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 : CompactionAlgorithm, CompactionAlgorithmSettings, DownloadRemoteLayersTaskInfo,
35 : DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy, InMemoryLayerInfo, LayerMapInfo,
36 : 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 : sync::gate::{Gate, GateGuard},
54 : };
55 : use wal_decoder::serialized_batch::SerializedValueBatch;
56 :
57 : use std::sync::atomic::Ordering as AtomicOrdering;
58 : use std::sync::{Arc, Mutex, RwLock, Weak};
59 : use std::time::{Duration, Instant, SystemTime};
60 : use std::{
61 : array,
62 : collections::{BTreeMap, HashMap, HashSet},
63 : sync::atomic::AtomicU64,
64 : };
65 : use std::{cmp::min, ops::ControlFlow};
66 : use std::{
67 : collections::btree_map::Entry,
68 : ops::{Deref, Range},
69 : };
70 : use std::{pin::pin, sync::OnceLock};
71 :
72 : use crate::{
73 : aux_file::AuxFileSizeEstimator,
74 : tenant::{
75 : config::AttachmentMode,
76 : layer_map::{LayerMap, SearchResult},
77 : metadata::TimelineMetadata,
78 : storage_layer::{inmemory_layer::IndexEntry, PersistentLayerDesc},
79 : },
80 : walingest::WalLagCooldown,
81 : walredo,
82 : };
83 : use crate::{
84 : context::{DownloadBehavior, RequestContext},
85 : disk_usage_eviction_task::DiskUsageEvictionInfo,
86 : pgdatadir_mapping::CollectKeySpaceError,
87 : };
88 : use crate::{
89 : disk_usage_eviction_task::finite_f32,
90 : tenant::storage_layer::{
91 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
92 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
93 : ValuesReconstructState,
94 : },
95 : };
96 : use crate::{
97 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
98 : };
99 : use crate::{
100 : l0_flush::{self, L0FlushGlobalState},
101 : metrics::GetKind,
102 : };
103 : use crate::{
104 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
105 : };
106 : use crate::{
107 : pgdatadir_mapping::DirectoryKind,
108 : virtual_file::{MaybeFatalIo, VirtualFile},
109 : };
110 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
111 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
112 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
113 :
114 : use crate::config::PageServerConf;
115 : use crate::keyspace::{KeyPartitioning, KeySpace};
116 : use crate::metrics::TimelineMetrics;
117 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
118 : use crate::tenant::config::TenantConfOpt;
119 : use pageserver_api::reltag::RelTag;
120 : use pageserver_api::shard::ShardIndex;
121 :
122 : use postgres_connection::PgConnectionConfig;
123 : use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
124 : use utils::{
125 : completion,
126 : generation::Generation,
127 : id::TimelineId,
128 : lsn::{AtomicLsn, Lsn, RecordLsn},
129 : seqwait::SeqWait,
130 : simple_rcu::{Rcu, RcuReadGuard},
131 : };
132 :
133 : use crate::task_mgr;
134 : use crate::task_mgr::TaskKind;
135 : use crate::tenant::gc_result::GcResult;
136 : use crate::ZERO_PAGE;
137 : use pageserver_api::key::Key;
138 :
139 : use self::delete::DeleteTimelineFlow;
140 : pub(super) use self::eviction_task::EvictionTaskTenantState;
141 : use self::eviction_task::EvictionTaskTimelineState;
142 : use self::layer_manager::LayerManager;
143 : use self::logical_size::LogicalSize;
144 : use self::walreceiver::{WalReceiver, WalReceiverConf};
145 :
146 : use super::{
147 : config::TenantConf, storage_layer::LayerVisibilityHint, upload_queue::NotInitialized,
148 : MaybeOffloaded,
149 : };
150 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
151 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
152 : use super::{
153 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
154 : storage_layer::ReadableLayer,
155 : };
156 : use super::{
157 : secondary::heatmap::{HeatMapLayer, HeatMapTimeline},
158 : GcError,
159 : };
160 :
161 : #[cfg(test)]
162 : use pageserver_api::value::Value;
163 :
164 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
165 : pub(crate) enum FlushLoopState {
166 : NotStarted,
167 : Running {
168 : #[cfg(test)]
169 : expect_initdb_optimization: bool,
170 : #[cfg(test)]
171 : initdb_optimization_count: usize,
172 : },
173 : Exited,
174 : }
175 :
176 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
177 : pub enum ImageLayerCreationMode {
178 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
179 : Try,
180 : /// Force creating the image layers if possible. For now, no image layers will be created
181 : /// for metadata keys. Used in compaction code path with force flag enabled.
