state.get_layers_visited();
1169 :
1170 667634 : for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
1171 667634 : match res {
1172 0 : Err(err) => {
1173 0 : results.insert(key, Err(err));
1174 0 : }
1175 667634 : Ok(state) => {
1176 667634 : let state = ValueReconstructState::from(state);
1177 :
1178 667634 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
1179 667634 : results.insert(key, reconstruct_res);
1180 : }
1181 : }
1182 : }
1183 627308 : reconstruct_timer.stop_and_record();
1184 627308 :
1185 627308 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1186 627308 : // when they're missing. Instead they are omitted from the resulting btree
1187 627308 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1188 627308 : // to avoid infinite results.
1189 627308 : if !results.is_empty() {
1190 627098 : let avg = layers_visited as f64 / results.len() as f64;
1191 627098 : if avg >= Self::VEC_GET_LAYERS_VISITED_WARN_THRESH {
1192 0 : use utils::rate_limit::RateLimit;
1193 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1194 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1195 0 : let mut rate_limit = LOGGED.lock().unwrap();
1196 0 : rate_limit.call(|| {
1197 0 : tracing::info!(
1198 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1199 0 : lsn = %lsn,
1200 0 : "Vectored read for {} visited {} layers on average per key and {} in total. {}/{} pages were returned",
1201 0 : keyspace, avg, layers_visited, results.len(), keyspace.total_raw_size());
1202 0 : });
1203 627098 : }
1204 :
1205 : // Note that this is an approximation. Tracking the exact number of layers visited
1206 : // per key requires virtually unbounded memory usage and is inefficient
1207 : // (i.e. segment tree tracking each range queried from a layer)
1208 627098 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED.observe(avg);
1209 210 : }
1210 :
1211 627308 : Ok(results)
1212 627324 : }
1213 :
1214 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1215 274326 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1216 274326 : self.last_record_lsn.load().last
1217 274326 : }
1218 :
1219 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1220 0 : self.last_record_lsn.load().prev
1221 0 : }
1222 :
1223 : /// Atomically get both last and prev.
1224 228 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1225 228 : self.last_record_lsn.load()
1226 228 : }
1227 :
1228 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1229 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1230 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1231 0 : self.last_record_lsn.status_receiver()
1232 0 : }
1233 :
1234 418 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1235 418 : self.disk_consistent_lsn.load()
1236 418 : }
1237 :
1238 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1239 : /// not validated with control plane yet.
1240 : /// See [`Self::get_remote_consistent_lsn_visible`].
1241 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1242 0 : self.remote_client.remote_consistent_lsn_projected()
1243 0 : }
1244 :
1245 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1246 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1247 : /// generation validation in the deletion queue.
1248 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1249 0 : self.remote_client.remote_consistent_lsn_visible()
1250 0 : }
1251 :
1252 : /// The sum of the file size of all historic layers in the layer map.
1253 : /// This method makes no distinction between local and remote layers.
1254 : /// Hence, the result **does not represent local filesystem usage**.
1255 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1256 0 : let guard = self.layers.read().await;
1257 0 : guard.layer_size_sum()
1258 0 : }
1259 :
1260 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1261 0 : self.metrics.resident_physical_size_get()
1262 0 : }
1263 :
1264 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1265 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1266 0 : }
1267 :
1268 : ///
1269 : /// Wait until WAL has been received and processed up to this LSN.
1270 : ///
1271 : /// You should call this before any of the other get_* or list_* functions. Calling
1272 : /// those functions with an LSN that has been processed yet is an error.
1273 : ///
1274 225196 : pub(crate) async fn wait_lsn(
1275 225196 : &self,
1276 225196 : lsn: Lsn,
1277 225196 : who_is_waiting: WaitLsnWaiter<'_>,
1278 225196 : ctx: &RequestContext, /* Prepare for use by cancellation */
1279 225196 : ) -> Result<(), WaitLsnError> {
1280 225196 : let state = self.current_state();
1281 225196 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1282 0 : return Err(WaitLsnError::Shutdown);
1283 225196 : } else if !matches!(state, TimelineState::Active) {
1284 0 : return Err(WaitLsnError::BadState(state));
1285 225196 : }
1286 225196 :
1287 225196 : if cfg!(debug_assertions) {
1288 225196 : match ctx.task_kind() {
1289 : TaskKind::WalReceiverManager
1290 : | TaskKind::WalReceiverConnectionHandler
1291 : | TaskKind::WalReceiverConnectionPoller => {
1292 0 : let is_myself = match who_is_waiting {
1293 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1294 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1295 : };
1296 0 : if is_myself {
1297 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1298 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1299 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1300 0 : }
1301 0 : } else {
1302 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1303 0 : // our walreceiver task can make progress independent of theirs
1304 0 : }
1305 : }
1306 225196 : _ => {}
1307 : }
1308 0 : }
1309 :
1310 225196 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1311 225196 :
1312 225196 : match self
1313 225196 : .last_record_lsn
1314 225196 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1315 225196 : .await
1316 : {
1317 225196 : Ok(()) => Ok(()),
1318 0 : Err(e) => {
1319 : use utils::seqwait::SeqWaitError::*;
1320 0 : match e {
1321 0 : Shutdown => Err(WaitLsnError::Shutdown),
1322 : Timeout => {
1323 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1324 0 : drop(_timer);
1325 0 : let walreceiver_status = self.walreceiver_status();
1326 0 : Err(WaitLsnError::Timeout(format!(
1327 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1328 0 : lsn,
1329 0 : self.get_last_record_lsn(),
1330 0 : self.get_disk_consistent_lsn(),
1331 0 : walreceiver_status,
1332 0 : )))
1333 : }
1334 : }
1335 : }
1336 : }
1337 225196 : }
1338 :
1339 0 : pub(crate) fn walreceiver_status(&self) -> String {
1340 0 : match &*self.walreceiver.lock().unwrap() {
1341 0 : None => "stopping or stopped".to_string(),
1342 0 : Some(walreceiver) => match walreceiver.status() {
1343 0 : Some(status) => status.to_human_readable_string(),
1344 0 : None => "Not active".to_string(),
1345 : },
1346 : }
1347 0 : }
1348 :
1349 : /// Check that it is valid to request operations with that lsn.
1350 232 : pub(crate) fn check_lsn_is_in_scope(
1351 232 : &self,
1352 232 : lsn: Lsn,
1353 232 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1354 232 : ) -> anyhow::Result<()> {
1355 232 : ensure!(
1356 232 : lsn >= **latest_gc_cutoff_lsn,
1357 4 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1358 4 : lsn,
1359 4 : **latest_gc_cutoff_lsn,
1360 : );
1361 228 : Ok(())
1362 232 : }
1363 :
1364 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1365 10 : pub(crate) fn init_lsn_lease(
1366 10 : &self,
1367 10 : lsn: Lsn,
1368 10 : length: Duration,
1369 10 : ctx: &RequestContext,
1370 10 : ) -> anyhow::Result<LsnLease> {
1371 10 : self.make_lsn_lease(lsn, length, true, ctx)
1372 10 : }
1373 :
1374 : /// 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.
1375 4 : pub(crate) fn renew_lsn_lease(
1376 4 : &self,
1377 4 : lsn: Lsn,
1378 4 : length: Duration,
1379 4 : ctx: &RequestContext,
1380 4 : ) -> anyhow::Result<LsnLease> {
1381 4 : self.make_lsn_lease(lsn, length, false, ctx)
1382 4 : }
1383 :
1384 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1385 : ///
1386 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1387 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1388 : ///
1389 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1390 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1391 14 : fn make_lsn_lease(
1392 14 : &self,
1393 14 : lsn: Lsn,
1394 14 : length: Duration,
1395 14 : init: bool,
1396 14 : _ctx: &RequestContext,
1397 14 : ) -> anyhow::Result<LsnLease> {
1398 12 : let lease = {
1399 : // Normalize the requested LSN to be aligned, and move to the first record
1400 : // if it points to the beginning of the page (header).
1401 14 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1402 14 :
1403 14 : let mut gc_info = self.gc_info.write().unwrap();
1404 14 :
1405 14 : let valid_until = SystemTime::now() + length;
1406 14 :
1407 14 : let entry = gc_info.leases.entry(lsn);
1408 14 :
1409 14 : match entry {
1410 6 : Entry::Occupied(mut occupied) => {
1411 6 : let existing_lease = occupied.get_mut();
1412 6 : if valid_until > existing_lease.valid_until {
1413 2 : existing_lease.valid_until = valid_until;
1414 2 : let dt: DateTime<Utc> = valid_until.into();
1415 2 : info!("lease extended to {}", dt);
1416 : } else {
1417 4 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1418 4 : info!("existing lease covers greater length, valid until {}", dt);
1419 : }
1420 :
1421 6 : existing_lease.clone()
1422 : }
1423 8 : Entry::Vacant(vacant) => {
1424 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
1425 : // not blocked by the lsn lease deadline.
1426 8 : let validate = {
1427 8 : let conf = self.tenant_conf.load();
1428 8 : conf.location.attach_mode == AttachmentMode::Single
1429 8 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1430 : };
1431 :
1432 8 : if init || validate {
1433 8 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1434 8 : if lsn < *latest_gc_cutoff_lsn {
1435 2 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1436 6 : }
1437 0 : }
1438 :
1439 6 : let dt: DateTime<Utc> = valid_until.into();
1440 6 : info!("lease created, valid until {}", dt);
1441 6 : vacant.insert(LsnLease { valid_until }).clone()
1442 : }
1443 : }
1444 : };
1445 :
1446 12 : Ok(lease)
1447 14 : }
1448 :
1449 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1450 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1451 0 : #[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))]
1452 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1453 : self.freeze0().await
1454 : }
1455 :
1456 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1457 1096 : #[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))]
1458 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1459 : self.freeze_and_flush0().await
1460 : }
1461 :
1462 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1463 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1464 1096 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1465 1096 : let mut g = self.write_lock.lock().await;
1466 1096 : let to_lsn = self.get_last_record_lsn();
1467 1096 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1468 1096 : }
1469 :
1470 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1471 : // polluting the span hierarchy.
1472 1096 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1473 1096 : let token = self.freeze0().await?;
1474 1096 : self.wait_flush_completion(token).await
1475 1096 : }
1476 :
1477 : // Check if an open ephemeral layer should be closed: this provides
1478 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1479 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1480 : // ephemeral layer bytes has been breached.
1481 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1482 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1483 : // If the write lock is held, there is an active wal receiver: rolling open layers
1484 : // is their responsibility while they hold this lock.
1485 0 : return;
1486 : };
1487 :
1488 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1489 : // time, and this was missed.
1490 : // if write_guard.is_none() { return; }
1491 :
1492 0 : let Ok(layers_guard) = self.layers.try_read() else {
1493 : // Don't block if the layer lock is busy
1494 0 : return;
1495 : };
1496 :
1497 0 : let Ok(lm) = layers_guard.layer_map() else {
1498 0 : return;
1499 : };
1500 :
1501 0 : let Some(open_layer) = &lm.open_layer else {
1502 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1503 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1504 : // that didn't result in writes to this shard.
1505 :
1506 : // Must not hold the layers lock while waiting for a flush.
1507 0 : drop(layers_guard);
1508 0 :
1509 0 : let last_record_lsn = self.get_last_record_lsn();
1510 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1511 0 : if last_record_lsn > disk_consistent_lsn {
1512 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1513 : // we are a sharded tenant and have skipped some WAL
1514 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1515 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1516 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1517 : // without any data ingested (yet) doesn't write a remote index as soon as it
1518 : // sees its LSN advance: we only do this if we've been layer-less
1519 : // for some time.
1520 0 : tracing::debug!(
1521 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1522 : disk_consistent_lsn,
1523 : last_record_lsn
1524 : );
1525 :
1526 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1527 : // We know there is no open layer, so we can request freezing without actually
1528 : // freezing anything. This is true even if we have dropped the layers_guard, we
1529 : // still hold the write_guard.
1530 0 : let _ = async {
1531 0 : let token = self
1532 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1533 0 : .await?;
1534 0 : self.wait_flush_completion(token).await
1535 0 : }
1536 0 : .await;
1537 0 : }
1538 0 : }
1539 :
1540 0 : return;
1541 : };
1542 :
1543 0 : let Some(current_size) = open_layer.try_len() else {
1544 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1545 : // read lock to get size should always succeed.
1546 0 : tracing::warn!("Lock conflict while reading size of open layer");
1547 0 : return;
1548 : };
1549 :
1550 0 : let current_lsn = self.get_last_record_lsn();
1551 :
1552 0 : let checkpoint_distance_override = open_layer.tick().await;
1553 :
1554 0 : if let Some(size_override) = checkpoint_distance_override {
1555 0 : if current_size > size_override {
1556 : // This is not harmful, but it only happens in relatively rare cases where
1557 : // time-based checkpoints are not happening fast enough to keep the amount of
1558 : // ephemeral data within configured limits. It's a sign of stress on the system.
1559 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1560 0 : }
1561 0 : }
1562 :
1563 0 : let checkpoint_distance =
1564 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1565 0 :
1566 0 : if self.should_roll(
1567 0 : current_size,
1568 0 : current_size,
1569 0 : checkpoint_distance,
1570 0 : self.get_last_record_lsn(),
1571 0 : self.last_freeze_at.load(),
1572 0 : open_layer.get_opened_at(),
1573 0 : ) {
1574 0 : match open_layer.info() {
1575 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1576 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1577 0 : // happens asynchronously in the background.
1578 0 : tracing::debug!(
1579 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1580 : );
1581 : }
1582 : InMemoryLayerInfo::Open { .. } => {
1583 : // Upgrade to a write lock and freeze the layer
1584 0 : drop(layers_guard);
1585 0 : let res = self
1586 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1587 0 : .await;
1588 :
1589 0 : if let Err(e) = res {
1590 0 : tracing::info!(
1591 0 : "failed to flush frozen layer after background freeze: {e:#}"
1592 : );
1593 0 : }
1594 : }
1595 : }
1596 0 : }
1597 0 : }
1598 :
1599 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1600 : ///
1601 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1602 : /// child timelines that are not offloaded yet.
1603 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1604 0 : if self.remote_client.is_archived() != Some(true) {
1605 0 : return (false, "the timeline is not archived");
1606 0 : }
1607 0 : if !self.remote_client.no_pending_work() {
1608 : // if the remote client is still processing some work, we can't offload
1609 0 : return (false, "the upload queue is not drained yet");
1610 0 : }
1611 0 :
1612 0 : (true, "ok")
1613 0 : }
1614 :
1615 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1616 : /// compaction tasks.
1617 364 : pub(crate) async fn compact(
1618 364 : self: &Arc<Self>,
1619 364 : cancel: &CancellationToken,
1620 364 : flags: EnumSet<CompactFlags>,
1621 364 : ctx: &RequestContext,
1622 364 : ) -> Result<bool, CompactionError> {
1623 364 : self.compact_with_options(
1624 364 : cancel,
1625 364 : CompactOptions {
1626 364 : flags,
1627 364 : compact_key_range: None,
1628 364 : compact_lsn_range: None,
1629 364 : sub_compaction: false,
1630 364 : sub_compaction_max_job_size_mb: None,
1631 364 : },
1632 364 : ctx,
1633 364 : )
1634 364 : .await
1635 364 : }
1636 :
1637 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1638 : /// compaction tasks.
1639 364 : pub(crate) async fn compact_with_options(
1640 364 : self: &Arc<Self>,
1641 364 : cancel: &CancellationToken,
1642 364 : options: CompactOptions,
1643 364 : ctx: &RequestContext,
1644 364 : ) -> Result<bool, CompactionError> {
1645 364 : // most likely the cancellation token is from background task, but in tests it could be the
1646 364 : // request task as well.
1647 364 :
1648 364 : let prepare = async move {
1649 364 : let guard = self.compaction_lock.lock().await;
1650 :
1651 364 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1652 364 : BackgroundLoopKind::Compaction,
1653 364 : ctx,
1654 364 : )
1655 364 : .await;
1656 :
1657 364 : (guard, permit)
1658 364 : };
1659 :
1660 : // this wait probably never needs any "long time spent" logging, because we already nag if
1661 : // compaction task goes over it's period (20s) which is quite often in production.
1662 364 : let (_guard, _permit) = tokio::select! {
1663 364 : tuple = prepare => { tuple },
1664 364 : _ = self.cancel.cancelled() => return Ok(false),
1665 364 : _ = cancel.cancelled() => return Ok(false),
1666 : };
1667 :
1668 364 : let last_record_lsn = self.get_last_record_lsn();
1669 364 :
1670 364 : // Last record Lsn could be zero in case the timeline was just created
1671 364 : if !last_record_lsn.is_valid() {
1672 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1673 0 : return Ok(false);
1674 364 : }
1675 364 :
1676 364 : match self.get_compaction_algorithm_settings().kind {
1677 : CompactionAlgorithm::Tiered => {
1678 0 : self.compact_tiered(cancel, ctx).await?;
1679 0 : Ok(false)
1680 : }
1681 364 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1682 : }
1683 364 : }
1684 :
1685 : /// Mutate the timeline with a [`TimelineWriter`].
1686 5133166 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1687 5133166 : TimelineWriter {
1688 5133166 : tl: self,
1689 5133166 : write_guard: self.write_lock.lock().await,
1690 : }
1691 5133166 : }
1692 :
1693 0 : pub(crate) fn activate(
1694 0 : self: &Arc<Self>,
1695 0 : parent: Arc<crate::tenant::Tenant>,
1696 0 : broker_client: BrokerClientChannel,
1697 0 : background_jobs_can_start: Option<&completion::Barrier>,
1698 0 : ctx: &RequestContext,
1699 0 : ) {
1700 0 : if self.tenant_shard_id.is_shard_zero() {
1701 0 : // Logical size is only maintained accurately on shard zero.
1702 0 : self.spawn_initial_logical_size_computation_task(ctx);
1703 0 : }
1704 0 : self.launch_wal_receiver(ctx, broker_client);
1705 0 : self.set_state(TimelineState::Active);
1706 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1707 0 : }
1708 :
1709 : /// After this function returns, there are no timeline-scoped tasks are left running.
1710 : ///
1711 : /// The preferred pattern for is:
1712 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1713 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1714 : /// go the extra mile and keep track of JoinHandles
1715 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1716 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1717 : ///
1718 : /// For legacy reasons, we still have multiple tasks spawned using
1719 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1720 : /// We refer to these as "timeline-scoped task_mgr tasks".
1721 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1722 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1723 : /// or [`task_mgr::shutdown_watcher`].
1724 : /// We want to gradually convert the code base away from these.
1725 : ///
1726 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1727 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1728 : /// ones that aren't mentioned here):
1729 : /// - [`TaskKind::TimelineDeletionWorker`]
1730 : /// - NB: also used for tenant deletion
1731 : /// - [`TaskKind::RemoteUploadTask`]`
1732 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1733 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1734 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1735 : /// - [`TaskKind::Eviction`]
1736 : /// - [`TaskKind::LayerFlushTask`]
1737 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1738 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1739 10 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1740 10 : debug_assert_current_span_has_tenant_and_timeline_id();
1741 10 :
1742 10 : // Regardless of whether we're going to try_freeze_and_flush
1743 10 : // or not, stop ingesting any more data. Walreceiver only provides
1744 10 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1745 10 : // So, only after the self.gate.close() below will we know for sure that
1746 10 : // no walreceiver tasks are left.
1747 10 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1748 10 : // data during the call to `self.freeze_and_flush()` below.
1749 10 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1750 10 : // which is what we'd need to properly model early shutdown of the walreceiver
1751 10 : // task sub-tree before the other Timeline task sub-trees.
1752 10 : let walreceiver = self.walreceiver.lock().unwrap().take();
1753 10 : tracing::debug!(
1754 0 : is_some = walreceiver.is_some(),
1755 0 : "Waiting for WalReceiverManager..."
1756 : );
1757 10 : if let Some(walreceiver) = walreceiver {
1758 0 : walreceiver.cancel();
1759 10 : }
1760 : // ... and inform any waiters for newer LSNs that there won't be any.
1761 10 : self.last_record_lsn.shutdown();
1762 10 :
1763 10 : if let ShutdownMode::FreezeAndFlush = mode {
1764 6 : if let Some((open, frozen)) = self
1765 6 : .layers
1766 6 : .read()
1767 6 : .await
1768 6 : .layer_map()
1769 6 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1770 6 : .ok()
1771 6 : .filter(|(open, frozen)| *open || *frozen > 0)
1772 : {
1773 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1774 6 : } else {
1775 6 : // this is double-shutdown, ignore it
1776 6 : }
1777 :
1778 : // we shut down walreceiver above, so, we won't add anything more
1779 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1780 : // to reach the disk & upload queue, then shut the upload queue and
1781 : // wait for it to drain.
1782 6 : match self.freeze_and_flush().await {
1783 : Ok(_) => {
1784 : // drain the upload queue
1785 : // if we did not wait for completion here, it might be our shutdown process
1786 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1787 : // be spawned.
1788 : //
1789 : // what is problematic is the shutting down of RemoteTimelineClient, because
1790 : // obviously it does not make sense to stop while we wait for it, but what
1791 : // about corner cases like s3 suddenly hanging up?
1792 6 : self.remote_client.shutdown().await;
1793 : }
1794 : Err(FlushLayerError::Cancelled) => {
1795 : // this is likely the second shutdown, ignore silently.
1796 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1797 0 : debug_assert!(self.cancel.is_cancelled());
1798 : }
1799 0 : Err(e) => {
1800 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1801 0 : // we have some extra WAL replay to do next time the timeline starts.
1802 0 : warn!("failed to freeze and flush: {e:#}");
1803 : }
1804 : }
1805 :
1806 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
1807 : // we also do a final check here to ensure that the queue is empty.
1808 6 : if !self.remote_client.no_pending_work() {
1809 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1810 6 : }
1811 4 : }
1812 :
1813 10 : if let ShutdownMode::Reload = mode {
1814 : // drain the upload queue
1815 2 : self.remote_client.shutdown().await;
1816 2 : if !self.remote_client.no_pending_work() {
1817 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1818 2 : }
1819 8 : }
1820 :
1821 : // Signal any subscribers to our cancellation token to drop out
1822 10 : tracing::debug!("Cancelling CancellationToken");
1823 10 : self.cancel.cancel();
1824 10 :
1825 10 : // Ensure Prevent new page service requests from starting.
1826 10 : self.handles.shutdown();
1827 10 :
1828 10 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1829 10 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1830 10 : self.remote_client.stop();
1831 10 :
1832 10 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1833 10 : // to shut down the upload queue tasks.
1834 10 : // TODO: fix that, task management should be encapsulated inside remote_client.
1835 10 : task_mgr::shutdown_tasks(
1836 10 : Some(TaskKind::RemoteUploadTask),
1837 10 : Some(self.tenant_shard_id),
1838 10 : Some(self.timeline_id),
1839 10 : )
1840 10 : .await;
1841 :
1842 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
1843 10 : tracing::debug!("Waiting for tasks...");
1844 10 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1845 :
1846 : {
1847 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
1848 : // open.
1849 10 : let mut write_guard = self.write_lock.lock().await;
1850 10 : self.layers.write().await.shutdown(&mut write_guard);
1851 10 : }
1852 10 :
1853 10 : // Finally wait until any gate-holders are complete.
1854 10 : //
1855 10 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1856 10 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1857 10 : self.gate.close().await;
1858 :
1859 10 : self.metrics.shutdown();
1860 10 : }
1861 :
1862 424 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1863 424 : match (self.current_state(), new_state) {
1864 424 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1865 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1866 : }
1867 0 : (st, TimelineState::Loading) => {
1868 0 : error!("ignoring transition from {st:?} into Loading state");
1869 : }
1870 0 : (TimelineState::Broken { .. }, new_state) => {
1871 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1872 : }
1873 : (TimelineState::Stopping, TimelineState::Active) => {
1874 0 : error!("Not activating a Stopping timeline");
1875 : }
1876 422 : (_, new_state) => {
1877 422 : self.state.send_replace(new_state);
1878 422 : }
1879 : }
1880 424 : }
1881 :
1882 2 : pub(crate) fn set_broken(&self, reason: String) {
1883 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1884 2 : let broken_state = TimelineState::Broken {
1885 2 : reason,
1886 2 : backtrace: backtrace_str,
1887 2 : };
1888 2 : self.set_state(broken_state);
1889 2 :
1890 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1891 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1892 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1893 2 : self.cancel.cancel();
1894 2 : }
1895 :
1896 226210 : pub(crate) fn current_state(&self) -> TimelineState {
1897 226210 : self.state.borrow().clone()
1898 226210 : }
1899 :
1900 6 : pub(crate) fn is_broken(&self) -> bool {
1901 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1902 6 : }
1903 :
1904 222 : pub(crate) fn is_active(&self) -> bool {
1905 222 : self.current_state() == TimelineState::Active
1906 222 : }
1907 :
1908 2 : pub(crate) fn is_archived(&self) -> Option<bool> {
1909 2 : self.remote_client.is_archived()
1910 2 : }
1911 :
1912 368 : pub(crate) fn is_stopping(&self) -> bool {
1913 368 : self.current_state() == TimelineState::Stopping
1914 368 : }
1915 :
1916 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1917 0 : self.state.subscribe()
1918 0 : }
1919 :
1920 225198 : pub(crate) async fn wait_to_become_active(
1921 225198 : &self,
1922 225198 : _ctx: &RequestContext, // Prepare for use by cancellation
1923 225198 : ) -> Result<(), TimelineState> {
1924 225198 : let mut receiver = self.state.subscribe();
1925 : loop {
1926 225198 : let current_state = receiver.borrow().clone();
1927 225198 : match current_state {
1928 : TimelineState::Loading => {
1929 0 : receiver
1930 0 : .changed()
1931 0 : .await
1932 0 : .expect("holding a reference to self");
1933 : }
1934 : TimelineState::Active { .. } => {
1935 225196 : return Ok(());
1936 : }
1937 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1938 : // There's no chance the timeline can transition back into ::Active
1939 2 : return Err(current_state);
1940 : }
1941 : }
1942 : }
1943 225198 : }
1944 :
1945 0 : pub(crate) async fn layer_map_info(
1946 0 : &self,
1947 0 : reset: LayerAccessStatsReset,
1948 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
1949 0 : let guard = self.layers.read().await;
1950 0 : let layer_map = guard.layer_map()?;
1951 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
1952 0 : if let Some(open_layer) = &layer_map.open_layer {
1953 0 : in_memory_layers.push(open_layer.info());
1954 0 : }
1955 0 : for frozen_layer in &layer_map.frozen_layers {
1956 0 : in_memory_layers.push(frozen_layer.info());
1957 0 : }
1958 :
1959 0 : let historic_layers = layer_map
1960 0 : .iter_historic_layers()
1961 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
1962 0 : .collect();
1963 0 :
1964 0 : Ok(LayerMapInfo {
1965 0 : in_memory_layers,
1966 0 : historic_layers,
1967 0 : })
1968 0 : }
1969 :
1970 0 : #[instrument(skip_all, fields(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id))]
1971 : pub(crate) async fn download_layer(
1972 : &self,
1973 : layer_file_name: &LayerName,
1974 : ) -> anyhow::Result<Option<bool>> {
1975 : let Some(layer) = self.find_layer(layer_file_name).await? else {
1976 : return Ok(None);
1977 : };
1978 :
1979 : layer.download().await?;
1980 :
1981 : Ok(Some(true))
1982 : }
1983 :
1984 : /// Evict just one layer.
1985 : ///
1986 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
1987 0 : pub(crate) async fn evict_layer(
1988 0 : &self,
1989 0 : layer_file_name: &LayerName,
1990 0 : ) -> anyhow::Result<Option<bool>> {
1991 0 : let _gate = self
1992 0 : .gate
1993 0 : .enter()
1994 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
1995 :
1996 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
1997 0 : return Ok(None);
1998 : };
1999 :
2000 : // curl has this by default
2001 0 : let timeout = std::time::Duration::from_secs(120);
2002 0 :
2003 0 : match local_layer.evict_and_wait(timeout).await {
2004 0 : Ok(()) => Ok(Some(true)),
2005 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2006 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2007 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2008 : }
2009 0 : }
2010 :
2011 4803010 : fn should_roll(
2012 4803010 : &self,
2013 4803010 : layer_size: u64,
2014 4803010 : projected_layer_size: u64,
2015 4803010 : checkpoint_distance: u64,
2016 4803010 : projected_lsn: Lsn,
2017 4803010 : last_freeze_at: Lsn,
2018 4803010 : opened_at: Instant,
2019 4803010 : ) -> bool {
2020 4803010 : let distance = projected_lsn.widening_sub(last_freeze_at);
2021 4803010 :
2022 4803010 : // Rolling the open layer can be triggered by:
2023 4803010 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2024 4803010 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2025 4803010 : // account for how writes are distributed across shards: we expect each node to consume
2026 4803010 : // 1/count of the LSN on average.
2027 4803010 : // 2. The size of the currently open layer.
2028 4803010 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2029 4803010 : // up and suspend activity.
2030 4803010 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2031 0 : info!(
2032 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2033 : projected_lsn, layer_size, distance
2034 : );
2035 :
2036 0 : true
2037 4803010 : } else if projected_layer_size >= checkpoint_distance {
2038 : // NB: this check is relied upon by:
2039 80 : let _ = IndexEntry::validate_checkpoint_distance;
2040 80 : info!(
2041 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2042 : projected_lsn, layer_size, projected_layer_size
2043 : );
2044 :
2045 80 : true
2046 4802930 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2047 0 : info!(
2048 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2049 0 : projected_lsn,
2050 0 : layer_size,
2051 0 : opened_at.elapsed()
2052 : );
2053 :
2054 0 : true
2055 : } else {
2056 4802930 : false
2057 : }
2058 4803010 : }
2059 : }
2060 :
2061 : /// Number of times we will compute partition within a checkpoint distance.
2062 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2063 :
2064 : // Private functions
2065 : impl Timeline {
2066 12 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2067 12 : let tenant_conf = self.tenant_conf.load();
2068 12 : tenant_conf
2069 12 : .tenant_conf
2070 12 : .lsn_lease_length
2071 12 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2072 12 : }
2073 :
2074 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2075 0 : let tenant_conf = self.tenant_conf.load();
2076 0 : tenant_conf
2077 0 : .tenant_conf
2078 0 : .lsn_lease_length_for_ts
2079 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2080 0 : }
2081 :
2082 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2083 0 : let tenant_conf = self.tenant_conf.load();
2084 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2085 0 : }
2086 :
2087 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2088 0 : let tenant_conf = self.tenant_conf.load();
2089 0 : tenant_conf
2090 0 : .tenant_conf
2091 0 : .lazy_slru_download
2092 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2093 0 : }
2094 :
2095 4804628 : fn get_checkpoint_distance(&self) -> u64 {
2096 4804628 : let tenant_conf = self.tenant_conf.load();
2097 4804628 : tenant_conf
2098 4804628 : .tenant_conf
2099 4804628 : .checkpoint_distance
2100 4804628 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2101 4804628 : }
2102 :
2103 4802930 : fn get_checkpoint_timeout(&self) -> Duration {
2104 4802930 : let tenant_conf = self.tenant_conf.load();
2105 4802930 : tenant_conf
2106 4802930 : .tenant_conf
2107 4802930 : .checkpoint_timeout
2108 4802930 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2109 4802930 : }
2110 :
2111 638 : fn get_compaction_target_size(&self) -> u64 {
2112 638 : let tenant_conf = self.tenant_conf.load();
2113 638 : tenant_conf
2114 638 : .tenant_conf
2115 638 : .compaction_target_size
2116 638 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2117 638 : }
2118 :
2119 1540 : fn get_compaction_threshold(&self) -> usize {
2120 1540 : let tenant_conf = self.tenant_conf.load();
2121 1540 : tenant_conf
2122 1540 : .tenant_conf
2123 1540 : .compaction_threshold
2124 1540 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2125 1540 : }
2126 :
2127 14 : fn get_image_creation_threshold(&self) -> usize {
2128 14 : let tenant_conf = self.tenant_conf.load();
2129 14 : tenant_conf
2130 14 : .tenant_conf
2131 14 : .image_creation_threshold
2132 14 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2133 14 : }
2134 :
2135 364 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2136 364 : let tenant_conf = &self.tenant_conf.load();
2137 364 : tenant_conf
2138 364 : .tenant_conf
2139 364 : .compaction_algorithm
2140 364 : .as_ref()
2141 364 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2142 364 : .clone()
2143 364 : }
2144 :
2145 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2146 0 : let tenant_conf = self.tenant_conf.load();
2147 0 : tenant_conf
2148 0 : .tenant_conf
2149 0 : .eviction_policy
2150 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2151 0 : }
2152 :
2153 422 : fn get_evictions_low_residence_duration_metric_threshold(
2154 422 : tenant_conf: &TenantConfOpt,
2155 422 : default_tenant_conf: &TenantConf,
2156 422 : ) -> Duration {
2157 422 : tenant_conf
2158 422 : .evictions_low_residence_duration_metric_threshold
2159 422 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2160 422 : }
2161 :
2162 724 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2163 724 : let tenant_conf = self.tenant_conf.load();
2164 724 : tenant_conf
2165 724 : .tenant_conf
2166 724 : .image_layer_creation_check_threshold
2167 724 : .unwrap_or(
2168 724 : self.conf
2169 724 : .default_tenant_conf
2170 724 : .image_layer_creation_check_threshold,
2171 724 : )
2172 724 : }
2173 :
2174 : /// Resolve the effective WAL receiver protocol to use for this tenant.
2175 : ///
2176 : /// Priority order is:
2177 : /// 1. Tenant config override
2178 : /// 2. Default value for tenant config override
2179 : /// 3. Pageserver config override
2180 : /// 4. Pageserver config default
2181 0 : pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
2182 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2183 0 : tenant_conf
2184 0 : .wal_receiver_protocol_override
2185 0 : .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
2186 0 : .unwrap_or(self.conf.wal_receiver_protocol)
2187 0 : }
2188 :
2189 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2190 0 : // NB: Most tenant conf options are read by background loops, so,
2191 0 : // changes will automatically be picked up.
2192 0 :
2193 0 : // The threshold is embedded in the metric. So, we need to update it.
2194 0 : {
2195 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2196 0 : &new_conf.tenant_conf,
2197 0 : &self.conf.default_tenant_conf,
2198 0 : );
2199 0 :
2200 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2201 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2202 0 :
2203 0 : let timeline_id_str = self.timeline_id.to_string();
2204 0 :
2205 0 : self.remote_client.update_config(&new_conf.location);
2206 0 :
2207 0 : self.metrics
2208 0 : .evictions_with_low_residence_duration
2209 0 : .write()
2210 0 : .unwrap()
2211 0 : .change_threshold(
2212 0 : &tenant_id_str,
2213 0 : &shard_id_str,
2214 0 : &timeline_id_str,
2215 0 : new_threshold,
2216 0 : );
2217 0 : }
2218 0 : }
2219 :
2220 : /// Open a Timeline handle.
2221 : ///
2222 : /// Loads the metadata for the timeline into memory, but not the layer map.
2223 : #[allow(clippy::too_many_arguments)]
2224 422 : pub(super) fn new(
2225 422 : conf: &'static PageServerConf,
2226 422 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2227 422 : metadata: &TimelineMetadata,
2228 422 : ancestor: Option<Arc<Timeline>>,
2229 422 : timeline_id: TimelineId,
2230 422 : tenant_shard_id: TenantShardId,
2231 422 : generation: Generation,
2232 422 : shard_identity: ShardIdentity,
2233 422 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2234 422 : resources: TimelineResources,
2235 422 : pg_version: u32,
2236 422 : state: TimelineState,
2237 422 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2238 422 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2239 422 : cancel: CancellationToken,
2240 422 : ) -> Arc<Self> {
2241 422 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2242 422 : let (state, _) = watch::channel(state);
2243 422 :
2244 422 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2245 422 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2246 422 :
2247 422 : let evictions_low_residence_duration_metric_threshold = {
2248 422 : let loaded_tenant_conf = tenant_conf.load();
2249 422 : Self::get_evictions_low_residence_duration_metric_threshold(
2250 422 : &loaded_tenant_conf.tenant_conf,
2251 422 : &conf.default_tenant_conf,
2252 422 : )
2253 : };
2254 :
2255 422 : if let Some(ancestor) = &ancestor {
2256 230 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2257 230 : // If we construct an explicit timeline object, it's obviously not offloaded
2258 230 : let is_offloaded = MaybeOffloaded::No;
2259 230 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2260 230 : }
2261 :
2262 422 : Arc::new_cyclic(|myself| {
2263 422 : let metrics = TimelineMetrics::new(
2264 422 : &tenant_shard_id,
2265 422 : &timeline_id,
2266 422 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2267 422 : "mtime",
2268 422 : evictions_low_residence_duration_metric_threshold,
2269 422 : ),
2270 422 : );
2271 422 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2272 :
2273 422 : let mut result = Timeline {
2274 422 : conf,
2275 422 : tenant_conf,
2276 422 : myself: myself.clone(),
2277 422 : timeline_id,
2278 422 : tenant_shard_id,
2279 422 : generation,
2280 422 : shard_identity,
2281 422 : pg_version,
2282 422 : layers: Default::default(),
2283 422 :
2284 422 : walredo_mgr,
2285 422 : walreceiver: Mutex::new(None),
2286 422 :
2287 422 : remote_client: Arc::new(resources.remote_client),
2288 422 :
2289 422 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2290 422 : last_record_lsn: SeqWait::new(RecordLsn {
2291 422 : last: disk_consistent_lsn,
2292 422 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2293 422 : }),
2294 422 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2295 422 :
2296 422 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2297 422 : last_freeze_ts: RwLock::new(Instant::now()),
2298 422 :
2299 422 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2300 422 :
2301 422 : ancestor_timeline: ancestor,
2302 422 : ancestor_lsn: metadata.ancestor_lsn(),
2303 422 :
2304 422 : metrics,
2305 422 :
2306 422 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2307 422 : &tenant_shard_id,
2308 422 : &timeline_id,
2309 422 : ),
2310 422 :
2311 2954 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2312 422 :
2313 422 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2314 422 :
2315 422 : layer_flush_start_tx,
2316 422 : layer_flush_done_tx,
2317 422 :
2318 422 : write_lock: tokio::sync::Mutex::new(None),
2319 422 :
2320 422 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2321 422 :
2322 422 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2323 422 : initdb_lsn: metadata.initdb_lsn(),
2324 422 :
2325 422 : current_logical_size: if disk_consistent_lsn.is_valid() {
2326 : // we're creating timeline data with some layer files existing locally,
2327 : // need to recalculate timeline's logical size based on data in the layers.
2328 234 : LogicalSize::deferred_initial(disk_consistent_lsn)
2329 : } else {
2330 : // we're creating timeline data without any layers existing locally,
2331 : // initial logical size is 0.
2332 188 : LogicalSize::empty_initial()
2333 : },
2334 422 : partitioning: tokio::sync::Mutex::new((
2335 422 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2336 422 : Lsn(0),
2337 422 : )),
2338 422 : repartition_threshold: 0,
2339 422 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2340 422 : last_image_layer_creation_check_instant: Mutex::new(None),
2341 422 :
2342 422 : last_received_wal: Mutex::new(None),
2343 422 : rel_size_cache: RwLock::new(RelSizeCache {
2344 422 : complete_as_of: disk_consistent_lsn,
2345 422 : map: HashMap::new(),
2346 422 : }),
2347 422 :
2348 422 : download_all_remote_layers_task_info: RwLock::new(None),
2349 422 :
2350 422 : state,
2351 422 :
2352 422 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2353 422 : EvictionTaskTimelineState::default(),
2354 422 : ),
2355 422 : delete_progress: TimelineDeleteProgress::default(),
2356 422 :
2357 422 : cancel,
2358 422 : gate: Gate::default(),
2359 422 :
2360 422 : compaction_lock: tokio::sync::Mutex::default(),
2361 422 : gc_lock: tokio::sync::Mutex::default(),
2362 422 :
2363 422 : standby_horizon: AtomicLsn::new(0),
2364 422 :
2365 422 : pagestream_throttle: resources.pagestream_throttle,
2366 422 :
2367 422 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2368 422 :
2369 422 : #[cfg(test)]
2370 422 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2371 422 :
2372 422 : l0_flush_global_state: resources.l0_flush_global_state,
2373 422 :
2374 422 : handles: Default::default(),
2375 422 :
2376 422 : attach_wal_lag_cooldown,
2377 422 :
2378 422 : create_idempotency,
2379 422 : };
2380 422 :
2381 422 : result.repartition_threshold =
2382 422 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2383 422 :
2384 422 : result
2385 422 : .metrics
2386 422 : .last_record_lsn_gauge
2387 422 : .set(disk_consistent_lsn.0 as i64);
2388 422 : result
2389 422 : })
2390 422 : }
2391 :
2392 596 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2393 596 : let Ok(guard) = self.gate.enter() else {
2394 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2395 0 : return;
2396 : };
2397 596 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2398 596 : match *flush_loop_state {
2399 416 : FlushLoopState::NotStarted => (),
2400 : FlushLoopState::Running { .. } => {
2401 180 : info!(
2402 0 : "skipping attempt to start flush_loop twice {}/{}",
2403 0 : self.tenant_shard_id, self.timeline_id
2404 : );
2405 180 : return;
2406 : }
2407 : FlushLoopState::Exited => {
2408 0 : warn!(
2409 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2410 0 : self.tenant_shard_id, self.timeline_id
2411 : );
2412 0 : return;
2413 : }
2414 : }
2415 :
2416 416 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2417 416 : let self_clone = Arc::clone(self);
2418 416 :
2419 416 : debug!("spawning flush loop");
2420 416 : *flush_loop_state = FlushLoopState::Running {
2421 416 : #[cfg(test)]
2422 416 : expect_initdb_optimization: false,
2423 416 : #[cfg(test)]
2424 416 : initdb_optimization_count: 0,
2425 416 : };
2426 416 : task_mgr::spawn(
2427 416 : task_mgr::BACKGROUND_RUNTIME.handle(),
2428 416 : task_mgr::TaskKind::LayerFlushTask,
2429 416 : self.tenant_shard_id,
2430 416 : Some(self.timeline_id),
2431 416 : "layer flush task",
2432 416 : async move {
2433 416 : let _guard = guard;
2434 416 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2435 416 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2436 10 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2437 10 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2438 10 : *flush_loop_state = FlushLoopState::Exited;
2439 10 : Ok(())
2440 10 : }
2441 416 : .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))
2442 : );
2443 596 : }
2444 :
2445 : /// Creates and starts the wal receiver.