182 : Force,
183 : /// Initial ingestion of the data, and no data should be dropped in this function. This
184 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
185 : /// code path.
186 : Initial,
187 : }
188 :
189 : impl std::fmt::Display for ImageLayerCreationMode {
190 716 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
191 716 : write!(f, "{:?}", self)
192 716 : }
193 : }
194 :
195 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
196 : /// Can be removed after all refactors are done.
197 28 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
198 28 : drop(rlock)
199 28 : }
200 :
201 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
202 : /// Can be removed after all refactors are done.
203 744 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
204 744 : drop(rlock)
205 744 : }
206 :
207 : /// The outward-facing resources required to build a Timeline
208 : pub struct TimelineResources {
209 : pub remote_client: RemoteTimelineClient,
210 : pub timeline_get_throttle:
211 : Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
212 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
213 : }
214 :
215 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
216 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
217 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
218 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
219 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
220 : pub(crate) struct RelSizeCache {
221 : pub(crate) complete_as_of: Lsn,
222 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
223 : }
224 :
225 : pub struct Timeline {
226 : pub(crate) conf: &'static PageServerConf,
227 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
228 :
229 : myself: Weak<Self>,
230 :
231 : pub(crate) tenant_shard_id: TenantShardId,
232 : pub timeline_id: TimelineId,
233 :
234 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
235 : /// Never changes for the lifetime of this [`Timeline`] object.
236 : ///
237 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
238 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
239 : pub(crate) generation: Generation,
240 :
241 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
242 : /// to shards, and is constant through the lifetime of this Timeline.
243 : shard_identity: ShardIdentity,
244 :
245 : pub pg_version: u32,
246 :
247 : /// The tuple has two elements.
248 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
249 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
250 : ///
251 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
252 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
253 : ///
254 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
255 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
256 : /// `PersistentLayerDesc`'s.
257 : ///
258 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
259 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
260 : /// runtime, e.g., during page reconstruction.
261 : ///
262 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
263 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
264 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
265 :
266 : last_freeze_at: AtomicLsn,
267 : // Atomic would be more appropriate here.
268 : last_freeze_ts: RwLock<Instant>,
269 :
270 : pub(crate) standby_horizon: AtomicLsn,
271 :
272 : // WAL redo manager. `None` only for broken tenants.
273 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
274 :
275 : /// Remote storage client.
276 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
277 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
278 :
279 : // What page versions do we hold in the repository? If we get a
280 : // request > last_record_lsn, we need to wait until we receive all
281 : // the WAL up to the request. The SeqWait provides functions for
282 : // that. TODO: If we get a request for an old LSN, such that the
283 : // versions have already been garbage collected away, we should
284 : // throw an error, but we don't track that currently.
285 : //
286 : // last_record_lsn.load().last points to the end of last processed WAL record.
287 : //
288 : // We also remember the starting point of the previous record in
289 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
290 : // first WAL record when the node is started up. But here, we just
291 : // keep track of it.
292 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
293 :
294 : // All WAL records have been processed and stored durably on files on
295 : // local disk, up to this LSN. On crash and restart, we need to re-process
296 : // the WAL starting from this point.
297 : //
298 : // Some later WAL records might have been processed and also flushed to disk
299 : // already, so don't be surprised to see some, but there's no guarantee on
300 : // them yet.
301 : disk_consistent_lsn: AtomicLsn,
302 :
303 : // Parent timeline that this timeline was branched from, and the LSN
304 : // of the branch point.
305 : ancestor_timeline: Option<Arc<Timeline>>,
306 : ancestor_lsn: Lsn,
307 :
308 : pub(super) metrics: TimelineMetrics,
309 :
310 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
311 : // in `crate::page_service` writes these metrics.
312 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
313 :
314 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
315 :
316 : /// Ensures layers aren't frozen by checkpointer between
317 : /// [`Timeline::get_layer_for_write`] and layer reads.
318 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
319 : /// Must always be acquired before the layer map/individual layer lock
320 : /// to avoid deadlock.
321 : ///
322 : /// The state is cleared upon freezing.
323 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
324 :
325 : /// Used to avoid multiple `flush_loop` tasks running
326 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
327 :
328 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
329 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
330 : /// The flush cycle counter is sent back on the layer_flush_done channel when
331 : /// the flush finishes. You can use that to wait for the flush to finish.
332 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
333 : /// read by whoever sends an update
334 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
335 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
336 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
337 :
338 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
339 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
340 :
341 : // List of child timelines and their branch points. This is needed to avoid
342 : // garbage collecting data that is still needed by the child timelines.
343 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
344 :
345 : // It may change across major versions so for simplicity
346 : // keep it after running initdb for a timeline.
347 : // It is needed in checks when we want to error on some operations
348 : // when they are requested for pre-initdb lsn.
349 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
350 : // though let's keep them both for better error visibility.
351 : pub initdb_lsn: Lsn,
352 :
353 : /// When did we last calculate the partitioning? Make it pub to test cases.
354 : pub(super) partitioning: tokio::sync::Mutex<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
355 :
356 : /// Configuration: how often should the partitioning be recalculated.
357 : repartition_threshold: u64,
358 :
359 : last_image_layer_creation_check_at: AtomicLsn,
360 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
361 :
362 : /// Current logical size of the "datadir", at the last LSN.
363 : current_logical_size: LogicalSize,
364 :
365 : /// Information about the last processed message by the WAL receiver,
366 : /// or None if WAL receiver has not received anything for this timeline
367 : /// yet.
368 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
369 : pub walreceiver: Mutex<Option<WalReceiver>>,
370 :
371 : /// Relation size cache
372 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
373 :
374 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
375 :
376 : state: watch::Sender<TimelineState>,
377 :
378 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
379 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
380 : pub delete_progress: TimelineDeleteProgress,
381 :
382 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
383 :
384 : /// Load or creation time information about the disk_consistent_lsn and when the loading
385 : /// happened. Used for consumption metrics.
386 : pub(crate) loaded_at: (Lsn, SystemTime),
387 :
388 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
389 : pub(crate) gate: Gate,
390 :
391 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
392 : /// to the timeline should drop out when this token fires.
393 : pub(crate) cancel: CancellationToken,
394 :
395 : /// Make sure we only have one running compaction at a time in tests.
396 : ///
397 : /// Must only be taken in two places:
398 : /// - [`Timeline::compact`] (this file)
399 : /// - [`delete::delete_local_timeline_directory`]
400 : ///
401 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
402 : compaction_lock: tokio::sync::Mutex<()>,
403 :
404 : /// Make sure we only have one running gc at a time.
405 : ///
406 : /// Must only be taken in two places:
407 : /// - [`Timeline::gc`] (this file)
408 : /// - [`delete::delete_local_timeline_directory`]
409 : ///
410 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
411 : gc_lock: tokio::sync::Mutex<()>,
412 :
413 : /// Cloned from [`super::Tenant::timeline_get_throttle`] on construction.
414 : timeline_get_throttle:
415 : Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
416 :
417 : /// Size estimator for aux file v2
418 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
419 :
420 : /// Some test cases directly place keys into the timeline without actually modifying the directory
421 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
422 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
423 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
424 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
425 : #[cfg(test)]
426 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
427 :
428 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
429 :
430 : pub(crate) handles: handle::PerTimelineState<crate::page_service::TenantManagerTypes>,
431 :
432 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
433 :
434 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
435 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
436 : }
437 :
438 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
439 :
440 : pub struct WalReceiverInfo {
441 : pub wal_source_connconf: PgConnectionConfig,
442 : pub last_received_msg_lsn: Lsn,
443 : pub last_received_msg_ts: u128,
444 : }
445 :
446 : /// Information about how much history needs to be retained, needed by
447 : /// Garbage Collection.
448 : #[derive(Default)]
449 : pub(crate) struct GcInfo {
450 : /// Specific LSNs that are needed.
451 : ///
452 : /// Currently, this includes all points where child branches have
453 : /// been forked off from. In the future, could also include
454 : /// explicit user-defined snapshot points.
455 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
456 :
457 : /// The cutoff coordinates, which are combined by selecting the minimum.