2446 : ///
2447 : /// This function is expected to be called at most once per Timeline's lifecycle
2448 : /// when the timeline is activated.
2449 0 : fn launch_wal_receiver(
2450 0 : self: &Arc<Self>,
2451 0 : ctx: &RequestContext,
2452 0 : broker_client: BrokerClientChannel,
2453 0 : ) {
2454 0 : info!(
2455 0 : "launching WAL receiver for timeline {} of tenant {}",
2456 0 : self.timeline_id, self.tenant_shard_id
2457 : );
2458 :
2459 0 : let tenant_conf = self.tenant_conf.load();
2460 0 : let wal_connect_timeout = tenant_conf
2461 0 : .tenant_conf
2462 0 : .walreceiver_connect_timeout
2463 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2464 0 : let lagging_wal_timeout = tenant_conf
2465 0 : .tenant_conf
2466 0 : .lagging_wal_timeout
2467 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2468 0 : let max_lsn_wal_lag = tenant_conf
2469 0 : .tenant_conf
2470 0 : .max_lsn_wal_lag
2471 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2472 0 :
2473 0 : let mut guard = self.walreceiver.lock().unwrap();
2474 0 : assert!(
2475 0 : guard.is_none(),
2476 0 : "multiple launches / re-launches of WAL receiver are not supported"
2477 : );
2478 0 : *guard = Some(WalReceiver::start(
2479 0 : Arc::clone(self),
2480 0 : WalReceiverConf {
2481 0 : protocol: self.resolve_wal_receiver_protocol(),
2482 0 : wal_connect_timeout,
2483 0 : lagging_wal_timeout,
2484 0 : max_lsn_wal_lag,
2485 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2486 0 : availability_zone: self.conf.availability_zone.clone(),
2487 0 : ingest_batch_size: self.conf.ingest_batch_size,
2488 0 : },
2489 0 : broker_client,
2490 0 : ctx,
2491 0 : ));
2492 0 : }
2493 :
2494 : /// Initialize with an empty layer map. Used when creating a new timeline.
2495 416 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2496 416 : let mut layers = self.layers.try_write().expect(
2497 416 : "in the context where we call this function, no other task has access to the object",
2498 416 : );
2499 416 : layers
2500 416 : .open_mut()
2501 416 : .expect("in this context the LayerManager must still be open")
2502 416 : .initialize_empty(Lsn(start_lsn.0));
2503 416 : }
2504 :
2505 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2506 : /// files.
2507 6 : pub(super) async fn load_layer_map(
2508 6 : &self,
2509 6 : disk_consistent_lsn: Lsn,
2510 6 : index_part: IndexPart,
2511 6 : ) -> anyhow::Result<()> {
2512 : use init::{Decision::*, Discovered, DismissedLayer};
2513 : use LayerName::*;
2514 :
2515 6 : let mut guard = self.layers.write().await;
2516 :
2517 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2518 6 :
2519 6 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2520 6 : // structs representing all files on disk
2521 6 : let timeline_path = self
2522 6 : .conf
2523 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2524 6 : let conf = self.conf;
2525 6 : let span = tracing::Span::current();
2526 6 :
2527 6 : // Copy to move into the task we're about to spawn
2528 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2529 :
2530 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2531 6 : move || {
2532 6 : let _g = span.entered();
2533 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2534 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2535 6 : let mut unrecognized_files = Vec::new();
2536 6 :
2537 6 : let mut path = timeline_path;
2538 :
2539 22 : for discovered in discovered {
2540 16 : let (name, kind) = match discovered {
2541 16 : Discovered::Layer(layer_file_name, local_metadata) => {
2542 16 : discovered_layers.push((layer_file_name, local_metadata));
2543 16 : continue;
2544 : }
2545 0 : Discovered::IgnoredBackup(path) => {
2546 0 : std::fs::remove_file(path)
2547 0 : .or_else(fs_ext::ignore_not_found)
2548 0 : .fatal_err("Removing .old file");
2549 0 : continue;
2550 : }
2551 0 : Discovered::Unknown(file_name) => {
2552 0 : // we will later error if there are any
2553 0 : unrecognized_files.push(file_name);
2554 0 : continue;
2555 : }
2556 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2557 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2558 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2559 : };
2560 0 : path.push(Utf8Path::new(&name));
2561 0 : init::cleanup(&path, kind)?;
2562 0 : path.pop();
2563 : }
2564 :
2565 6 : if !unrecognized_files.is_empty() {
2566 : // assume that if there are any there are many many.
2567 0 : let n = unrecognized_files.len();
2568 0 : let first = &unrecognized_files[..n.min(10)];
2569 0 : anyhow::bail!(
2570 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2571 0 : );
2572 6 : }
2573 6 :
2574 6 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2575 6 :
2576 6 : let mut loaded_layers = Vec::new();
2577 6 : let mut needs_cleanup = Vec::new();
2578 6 : let mut total_physical_size = 0;
2579 :
2580 22 : for (name, decision) in decided {
2581 16 : let decision = match decision {
2582 16 : Ok(decision) => decision,
2583 0 : Err(DismissedLayer::Future { local }) => {
2584 0 : if let Some(local) = local {
2585 0 : init::cleanup_future_layer(
2586 0 : &local.local_path,
2587 0 : &name,
2588 0 : disk_consistent_lsn,
2589 0 : )?;
2590 0 : }
2591 0 : needs_cleanup.push(name);
2592 0 : continue;
2593 : }
2594 0 : Err(DismissedLayer::LocalOnly(local)) => {
2595 0 : init::cleanup_local_only_file(&name, &local)?;
2596 : // this file never existed remotely, we will have to do rework
2597 0 : continue;
2598 : }
2599 0 : Err(DismissedLayer::BadMetadata(local)) => {
2600 0 : init::cleanup_local_file_for_remote(&local)?;
2601 : // this file never existed remotely, we will have to do rework
2602 0 : continue;
2603 : }
2604 : };
2605 :
2606 16 : match &name {
2607 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2608 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2609 : }
2610 :
2611 16 : tracing::debug!(layer=%name, ?decision, "applied");
2612 :
2613 16 : let layer = match decision {
2614 16 : Resident { local, remote } => {
2615 16 : total_physical_size += local.file_size;
2616 16 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2617 16 : .drop_eviction_guard()
2618 : }
2619 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2620 : };
2621 :
2622 16 : loaded_layers.push(layer);
2623 : }
2624 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2625 6 : }
2626 6 : })
2627 6 : .await
2628 6 : .map_err(anyhow::Error::new)
2629 6 : .and_then(|x| x)?;
2630 :
2631 6 : let num_layers = loaded_layers.len();
2632 6 :
2633 6 : guard
2634 6 : .open_mut()
2635 6 : .expect("layermanager must be open during init")
2636 6 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2637 6 :
2638 6 : self.remote_client
2639 6 : .schedule_layer_file_deletion(&needs_cleanup)?;
2640 6 : self.remote_client
2641 6 : .schedule_index_upload_for_file_changes()?;
2642 : // This barrier orders above DELETEs before any later operations.
2643 : // This is critical because code executing after the barrier might
2644 : // create again objects with the same key that we just scheduled for deletion.
2645 : // For example, if we just scheduled deletion of an image layer "from the future",
2646 : // later compaction might run again and re-create the same image layer.
2647 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2648 : // "same" here means same key range and LSN.
2649 : //
2650 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2651 : // the upload queue may execute the PUT first, then the DELETE.
2652 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2653 : //
2654 : // 1. a future image layer is created and uploaded
2655 : // 2. ps restart
2656 : // 3. the future layer from (1) is deleted during load layer map
2657 : // 4. image layer is re-created and uploaded
2658 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2659 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2660 : //
2661 : // See https://github.com/neondatabase/neon/issues/5878
2662 : //
2663 : // NB: generation numbers naturally protect against this because they disambiguate
2664 : // (1) and (4)
2665 : // TODO: this is basically a no-op now, should we remove it?
2666 6 : self.remote_client.schedule_barrier()?;
2667 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2668 : // on retry.
2669 :
2670 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2671 6 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2672 6 : self.update_layer_visibility().await?;
2673 :
2674 6 : info!(
2675 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2676 : num_layers, disk_consistent_lsn, total_physical_size
2677 : );
2678 :
2679 6 : timer.stop_and_record();
2680 6 : Ok(())
2681 6 : }
2682 :
2683 : /// Retrieve current logical size of the timeline.
2684 : ///
2685 : /// The size could be lagging behind the actual number, in case
2686 : /// the initial size calculation has not been run (gets triggered on the first size access).
2687 : ///
2688 : /// return size and boolean flag that shows if the size is exact
2689 0 : pub(crate) fn get_current_logical_size(
2690 0 : self: &Arc<Self>,
2691 0 : priority: GetLogicalSizePriority,
2692 0 : ctx: &RequestContext,
2693 0 : ) -> logical_size::CurrentLogicalSize {
2694 0 : if !self.tenant_shard_id.is_shard_zero() {
2695 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2696 : // when HTTP API is serving a GET for timeline zero, return zero
2697 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2698 0 : }
2699 0 :
2700 0 : let current_size = self.current_logical_size.current_size();
2701 0 : debug!("Current size: {current_size:?}");
2702 :
2703 0 : match (current_size.accuracy(), priority) {
2704 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2705 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2706 0 : // background task will eventually deliver an exact value, we're in no rush
2707 0 : }
2708 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2709 : // background task is not ready, but user is asking for it now;
2710 : // => make the background task skip the line
2711 : // (The alternative would be to calculate the size here, but,
2712 : // it can actually take a long time if the user has a lot of rels.
2713 : // And we'll inevitable need it again; So, let the background task do the work.)
2714 0 : match self
2715 0 : .current_logical_size
2716 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2717 0 : .get()
2718 : {
2719 0 : Some(cancel) => cancel.cancel(),
2720 : None => {
2721 0 : match self.current_state() {
2722 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2723 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2724 0 : // Don't make noise.
2725 0 : }
2726 : TimelineState::Loading => {
2727 : // Import does not return an activated timeline.
2728 0 : info!("discarding priority boost for logical size calculation because timeline is not yet active");
2729 : }
2730 : TimelineState::Active => {
2731 : // activation should be setting the once cell
2732 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2733 0 : debug_assert!(false);
2734 : }
2735 : }
2736 : }
2737 : }
2738 : }
2739 : }
2740 :
2741 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2742 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2743 0 : let first = self
2744 0 : .current_logical_size
2745 0 : .did_return_approximate_to_walreceiver
2746 0 : .compare_exchange(
2747 0 : false,
2748 0 : true,
2749 0 : AtomicOrdering::Relaxed,
2750 0 : AtomicOrdering::Relaxed,
2751 0 : )
2752 0 : .is_ok();
2753 0 : if first {
2754 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2755 0 : }
2756 0 : }
2757 0 : }
2758 :
2759 0 : current_size
2760 0 : }
2761 :
2762 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2763 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2764 : // nothing to do for freshly created timelines;
2765 0 : assert_eq!(
2766 0 : self.current_logical_size.current_size().accuracy(),
2767 0 : logical_size::Accuracy::Exact,
2768 0 : );
2769 0 : self.current_logical_size.initialized.add_permits(1);
2770 0 : return;
2771 : };
2772 :
2773 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2774 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2775 0 : self.current_logical_size
2776 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2777 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2778 0 :
2779 0 : let self_clone = Arc::clone(self);
2780 0 : let background_ctx = ctx.detached_child(
2781 0 : TaskKind::InitialLogicalSizeCalculation,
2782 0 : DownloadBehavior::Download,
2783 0 : );
2784 0 : task_mgr::spawn(
2785 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2786 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2787 0 : self.tenant_shard_id,
2788 0 : Some(self.timeline_id),
2789 0 : "initial size calculation",
2790 : // NB: don't log errors here, task_mgr will do that.
2791 0 : async move {
2792 0 : let cancel = task_mgr::shutdown_token();
2793 0 : self_clone
2794 0 : .initial_logical_size_calculation_task(
2795 0 : initial_part_end,
2796 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2797 0 : cancel,
2798 0 : background_ctx,
2799 0 : )
2800 0 : .await;
2801 0 : Ok(())
2802 0 : }
2803 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)),
2804 : );
2805 0 : }
2806 :
2807 0 : async fn initial_logical_size_calculation_task(
2808 0 : self: Arc<Self>,
2809 0 : initial_part_end: Lsn,
2810 0 : skip_concurrency_limiter: CancellationToken,
2811 0 : cancel: CancellationToken,
2812 0 : background_ctx: RequestContext,
2813 0 : ) {
2814 0 : scopeguard::defer! {
2815 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2816 0 : self.current_logical_size.initialized.add_permits(1);
2817 0 : }
2818 0 :
2819 0 : let try_once = |attempt: usize| {
2820 0 : let background_ctx = &background_ctx;
2821 0 : let self_ref = &self;
2822 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2823 0 : async move {
2824 0 : let cancel = task_mgr::shutdown_token();
2825 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2826 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2827 0 : background_ctx,
2828 0 : );
2829 :
2830 : use crate::metrics::initial_logical_size::StartCircumstances;
2831 0 : let (_maybe_permit, circumstances) = tokio::select! {
2832 0 : permit = wait_for_permit => {
2833 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2834 : }
2835 0 : _ = self_ref.cancel.cancelled() => {
2836 0 : return Err(CalculateLogicalSizeError::Cancelled);
2837 : }
2838 0 : _ = cancel.cancelled() => {
2839 0 : return Err(CalculateLogicalSizeError::Cancelled);
2840 : },
2841 0 : () = skip_concurrency_limiter.cancelled() => {
2842 : // Some action that is part of a end user interaction requested logical size
2843 : // => break out of the rate limit
2844 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2845 : // but then again what happens if they cancel; also, we should just be using
2846 : // one runtime across the entire process, so, let's leave this for now.
2847 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2848 : }
2849 : };
2850 :
2851 0 : let metrics_guard = if attempt == 1 {
2852 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2853 : } else {
2854 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2855 : };
2856 :
2857 0 : let calculated_size = self_ref
2858 0 : .logical_size_calculation_task(
2859 0 : initial_part_end,
2860 0 : LogicalSizeCalculationCause::Initial,
2861 0 : background_ctx,
2862 0 : )
2863 0 : .await?;
2864 :
2865 0 : self_ref
2866 0 : .trigger_aux_file_size_computation(initial_part_end, background_ctx)
2867 0 : .await?;
2868 :
2869 : // TODO: add aux file size to logical size
2870 :
2871 0 : Ok((calculated_size, metrics_guard))
2872 0 : }
2873 0 : };
2874 :
2875 0 : let retrying = async {
2876 0 : let mut attempt = 0;
2877 : loop {
2878 0 : attempt += 1;
2879 0 :
2880 0 : match try_once(attempt).await {
2881 0 : Ok(res) => return ControlFlow::Continue(res),
2882 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
2883 : Err(
2884 0 : e @ (CalculateLogicalSizeError::Decode(_)
2885 0 : | CalculateLogicalSizeError::PageRead(_)),
2886 0 : ) => {
2887 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2888 : // exponential back-off doesn't make sense at these long intervals;
2889 : // use fixed retry interval with generous jitter instead
2890 0 : let sleep_duration = Duration::from_secs(
2891 0 : u64::try_from(
2892 0 : // 1hour base
2893 0 : (60_i64 * 60_i64)
2894 0 : // 10min jitter
2895 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2896 0 : )
2897 0 : .expect("10min < 1hour"),
2898 0 : );
2899 0 : tokio::time::sleep(sleep_duration).await;
2900 : }
2901 : }
2902 : }
2903 0 : };
2904 :
2905 0 : let (calculated_size, metrics_guard) = tokio::select! {
2906 0 : res = retrying => {
2907 0 : match res {
2908 0 : ControlFlow::Continue(calculated_size) => calculated_size,
2909 0 : ControlFlow::Break(()) => return,
2910 : }
2911 : }
2912 0 : _ = cancel.cancelled() => {
2913 0 : return;
2914 : }
2915 : };
2916 :
2917 : // we cannot query current_logical_size.current_size() to know the current
2918 : // *negative* value, only truncated to u64.
2919 0 : let added = self
2920 0 : .current_logical_size
2921 0 : .size_added_after_initial
2922 0 : .load(AtomicOrdering::Relaxed);
2923 0 :
2924 0 : let sum = calculated_size.saturating_add_signed(added);
2925 0 :
2926 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2927 0 : self.metrics.current_logical_size_gauge.set(sum);
2928 0 :
2929 0 : self.current_logical_size
2930 0 : .initial_logical_size
2931 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
2932 0 : .ok()
2933 0 : .expect("only this task sets it");
2934 0 : }
2935 :
2936 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
2937 0 : self: &Arc<Self>,
2938 0 : lsn: Lsn,
2939 0 : cause: LogicalSizeCalculationCause,
2940 0 : ctx: RequestContext,
2941 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
2942 0 : let (sender, receiver) = oneshot::channel();
2943 0 : let self_clone = Arc::clone(self);
2944 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
2945 0 : // we should stop the size calculation work and return an error.
2946 0 : // That would require restructuring this function's API to
2947 0 : // return the result directly, instead of a Receiver for the result.
2948 0 : let ctx = ctx.detached_child(
2949 0 : TaskKind::OndemandLogicalSizeCalculation,
2950 0 : DownloadBehavior::Download,
2951 0 : );
2952 0 : task_mgr::spawn(
2953 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2954 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
2955 0 : self.tenant_shard_id,
2956 0 : Some(self.timeline_id),
2957 0 : "ondemand logical size calculation",
2958 0 : async move {
2959 0 : let res = self_clone
2960 0 : .logical_size_calculation_task(lsn, cause, &ctx)
2961 0 : .await;
2962 0 : let _ = sender.send(res).ok();
2963 0 : Ok(()) // Receiver is responsible for handling errors
2964 0 : }
2965 0 : .in_current_span(),
2966 0 : );
2967 0 : receiver
2968 0 : }
2969 :
2970 : /// # Cancel-Safety
2971 : ///
2972 : /// This method is cancellation-safe.
2973 0 : #[instrument(skip_all)]
2974 : async fn logical_size_calculation_task(
2975 : self: &Arc<Self>,
2976 : lsn: Lsn,
2977 : cause: LogicalSizeCalculationCause,
2978 : ctx: &RequestContext,
2979 : ) -> Result<u64, CalculateLogicalSizeError> {
2980 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
2981 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
2982 : // accurate relation sizes, and they do not emit consumption metrics.
2983 : debug_assert!(self.tenant_shard_id.is_shard_zero());
2984 :
2985 : let guard = self
2986 : .gate
2987 : .enter()
2988 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
2989 :
2990 : let self_calculation = Arc::clone(self);
2991 :
2992 0 : let mut calculation = pin!(async {
2993 0 : let ctx = ctx.attached_child();
2994 0 : self_calculation
2995 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
2996 0 : .await
2997 0 : });
2998 :
2999 : tokio::select! {
3000 : res = &mut calculation => { res }
3001 : _ = self.cancel.cancelled() => {
3002 : debug!("cancelling logical size calculation for timeline shutdown");
3003 : calculation.await
3004 : }
3005 : }
3006 : }
3007 :
3008 : /// Calculate the logical size of the database at the latest LSN.
3009 : ///
3010 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3011 : /// especially if we need to download remote layers.
3012 : ///
3013 : /// # Cancel-Safety
3014 : ///
3015 : /// This method is cancellation-safe.
3016 0 : async fn calculate_logical_size(
3017 0 : &self,
3018 0 : up_to_lsn: Lsn,
3019 0 : cause: LogicalSizeCalculationCause,
3020 0 : _guard: &GateGuard,
3021 0 : ctx: &RequestContext,
3022 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3023 0 : info!(
3024 0 : "Calculating logical size for timeline {} at {}",
3025 : self.timeline_id, up_to_lsn
3026 : );
3027 :
3028 0 : pausable_failpoint!("timeline-calculate-logical-size-pause");
3029 :
3030 : // See if we've already done the work for initial size calculation.
3031 : // This is a short-cut for timelines that are mostly unused.
3032 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3033 0 : return Ok(size);
3034 0 : }
3035 0 : let storage_time_metrics = match cause {
3036 : LogicalSizeCalculationCause::Initial
3037 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3038 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3039 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3040 0 : &self.metrics.imitate_logical_size_histo
3041 : }
3042 : };
3043 0 : let timer = storage_time_metrics.start_timer();
3044 0 : let logical_size = self
3045 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3046 0 : .await?;
3047 0 : debug!("calculated logical size: {logical_size}");
3048 0 : timer.stop_and_record();
3049 0 : Ok(logical_size)
3050 0 : }
3051 :
3052 : /// Update current logical size, adding `delta' to the old value.
3053 270570 : fn update_current_logical_size(&self, delta: i64) {
3054 270570 : let logical_size = &self.current_logical_size;
3055 270570 : logical_size.increment_size(delta);
3056 270570 :
3057 270570 : // Also set the value in the prometheus gauge. Note that
3058 270570 : // there is a race condition here: if this is is called by two
3059 270570 : // threads concurrently, the prometheus gauge might be set to
3060 270570 : // one value while current_logical_size is set to the
3061 270570 : // other.
3062 270570 : match logical_size.current_size() {
3063 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
3064 270570 : .metrics
3065 270570 : .current_logical_size_gauge
3066 270570 : .set(new_current_size.into()),
3067 0 : CurrentLogicalSize::Approximate(_) => {
3068 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3069 0 : // forth between the initial size calculation task.