458 : pub(crate) cutoffs: GcCutoffs,
459 :
460 : /// Leases granted to particular LSNs.
461 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
462 :
463 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
464 : pub(crate) within_ancestor_pitr: bool,
465 : }
466 :
467 : impl GcInfo {
468 228 : pub(crate) fn min_cutoff(&self) -> Lsn {
469 228 : self.cutoffs.select_min()
470 228 : }
471 :
472 232 : pub(super) fn insert_child(
473 232 : &mut self,
474 232 : child_id: TimelineId,
475 232 : child_lsn: Lsn,
476 232 : is_offloaded: MaybeOffloaded,
477 232 : ) {
478 232 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
479 232 : self.retain_lsns.sort_by_key(|i| i.0);
480 232 : }
481 :
482 4 : pub(super) fn remove_child_maybe_offloaded(
483 4 : &mut self,
484 4 : child_id: TimelineId,
485 4 : maybe_offloaded: MaybeOffloaded,
486 4 : ) -> bool {
487 4 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
488 4 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
489 4 : let mut removed = false;
490 6 : self.retain_lsns.retain(|i| {
491 6 : if removed {
492 2 : return true;
493 4 : }
494 4 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
495 4 : removed |= remove;
496 4 : !remove
497 6 : });
498 4 : removed
499 4 : }
500 :
501 4 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
502 4 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
503 4 : }
504 :
505 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
506 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
507 0 : }
508 : }
509 :
510 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
511 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
512 : /// between time-based and space-based retention for observability and consumption metrics purposes.
513 : #[derive(Debug, Clone)]
514 : pub(crate) struct GcCutoffs {
515 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
516 : /// history we must keep to retain a specified number of bytes of WAL.
517 : pub(crate) space: Lsn,
518 :
519 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
520 : /// history we must keep to enable reading back at least the PITR interval duration.
521 : pub(crate) time: Lsn,
522 : }
523 :
524 : impl Default for GcCutoffs {
525 418 : fn default() -> Self {
526 418 : Self {
527 418 : space: Lsn::INVALID,
528 418 : time: Lsn::INVALID,
529 418 : }
530 418 : }
531 : }
532 :
533 : impl GcCutoffs {
534 268 : fn select_min(&self) -> Lsn {
535 268 : std::cmp::min(self.space, self.time)
536 268 : }
537 : }
538 :
539 : pub(crate) struct TimelineVisitOutcome {
540 : completed_keyspace: KeySpace,
541 : image_covered_keyspace: KeySpace,
542 : }
543 :
544 : /// An error happened in a get() operation.
545 2 : #[derive(thiserror::Error, Debug)]
546 : pub(crate) enum PageReconstructError {
547 : #[error(transparent)]
548 : Other(anyhow::Error),
549 :
550 : #[error("Ancestor LSN wait error: {0}")]
551 : AncestorLsnTimeout(WaitLsnError),
552 :
553 : #[error("timeline shutting down")]
554 : Cancelled,
555 :
556 : /// An error happened replaying WAL records
557 : #[error(transparent)]
558 : WalRedo(anyhow::Error),
559 :
560 : #[error("{0}")]
561 : MissingKey(MissingKeyError),
562 : }
563 :
564 : impl From<anyhow::Error> for PageReconstructError {
565 0 : fn from(value: anyhow::Error) -> Self {
566 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
567 0 : match value.downcast::<PageReconstructError>() {
568 0 : Ok(pre) => pre,
569 0 : Err(other) => PageReconstructError::Other(other),
570 : }
571 0 : }
572 : }
573 :
574 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
575 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
576 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
577 0 : }
578 : }
579 :
580 : impl From<layer_manager::Shutdown> for PageReconstructError {
581 0 : fn from(_: layer_manager::Shutdown) -> Self {
582 0 : PageReconstructError::Cancelled
583 0 : }
584 : }
585 :
586 : impl GetVectoredError {
587 : #[cfg(test)]
588 6 : pub(crate) fn is_missing_key_error(&self) -> bool {
589 6 : matches!(self, Self::MissingKey(_))
590 6 : }
591 : }
592 :
593 : impl From<layer_manager::Shutdown> for GetVectoredError {
594 0 : fn from(_: layer_manager::Shutdown) -> Self {
595 0 : GetVectoredError::Cancelled
596 0 : }
597 : }
598 :
599 : #[derive(thiserror::Error)]
600 : pub struct MissingKeyError {
601 : key: Key,
602 : shard: ShardNumber,
603 : cont_lsn: Lsn,
604 : request_lsn: Lsn,
605 : ancestor_lsn: Option<Lsn>,
606 : backtrace: Option<std::backtrace::Backtrace>,
607 : }
608 :
609 : impl std::fmt::Debug for MissingKeyError {
610 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
611 0 : write!(f, "{}", self)
612 0 : }
613 : }
614 :
615 : impl std::fmt::Display for MissingKeyError {
616 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
617 0 : write!(
618 0 : f,
619 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
620 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
621 0 : )?;
622 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
623 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
624 0 : }
625 :
626 0 : if let Some(ref backtrace) = self.backtrace {
627 0 : write!(f, "\n{}", backtrace)?;
628 0 : }
629 :
630 0 : Ok(())
631 0 : }
632 : }
633 :
634 : impl PageReconstructError {
635 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
636 0 : pub(crate) fn is_stopping(&self) -> bool {
637 : use PageReconstructError::*;
638 0 : match self {
639 0 : Cancelled => true,
640 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
641 : }
642 0 : }
643 : }
644 :
645 0 : #[derive(thiserror::Error, Debug)]
646 : pub(crate) enum CreateImageLayersError {
647 : #[error("timeline shutting down")]
648 : Cancelled,
649 :
650 : #[error("read failed")]
651 : GetVectoredError(#[source] GetVectoredError),
652 :
653 : #[error("reconstruction failed")]
654 : PageReconstructError(#[source] PageReconstructError),
655 :
656 : #[error(transparent)]
657 : Other(#[from] anyhow::Error),
658 : }
659 :
660 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
661 0 : fn from(_: layer_manager::Shutdown) -> Self {
662 0 : CreateImageLayersError::Cancelled
663 0 : }
664 : }
665 :
666 0 : #[derive(thiserror::Error, Debug, Clone)]
667 : pub(crate) enum FlushLayerError {
668 : /// Timeline cancellation token was cancelled
669 : #[error("timeline shutting down")]
670 : Cancelled,
671 :
672 : /// We tried to flush a layer while the Timeline is in an unexpected state
673 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
674 : NotRunning(FlushLoopState),
675 :
676 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
677 : // loop via a watch channel, where we can only borrow it.
678 : #[error("create image layers (shared)")]
679 : CreateImageLayersError(Arc<CreateImageLayersError>),
680 :
681 : #[error("other (shared)")]
682 : Other(#[from] Arc<anyhow::Error>),
683 : }
684 :
685 : impl FlushLayerError {
686 : // When crossing from generic anyhow errors to this error type, we explicitly check
687 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
688 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
689 0 : let cancelled = timeline.cancel.is_cancelled()
690 : // The upload queue might have been shut down before the official cancellation of the timeline.