3070 0 : }
3071 : }
3072 270570 : }
3073 :
3074 2836 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3075 2836 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3076 2836 : let aux_metric =
3077 2836 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3078 2836 :
3079 2836 : let sum_of_entries = self
3080 2836 : .directory_metrics
3081 2836 : .iter()
3082 19852 : .map(|v| v.load(AtomicOrdering::Relaxed))
3083 2836 : .sum();
3084 : // Set a high general threshold and a lower threshold for the auxiliary files,
3085 : // as we can have large numbers of relations in the db directory.
3086 : const SUM_THRESHOLD: u64 = 5000;
3087 : const AUX_THRESHOLD: u64 = 1000;
3088 2836 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3089 0 : self.metrics
3090 0 : .directory_entries_count_gauge
3091 0 : .set(sum_of_entries);
3092 2836 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3093 0 : metric.set(sum_of_entries);
3094 2836 : }
3095 2836 : }
3096 :
3097 0 : async fn find_layer(
3098 0 : &self,
3099 0 : layer_name: &LayerName,
3100 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3101 0 : let guard = self.layers.read().await;
3102 0 : let layer = guard
3103 0 : .layer_map()?
3104 0 : .iter_historic_layers()
3105 0 : .find(|l| &l.layer_name() == layer_name)
3106 0 : .map(|found| guard.get_from_desc(&found));
3107 0 : Ok(layer)
3108 0 : }
3109 :
3110 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3111 : /// indicating which layers are currently on-disk on the primary.
3112 : ///
3113 : /// None is returned if the Timeline is in a state where uploading a heatmap
3114 : /// doesn't make sense, such as shutting down or initializing. The caller
3115 : /// should treat this as a cue to simply skip doing any heatmap uploading
3116 : /// for this timeline.
3117 2 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3118 2 : if !self.is_active() {
3119 0 : return None;
3120 2 : }
3121 :
3122 2 : let guard = self.layers.read().await;
3123 :
3124 10 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3125 10 : match layer.visibility() {
3126 : LayerVisibilityHint::Visible => {
3127 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3128 8 : let last_activity_ts = layer.latest_activity();
3129 8 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
3130 : }
3131 : LayerVisibilityHint::Covered => {
3132 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3133 2 : None
3134 : }
3135 : }
3136 10 : });
3137 2 :
3138 2 : let mut layers = resident.collect::<Vec<_>>();
3139 2 :
3140 2 : // Sort layers in order of which to download first. For a large set of layers to download, we
3141 2 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3142 2 : // or hours later:
3143 2 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3144 2 : // only exist for a few minutes before being compacted into L1s.
3145 2 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3146 2 : // the layer is likely to be covered by an image layer during compaction.
3147 22 : layers.sort_by_key(|(desc, _meta, _atime)| {
3148 22 : std::cmp::Reverse((
3149 22 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3150 22 : desc.lsn_range.end,
3151 22 : ))
3152 22 : });
3153 2 :
3154 2 : let layers = layers
3155 2 : .into_iter()
3156 8 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3157 2 : .collect();
3158 2 :
3159 2 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3160 2 : }
3161 :
3162 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3163 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3164 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3165 0 : // branchpoint in the value in IndexPart::lineage
3166 0 : self.ancestor_lsn == lsn
3167 0 : || (self.ancestor_lsn == Lsn::INVALID
3168 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3169 0 : }
3170 : }
3171 :
3172 : impl Timeline {
3173 : #[allow(clippy::doc_lazy_continuation)]
3174 : /// Get the data needed to reconstruct all keys in the provided keyspace
3175 : ///
3176 : /// The algorithm is as follows:
3177 : /// 1. While some keys are still not done and there's a timeline to visit:
3178 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3179 : /// 2.1: Build the fringe for the current keyspace
3180 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3181 : /// intersects
3182 : /// 2.3. Pop the timeline from the fringe
3183 : /// 2.4. If the fringe is empty, go back to 1
3184 627324 : async fn get_vectored_reconstruct_data(
3185 627324 : &self,
3186 627324 : mut keyspace: KeySpace,
3187 627324 : request_lsn: Lsn,
3188 627324 : reconstruct_state: &mut ValuesReconstructState,
3189 627324 : ctx: &RequestContext,
3190 627324 : ) -> Result<(), GetVectoredError> {
3191 627324 : let mut timeline_owned: Arc<Timeline>;
3192 627324 : let mut timeline = self;
3193 627324 :
3194 627324 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3195 :
3196 627322 : let missing_keyspace = loop {
3197 852520 : if self.cancel.is_cancelled() {
3198 0 : return Err(GetVectoredError::Cancelled);
3199 852520 : }
3200 :
3201 : let TimelineVisitOutcome {
3202 852520 : completed_keyspace: completed,
3203 852520 : image_covered_keyspace,
3204 852520 : } = Self::get_vectored_reconstruct_data_timeline(
3205 852520 : timeline,
3206 852520 : keyspace.clone(),
3207 852520 : cont_lsn,
3208 852520 : reconstruct_state,
3209 852520 : &self.cancel,
3210 852520 : ctx,
3211 852520 : )
3212 852520 : .await?;
3213 :
3214 852520 : keyspace.remove_overlapping_with(&completed);
3215 852520 :
3216 852520 : // Do not descend into the ancestor timeline for aux files.
3217 852520 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3218 852520 : // stalling compaction.
3219 852520 : keyspace.remove_overlapping_with(&KeySpace {
3220 852520 : ranges: vec![NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE],
3221 852520 : });
3222 852520 :
3223 852520 : // Keyspace is fully retrieved
3224 852520 : if keyspace.is_empty() {
3225 627308 : break None;
3226 225212 : }
3227 :
3228 225212 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3229 : // Not fully retrieved but no ancestor timeline.
3230 14 : break Some(keyspace);
3231 : };
3232 :
3233 : // Now we see if there are keys covered by the image layer but does not exist in the
3234 : // image layer, which means that the key does not exist.
3235 :
3236 : // The block below will stop the vectored search if any of the keys encountered an image layer
3237 : // which did not contain a snapshot for said key. Since we have already removed all completed
3238 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3239 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3240 : // and stop the search as a result of that.
3241 225198 : let removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3242 225198 : if !removed.is_empty() {
3243 0 : break Some(removed);
3244 225198 : }
3245 225198 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3246 225198 : // keyspace.
3247 225198 :
3248 225198 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3249 225198 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3250 225198 : timeline_owned = timeline
3251 225198 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3252 225198 : .await?;
3253 225196 : timeline = &*timeline_owned;
3254 : };
3255 :
3256 627322 : if let Some(missing_keyspace) = missing_keyspace {
3257 14 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3258 14 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3259 14 : shard: self
3260 14 : .shard_identity
3261 14 : .get_shard_number(&missing_keyspace.start().unwrap()),
3262 14 : cont_lsn,
3263 14 : request_lsn,
3264 14 : ancestor_lsn: Some(timeline.ancestor_lsn),
3265 14 : backtrace: None,
3266 14 : }));
3267 627308 : }
3268 627308 :
3269 627308 : Ok(())
3270 627324 : }
3271 :
3272 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3273 : ///
3274 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3275 : /// the current keyspace. The current keyspace of the search at any given timeline
3276 : /// is the original keyspace minus all the keys that have been completed minus
3277 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3278 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3279 : ///
3280 : /// This is basically a depth-first search visitor implementation where a vertex
3281 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3282 : ///
3283 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3284 : /// and get all the required reconstruct data from the layer in one go.
3285 : ///
3286 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3287 : /// decides how to deal with these two keyspaces.
3288 852520 : async fn get_vectored_reconstruct_data_timeline(
3289 852520 : timeline: &Timeline,
3290 852520 : keyspace: KeySpace,
3291 852520 : mut cont_lsn: Lsn,
3292 852520 : reconstruct_state: &mut ValuesReconstructState,
3293 852520 : cancel: &CancellationToken,
3294 852520 : ctx: &RequestContext,
3295 852520 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3296 852520 : let mut unmapped_keyspace = keyspace.clone();
3297 852520 : let mut fringe = LayerFringe::new();
3298 852520 :
3299 852520 : let mut completed_keyspace = KeySpace::default();
3300 852520 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3301 :
3302 : loop {
3303 1699157 : if cancel.is_cancelled() {
3304 0 : return Err(GetVectoredError::Cancelled);
3305 1699157 : }
3306 1699157 :
3307 1699157 : let (keys_done_last_step, keys_with_image_coverage) =
3308 1699157 : reconstruct_state.consume_done_keys();
3309 1699157 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3310 1699157 : completed_keyspace.merge(&keys_done_last_step);
3311 1699157 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3312 22512 : unmapped_keyspace
3313 22512 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3314 22512 : image_covered_keyspace.add_range(keys_with_image_coverage);
3315 1676645 : }
3316 :
3317 : // Do not descent any further if the last layer we visited
3318 : // completed all keys in the keyspace it inspected. This is not
3319 : // required for correctness, but avoids visiting extra layers
3320 : // which turns out to be a perf bottleneck in some cases.
3321 1699157 : if !unmapped_keyspace.is_empty() {
3322 1074097 : let guard = timeline.layers.read().await;
3323 1074097 : let layers = guard.layer_map()?;
3324 :
3325 1074097 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3326 913591 : let start_lsn = l.get_lsn_range().start;
3327 913591 : cont_lsn > start_lsn
3328 1074097 : });
3329 1074097 :
3330 1074097 : match in_memory_layer {
3331 606343 : Some(l) => {
3332 606343 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3333 606343 : fringe.update(
3334 606343 : ReadableLayer::InMemoryLayer(l),
3335 606343 : unmapped_keyspace.clone(),
3336 606343 : lsn_range,
3337 606343 : );
3338 606343 : }
3339 : None => {
3340 467776 : for range in unmapped_keyspace.ranges.iter() {
3341 467776 : let results = layers.range_search(range.clone(), cont_lsn);
3342 467776 :
3343 467776 : results
3344 467776 : .found
3345 467776 : .into_iter()
3346 467776 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3347 240308 : (
3348 240308 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3349 240308 : keyspace_accum.to_keyspace(),
3350 240308 : lsn_floor..cont_lsn,
3351 240308 : )
3352 467776 : })
3353 467776 : .for_each(|(layer, keyspace, lsn_range)| {
3354 240308 : fringe.update(layer, keyspace, lsn_range)
3355 467776 : });
3356 467776 : }
3357 : }
3358 : }
3359 :
3360 : // It's safe to drop the layer map lock after planning the next round of reads.
3361 : // The fringe keeps readable handles for the layers which are safe to read even
3362 : // if layers were compacted or flushed.
3363 : //
3364 : // The more interesting consideration is: "Why is the read algorithm still correct
3365 : // if the layer map changes while it is operating?". Doing a vectored read on a
3366 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3367 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3368 : // range at *a particular point in time*. It is fine for the answer to be different
3369 : // at two different time points.
3370 1074097 : drop(guard);
3371 625060 : }
3372 :
3373 1699157 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3374 846637 : let next_cont_lsn = lsn_range.start;
3375 846637 : layer_to_read
3376 846637 : .get_values_reconstruct_data(
3377 846637 : keyspace_to_read.clone(),
3378 846637 : lsn_range,
3379 846637 : reconstruct_state,
3380 846637 : ctx,
3381 846637 : )
3382 846637 : .await?;
3383 :
3384 846637 : unmapped_keyspace = keyspace_to_read;
3385 846637 : cont_lsn = next_cont_lsn;
3386 846637 :
3387 846637 : reconstruct_state.on_layer_visited(&layer_to_read);
3388 : } else {
3389 852520 : break;
3390 852520 : }
3391 852520 : }
3392 852520 :
3393 852520 : Ok(TimelineVisitOutcome {
3394 852520 : completed_keyspace,
3395 852520 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3396 852520 : })
3397 852520 : }
3398 :
3399 225198 : async fn get_ready_ancestor_timeline(
3400 225198 : &self,
3401 225198 : ancestor: &Arc<Timeline>,
3402 225198 : ctx: &RequestContext,
3403 225198 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3404 225198 : // It's possible that the ancestor timeline isn't active yet, or
3405 225198 : // is active but hasn't yet caught up to the branch point. Wait
3406 225198 : // for it.
3407 225198 : //
3408 225198 : // This cannot happen while the pageserver is running normally,
3409 225198 : // because you cannot create a branch from a point that isn't
3410 225198 : // present in the pageserver yet. However, we don't wait for the
3411 225198 : // branch point to be uploaded to cloud storage before creating
3412 225198 : // a branch. I.e., the branch LSN need not be remote consistent
3413 225198 : // for the branching operation to succeed.
3414 225198 : //
3415 225198 : // Hence, if we try to load a tenant in such a state where
3416 225198 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3417 225198 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3418 225198 : // then we will need to wait for the ancestor timeline to
3419 225198 : // re-stream WAL up to branch_lsn before we access it.
3420 225198 : //
3421 225198 : // How can a tenant get in such a state?
3422 225198 : // - ungraceful pageserver process exit
3423 225198 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3424 225198 : //
3425 225198 : // NB: this could be avoided by requiring
3426 225198 : // branch_lsn >= remote_consistent_lsn
3427 225198 : // during branch creation.
3428 225198 : match ancestor.wait_to_become_active(ctx).await {
3429 225196 : Ok(()) => {}
3430 : Err(TimelineState::Stopping) => {
3431 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3432 0 : return Err(GetReadyAncestorError::Cancelled);
3433 : }
3434 2 : Err(state) => {
3435 2 : return Err(GetReadyAncestorError::BadState {
3436 2 : timeline_id: ancestor.timeline_id,
3437 2 : state,
3438 2 : });
3439 : }
3440 : }
3441 225196 : ancestor
3442 225196 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3443 225196 : .await
3444 225196 : .map_err(|e| match e {
3445 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3446 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3447 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3448 0 : timeline_id: ancestor.timeline_id,
3449 0 : state,
3450 0 : },
3451 225196 : })?;
3452 :
3453 225196 : Ok(ancestor.clone())
3454 225198 : }
3455 :
3456 151304 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3457 151304 : &self.shard_identity
3458 151304 : }
3459 :
3460 : #[inline(always)]
3461 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3462 0 : ShardTimelineId {
3463 0 : shard_index: ShardIndex {
3464 0 : shard_number: self.shard_identity.number,
3465 0 : shard_count: self.shard_identity.count,
3466 0 : },
3467 0 : timeline_id: self.timeline_id,
3468 0 : }
3469 0 : }
3470 :
3471 : /// Returns a non-frozen open in-memory layer for ingestion.
3472 : ///
3473 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3474 : /// this function without holding the mutex.
3475 1272 : async fn get_layer_for_write(
3476 1272 : &self,
3477 1272 : lsn: Lsn,
3478 1272 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3479 1272 : ctx: &RequestContext,
3480 1272 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3481 1272 : let mut guard = self.layers.write().await;
3482 :
3483 1272 : let last_record_lsn = self.get_last_record_lsn();
3484 1272 : ensure!(
3485 1272 : lsn > last_record_lsn,
3486 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3487 : lsn,
3488 : last_record_lsn,
3489 : );
3490 :
3491 1272 : let layer = guard
3492 1272 : .open_mut()?
3493 1272 : .get_layer_for_write(
3494 1272 : lsn,
3495 1272 : self.conf,
3496 1272 : self.timeline_id,
3497 1272 : self.tenant_shard_id,
3498 1272 : &self.gate,
3499 1272 : ctx,
3500 1272 : )
3501 1272 : .await?;
3502 1272 : Ok(layer)
3503 1272 : }
3504 :
3505 5279068 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3506 5279068 : assert!(new_lsn.is_aligned());
3507 :
3508 5279068 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
3509 5279068 : self.last_record_lsn.advance(new_lsn);
3510 5279068 : }
3511 :
3512 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3513 : ///
3514 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3515 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3516 1176 : async fn freeze_inmem_layer_at(
3517 1176 : &self,
3518 1176 : at: Lsn,
3519 1176 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3520 1176 : ) -> Result<u64, FlushLayerError> {
3521 1176 : let frozen = {
3522 1176 : let mut guard = self.layers.write().await;
3523 1176 : guard
3524 1176 : .open_mut()?
3525 1176 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock)
3526 1176 : .await
3527 : };
3528 :
3529 1176 : if frozen {
3530 1148 : let now = Instant::now();
3531 1148 : *(self.last_freeze_ts.write().unwrap()) = now;
3532 1148 : }
3533 :
3534 : // Increment the flush cycle counter and wake up the flush task.
3535 : // Remember the new value, so that when we listen for the flush
3536 : // to finish, we know when the flush that we initiated has
3537 : // finished, instead of some other flush that was started earlier.
3538 1176 : let mut my_flush_request = 0;
3539 1176 :
3540 1176 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3541 1176 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3542 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3543 1176 : }
3544 1176 :
3545 1176 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3546 1176 : my_flush_request = *counter + 1;
3547 1176 : *counter = my_flush_request;
3548 1176 : *lsn = std::cmp::max(at, *lsn);
3549 1176 : });
3550 1176 :
3551 1176 : assert_ne!(my_flush_request, 0);
3552 :
3553 1176 : Ok(my_flush_request)
3554 1176 : }
3555 :
3556 : /// Layer flusher task's main loop.
3557 416 : async fn flush_loop(
3558 416 : self: &Arc<Self>,
3559 416 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3560 416 : ctx: &RequestContext,
3561 416 : ) {
3562 416 : info!("started flush loop");
3563 : loop {
3564 1564 : tokio::select! {
3565 1564 : _ = self.cancel.cancelled() => {
3566 10 : info!("shutting down layer flush task due to Timeline::cancel");
3567 10 : break;
3568 : },
3569 1564 : _ = layer_flush_start_rx.changed() => {}
3570 1148 : }
3571 1148 : trace!("waking up");
3572 1148 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3573 1148 :
3574 1148 : // The highest LSN to which we flushed in the loop over frozen layers
3575 1148 : let mut flushed_to_lsn = Lsn(0);
3576 :
3577 1148 : let result = loop {
3578 2296 : if self.cancel.is_cancelled() {
3579 0 : info!("dropping out of flush loop for timeline shutdown");
3580 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3581 : // anyone waiting on that will respect self.cancel as well: they will stop
3582 : // waiting at the same time we as drop out of this loop.
3583 0 : return;
3584 2296 : }
3585 2296 :
3586 2296 : let timer = self.metrics.flush_time_histo.start_timer();
3587 :
3588 : let num_frozen_layers;
3589 : let frozen_layer_total_size;
3590 2296 : let layer_to_flush = {
3591 2296 : let guard = self.layers.read().await;
3592 2296 : let Ok(lm) = guard.layer_map() else {
3593 0 : info!("dropping out of flush loop for timeline shutdown");
3594 0 : return;
3595 : };
3596 2296 : num_frozen_layers = lm.frozen_layers.len();
3597 2296 : frozen_layer_total_size = lm
3598 2296 : .frozen_layers
3599 2296 : .iter()
3600 2296 : .map(|l| l.estimated_in_mem_size())
3601 2296 : .sum::<u64>();
3602 2296 : lm.frozen_layers.front().cloned()
3603 : // drop 'layers' lock to allow concurrent reads and writes
3604 : };
3605 2296 : let Some(layer_to_flush) = layer_to_flush else {
3606 1148 : break Ok(());
3607 : };
3608 1148 : if num_frozen_layers
3609 1148 : > std::cmp::max(
3610 1148 : self.get_compaction_threshold(),
3611 1148 : DEFAULT_COMPACTION_THRESHOLD,
3612 1148 : )
3613 0 : && frozen_layer_total_size >= /* 128 MB */ 128000000
3614 : {
3615 0 : tracing::warn!(
3616 0 : "too many frozen layers: {num_frozen_layers} layers with estimated in-mem size of {frozen_layer_total_size} bytes",
3617 : );
3618 1148 : }
3619 1148 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3620 1148 : Ok(this_layer_to_lsn) => {
3621 1148 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3622 1148 : }
3623 : Err(FlushLayerError::Cancelled) => {
3624 0 : info!("dropping out of flush loop for timeline shutdown");
3625 0 : return;
3626 : }
3627 0 : err @ Err(
3628 0 : FlushLayerError::NotRunning(_)
3629 0 : | FlushLayerError::Other(_)
3630 0 : | FlushLayerError::CreateImageLayersError(_),
3631 0 : ) => {
3632 0 : error!("could not flush frozen layer: {err:?}");
3633 0 : break err.map(|_| ());
3634 : }
3635 : }
3636 1148 : timer.stop_and_record();
3637 : };
3638 :
3639 : // Unsharded tenants should never advance their LSN beyond the end of the
3640 : // highest layer they write: such gaps between layer data and the frozen LSN
3641 : // are only legal on sharded tenants.
3642 1148 : debug_assert!(
3643 1148 : self.shard_identity.count.count() > 1
3644 1148 : || flushed_to_lsn >= frozen_to_lsn
3645 68 : || !flushed_to_lsn.is_valid()
3646 : );
3647 :
3648 1148 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3649 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3650 : // to us via layer_flush_start_rx, then advance it here.
3651 : //
3652 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3653 : // never encounter a gap in the wal.
3654 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3655 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3656 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3657 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3658 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3659 0 : }
3660 0 : }
3661 1148 : }
3662 :
3663 : // Notify any listeners that we're done
3664 1148 : let _ = self
3665 1148 : .layer_flush_done_tx
3666 1148 : .send_replace((flush_counter, result));
3667 : }
3668 10 : }
3669 :
3670 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
3671 1096 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3672 1096 : let mut rx = self.layer_flush_done_tx.subscribe();
3673 : loop {
3674 : {
3675 2191 : let (last_result_counter, last_result) = &*rx.borrow();
3676 2191 : if *last_result_counter >= request {
3677 1096 : if let Err(err) = last_result {
3678 : // We already logged the original error in
3679 : // flush_loop. We cannot propagate it to the caller
3680 : // here, because it might not be Cloneable
3681 0 : return Err(err.clone());
3682 : } else {
3683 1096 : return Ok(());
3684 : }
3685 1095 : }
3686 1095 : }
3687 1095 : trace!("waiting for flush to complete");
3688 1095 : tokio::select! {
3689 1095 : rx_e = rx.changed() => {
3690 1095 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3691 : },
3692 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3693 : // the notification from [`flush_loop`] that it completed.
3694 1095 : _ = self.cancel.cancelled() => {
3695 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3696 0 : return Ok(())
3697 : }
3698 : };
3699 1095 : trace!("done")
3700 : }
3701 1096 : }
3702 :
3703 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3704 : ///
3705 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3706 1148 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3707 : async fn flush_frozen_layer(
3708 : self: &Arc<Self>,
3709 : frozen_layer: Arc<InMemoryLayer>,
3710 : ctx: &RequestContext,
3711 : ) -> Result<Lsn, FlushLayerError> {
3712 : debug_assert_current_span_has_tenant_and_timeline_id();
3713 :
3714 : // As a special case, when we have just imported an image into the repository,
3715 : // instead of writing out a L0 delta layer, we directly write out image layer
3716 : // files instead. This is possible as long as *all* the data imported into the
3717 : // repository have the same LSN.
3718 : let lsn_range = frozen_layer.get_lsn_range();
3719 :
3720 : // Whether to directly create image layers for this flush, or flush them as delta layers
3721 : let create_image_layer =
3722 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3723 :
3724 : #[cfg(test)]
3725 : {
3726 : match &mut *self.flush_loop_state.lock().unwrap() {
3727 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3728 : panic!("flush loop not running")
3729 : }
3730 : FlushLoopState::Running {
3731 : expect_initdb_optimization,
3732 : initdb_optimization_count,
3733 : ..
3734 : } => {
3735 : if create_image_layer {
3736 : *initdb_optimization_count += 1;
3737 : } else {
3738 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3739 : }
3740 : }
3741 : }
3742 : }
3743 :
3744 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3745 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3746 : // require downloading anything during initial import.
3747 : let ((rel_partition, metadata_partition), _lsn) = self
3748 : .repartition(
3749 : self.initdb_lsn,
3750 : self.get_compaction_target_size(),
3751 : EnumSet::empty(),
3752 : ctx,
3753 : )
3754 : .await
3755 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
3756 :
3757 : if self.cancel.is_cancelled() {
3758 : return Err(FlushLayerError::Cancelled);
3759 : }
3760 :
3761 : let mut layers_to_upload = Vec::new();
3762 : layers_to_upload.extend(
3763 : self.create_image_layers(
3764 : &rel_partition,
3765 : self.initdb_lsn,
3766 : ImageLayerCreationMode::Initial,
3767 : ctx,
3768 : )
3769 : .await?,
3770 : );
3771 : if !metadata_partition.parts.is_empty() {
3772 : assert_eq!(
3773 : metadata_partition.parts.len(),
3774 : 1,
3775 : "currently sparse keyspace should only contain a single metadata keyspace"
3776 : );
3777 : layers_to_upload.extend(
3778 : self.create_image_layers(
3779 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
3780 : // every single key within the keyspace, and therefore, it's safe to force converting it
3781 : // into a dense keyspace before calling this function.
3782 : &metadata_partition.into_dense(),
3783 : self.initdb_lsn,
3784 : ImageLayerCreationMode::Initial,
3785 : ctx,
3786 : )
3787 : .await?,
3788 : );
3789 : }
3790 :
3791 : (layers_to_upload, None)
3792 : } else {
3793 : // Normal case, write out a L0 delta layer file.
3794 : // `create_delta_layer` will not modify the layer map.
3795 : // We will remove frozen layer and add delta layer in one atomic operation later.
3796 : let Some(layer) = self
3797 : .create_delta_layer(&frozen_layer, None, ctx)
3798 : .await
3799 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
3800 : else {
3801 : panic!("delta layer cannot be empty if no filter is applied");
3802 : };
3803 : (
3804 : // FIXME: even though we have a single image and single delta layer assumption
3805 : // we push them to vec
3806 : vec![layer.clone()],
3807 : Some(layer),
3808 : )
3809 : };
3810 :
3811 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
3812 :
3813 : if self.cancel.is_cancelled() {
3814 : return Err(FlushLayerError::Cancelled);
3815 : }
3816 :
3817 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
3818 :
3819 : // The new on-disk layers are now in the layer map. We can remove the
3820 : // in-memory layer from the map now. The flushed layer is stored in
3821 : // the mapping in `create_delta_layer`.
3822 : {
3823 : let mut guard = self.layers.write().await;
3824 :
3825 : guard.open_mut()?.finish_flush_l0_layer(
3826 : delta_layer_to_add.as_ref(),
3827 : &frozen_layer,
3828 : &self.metrics,
3829 : );
3830 :
3831 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
3832 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
3833 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
3834 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
3835 : }
3836 : // release lock on 'layers'
3837 : };
3838 :
3839 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
3840 : // a compaction can delete the file and then it won't be available for uploads any more.
3841 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
3842 : // race situation.
3843 : // See https://github.com/neondatabase/neon/issues/4526
3844 : pausable_failpoint!("flush-frozen-pausable");
3845 :
3846 : // This failpoint is used by another test case `test_pageserver_recovery`.
3847 : fail_point!("flush-frozen-exit");
3848 :
3849 : Ok(Lsn(lsn_range.end.0 - 1))
3850 : }
3851 :
3852 : /// Return true if the value changed
3853 : ///
3854 : /// This function must only be used from the layer flush task.
3855 1148 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
3856 1148 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
3857 1148 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
3858 :
3859 1148 : self.metrics
3860 1148 : .disk_consistent_lsn_gauge
3861 1148 : .set(new_value.0 as i64);
3862 1148 : new_value != old_value
3863 1148 : }
3864 :
3865 : /// Update metadata file
3866 1150 : fn schedule_uploads(
3867 1150 : &self,
3868 1150 : disk_consistent_lsn: Lsn,
3869 1150 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
3870 1150 : ) -> anyhow::Result<()> {
3871 1150 : // We can only save a valid 'prev_record_lsn' value on disk if we
3872 1150 : // flushed *all* in-memory changes to disk. We only track
3873 1150 : // 'prev_record_lsn' in memory for the latest processed record, so we
3874 1150 : // don't remember what the correct value that corresponds to some old
3875 1150 : // LSN is. But if we flush everything, then the value corresponding
3876 1150 : // current 'last_record_lsn' is correct and we can store it on disk.
3877 1150 : let RecordLsn {
3878 1150 : last: last_record_lsn,
3879 1150 : prev: prev_record_lsn,
3880 1150 : } = self.last_record_lsn.load();
3881 1150 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
3882 1070 : Some(prev_record_lsn)
3883 : } else {
3884 80 : None
3885 : };
3886 :
3887 1150 : let update = crate::tenant::metadata::MetadataUpdate::new(
3888 1150 : disk_consistent_lsn,
3889 1150 : ondisk_prev_record_lsn,
3890 1150 : *self.latest_gc_cutoff_lsn.read(),
3891 1150 : );
3892 1150 :
3893 1150 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
3894 0 : "{}",
3895 0 : x.unwrap()
3896 1150 : ));
3897 :
3898 2310 : for layer in layers_to_upload {
3899 1160 : self.remote_client.schedule_layer_file_upload(layer)?;
3900 : }
3901 1150 : self.remote_client
3902 1150 : .schedule_index_upload_for_metadata_update(&update)?;
3903 :
3904 1150 : Ok(())
3905 1150 : }
3906 :
3907 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
3908 0 : self.remote_client
3909 0 : .preserve_initdb_archive(
3910 0 : &self.tenant_shard_id.tenant_id,
3911 0 : &self.timeline_id,
3912 0 : &self.cancel,
3913 0 : )
3914 0 : .await
3915 0 : }
3916 :
3917 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
3918 : // in layer map immediately. The caller is responsible to put it into the layer map.
3919 968 : async fn create_delta_layer(
3920 968 : self: &Arc<Self>,
3921 968 : frozen_layer: &Arc<InMemoryLayer>,
3922 968 : key_range: Option<Range<Key>>,
3923 968 : ctx: &RequestContext,
3924 968 : ) -> anyhow::Result<Option<ResidentLayer>> {
3925 968 : let self_clone = Arc::clone(self);
3926 968 : let frozen_layer = Arc::clone(frozen_layer);
3927 968 : let ctx = ctx.attached_child();
3928 968 : let work = async move {
3929 968 : let Some((desc, path)) = frozen_layer
3930 968 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
3931 968 : .await?
3932 : else {
3933 0 : return Ok(None);
3934 : };
3935 968 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
3936 :
3937 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
3938 : // We just need to fsync the directory in which these inodes are linked,
3939 : // which we know to be the timeline directory.
3940 : //
3941 : // We use fatal_err() below because the after write_to_disk returns with success,
3942 : // the in-memory state of the filesystem already has the layer file in its final place,
3943 : // and subsequent pageserver code could think it's durable while it really isn't.
3944 968 : let timeline_dir = VirtualFile::open(
3945 968 : &self_clone
3946 968 : .conf
3947 968 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
3948 968 : &ctx,
3949 968 : )
3950 968 : .await
3951 968 : .fatal_err("VirtualFile::open for timeline dir fsync");
3952 968 : timeline_dir
3953 968 : .sync_all()
3954 968 : .await
3955 968 : .fatal_err("VirtualFile::sync_all timeline dir");
3956 968 : anyhow::Ok(Some(new_delta))
3957 968 : };
3958 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
3959 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
3960 : use crate::virtual_file::io_engine::IoEngine;
3961 968 : match crate::virtual_file::io_engine::get() {
3962 0 : IoEngine::NotSet => panic!("io engine not set"),
3963 : IoEngine::StdFs => {
3964 484 : let span = tracing::info_span!("blocking");
3965 484 : tokio::task::spawn_blocking({
3966 484 : move || Handle::current().block_on(work.instrument(span))
3967 484 : })
3968 484 : .await
3969 484 : .context("spawn_blocking")
3970 484 : .and_then(|x| x)
3971 : }
3972 : #[cfg(target_os = "linux")]
3973 484 : IoEngine::TokioEpollUring => work.await,
3974 : }
3975 968 : }
3976 :
3977 544 : async fn repartition(
3978 544 : &self,
3979 544 : lsn: Lsn,
3980 544 : partition_size: u64,
3981 544 : flags: EnumSet<CompactFlags>,
3982 544 : ctx: &RequestContext,
3983 544 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
3984 544 : let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
3985 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
3986 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
3987 : // and hence before the compaction task starts.
3988 : // Note that there are a third "caller" that will take the `partitioning` lock. It is `gc_compaction_split_jobs` for
3989 : // gc-compaction where it uses the repartition data to determine the split jobs. In the future, it might use its own
3990 : // heuristics, but for now, we should allow concurrent access to it and let the caller retry compaction.
3991 0 : return Err(CompactionError::Other(anyhow!(
3992 0 : "repartition() called concurrently, this is rare and a retry should be fine"
3993 0 : )));
3994 : };
3995 544 : let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
3996 544 : if lsn < *partition_lsn {
3997 0 : return Err(CompactionError::Other(anyhow!(
3998 0 : "repartition() called with LSN going backwards, this should not happen"
3999 0 : )));
4000 544 : }
4001 544 :
4002 544 : let distance = lsn.0 - partition_lsn.0;
4003 544 : if *partition_lsn != Lsn(0)
4004 262 : && distance <= self.repartition_threshold
4005 262 : && !flags.contains(CompactFlags::ForceRepartition)
4006 : {
4007 248 : debug!(
4008 : distance,
4009 : threshold = self.repartition_threshold,
4010 0 : "no repartitioning needed"
4011 : );
4012 248 : return Ok((
4013 248 : (dense_partition.clone(), sparse_partition.clone()),
4014 248 : *partition_lsn,
4015 248 : ));
4016 296 : }
4017 :
4018 296 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4019 296 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4020 296 : let sparse_partitioning = SparseKeyPartitioning {
4021 296 : parts: vec![sparse_ks],
4022 296 : }; // no partitioning for metadata keys for now
4023 296 : *partitioning_guard = ((dense_partitioning, sparse_partitioning), lsn);
4024 296 :
4025 296 : Ok((partitioning_guard.0.clone(), partitioning_guard.1))
4026 544 : }
4027 :
4028 : // Is it time to create a new image layer for the given partition?
4029 14 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4030 14 : let threshold = self.get_image_creation_threshold();
4031 :
4032 14 : let guard = self.layers.read().await;
4033 14 : let Ok(layers) = guard.layer_map() else {
4034 0 : return false;
4035 : };
4036 :
4037 14 : let mut max_deltas = 0;
4038 28 : for part_range in &partition.ranges {
4039 14 : let image_coverage = layers.image_coverage(part_range, lsn);
4040 28 : for (img_range, last_img) in image_coverage {
4041 14 : let img_lsn = if let Some(last_img) = last_img {
4042 0 : last_img.get_lsn_range().end
4043 : } else {
4044 14 : Lsn(0)
4045 : };
4046 : // Let's consider an example:
4047 : //
4048 : // delta layer with LSN range 71-81
4049 : // delta layer with LSN range 81-91
4050 : // delta layer with LSN range 91-101
4051 : // image layer at LSN 100
4052 : //
4053 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4054 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4055 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4056 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4057 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4058 14 : if img_lsn < lsn {
4059 14 : let num_deltas =
4060 14 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4061 14 :
4062 14 : max_deltas = max_deltas.max(num_deltas);
4063 14 : if num_deltas >= threshold {
4064 0 : debug!(
4065 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4066 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4067 : );
4068 0 : return true;
4069 14 : }
4070 0 : }
4071 : }
4072 : }
4073 :
4074 14 : debug!(
4075 : max_deltas,
4076 0 : "none of the partitioned ranges had >= {threshold} deltas"
4077 : );
4078 14 : false
4079 14 : }
4080 :
4081 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4082 : /// so that at most one image layer will be produced from this function.
4083 206 : async fn create_image_layer_for_rel_blocks(
4084 206 : self: &Arc<Self>,
4085 206 : partition: &KeySpace,
4086 206 : mut image_layer_writer: ImageLayerWriter,
4087 206 : lsn: Lsn,
4088 206 : ctx: &RequestContext,
4089 206 : img_range: Range<Key>,
4090 206 : start: Key,
4091 206 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4092 206 : let mut wrote_keys = false;
4093 206 :
4094 206 : let mut key_request_accum = KeySpaceAccum::new();
4095 1354 : for range in &partition.ranges {
4096 1148 : let mut key = range.start;
4097 2488 : while key < range.end {
4098 : // Decide whether to retain this key: usually we do, but sharded tenants may
4099 : // need to drop keys that don't belong to them. If we retain the key, add it
4100 : // to `key_request_accum` for later issuing a vectored get
4101 1340 : if self.shard_identity.is_key_disposable(&key) {
4102 0 : debug!(
4103 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4104 0 : key,
4105 0 : self.shard_identity.get_shard_number(&key)
4106 : );
4107 1340 : } else {
4108 1340 : key_request_accum.add_key(key);
4109 1340 : }
4110 :
4111 1340 : let last_key_in_range = key.next() == range.end;
4112 1340 : key = key.next();
4113 1340 :
4114 1340 : // Maybe flush `key_rest_accum`
4115 1340 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4116 1340 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4117 : {
4118 1148 : let results = self
4119 1148 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
4120 1148 : .await?;
4121 :
4122 1148 : if self.cancel.is_cancelled() {
4123 0 : return Err(CreateImageLayersError::Cancelled);
4124 1148 : }
4125 :
4126 2488 : for (img_key, img) in results {
4127 1340 : let img = match img {
4128 1340 : Ok(img) => img,
4129 0 : Err(err) => {
4130 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4131 0 : // page without losing any actual user data. That seems better
4132 0 : // than failing repeatedly and getting stuck.
4133 0 : //
4134 0 : // We had a bug at one point, where we truncated the FSM and VM
4135 0 : // in the pageserver, but the Postgres didn't know about that
4136 0 : // and continued to generate incremental WAL records for pages
4137 0 : // that didn't exist in the pageserver. Trying to replay those
4138 0 : // WAL records failed to find the previous image of the page.
4139 0 : // This special case allows us to recover from that situation.
4140 0 : // See https://github.com/neondatabase/neon/issues/2601.
4141 0 : //
4142 0 : // Unfortunately we cannot do this for the main fork, or for
4143 0 : // any metadata keys, keys, as that would lead to actual data
4144 0 : // loss.
4145 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4146 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4147 0 : ZERO_PAGE.clone()
4148 : } else {
4149 0 : return Err(CreateImageLayersError::from(err));
4150 : }
4151 : }
4152 : };
4153 :
4154 : // Write all the keys we just read into our new image layer.
4155 1340 : image_layer_writer.put_image(img_key, img, ctx).await?;
4156 1340 : wrote_keys = true;
4157 : }
4158 192 : }
4159 : }
4160 : }
4161 :
4162 206 : if wrote_keys {
4163 : // Normal path: we have written some data into the new image layer for this
4164 : // partition, so flush it to disk.
4165 206 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4166 206 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4167 206 : info!("created image layer for rel {}", image_layer.local_path());
4168 206 : Ok(ImageLayerCreationOutcome {
4169 206 : image: Some(image_layer),
4170 206 : next_start_key: img_range.end,
4171 206 : })
4172 : } else {
4173 : // Special case: the image layer may be empty if this is a sharded tenant and the
4174 : // partition does not cover any keys owned by this shard. In this case, to ensure
4175 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4176 : // layer we write will cover the key range that we just scanned.
4177 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4178 0 : Ok(ImageLayerCreationOutcome {
4179 0 : image: None,
4180 0 : next_start_key: start,
4181 0 : })
4182 : }
4183 206 : }
4184 :
4185 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4186 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4187 : /// would not be too large to fit in a single image layer.
4188 : #[allow(clippy::too_many_arguments)]
4189 196 : async fn create_image_layer_for_metadata_keys(
4190 196 : self: &Arc<Self>,
4191 196 : partition: &KeySpace,
4192 196 : mut image_layer_writer: ImageLayerWriter,
4193 196 : lsn: Lsn,
4194 196 : ctx: &RequestContext,
4195 196 : img_range: Range<Key>,
4196 196 : mode: ImageLayerCreationMode,
4197 196 : start: Key,
4198 196 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4199 196 : // Metadata keys image layer creation.
4200 196 : let mut reconstruct_state = ValuesReconstructState::default();
4201 196 : let begin = Instant::now();
4202 196 : let data = self
4203 196 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4204 196 : .await?;
4205 196 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4206 196 : let mut new_data = BTreeMap::new();
4207 196 : let mut total_kb_retrieved = 0;
4208 196 : let mut total_keys_retrieved = 0;
4209 10208 : for (k, v) in data {
4210 10012 : let v = v?;
4211 10012 : total_kb_retrieved += KEY_SIZE + v.len();
4212 10012 : total_keys_retrieved += 1;
4213 10012 : new_data.insert(k, v);
4214 : }
4215 196 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4216 196 : };
4217 196 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4218 196 : let elapsed = begin.elapsed();
4219 196 :
4220 196 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4221 196 : info!(
4222 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()
4223 : );
4224 :
4225 196 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4226 2 : return Ok(ImageLayerCreationOutcome {
4227 2 : image: None,
4228 2 : next_start_key: img_range.end,
4229 2 : });
4230 194 : }
4231 194 : if self.cancel.is_cancelled() {
4232 0 : return Err(CreateImageLayersError::Cancelled);
4233 194 : }
4234 194 : let mut wrote_any_image = false;
4235 10206 : for (k, v) in data {
4236 10012 : if v.is_empty() {
4237 : // the key has been deleted, it does not need an image
4238 : // in metadata keyspace, an empty image == tombstone
4239 8 : continue;
4240 10004 : }
4241 10004 : wrote_any_image = true;
4242 10004 :
4243 10004 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4244 10004 :
4245 10004 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4246 10004 : // on the normal data path either.
4247 10004 : image_layer_writer.put_image(k, v, ctx).await?;
4248 : }
4249 :
4250 194 : if wrote_any_image {
4251 : // Normal path: we have written some data into the new image layer for this
4252 : // partition, so flush it to disk.
4253 12 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4254 12 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4255 12 : info!(
4256 0 : "created image layer for metadata {}",
4257 0 : image_layer.local_path()
4258 : );
4259 12 : Ok(ImageLayerCreationOutcome {
4260 12 : image: Some(image_layer),
4261 12 : next_start_key: img_range.end,
4262 12 : })
4263 : } else {
4264 : // Special case: the image layer may be empty if this is a sharded tenant and the
4265 : // partition does not cover any keys owned by this shard. In this case, to ensure
4266 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4267 : // layer we write will cover the key range that we just scanned.