691 0 : || err
692 0 : .downcast_ref::<NotInitialized>()
693 0 : .map(NotInitialized::is_stopping)
694 0 : .unwrap_or_default();
695 0 : if cancelled {
696 0 : Self::Cancelled
697 : } else {
698 0 : Self::Other(Arc::new(err))
699 : }
700 0 : }
701 : }
702 :
703 : impl From<layer_manager::Shutdown> for FlushLayerError {
704 0 : fn from(_: layer_manager::Shutdown) -> Self {
705 0 : FlushLayerError::Cancelled
706 0 : }
707 : }
708 :
709 0 : #[derive(thiserror::Error, Debug)]
710 : pub(crate) enum GetVectoredError {
711 : #[error("timeline shutting down")]
712 : Cancelled,
713 :
714 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
715 : Oversized(u64),
716 :
717 : #[error("requested at invalid LSN: {0}")]
718 : InvalidLsn(Lsn),
719 :
720 : #[error("requested key not found: {0}")]
721 : MissingKey(MissingKeyError),
722 :
723 : #[error("ancestry walk")]
724 : GetReadyAncestorError(#[source] GetReadyAncestorError),
725 :
726 : #[error(transparent)]
727 : Other(#[from] anyhow::Error),
728 : }
729 :
730 : impl From<GetReadyAncestorError> for GetVectoredError {
731 2 : fn from(value: GetReadyAncestorError) -> Self {
732 : use GetReadyAncestorError::*;
733 2 : match value {
734 0 : Cancelled => GetVectoredError::Cancelled,
735 : AncestorLsnTimeout(_) | BadState { .. } => {
736 2 : GetVectoredError::GetReadyAncestorError(value)
737 : }
738 : }
739 2 : }
740 : }
741 :
742 2 : #[derive(thiserror::Error, Debug)]
743 : pub(crate) enum GetReadyAncestorError {
744 : #[error("ancestor LSN wait error")]
745 : AncestorLsnTimeout(#[from] WaitLsnError),
746 :
747 : #[error("bad state on timeline {timeline_id}: {state:?}")]
748 : BadState {
749 : timeline_id: TimelineId,
750 : state: TimelineState,
751 : },
752 :
753 : #[error("cancelled")]
754 : Cancelled,
755 : }
756 :
757 : #[derive(Clone, Copy)]
758 : pub enum LogicalSizeCalculationCause {
759 : Initial,
760 : ConsumptionMetricsSyntheticSize,
761 : EvictionTaskImitation,
762 : TenantSizeHandler,
763 : }
764 :
765 : pub enum GetLogicalSizePriority {
766 : User,
767 : Background,
768 : }
769 :
770 0 : #[derive(enumset::EnumSetType)]
771 : pub(crate) enum CompactFlags {
772 : ForceRepartition,
773 : ForceImageLayerCreation,
774 : ForceL0Compaction,
775 : EnhancedGcBottomMostCompaction,
776 : DryRun,
777 : }
778 :
779 : #[serde_with::serde_as]
780 0 : #[derive(Debug, Clone, serde::Deserialize)]
781 : pub(crate) struct CompactRange {
782 : #[serde_as(as = "serde_with::DisplayFromStr")]
783 : pub start: Key,
784 : #[serde_as(as = "serde_with::DisplayFromStr")]
785 : pub end: Key,
786 : }
787 :
788 : #[derive(Clone, Default)]
789 : pub(crate) struct CompactOptions {
790 : pub flags: EnumSet<CompactFlags>,
791 : pub compact_range: Option<CompactRange>,
792 : }
793 :
794 : impl std::fmt::Debug for Timeline {
795 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
796 0 : write!(f, "Timeline<{}>", self.timeline_id)
797 0 : }
798 : }
799 :
800 0 : #[derive(thiserror::Error, Debug)]
801 : pub(crate) enum WaitLsnError {
802 : // Called on a timeline which is shutting down
803 : #[error("Shutdown")]
804 : Shutdown,
805 :
806 : // Called on an timeline not in active state or shutting down
807 : #[error("Bad timeline state: {0:?}")]
808 : BadState(TimelineState),
809 :
810 : // Timeout expired while waiting for LSN to catch up with goal.
811 : #[error("{0}")]
812 : Timeout(String),
813 : }
814 :
815 : // The impls below achieve cancellation mapping for errors.
816 : // Perhaps there's a way of achieving this with less cruft.
817 :
818 : impl From<CreateImageLayersError> for CompactionError {
819 0 : fn from(e: CreateImageLayersError) -> Self {
820 0 : match e {
821 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
822 0 : CreateImageLayersError::Other(e) => {
823 0 : CompactionError::Other(e.context("create image layers"))
824 : }
825 0 : _ => CompactionError::Other(e.into()),
826 : }
827 0 : }
828 : }
829 :
830 : impl From<CreateImageLayersError> for FlushLayerError {
831 0 : fn from(e: CreateImageLayersError) -> Self {
832 0 : match e {
833 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
834 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
835 : }
836 0 : }
837 : }
838 :
839 : impl From<PageReconstructError> for CreateImageLayersError {
840 0 : fn from(e: PageReconstructError) -> Self {
841 0 : match e {
842 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
843 0 : _ => CreateImageLayersError::PageReconstructError(e),
844 : }
845 0 : }
846 : }
847 :
848 : impl From<GetVectoredError> for CreateImageLayersError {
849 0 : fn from(e: GetVectoredError) -> Self {
850 0 : match e {
851 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
852 0 : _ => CreateImageLayersError::GetVectoredError(e),
853 : }
854 0 : }
855 : }
856 :
857 : impl From<GetVectoredError> for PageReconstructError {
858 6 : fn from(e: GetVectoredError) -> Self {
859 6 : match e {
860 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
861 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
862 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
863 4 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
864 2 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
865 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
866 : }
867 6 : }
868 : }
869 :
870 : impl From<GetReadyAncestorError> for PageReconstructError {
871 2 : fn from(e: GetReadyAncestorError) -> Self {
872 : use GetReadyAncestorError::*;
873 2 : match e {
874 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
875 2 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
876 0 : Cancelled => PageReconstructError::Cancelled,
877 : }
878 2 : }
879 : }
880 :
881 : pub(crate) enum WaitLsnWaiter<'a> {
882 : Timeline(&'a Timeline),
883 : Tenant,
884 : PageService,
885 : }
886 :
887 : /// Argument to [`Timeline::shutdown`].