4268 182 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4269 182 : Ok(ImageLayerCreationOutcome {
4270 182 : image: None,
4271 182 : next_start_key: start,
4272 182 : })
4273 : }
4274 196 : }
4275 :
4276 : /// Predicate function which indicates whether we should check if new image layers
4277 : /// are required. Since checking if new image layers are required is expensive in
4278 : /// terms of CPU, we only do it in the following cases:
4279 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4280 : /// 2. If enough time has passed since the last check:
4281 : /// 1. For large tenants, we wish to perform the check more often since they
4282 : /// suffer from the lack of image layers
4283 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4284 724 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4285 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4286 :
4287 724 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4288 724 : let distance = lsn
4289 724 : .checked_sub(last_checks_at)
4290 724 : .expect("Attempt to compact with LSN going backwards");
4291 724 : let min_distance =
4292 724 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4293 724 :
4294 724 : let distance_based_decision = distance.0 >= min_distance;
4295 724 :
4296 724 : let mut time_based_decision = false;
4297 724 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4298 724 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4299 622 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4300 : {
4301 0 : self.get_checkpoint_timeout()
4302 : } else {
4303 622 : Duration::from_secs(3600 * 48)
4304 : };
4305 :
4306 622 : time_based_decision = match *last_check_instant {
4307 442 : Some(last_check) => {
4308 442 : let elapsed = last_check.elapsed();
4309 442 : elapsed >= check_required_after
4310 : }
4311 180 : None => true,
4312 : };
4313 102 : }
4314 :
4315 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4316 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4317 : // check.
4318 724 : let decision = distance_based_decision || time_based_decision;
4319 :
4320 724 : if decision {
4321 182 : self.last_image_layer_creation_check_at.store(lsn);
4322 182 : *last_check_instant = Some(Instant::now());
4323 542 : }
4324 :
4325 724 : decision
4326 724 : }
4327 :
4328 724 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4329 : async fn create_image_layers(
4330 : self: &Arc<Timeline>,
4331 : partitioning: &KeyPartitioning,
4332 : lsn: Lsn,
4333 : mode: ImageLayerCreationMode,
4334 : ctx: &RequestContext,
4335 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4336 : let timer = self.metrics.create_images_time_histo.start_timer();
4337 : let mut image_layers = Vec::new();
4338 :
4339 : // We need to avoid holes between generated image layers.
4340 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4341 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4342 : //
4343 : // How such hole between partitions can appear?
4344 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4345 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4346 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4347 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4348 : let mut start = Key::MIN;
4349 :
4350 : let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
4351 :
4352 : for partition in partitioning.parts.iter() {
4353 : if self.cancel.is_cancelled() {
4354 : return Err(CreateImageLayersError::Cancelled);
4355 : }
4356 :
4357 : let img_range = start..partition.ranges.last().unwrap().end;
4358 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4359 : if compact_metadata {
4360 : for range in &partition.ranges {
4361 : assert!(
4362 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4363 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4364 : "metadata keys must be partitioned separately"
4365 : );
4366 : }
4367 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4368 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4369 : // might mess up with evictions.
4370 : start = img_range.end;
4371 : continue;
4372 : }
4373 : // For initial and force modes, we always generate image layers for metadata keys.
4374 : } else if let ImageLayerCreationMode::Try = mode {
4375 : // check_for_image_layers = false -> skip
4376 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4377 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4378 : start = img_range.end;
4379 : continue;
4380 : }
4381 : }
4382 : if let ImageLayerCreationMode::Force = mode {
4383 : // When forced to create image layers, we might try and create them where they already
4384 : // exist. This mode is only used in tests/debug.
4385 : let layers = self.layers.read().await;
4386 : if layers.contains_key(&PersistentLayerKey {
4387 : key_range: img_range.clone(),
4388 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4389 : is_delta: false,
4390 : }) {
4391 : tracing::info!(
4392 : "Skipping image layer at {lsn} {}..{}, already exists",
4393 : img_range.start,
4394 : img_range.end
4395 : );
4396 : start = img_range.end;
4397 : continue;
4398 : }
4399 : }
4400 :
4401 : let image_layer_writer = ImageLayerWriter::new(
4402 : self.conf,
4403 : self.timeline_id,
4404 : self.tenant_shard_id,
4405 : &img_range,
4406 : lsn,
4407 : ctx,
4408 : )
4409 : .await?;
4410 :
4411 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4412 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4413 0 : "failpoint image-layer-writer-fail-before-finish"
4414 0 : )))
4415 0 : });
4416 :
4417 : if !compact_metadata {
4418 : let ImageLayerCreationOutcome {
4419 : image,
4420 : next_start_key,
4421 : } = self
4422 : .create_image_layer_for_rel_blocks(
4423 : partition,
4424 : image_layer_writer,
4425 : lsn,
4426 : ctx,
4427 : img_range,
4428 : start,
4429 : )
4430 : .await?;
4431 :
4432 : start = next_start_key;
4433 : image_layers.extend(image);
4434 : } else {
4435 : let ImageLayerCreationOutcome {
4436 : image,
4437 : next_start_key,
4438 : } = self
4439 : .create_image_layer_for_metadata_keys(
4440 : partition,
4441 : image_layer_writer,
4442 : lsn,
4443 : ctx,
4444 : img_range,
4445 : mode,
4446 : start,
4447 : )
4448 : .await?;
4449 : start = next_start_key;
4450 : image_layers.extend(image);
4451 : }
4452 : }
4453 :
4454 : let mut guard = self.layers.write().await;
4455 :
4456 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4457 : // now they are being scheduled outside of write lock; current way is inconsistent with
4458 : // compaction lock order.
4459 : guard
4460 : .open_mut()?
4461 : .track_new_image_layers(&image_layers, &self.metrics);
4462 : drop_wlock(guard);
4463 : timer.stop_and_record();
4464 :
4465 : // Creating image layers may have caused some previously visible layers to be covered
4466 : if !image_layers.is_empty() {
4467 : self.update_layer_visibility().await?;
4468 : }
4469 :
4470 : Ok(image_layers)
4471 : }
4472 :
4473 : /// Wait until the background initial logical size calculation is complete, or
4474 : /// this Timeline is shut down. Calling this function will cause the initial
4475 : /// logical size calculation to skip waiting for the background jobs barrier.
4476 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4477 0 : if !self.shard_identity.is_shard_zero() {
4478 : // We don't populate logical size on shard >0: skip waiting for it.
4479 0 : return;
4480 0 : }
4481 0 :
4482 0 : if self.remote_client.is_deleting() {
4483 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4484 0 : return;
4485 0 : }
4486 0 :
4487 0 : if self.current_logical_size.current_size().is_exact() {
4488 : // root timelines are initialized with exact count, but never start the background
4489 : // calculation
4490 0 : return;
4491 0 : }
4492 :
4493 0 : if let Some(await_bg_cancel) = self
4494 0 : .current_logical_size
4495 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4496 0 : .get()
4497 0 : {
4498 0 : await_bg_cancel.cancel();
4499 0 : } else {
4500 : // We should not wait if we were not able to explicitly instruct
4501 : // the logical size cancellation to skip the concurrency limit semaphore.
4502 : // TODO: this is an unexpected case. We should restructure so that it
4503 : // can't happen.
4504 0 : tracing::warn!(
4505 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4506 : );
4507 0 : debug_assert!(false);
4508 : }
4509 :
4510 0 : tokio::select!(
4511 0 : _ = self.current_logical_size.initialized.acquire() => {},
4512 0 : _ = self.cancel.cancelled() => {}
4513 : )
4514 0 : }
4515 :
4516 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4517 : /// Timelines layers up to the ancestor_lsn.
4518 : ///
4519 : /// Requires a timeline that:
4520 : /// - has an ancestor to detach from
4521 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4522 : /// a technical requirement
4523 : ///
4524 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4525 : /// polled again until completion.
4526 : ///
4527 : /// During the operation all timelines sharing the data with this timeline will be reparented
4528 : /// from our ancestor to be branches of this timeline.
4529 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4530 0 : self: &Arc<Timeline>,
4531 0 : tenant: &crate::tenant::Tenant,
4532 0 : options: detach_ancestor::Options,
4533 0 : ctx: &RequestContext,
4534 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
4535 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4536 0 : }
4537 :
4538 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
4539 : /// reparents any reparentable children of previous ancestor.
4540 : ///
4541 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
4542 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
4543 : /// successfully, tenant must be reloaded.
4544 : ///
4545 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
4546 : /// resetting the tenant.
4547 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
4548 0 : self: &Arc<Timeline>,
4549 0 : tenant: &crate::tenant::Tenant,
4550 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4551 0 : ctx: &RequestContext,
4552 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
4553 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
4554 0 : }
4555 :
4556 : /// Final step which unblocks the GC.
4557 : ///
4558 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
4559 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4560 0 : self: &Arc<Timeline>,
4561 0 : tenant: &crate::tenant::Tenant,
4562 0 : attempt: detach_ancestor::Attempt,
4563 0 : ctx: &RequestContext,
4564 0 : ) -> Result<(), detach_ancestor::Error> {
4565 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
4566 0 : }
4567 : }
4568 :
4569 : impl Drop for Timeline {
4570 10 : fn drop(&mut self) {
4571 10 : if let Some(ancestor) = &self.ancestor_timeline {
4572 : // This lock should never be poisoned, but in case it is we do a .map() instead of
4573 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
4574 4 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
4575 4 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
4576 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
4577 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
4578 4 : }
4579 0 : }
4580 6 : }
4581 10 : }
4582 : }
4583 :
4584 : /// Top-level failure to compact.
4585 : #[derive(Debug, thiserror::Error)]
4586 : pub(crate) enum CompactionError {
4587 : #[error("The timeline or pageserver is shutting down")]
4588 : ShuttingDown,
4589 : /// Compaction tried to offload a timeline and failed
4590 : #[error("Failed to offload timeline: {0}")]
4591 : Offload(OffloadError),
4592 : /// Compaction cannot be done right now; page reconstruction and so on.
4593 : #[error(transparent)]
4594 : Other(anyhow::Error),
4595 : }
4596 :
4597 : impl From<OffloadError> for CompactionError {
4598 0 : fn from(e: OffloadError) -> Self {
4599 0 : match e {
4600 0 : OffloadError::Cancelled => Self::ShuttingDown,
4601 0 : _ => Self::Offload(e),
4602 : }
4603 0 : }
4604 : }
4605 :
4606 : impl CompactionError {
4607 0 : pub fn is_cancelled(&self) -> bool {
4608 0 : matches!(self, CompactionError::ShuttingDown)
4609 0 : }
4610 : }
4611 :
4612 : impl From<CollectKeySpaceError> for CompactionError {
4613 0 : fn from(err: CollectKeySpaceError) -> Self {
4614 0 : match err {
4615 : CollectKeySpaceError::Cancelled
4616 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4617 0 : CompactionError::ShuttingDown
4618 : }
4619 0 : e => CompactionError::Other(e.into()),
4620 : }
4621 0 : }
4622 : }
4623 :
4624 : impl From<super::upload_queue::NotInitialized> for CompactionError {
4625 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
4626 0 : match value {
4627 : super::upload_queue::NotInitialized::Uninitialized => {
4628 0 : CompactionError::Other(anyhow::anyhow!(value))
4629 : }
4630 : super::upload_queue::NotInitialized::ShuttingDown
4631 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
4632 : }
4633 0 : }
4634 : }
4635 :
4636 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
4637 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
4638 0 : match e {
4639 : super::storage_layer::layer::DownloadError::TimelineShutdown
4640 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
4641 0 : CompactionError::ShuttingDown
4642 : }
4643 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
4644 : | super::storage_layer::layer::DownloadError::DownloadRequired
4645 : | super::storage_layer::layer::DownloadError::NotFile(_)
4646 : | super::storage_layer::layer::DownloadError::DownloadFailed
4647 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
4648 0 : CompactionError::Other(anyhow::anyhow!(e))
4649 : }
4650 : #[cfg(test)]
4651 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
4652 0 : CompactionError::Other(anyhow::anyhow!(e))
4653 : }
4654 : }
4655 0 : }
4656 : }
4657 :
4658 : impl From<layer_manager::Shutdown> for CompactionError {
4659 0 : fn from(_: layer_manager::Shutdown) -> Self {
4660 0 : CompactionError::ShuttingDown
4661 0 : }
4662 : }
4663 :
4664 : #[serde_as]
4665 196 : #[derive(serde::Serialize)]
4666 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4667 :
4668 : #[derive(Default)]
4669 : enum DurationRecorder {
4670 : #[default]
4671 : NotStarted,
4672 : Recorded(RecordedDuration, tokio::time::Instant),
4673 : }
4674 :
4675 : impl DurationRecorder {
4676 504 : fn till_now(&self) -> DurationRecorder {
4677 504 : match self {
4678 : DurationRecorder::NotStarted => {
4679 0 : panic!("must only call on recorded measurements")
4680 : }
4681 504 : DurationRecorder::Recorded(_, ended) => {
4682 504 : let now = tokio::time::Instant::now();
4683 504 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4684 504 : }
4685 504 : }
4686 504 : }
4687 196 : fn into_recorded(self) -> Option<RecordedDuration> {
4688 196 : match self {
4689 0 : DurationRecorder::NotStarted => None,
4690 196 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4691 : }
4692 196 : }
4693 : }
4694 :
4695 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
4696 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
4697 : /// the layer descriptor requires the user to provide the ranges, which should cover all
4698 : /// keys specified in the `data` field.
4699 : #[cfg(test)]
4700 : #[derive(Clone)]
4701 : pub struct DeltaLayerTestDesc {
4702 : pub lsn_range: Range<Lsn>,
4703 : pub key_range: Range<Key>,
4704 : pub data: Vec<(Key, Lsn, Value)>,
4705 : }
4706 :
4707 : #[cfg(test)]
4708 : impl DeltaLayerTestDesc {
4709 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
4710 2 : Self {
4711 2 : lsn_range,
4712 2 : key_range,
4713 2 : data,
4714 2 : }
4715 2 : }
4716 :
4717 88 : pub fn new_with_inferred_key_range(
4718 88 : lsn_range: Range<Lsn>,
4719 88 : data: Vec<(Key, Lsn, Value)>,
4720 88 : ) -> Self {
4721 220 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
4722 220 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
4723 88 : Self {
4724 88 : key_range: (*key_min)..(key_max.next()),
4725 88 : lsn_range,
4726 88 : data,
4727 88 : }
4728 88 : }
4729 :
4730 10 : pub(crate) fn layer_name(&self) -> LayerName {
4731 10 : LayerName::Delta(super::storage_layer::DeltaLayerName {
4732 10 : key_range: self.key_range.clone(),
4733 10 : lsn_range: self.lsn_range.clone(),
4734 10 : })
4735 10 : }
4736 : }
4737 :
4738 : impl Timeline {
4739 28 : async fn finish_compact_batch(
4740 28 : self: &Arc<Self>,
4741 28 : new_deltas: &[ResidentLayer],
4742 28 : new_images: &[ResidentLayer],
4743 28 : layers_to_remove: &[Layer],
4744 28 : ) -> Result<(), CompactionError> {
4745 28 : let mut guard = tokio::select! {
4746 28 : guard = self.layers.write() => guard,
4747 28 : _ = self.cancel.cancelled() => {
4748 0 : return Err(CompactionError::ShuttingDown);
4749 : }
4750 : };
4751 :
4752 28 : let mut duplicated_layers = HashSet::new();
4753 28 :
4754 28 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
4755 :
4756 336 : for l in new_deltas {
4757 308 : if guard.contains(l.as_ref()) {
4758 : // expected in tests
4759 0 : tracing::error!(layer=%l, "duplicated L1 layer");
4760 :
4761 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
4762 : // `guard` on self.layers. as of writing this, there are no error returns except
4763 : // for compact_level0_phase1 creating an L0, which does not happen in practice
4764 : // because we have not implemented L0 => L0 compaction.
4765 0 : duplicated_layers.insert(l.layer_desc().key());
4766 308 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
4767 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
4768 308 : } else {
4769 308 : insert_layers.push(l.clone());
4770 308 : }
4771 : }
4772 :
4773 : // only remove those inputs which were not outputs
4774 28 : let remove_layers: Vec<Layer> = layers_to_remove
4775 28 : .iter()
4776 402 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
4777 28 : .cloned()
4778 28 : .collect();
4779 28 :
4780 28 : if !new_images.is_empty() {
4781 0 : guard
4782 0 : .open_mut()?
4783 0 : .track_new_image_layers(new_images, &self.metrics);
4784 28 : }
4785 :
4786 28 : guard
4787 28 : .open_mut()?
4788 28 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
4789 28 :
4790 28 : self.remote_client
4791 28 : .schedule_compaction_update(&remove_layers, new_deltas)?;
4792 :
4793 28 : drop_wlock(guard);
4794 28 :
4795 28 : Ok(())
4796 28 : }
4797 :
4798 0 : async fn rewrite_layers(
4799 0 : self: &Arc<Self>,
4800 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
4801 0 : mut drop_layers: Vec<Layer>,
4802 0 : ) -> Result<(), CompactionError> {
4803 0 : let mut guard = self.layers.write().await;
4804 :
4805 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
4806 : // to avoid double-removing, and avoid rewriting something that was removed.
4807 0 : replace_layers.retain(|(l, _)| guard.contains(l));
4808 0 : drop_layers.retain(|l| guard.contains(l));
4809 0 :
4810 0 : guard
4811 0 : .open_mut()?
4812 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
4813 0 :
4814 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
4815 0 :
4816 0 : self.remote_client
4817 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
4818 :
4819 0 : Ok(())
4820 0 : }
4821 :
4822 : /// Schedules the uploads of the given image layers
4823 364 : fn upload_new_image_layers(
4824 364 : self: &Arc<Self>,
4825 364 : new_images: impl IntoIterator<Item = ResidentLayer>,
4826 364 : ) -> Result<(), super::upload_queue::NotInitialized> {
4827 390 : for layer in new_images {
4828 26 : self.remote_client.schedule_layer_file_upload(layer)?;
4829 : }
4830 : // should any new image layer been created, not uploading index_part will
4831 : // result in a mismatch between remote_physical_size and layermap calculated
4832 : // size, which will fail some tests, but should not be an issue otherwise.
4833 364 : self.remote_client
4834 364 : .schedule_index_upload_for_file_changes()?;
4835 364 : Ok(())
4836 364 : }
4837 :
4838 0 : async fn find_gc_time_cutoff(
4839 0 : &self,
4840 0 : pitr: Duration,
4841 0 : cancel: &CancellationToken,
4842 0 : ctx: &RequestContext,
4843 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
4844 0 : debug_assert_current_span_has_tenant_and_timeline_id();
4845 0 : if self.shard_identity.is_shard_zero() {
4846 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
4847 0 : let now = SystemTime::now();
4848 0 : let time_range = if pitr == Duration::ZERO {
4849 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
4850 : } else {
4851 0 : pitr
4852 : };
4853 :
4854 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
4855 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
4856 0 : let timestamp = to_pg_timestamp(time_cutoff);
4857 :
4858 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
4859 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
4860 0 : LsnForTimestamp::Future(lsn) => {
4861 0 : // The timestamp is in the future. That sounds impossible,
4862 0 : // but what it really means is that there hasn't been
4863 0 : // any commits since the cutoff timestamp.
4864 0 : //
4865 0 : // In this case we should use the LSN of the most recent commit,
4866 0 : // which is implicitly the last LSN in the log.
4867 0 : debug!("future({})", lsn);
4868 0 : Some(self.get_last_record_lsn())
4869 : }
4870 0 : LsnForTimestamp::Past(lsn) => {
4871 0 : debug!("past({})", lsn);
4872 0 : None
4873 : }
4874 0 : LsnForTimestamp::NoData(lsn) => {
4875 0 : debug!("nodata({})", lsn);
4876 0 : None
4877 : }
4878 : };
4879 0 : Ok(time_cutoff)
4880 : } else {
4881 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
4882 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
4883 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
4884 : // space cutoff that we would also have respected ourselves.
4885 0 : match self
4886 0 : .remote_client
4887 0 : .download_foreign_index(ShardNumber(0), cancel)
4888 0 : .await
4889 : {
4890 0 : Ok((index_part, index_generation, _index_mtime)) => {
4891 0 : tracing::info!("GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
4892 0 : index_part.metadata.latest_gc_cutoff_lsn());
4893 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
4894 : }
4895 : Err(DownloadError::NotFound) => {
4896 : // This is unexpected, because during timeline creations shard zero persists to remote
4897 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
4898 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
4899 : // state, then shard zero should _eventually_ write an index when it GCs.
4900 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
4901 0 : Ok(None)
4902 : }
4903 0 : Err(e) => {
4904 0 : // TODO: this function should return a different error type than page reconstruct error
4905 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
4906 : }
4907 : }
4908 :
4909 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
4910 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
4911 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
4912 : // new location. This is legal in principle, but problematic in practice because it might result
4913 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
4914 : // because they have GC'd past the branch point.
4915 : }
4916 0 : }
4917 :
4918 : /// Find the Lsns above which layer files need to be retained on
4919 : /// garbage collection.
4920 : ///
4921 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
4922 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
4923 : /// the space-based retention.
4924 : ///
4925 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
4926 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
4927 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
4928 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
4929 : /// in the response as the GC cutoff point for the timeline.
4930 4 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
4931 : pub(super) async fn find_gc_cutoffs(
4932 : &self,
4933 : space_cutoff: Lsn,
4934 : pitr: Duration,
4935 : cancel: &CancellationToken,
4936 : ctx: &RequestContext,
4937 : ) -> Result<GcCutoffs, PageReconstructError> {
4938 : let _timer = self
4939 : .metrics
4940 : .find_gc_cutoffs_histo
4941 : .start_timer()
4942 : .record_on_drop();
4943 :
4944 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
4945 :
4946 : if cfg!(test) {
4947 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
4948 : if pitr == Duration::ZERO {
4949 : return Ok(GcCutoffs {
4950 : time: self.get_last_record_lsn(),
4951 : space: space_cutoff,
4952 : });
4953 : }
4954 : }
4955 :
4956 : // Calculate a time-based limit on how much to retain:
4957 : // - if PITR interval is set, then this is our cutoff.
4958 : // - if PITR interval is not set, then we do a lookup
4959 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
4960 : let time_cutoff = self.find_gc_time_cutoff(pitr, cancel, ctx).await?;
4961 :
4962 : Ok(match (pitr, time_cutoff) {
4963 : (Duration::ZERO, Some(time_cutoff)) => {
4964 : // PITR is not set. Retain the size-based limit, or the default time retention,
4965 : // whichever requires less data.
4966 : GcCutoffs {
4967 : time: self.get_last_record_lsn(),
4968 : space: std::cmp::max(time_cutoff, space_cutoff),
4969 : }
4970 : }
4971 : (Duration::ZERO, None) => {
4972 : // PITR is not set, and time lookup failed
4973 : GcCutoffs {
4974 : time: self.get_last_record_lsn(),
4975 : space: space_cutoff,
4976 : }
4977 : }
4978 : (_, None) => {
4979 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
4980 : // cannot advance beyond what was already GC'd, and respect space-based retention
4981 : GcCutoffs {
4982 : time: *self.get_latest_gc_cutoff_lsn(),
4983 : space: space_cutoff,
4984 : }
4985 : }
4986 : (_, Some(time_cutoff)) => {
4987 : // PITR interval is set and we looked up timestamp successfully. Ignore
4988 : // size based retention and make time cutoff authoritative
4989 : GcCutoffs {
4990 : time: time_cutoff,
4991 : space: time_cutoff,
4992 : }
4993 : }
4994 : })
4995 : }
4996 :
4997 : /// Garbage collect layer files on a timeline that are no longer needed.
4998 : ///
4999 : /// Currently, we don't make any attempt at removing unneeded page versions
5000 : /// within a layer file. We can only remove the whole file if it's fully
5001 : /// obsolete.
5002 4 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
5003 : // this is most likely the background tasks, but it might be the spawned task from
5004 : // immediate_gc
5005 4 : let _g = tokio::select! {
5006 4 : guard = self.gc_lock.lock() => guard,
5007 4 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
5008 : };
5009 4 : let timer = self.metrics.garbage_collect_histo.start_timer();
5010 4 :
5011 4 : fail_point!("before-timeline-gc");
5012 4 :
5013 4 : // Is the timeline being deleted?
5014 4 : if self.is_stopping() {
5015 0 : return Err(GcError::TimelineCancelled);
5016 4 : }
5017 4 :
5018 4 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
5019 4 : let gc_info = self.gc_info.read().unwrap();
5020 4 :
5021 4 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
5022 4 : let time_cutoff = gc_info.cutoffs.time;
5023 4 : let retain_lsns = gc_info
5024 4 : .retain_lsns
5025 4 : .iter()
5026 4 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
5027 4 : .collect();
5028 4 :
5029 4 : // Gets the maximum LSN that holds the valid lease.
5030 4 : //
5031 4 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
5032 4 : // Here, we do not check for stale leases again.
5033 4 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
5034 4 :
5035 4 : (
5036 4 : space_cutoff,
5037 4 : time_cutoff,
5038 4 : retain_lsns,
5039 4 : max_lsn_with_valid_lease,
5040 4 : )
5041 4 : };
5042 4 :
5043 4 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
5044 4 : let standby_horizon = self.standby_horizon.load();
5045 4 : // Hold GC for the standby, but as a safety guard do it only within some
5046 4 : // reasonable lag.
5047 4 : if standby_horizon != Lsn::INVALID {
5048 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
5049 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
5050 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
5051 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
5052 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
5053 : } else {
5054 0 : warn!(
5055 0 : "standby is lagging for more than {}MB, not holding gc for it",
5056 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
5057 : )
5058 : }
5059 0 : }
5060 4 : }
5061 :
5062 : // Reset standby horizon to ignore it if it is not updated till next GC.
5063 : // It is an easy way to unset it when standby disappears without adding
5064 : // more conf options.
5065 4 : self.standby_horizon.store(Lsn::INVALID);
5066 4 : self.metrics
5067 4 : .standby_horizon_gauge
5068 4 : .set(Lsn::INVALID.0 as i64);
5069 :
5070 4 : let res = self
5071 4 : .gc_timeline(
5072 4 : space_cutoff,
5073 4 : time_cutoff,
5074 4 : retain_lsns,
5075 4 : max_lsn_with_valid_lease,
5076 4 : new_gc_cutoff,
5077 4 : )
5078 4 : .instrument(
5079 4 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5080 : )
5081 4 : .await?;
5082 :
5083 : // only record successes
5084 4 : timer.stop_and_record();
5085 4 :
5086 4 : Ok(res)
5087 4 : }
5088 :
5089 4 : async fn gc_timeline(
5090 4 : &self,
5091 4 : space_cutoff: Lsn,
5092 4 : time_cutoff: Lsn,
5093 4 : retain_lsns: Vec<Lsn>,
5094 4 : max_lsn_with_valid_lease: Option<Lsn>,
5095 4 : new_gc_cutoff: Lsn,
5096 4 : ) -> Result<GcResult, GcError> {
5097 4 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5098 4 :
5099 4 : let now = SystemTime::now();
5100 4 : let mut result: GcResult = GcResult::default();
5101 4 :
5102 4 : // Nothing to GC. Return early.
5103 4 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
5104 4 : if latest_gc_cutoff >= new_gc_cutoff {
5105 0 : info!(
5106 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5107 : );
5108 0 : return Ok(result);
5109 4 : }
5110 :
5111 : // We need to ensure that no one tries to read page versions or create
5112 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5113 : // for details. This will block until the old value is no longer in use.
5114 : //
5115 : // The GC cutoff should only ever move forwards.
5116 4 : let waitlist = {
5117 4 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
5118 4 : if *write_guard > new_gc_cutoff {
5119 0 : return Err(GcError::BadLsn {
5120 0 : why: format!(
5121 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5122 0 : *write_guard, new_gc_cutoff
5123 0 : ),
5124 0 : });
5125 4 : }
5126 4 :
5127 4 : write_guard.store_and_unlock(new_gc_cutoff)
5128 4 : };
5129 4 : waitlist.wait().await;
5130 :
5131 4 : info!("GC starting");
5132 :
5133 4 : debug!("retain_lsns: {:?}", retain_lsns);
5134 :
5135 4 : let mut layers_to_remove = Vec::new();
5136 :
5137 : // Scan all layers in the timeline (remote or on-disk).
5138 : //
5139 : // Garbage collect the layer if all conditions are satisfied:
5140 : // 1. it is older than cutoff LSN;
5141 : // 2. it is older than PITR interval;
5142 : // 3. it doesn't need to be retained for 'retain_lsns';
5143 : // 4. it does not need to be kept for LSNs holding valid leases.
5144 : // 5. newer on-disk image layers cover the layer's whole key range
5145 : //
5146 : // TODO holding a write lock is too agressive and avoidable
5147 4 : let mut guard = self.layers.write().await;
5148 4 : let layers = guard.layer_map()?;
5149 24 : 'outer: for l in layers.iter_historic_layers() {
5150 24 : result.layers_total += 1;
5151 24 :
5152 24 : // 1. Is it newer than GC horizon cutoff point?
5153 24 : if l.get_lsn_range().end > space_cutoff {
5154 2 : info!(
5155 0 : "keeping {} because it's newer than space_cutoff {}",
5156 0 : l.layer_name(),
5157 : space_cutoff,
5158 : );
5159 2 : result.layers_needed_by_cutoff += 1;
5160 2 : continue 'outer;
5161 22 : }
5162 22 :
5163 22 : // 2. It is newer than PiTR cutoff point?
5164 22 : if l.get_lsn_range().end > time_cutoff {
5165 0 : info!(
5166 0 : "keeping {} because it's newer than time_cutoff {}",
5167 0 : l.layer_name(),
5168 : time_cutoff,
5169 : );
5170 0 : result.layers_needed_by_pitr += 1;
5171 0 : continue 'outer;
5172 22 : }
5173 :
5174 : // 3. Is it needed by a child branch?
5175 : // NOTE With that we would keep data that
5176 : // might be referenced by child branches forever.
5177 : // We can track this in child timeline GC and delete parent layers when
5178 : // they are no longer needed. This might be complicated with long inheritance chains.
5179 : //
5180 : // TODO Vec is not a great choice for `retain_lsns`
5181 22 : for retain_lsn in &retain_lsns {
5182 : // start_lsn is inclusive
5183 0 : if &l.get_lsn_range().start <= retain_lsn {
5184 0 : info!(
5185 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5186 0 : l.layer_name(),
5187 0 : retain_lsn,
5188 0 : l.is_incremental(),
5189 : );
5190 0 : result.layers_needed_by_branches += 1;
5191 0 : continue 'outer;
5192 0 : }
5193 : }
5194 :
5195 : // 4. Is there a valid lease that requires us to keep this layer?
5196 22 : if let Some(lsn) = &max_lsn_with_valid_lease {
5197 : // keep if layer start <= any of the lease
5198 18 : if &l.get_lsn_range().start <= lsn {
5199 14 : info!(
5200 0 : "keeping {} because there is a valid lease preventing GC at {}",
5201 0 : l.layer_name(),
5202 : lsn,
5203 : );
5204 14 : result.layers_needed_by_leases += 1;
5205 14 : continue 'outer;
5206 4 : }
5207 4 : }
5208 :
5209 : // 5. Is there a later on-disk layer for this relation?
5210 : //
5211 : // The end-LSN is exclusive, while disk_consistent_lsn is
5212 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5213 : // OK for a delta layer to have end LSN 101, but if the end LSN
5214 : // is 102, then it might not have been fully flushed to disk
5215 : // before crash.
5216 : //
5217 : // For example, imagine that the following layers exist:
5218 : //
5219 : // 1000 - image (A)
5220 : // 1000-2000 - delta (B)
5221 : // 2000 - image (C)
5222 : // 2000-3000 - delta (D)
5223 : // 3000 - image (E)
5224 : //
5225 : // If GC horizon is at 2500, we can remove layers A and B, but
5226 : // we cannot remove C, even though it's older than 2500, because
5227 : // the delta layer 2000-3000 depends on it.
5228 8 : if !layers
5229 8 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5230 : {
5231 6 : info!("keeping {} because it is the latest layer", l.layer_name());
5232 6 : result.layers_not_updated += 1;
5233 6 : continue 'outer;
5234 2 : }
5235 2 :
5236 2 : // We didn't find any reason to keep this file, so remove it.
5237 2 : info!(
5238 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5239 0 : l.layer_name(),
5240 0 : l.is_incremental(),
5241 : );
5242 2 : layers_to_remove.push(l);
5243 : }
5244 :
5245 4 : if !layers_to_remove.is_empty() {
5246 : // Persist the new GC cutoff value before we actually remove anything.
5247 : // This unconditionally schedules also an index_part.json update, even though, we will
5248 : // be doing one a bit later with the unlinked gc'd layers.
5249 2 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5250 2 : self.schedule_uploads(disk_consistent_lsn, None)
5251 2 : .map_err(|e| {
5252 0 : if self.cancel.is_cancelled() {
5253 0 : GcError::TimelineCancelled
5254 : } else {
5255 0 : GcError::Remote(e)
5256 : }
5257 2 : })?;
5258 :
5259 2 : let gc_layers = layers_to_remove
5260 2 : .iter()
5261 2 : .map(|x| guard.get_from_desc(x))
5262 2 : .collect::<Vec<Layer>>();
5263 2 :
5264 2 : result.layers_removed = gc_layers.len() as u64;
5265 2 :
5266 2 : self.remote_client.schedule_gc_update(&gc_layers)?;
5267 :
5268 2 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5269 2 :
5270 2 : #[cfg(feature = "testing")]
5271 2 : {
5272 2 : result.doomed_layers = gc_layers;
5273 2 : }
5274 2 : }
5275 :
5276 4 : info!(
5277 0 : "GC completed removing {} layers, cutoff {}",
5278 : result.layers_removed, new_gc_cutoff
5279 : );
5280 :
5281 4 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5282 4 : Ok(result)
5283 4 : }
5284 :
5285 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5286 668228 : async fn reconstruct_value(
5287 668228 : &self,
5288 668228 : key: Key,
5289 668228 : request_lsn: Lsn,
5290 668228 : mut data: ValueReconstructState,
5291 668228 : ) -> Result<Bytes, PageReconstructError> {
5292 668228 : // Perform WAL redo if needed
5293 668228 : data.records.reverse();
5294 668228 :
5295 668228 : // If we have a page image, and no WAL, we're all set
5296 668228 : if data.records.is_empty() {
5297 667708 : if let Some((img_lsn, img)) = &data.img {
5298 667708 : trace!(
5299 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5300 : key,
5301 : img_lsn,
5302 : request_lsn,
5303 : );
5304 667708 : Ok(img.clone())
5305 : } else {
5306 0 : Err(PageReconstructError::from(anyhow!(
5307 0 : "base image for {key} at {request_lsn} not found"
5308 0 : )))
5309 : }
5310 : } else {
5311 : // We need to do WAL redo.
5312 : //
5313 : // If we don't have a base image, then the oldest WAL record better initialize
5314 : // the page
5315 520 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5316 0 : Err(PageReconstructError::from(anyhow!(
5317 0 : "Base image for {} at {} not found, but got {} WAL records",
5318 0 : key,
5319 0 : request_lsn,
5320 0 : data.records.len()
5321 0 : )))
5322 : } else {
5323 520 : if data.img.is_some() {
5324 454 : trace!(
5325 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5326 0 : data.records.len(),
5327 : key,
5328 : request_lsn
5329 : );
5330 : } else {
5331 66 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5332 : };
5333 520 : let res = self
5334 520 : .walredo_mgr
5335 520 : .as_ref()
5336 520 : .context("timeline has no walredo manager")
5337 520 : .map_err(PageReconstructError::WalRedo)?
5338 520 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5339 520 : .await;
5340 520 : let img = match res {
5341 520 : Ok(img) => img,
5342 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5343 0 : Err(walredo::Error::Other(e)) => {
5344 0 : return Err(PageReconstructError::WalRedo(
5345 0 : e.context("reconstruct a page image"),
5346 0 : ))
5347 : }
5348 : };
5349 520 : Ok(img)
5350 : }
5351 : }
5352 668228 : }
5353 :
5354 0 : pub(crate) async fn spawn_download_all_remote_layers(
5355 0 : self: Arc<Self>,
5356 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5357 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5358 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5359 :
5360 : // this is not really needed anymore; it has tests which really check the return value from
5361 : // http api. it would be better not to maintain this anymore.
5362 :
5363 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5364 0 : if let Some(st) = &*status_guard {
5365 0 : match &st.state {
5366 : DownloadRemoteLayersTaskState::Running => {
5367 0 : return Err(st.clone());
5368 : }
5369 : DownloadRemoteLayersTaskState::ShutDown
5370 0 : | DownloadRemoteLayersTaskState::Completed => {
5371 0 : *status_guard = None;
5372 0 : }
5373 : }
5374 0 : }
5375 :
5376 0 : let self_clone = Arc::clone(&self);
5377 0 : let task_id = task_mgr::spawn(
5378 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5379 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5380 0 : self.tenant_shard_id,
5381 0 : Some(self.timeline_id),
5382 0 : "download all remote layers task",
5383 0 : async move {
5384 0 : self_clone.download_all_remote_layers(request).await;
5385 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5386 0 : match &mut *status_guard {
5387 : None => {
5388 0 : warn!("tasks status is supposed to be Some(), since we are running");
5389 : }
5390 0 : Some(st) => {
5391 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5392 0 : if st.task_id != exp_task_id {
5393 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5394 0 : } else {
5395 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5396 0 : }
5397 : }
5398 : };
5399 0 : Ok(())
5400 0 : }
5401 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))
5402 : );
5403 :
5404 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5405 0 : task_id: format!("{task_id}"),
5406 0 : state: DownloadRemoteLayersTaskState::Running,
5407 0 : total_layer_count: 0,
5408 0 : successful_download_count: 0,
5409 0 : failed_download_count: 0,
5410 0 : };
5411 0 : *status_guard = Some(initial_info.clone());
5412 0 :
5413 0 : Ok(initial_info)
5414 0 : }
5415 :
5416 0 : async fn download_all_remote_layers(
5417 0 : self: &Arc<Self>,
5418 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5419 0 : ) {
5420 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5421 :
5422 0 : let remaining = {
5423 0 : let guard = self.layers.read().await;
5424 0 : let Ok(lm) = guard.layer_map() else {
5425 : // technically here we could look into iterating accessible layers, but downloading
5426 : // all layers of a shutdown timeline makes no sense regardless.
5427 0 : tracing::info!("attempted to download all layers of shutdown timeline");
5428 0 : return;
5429 : };
5430 0 : lm.iter_historic_layers()
5431 0 : .map(|desc| guard.get_from_desc(&desc))
5432 0 : .collect::<Vec<_>>()
5433 0 : };
5434 0 : let total_layer_count = remaining.len();
5435 :
5436 : macro_rules! lock_status {
5437 : ($st:ident) => {
5438 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5439 : let st = st
5440 : .as_mut()
5441 : .expect("this function is only called after the task has been spawned");
5442 : assert_eq!(
5443 : st.task_id,
5444 : format!(
5445 : "{}",
5446 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5447 : )
5448 : );
5449 : let $st = st;
5450 : };
5451 : }
5452 :
5453 : {
5454 0 : lock_status!(st);
5455 0 : st.total_layer_count = total_layer_count as u64;
5456 0 : }
5457 0 :
5458 0 : let mut remaining = remaining.into_iter();
5459 0 : let mut have_remaining = true;
5460 0 : let mut js = tokio::task::JoinSet::new();
5461 0 :
5462 0 : let cancel = task_mgr::shutdown_token();
5463 0 :
5464 0 : let limit = request.max_concurrent_downloads;
5465 :
5466 : loop {
5467 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5468 0 : let Some(next) = remaining.next() else {
5469 0 : have_remaining = false;
5470 0 : break;
5471 : };
5472 :
5473 0 : let span = tracing::info_span!("download", layer = %next);
5474 :
5475 0 : js.spawn(
5476 0 : async move {
5477 0 : let res = next.download().await;
5478 0 : (next, res)
5479 0 : }
5480 0 : .instrument(span),
5481 0 : );
5482 0 : }
5483 :
5484 0 : while let Some(res) = js.join_next().await {
5485 0 : match res {
5486 : Ok((_, Ok(_))) => {
5487 0 : lock_status!(st);
5488 0 : st.successful_download_count += 1;
5489 : }
5490 0 : Ok((layer, Err(e))) => {
5491 0 : tracing::error!(%layer, "download failed: {e:#}");
5492 0 : lock_status!(st);
5493 0 : st.failed_download_count += 1;
5494 : }
5495 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5496 0 : Err(je) if je.is_panic() => {
5497 0 : lock_status!(st);
5498 0 : st.failed_download_count += 1;
5499 : }
5500 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5501 : }
5502 : }
5503 :
5504 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5505 0 : break;
5506 0 : }
5507 : }
5508 :
5509 : {
5510 0 : lock_status!(st);
5511 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5512 : }
5513 0 : }
5514 :
5515 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5516 0 : &self,
5517 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5518 0 : self.download_all_remote_layers_task_info
5519 0 : .read()
5520 0 : .unwrap()
5521 0 : .clone()
5522 0 : }
5523 : }
5524 :
5525 : impl Timeline {
5526 : /// Returns non-remote layers for eviction.
5527 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5528 0 : let guard = self.layers.read().await;
5529 0 : let mut max_layer_size: Option<u64> = None;
5530 0 :
5531 0 : let resident_layers = guard
5532 0 : .likely_resident_layers()
5533 0 : .map(|layer| {
5534 0 : let file_size = layer.layer_desc().file_size;
5535 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5536 0 :
5537 0 : let last_activity_ts = layer.latest_activity();
5538 0 :
5539 0 : EvictionCandidate {
5540 0 : layer: layer.to_owned().into(),
5541 0 : last_activity_ts,
5542 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5543 0 : visibility: layer.visibility(),
5544 0 : }
5545 0 : })
5546 0 : .collect();
5547 0 :
5548 0 : DiskUsageEvictionInfo {
5549 0 : max_layer_size,
5550 0 : resident_layers,
5551 0 : }
5552 0 : }
5553 :
5554 1768 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5555 1768 : ShardIndex {
5556 1768 : shard_number: self.tenant_shard_id.shard_number,
5557 1768 : shard_count: self.tenant_shard_id.shard_count,
5558 1768 : }
5559 1768 : }
5560 :
5561 : /// Persistently blocks gc for `Manual` reason.
5562 : ///
5563 : /// Returns true if no such block existed before, false otherwise.
5564 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
5565 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5566 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5567 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
5568 0 : }
5569 :
5570 : /// Persistently unblocks gc for `Manual` reason.