888 : #[derive(Debug, Clone, Copy)]
889 : pub(crate) enum ShutdownMode {
890 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
891 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
892 : ///
893 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
894 : /// the call to [`Timeline::shutdown`].
895 : FreezeAndFlush,
896 : /// Only flush the layers to the remote storage without freezing any open layers. This is the
897 : /// mode used by ancestor detach and any other operations that reloads a tenant but not increasing
898 : /// the generation number.
899 : Flush,
900 : /// Shut down immediately, without waiting for any open layers to flush.
901 : Hard,
902 : }
903 :
904 : struct ImageLayerCreationOutcome {
905 : image: Option<ResidentLayer>,
906 : next_start_key: Key,
907 : }
908 :
909 : /// Public interface functions
910 : impl Timeline {
911 : /// Get the LSN where this branch was created
912 4 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
913 4 : self.ancestor_lsn
914 4 : }
915 :
916 : /// Get the ancestor's timeline id
917 12 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
918 12 : self.ancestor_timeline
919 12 : .as_ref()
920 12 : .map(|ancestor| ancestor.timeline_id)
921 12 : }
922 :
923 : /// Get the ancestor timeline
924 2 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
925 2 : self.ancestor_timeline.as_ref()
926 2 : }
927 :
928 : /// Get the bytes written since the PITR cutoff on this branch, and
929 : /// whether this branch's ancestor_lsn is within its parent's PITR.
930 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
931 0 : let gc_info = self.gc_info.read().unwrap();
932 0 : let history = self
933 0 : .get_last_record_lsn()
934 0 : .checked_sub(gc_info.cutoffs.time)
935 0 : .unwrap_or(Lsn(0))
936 0 : .0;
937 0 : (history, gc_info.within_ancestor_pitr)
938 0 : }
939 :
940 : /// Lock and get timeline's GC cutoff
941 248 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
942 248 : self.latest_gc_cutoff_lsn.read()
943 248 : }
944 :
945 : /// Look up given page version.
946 : ///
947 : /// If a remote layer file is needed, it is downloaded as part of this
948 : /// call.
949 : ///
950 : /// This method enforces [`Self::timeline_get_throttle`] internally.
951 : ///
952 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
953 : /// abstraction above this needs to store suitable metadata to track what
954 : /// data exists with what keys, in separate metadata entries. If a
955 : /// non-existent key is requested, we may incorrectly return a value from
956 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
957 : /// non-existing key.
958 : ///
959 : /// # Cancel-Safety
960 : ///
961 : /// This method is cancellation-safe.
962 : #[inline(always)]
963 606749 : pub(crate) async fn get(
964 606749 : &self,
965 606749 : key: Key,
966 606749 : lsn: Lsn,
967 606749 : ctx: &RequestContext,
968 606749 : ) -> Result<Bytes, PageReconstructError> {
969 606749 : if !lsn.is_valid() {
970 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
971 606749 : }
972 606749 :
973 606749 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
974 606749 : // already checked the key against the shard_identity when looking up the Timeline from
975 606749 : // page_service.
976 606749 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
977 :
978 606749 : self.timeline_get_throttle.throttle(ctx, 1).await;
979 :
980 606749 : let keyspace = KeySpace {
981 606749 : ranges: vec![key..key.next()],
982 606749 : };
983 606749 :
984 606749 : // Initialise the reconstruct state for the key with the cache
985 606749 : // entry returned above.