5571 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
5572 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5573 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5574 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
5575 0 : }
5576 :
5577 : #[cfg(test)]
5578 44 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5579 44 : self.last_record_lsn.advance(new_lsn);
5580 44 : }
5581 :
5582 : #[cfg(test)]
5583 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
5584 2 : self.disk_consistent_lsn.store(new_value);
5585 2 : }
5586 :
5587 : /// Force create an image layer and place it into the layer map.
5588 : ///
5589 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5590 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5591 : /// placed into the layer map in one run AND be validated.
5592 : #[cfg(test)]
5593 56 : pub(super) async fn force_create_image_layer(
5594 56 : self: &Arc<Timeline>,
5595 56 : lsn: Lsn,
5596 56 : mut images: Vec<(Key, Bytes)>,
5597 56 : check_start_lsn: Option<Lsn>,
5598 56 : ctx: &RequestContext,
5599 56 : ) -> anyhow::Result<()> {
5600 56 : let last_record_lsn = self.get_last_record_lsn();
5601 56 : assert!(
5602 56 : lsn <= last_record_lsn,
5603 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5604 : );
5605 56 : if let Some(check_start_lsn) = check_start_lsn {
5606 56 : assert!(lsn >= check_start_lsn);
5607 0 : }
5608 162 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5609 56 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5610 56 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5611 56 : let mut image_layer_writer = ImageLayerWriter::new(
5612 56 : self.conf,
5613 56 : self.timeline_id,
5614 56 : self.tenant_shard_id,
5615 56 : &(min_key..end_key),
5616 56 : lsn,
5617 56 : ctx,
5618 56 : )
5619 56 : .await?;
5620 274 : for (key, img) in images {
5621 218 : image_layer_writer.put_image(key, img, ctx).await?;
5622 : }
5623 56 : let (desc, path) = image_layer_writer.finish(ctx).await?;
5624 56 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5625 56 : info!("force created image layer {}", image_layer.local_path());
5626 : {
5627 56 : let mut guard = self.layers.write().await;
5628 56 : guard.open_mut().unwrap().force_insert_layer(image_layer);
5629 56 : }
5630 56 :
5631 56 : Ok(())
5632 56 : }
5633 :
5634 : /// Force create a delta layer and place it into the layer map.
5635 : ///
5636 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5637 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5638 : /// placed into the layer map in one run AND be validated.
5639 : #[cfg(test)]
5640 90 : pub(super) async fn force_create_delta_layer(
5641 90 : self: &Arc<Timeline>,
5642 90 : mut deltas: DeltaLayerTestDesc,
5643 90 : check_start_lsn: Option<Lsn>,
5644 90 : ctx: &RequestContext,
5645 90 : ) -> anyhow::Result<()> {
5646 90 : let last_record_lsn = self.get_last_record_lsn();
5647 90 : deltas
5648 90 : .data
5649 132 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5650 90 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5651 90 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5652 312 : for (_, lsn, _) in &deltas.data {
5653 222 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5654 : }
5655 90 : assert!(
5656 90 : deltas.lsn_range.end <= last_record_lsn,
5657 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
5658 : deltas.lsn_range.end,
5659 : last_record_lsn
5660 : );
5661 90 : if let Some(check_start_lsn) = check_start_lsn {
5662 90 : assert!(deltas.lsn_range.start >= check_start_lsn);
5663 0 : }
5664 90 : let mut delta_layer_writer = DeltaLayerWriter::new(
5665 90 : self.conf,
5666 90 : self.timeline_id,
5667 90 : self.tenant_shard_id,
5668 90 : deltas.key_range.start,
5669 90 : deltas.lsn_range,
5670 90 : ctx,
5671 90 : )
5672 90 : .await?;
5673 312 : for (key, lsn, val) in deltas.data {
5674 222 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5675 : }
5676 90 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
5677 90 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5678 90 : info!("force created delta layer {}", delta_layer.local_path());
5679 : {
5680 90 : let mut guard = self.layers.write().await;
5681 90 : guard.open_mut().unwrap().force_insert_layer(delta_layer);
5682 90 : }
5683 90 :
5684 90 : Ok(())
5685 90 : }
5686 :
5687 : /// Return all keys at the LSN in the image layers
5688 : #[cfg(test)]
5689 6 : pub(crate) async fn inspect_image_layers(
5690 6 : self: &Arc<Timeline>,
5691 6 : lsn: Lsn,
5692 6 : ctx: &RequestContext,
5693 6 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
5694 6 : let mut all_data = Vec::new();
5695 6 : let guard = self.layers.read().await;
5696 34 : for layer in guard.layer_map()?.iter_historic_layers() {
5697 34 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
5698 8 : let layer = guard.get_from_desc(&layer);
5699 8 : let mut reconstruct_data = ValuesReconstructState::default();
5700 8 : layer
5701 8 : .get_values_reconstruct_data(
5702 8 : KeySpace::single(Key::MIN..Key::MAX),
5703 8 : lsn..Lsn(lsn.0 + 1),
5704 8 : &mut reconstruct_data,
5705 8 : ctx,
5706 8 : )
5707 8 : .await?;
5708 74 : for (k, v) in reconstruct_data.keys {
5709 66 : all_data.push((k, v?.img.unwrap().1));
5710 : }
5711 26 : }
5712 : }
5713 6 : all_data.sort();
5714 6 : Ok(all_data)
5715 6 : }
5716 :
5717 : /// Get all historic layer descriptors in the layer map
5718 : #[cfg(test)]
5719 24 : pub(crate) async fn inspect_historic_layers(
5720 24 : self: &Arc<Timeline>,
5721 24 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
5722 24 : let mut layers = Vec::new();
5723 24 : let guard = self.layers.read().await;
5724 114 : for layer in guard.layer_map()?.iter_historic_layers() {
5725 114 : layers.push(layer.key());
5726 114 : }
5727 24 : Ok(layers)
5728 24 : }
5729 :
5730 : #[cfg(test)]
5731 10 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
5732 10 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
5733 10 : keyspace.merge(&ks);
5734 10 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
5735 10 : }
5736 : }
5737 :
5738 : /// Tracking writes ingestion does to a particular in-memory layer.
5739 : ///
5740 : /// Cleared upon freezing a layer.
5741 : pub(crate) struct TimelineWriterState {
5742 : open_layer: Arc<InMemoryLayer>,
5743 : current_size: u64,
5744 : // Previous Lsn which passed through
5745 : prev_lsn: Option<Lsn>,
5746 : // Largest Lsn which passed through the current writer
5747 : max_lsn: Option<Lsn>,
5748 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
5749 : cached_last_freeze_at: Lsn,
5750 : }
5751 :
5752 : impl TimelineWriterState {
5753 1272 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
5754 1272 : Self {
5755 1272 : open_layer,
5756 1272 : current_size,
5757 1272 : prev_lsn: None,
5758 1272 : max_lsn: None,
5759 1272 : cached_last_freeze_at: last_freeze_at,
5760 1272 : }
5761 1272 : }
5762 : }
5763 :
5764 : /// Various functions to mutate the timeline.
5765 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
5766 : // This is probably considered a bad practice in Rust and should be fixed eventually,
5767 : // but will cause large code changes.
5768 : pub(crate) struct TimelineWriter<'a> {
5769 : tl: &'a Timeline,
5770 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
5771 : }
5772 :
5773 : impl Deref for TimelineWriter<'_> {
5774 : type Target = Timeline;
5775 :
5776 9897696 : fn deref(&self) -> &Self::Target {
5777 9897696 : self.tl
5778 9897696 : }
5779 : }
5780 :
5781 : #[derive(PartialEq)]
5782 : enum OpenLayerAction {
5783 : Roll,
5784 : Open,
5785 : None,
5786 : }
5787 :
5788 : impl TimelineWriter<'_> {
5789 4804208 : async fn handle_open_layer_action(
5790 4804208 : &mut self,
5791 4804208 : at: Lsn,
5792 4804208 : action: OpenLayerAction,
5793 4804208 : ctx: &RequestContext,
5794 4804208 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
5795 4804208 : match action {
5796 : OpenLayerAction::Roll => {
5797 80 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
5798 80 : self.roll_layer(freeze_at).await?;
5799 80 : self.open_layer(at, ctx).await?;
5800 : }
5801 1192 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
5802 : OpenLayerAction::None => {
5803 4802936 : assert!(self.write_guard.is_some());
5804 : }
5805 : }
5806 :
5807 4804208 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
5808 4804208 : }
5809 :
5810 1272 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
5811 1272 : let layer = self
5812 1272 : .tl
5813 1272 : .get_layer_for_write(at, &self.write_guard, ctx)
5814 1272 : .await?;
5815 1272 : let initial_size = layer.size().await?;
5816 :
5817 1272 : let last_freeze_at = self.last_freeze_at.load();
5818 1272 : self.write_guard.replace(TimelineWriterState::new(
5819 1272 : layer,
5820 1272 : initial_size,
5821 1272 : last_freeze_at,
5822 1272 : ));
5823 1272 :
5824 1272 : Ok(())
5825 1272 : }
5826 :
5827 80 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
5828 80 : let current_size = self.write_guard.as_ref().unwrap().current_size;
5829 80 :
5830 80 : // self.write_guard will be taken by the freezing
5831 80 : self.tl
5832 80 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
5833 80 : .await?;
5834 :
5835 80 : assert!(self.write_guard.is_none());
5836 :
5837 80 : if current_size >= self.get_checkpoint_distance() * 2 {
5838 0 : warn!("Flushed oversized open layer with size {}", current_size)
5839 80 : }
5840 :
5841 80 : Ok(())
5842 80 : }
5843 :
5844 4804208 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
5845 4804208 : let state = &*self.write_guard;
5846 4804208 : let Some(state) = &state else {
5847 1192 : return OpenLayerAction::Open;
5848 : };
5849 :
5850 : #[cfg(feature = "testing")]
5851 4803016 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
5852 : // this check and assertion are not really needed because
5853 : // LayerManager::try_freeze_in_memory_layer will always clear out the
5854 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
5855 : // is no TimelineWriterState.
5856 0 : assert!(
5857 0 : state.open_layer.end_lsn.get().is_some(),
5858 0 : "our open_layer must be outdated"
5859 : );
5860 :
5861 : // this would be a memory leak waiting to happen because the in-memory layer always has
5862 : // an index
5863 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
5864 4803016 : }
5865 4803016 :
5866 4803016 : if state.prev_lsn == Some(lsn) {
5867 : // Rolling mid LSN is not supported by [downstream code].
5868 : // Hence, only roll at LSN boundaries.
5869 : //
5870 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
5871 6 : return OpenLayerAction::None;
5872 4803010 : }
5873 4803010 :
5874 4803010 : if state.current_size == 0 {
5875 : // Don't roll empty layers
5876 0 : return OpenLayerAction::None;
5877 4803010 : }
5878 4803010 :
5879 4803010 : if self.tl.should_roll(
5880 4803010 : state.current_size,
5881 4803010 : state.current_size + new_value_size,
5882 4803010 : self.get_checkpoint_distance(),
5883 4803010 : lsn,
5884 4803010 : state.cached_last_freeze_at,
5885 4803010 : state.open_layer.get_opened_at(),
5886 4803010 : ) {
5887 80 : OpenLayerAction::Roll
5888 : } else {
5889 4802930 : OpenLayerAction::None
5890 : }
5891 4804208 : }
5892 :
5893 : /// Put a batch of keys at the specified Lsns.
5894 4804206 : pub(crate) async fn put_batch(
5895 4804206 : &mut self,
5896 4804206 : batch: SerializedValueBatch,
5897 4804206 : ctx: &RequestContext,
5898 4804206 : ) -> anyhow::Result<()> {
5899 4804206 : if !batch.has_data() {
5900 0 : return Ok(());
5901 4804206 : }
5902 4804206 :
5903 4804206 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
5904 4804206 : // We don't assert this in release builds, since key ownership policies may change over
5905 4804206 : // time. Stray keys will be removed during compaction.
5906 4804206 : if cfg!(debug_assertions) {
5907 9894704 : for metadata in &batch.metadata {
5908 5090498 : if let ValueMeta::Serialized(metadata) = metadata {
5909 5090498 : let key = Key::from_compact(metadata.key);
5910 5090498 : assert!(
5911 5090498 : self.shard_identity.is_key_local(&key)
5912 0 : || self.shard_identity.is_key_global(&key),
5913 0 : "key {key} does not belong on shard {}",
5914 0 : self.shard_identity.shard_index()
5915 : );
5916 0 : }
5917 : }
5918 0 : }
5919 :
5920 4804206 : let batch_max_lsn = batch.max_lsn;
5921 4804206 : let buf_size: u64 = batch.buffer_size() as u64;
5922 4804206 :
5923 4804206 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
5924 4804206 : let layer = self
5925 4804206 : .handle_open_layer_action(batch_max_lsn, action, ctx)
5926 4804206 : .await?;
5927 :
5928 4804206 : let res = layer.put_batch(batch, ctx).await;
5929 :
5930 4804206 : if res.is_ok() {
5931 4804206 : // Update the current size only when the entire write was ok.
5932 4804206 : // In case of failures, we may have had partial writes which
5933 4804206 : // render the size tracking out of sync. That's ok because
5934 4804206 : // the checkpoint distance should be significantly smaller
5935 4804206 : // than the S3 single shot upload limit of 5GiB.
5936 4804206 : let state = self.write_guard.as_mut().unwrap();
5937 4804206 :
5938 4804206 : state.current_size += buf_size;
5939 4804206 : state.prev_lsn = Some(batch_max_lsn);
5940 4804206 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
5941 4804206 : }
5942 :
5943 4804206 : res
5944 4804206 : }
5945 :
5946 : #[cfg(test)]
5947 : /// Test helper, for tests that would like to poke individual values without composing a batch
5948 4390154 : pub(crate) async fn put(
5949 4390154 : &mut self,
5950 4390154 : key: Key,
5951 4390154 : lsn: Lsn,
5952 4390154 : value: &Value,
5953 4390154 : ctx: &RequestContext,
5954 4390154 : ) -> anyhow::Result<()> {
5955 : use utils::bin_ser::BeSer;
5956 4390154 : if !key.is_valid_key_on_write_path() {
5957 0 : bail!(
5958 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
5959 0 : key
5960 0 : );
5961 4390154 : }
5962 4390154 : let val_ser_size = value.serialized_size().unwrap() as usize;
5963 4390154 : let batch = SerializedValueBatch::from_values(vec![(
5964 4390154 : key.to_compact(),
5965 4390154 : lsn,
5966 4390154 : val_ser_size,
5967 4390154 : value.clone(),
5968 4390154 : )]);
5969 4390154 :
5970 4390154 : self.put_batch(batch, ctx).await
5971 4390154 : }
5972 :
5973 2 : pub(crate) async fn delete_batch(
5974 2 : &mut self,
5975 2 : batch: &[(Range<Key>, Lsn)],
5976 2 : ctx: &RequestContext,
5977 2 : ) -> anyhow::Result<()> {
5978 2 : if let Some((_, lsn)) = batch.first() {
5979 2 : let action = self.get_open_layer_action(*lsn, 0);
5980 2 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
5981 2 : layer.put_tombstones(batch).await?;
5982 0 : }
5983 :
5984 2 : Ok(())
5985 2 : }
5986 :
5987 : /// Track the end of the latest digested WAL record.
5988 : /// Remember the (end of) last valid WAL record remembered in the timeline.
5989 : ///
5990 : /// Call this after you have finished writing all the WAL up to 'lsn'.
5991 : ///
5992 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
5993 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
5994 : /// the latest and can be read.
5995 5279068 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
5996 5279068 : self.tl.finish_write(new_lsn);
5997 5279068 : }
5998 :
5999 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
6000 270570 : self.tl.update_current_logical_size(delta)
6001 270570 : }
6002 : }
6003 :
6004 : // We need TimelineWriter to be send in upcoming conversion of
6005 : // Timeline::layers to tokio::sync::RwLock.
6006 : #[test]
6007 2 : fn is_send() {
6008 2 : fn _assert_send<T: Send>() {}
6009 2 : _assert_send::<TimelineWriter<'_>>();
6010 2 : }
6011 :
6012 : #[cfg(test)]
6013 : mod tests {
6014 : use pageserver_api::key::Key;
6015 : use pageserver_api::value::Value;
6016 : use utils::{id::TimelineId, lsn::Lsn};
6017 :
6018 : use crate::tenant::{
6019 : harness::{test_img, TenantHarness},
6020 : layer_map::LayerMap,
6021 : storage_layer::{Layer, LayerName},
6022 : timeline::{DeltaLayerTestDesc, EvictionError},
6023 : Timeline,
6024 : };
6025 :
6026 : #[tokio::test]
6027 2 : async fn test_heatmap_generation() {
6028 2 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
6029 2 :
6030 2 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6031 2 : Lsn(0x10)..Lsn(0x20),
6032 2 : vec![(
6033 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6034 2 : Lsn(0x11),
6035 2 : Value::Image(test_img("foo")),
6036 2 : )],
6037 2 : );
6038 2 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6039 2 : Lsn(0x10)..Lsn(0x20),
6040 2 : vec![(
6041 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6042 2 : Lsn(0x11),
6043 2 : Value::Image(test_img("foo")),
6044 2 : )],
6045 2 : );
6046 2 : let l0_delta = DeltaLayerTestDesc::new(
6047 2 : Lsn(0x20)..Lsn(0x30),
6048 2 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6049 2 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6050 2 : vec![(
6051 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6052 2 : Lsn(0x25),
6053 2 : Value::Image(test_img("foo")),
6054 2 : )],
6055 2 : );
6056 2 : let delta_layers = vec![
6057 2 : covered_delta.clone(),
6058 2 : visible_delta.clone(),
6059 2 : l0_delta.clone(),
6060 2 : ];
6061 2 :
6062 2 : let image_layer = (
6063 2 : Lsn(0x40),
6064 2 : vec![(
6065 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6066 2 : test_img("bar"),
6067 2 : )],
6068 2 : );
6069 2 : let image_layers = vec![image_layer];
6070 2 :
6071 2 : let (tenant, ctx) = harness.load().await;
6072 2 : let timeline = tenant
6073 2 : .create_test_timeline_with_layers(
6074 2 : TimelineId::generate(),
6075 2 : Lsn(0x10),
6076 2 : 14,
6077 2 : &ctx,
6078 2 : delta_layers,
6079 2 : image_layers,
6080 2 : Lsn(0x100),
6081 2 : )
6082 2 : .await
6083 2 : .unwrap();
6084 2 :
6085 2 : // Layer visibility is an input to heatmap generation, so refresh it first
6086 2 : timeline.update_layer_visibility().await.unwrap();
6087 2 :
6088 2 : let heatmap = timeline
6089 2 : .generate_heatmap()
6090 2 : .await
6091 2 : .expect("Infallible while timeline is not shut down");
6092 2 :
6093 2 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
6094 2 :
6095 2 : // L0 should come last
6096 2 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
6097 2 :
6098 2 : let mut last_lsn = Lsn::MAX;
6099 10 : for layer in heatmap.layers {
6100 2 : // Covered layer should be omitted
6101 8 : assert!(layer.name != covered_delta.layer_name());
6102 2 :
6103 8 : let layer_lsn = match &layer.name {
6104 4 : LayerName::Delta(d) => d.lsn_range.end,
6105 4 : LayerName::Image(i) => i.lsn,
6106 2 : };
6107 2 :
6108 2 : // Apart from L0s, newest Layers should come first
6109 8 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
6110 6 : assert!(layer_lsn <= last_lsn);
6111 6 : last_lsn = layer_lsn;
6112 2 : }
6113 2 : }
6114 2 : }
6115 :
6116 : #[tokio::test]
6117 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
6118 2 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6119 2 : .await
6120 2 : .unwrap();
6121 2 :
6122 2 : let (tenant, ctx) = harness.load().await;
6123 2 : let timeline = tenant
6124 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6125 2 : .await
6126 2 : .unwrap();
6127 2 :
6128 2 : let layer = find_some_layer(&timeline).await;
6129 2 : let layer = layer
6130 2 : .keep_resident()
6131 2 : .await
6132 2 : .expect("no download => no downloading errors")
6133 2 : .drop_eviction_guard();
6134 2 :
6135 2 : let forever = std::time::Duration::from_secs(120);
6136 2 :
6137 2 : let first = layer.evict_and_wait(forever);
6138 2 : let second = layer.evict_and_wait(forever);
6139 2 :
6140 2 : let (first, second) = tokio::join!(first, second);
6141 2 :
6142 2 : let res = layer.keep_resident().await;
6143 2 : assert!(res.is_none(), "{res:?}");
6144 2 :
6145 2 : match (first, second) {
6146 2 : (Ok(()), Ok(())) => {
6147 2 : // because there are no more timeline locks being taken on eviction path, we can
6148 2 : // witness all three outcomes here.
6149 2 : }
6150 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
6151 0 : // if one completes before the other, this is fine just as well.
6152 0 : }
6153 2 : other => unreachable!("unexpected {:?}", other),
6154 2 : }
6155 2 : }
6156 :
6157 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
6158 2 : let layers = timeline.layers.read().await;
6159 2 : let desc = layers
6160 2 : .layer_map()
6161 2 : .unwrap()
6162 2 : .iter_historic_layers()
6163 2 : .next()
6164 2 : .expect("must find one layer to evict");
6165 2 :
6166 2 : layers.get_from_desc(&desc)
6167 2 : }
6168 : }
|