986 606749 : let mut reconstruct_state = ValuesReconstructState::new();
987 :
988 606749 : let vectored_res = self
989 606749 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
990 183014 : .await;
991 :
992 606749 : let key_value = vectored_res?.pop_first();
993 606743 : match key_value {
994 606731 : Some((got_key, value)) => {
995 606731 : if got_key != key {
996 0 : error!(
997 0 : "Expected {}, but singular vectored get returned {}",
998 : key, got_key
999 : );
1000 0 : Err(PageReconstructError::Other(anyhow!(
1001 0 : "Singular vectored get returned wrong key"
1002 0 : )))
1003 : } else {
1004 606731 : value
1005 : }
1006 : }
1007 12 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1008 12 : key,
1009 12 : shard: self.shard_identity.get_shard_number(&key),
1010 12 : cont_lsn: Lsn(0),
1011 12 : request_lsn: lsn,
1012 12 : ancestor_lsn: None,
1013 12 : backtrace: None,
1014 12 : })),
1015 : }
1016 606749 : }
1017 :
1018 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1019 : pub(crate) const VEC_GET_LAYERS_VISITED_WARN_THRESH: f64 = 512.0;
1020 :
1021 : /// Look up multiple page versions at a given LSN
1022 : ///
1023 : /// This naive implementation will be replaced with a more efficient one
1024 : /// which actually vectorizes the read path.
1025 19528 : pub(crate) async fn get_vectored(
1026 19528 : &self,
1027 19528 : keyspace: KeySpace,
1028 19528 : lsn: Lsn,
1029 19528 : ctx: &RequestContext,
1030 19528 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1031 19528 : if !lsn.is_valid() {
1032 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1033 19528 : }
1034 19528 :
1035 19528 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1036 19528 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1037 0 : return Err(GetVectoredError::Oversized(key_count));
1038 19528 : }
1039 :
1040 39056 : for range in &keyspace.ranges {
1041 19528 : let mut key = range.start;
1042 39262 : while key != range.end {
1043 19734 : assert!(!self.shard_identity.is_key_disposable(&key));
1044 19734 : key = key.next();
1045 : }
1046 : }
1047 :
1048 19528 : trace!(
1049 0 : "get vectored request for {:?}@{} from task kind {:?}",
1050 0 : keyspace,
1051 0 : lsn,
1052 0 : ctx.task_kind(),
1053 : );
1054 :
1055 19528 : let start = crate::metrics::GET_VECTORED_LATENCY
1056 19528 : .for_task_kind(ctx.task_kind())
1057 19528 : .map(|metric| (metric, Instant::now()));
1058 :
1059 : // start counting after throttle so that throttle time
1060 : // is always less than observation time
1061 19528 : let throttled = self
1062 19528 : .timeline_get_throttle
1063 19528 : .throttle(ctx, key_count as usize)
1064 0 : .await;
1065 :
1066 19528 : let res = self
1067 19528 : .get_vectored_impl(
1068 19528 : keyspace.clone(),
1069 19528 : lsn,
1070 19528 : &mut ValuesReconstructState::new(),
1071 19528 : ctx,
1072 19528 : )
1073 7037 : .await;
1074 :
1075 19528 : if let Some((metric, start)) = start {
1076 0 : let elapsed = start.elapsed();
1077 0 : let ex_throttled = if let Some(throttled) = throttled {
1078 0 : elapsed.checked_sub(throttled)
1079 : } else {
1080 0 : Some(elapsed)
1081 : };
1082 :
1083 0 : if let Some(ex_throttled) = ex_throttled {
1084 0 : metric.observe(ex_throttled.as_secs_f64());
1085 0 : } else {
1086 0 : use utils::rate_limit::RateLimit;
1087 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1088 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
1089 0 : let mut rate_limit = LOGGED.lock().unwrap();
1090 0 : rate_limit.call(|| {
1091 0 : warn!("error deducting time spent throttled; this message is logged at a global rate limit");
1092 0 : });
1093 0 : }
1094 19528 : }
1095 :
1096 19528 : res
1097 19528 : }
1098 :
1099 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1100 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1101 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1102 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1103 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1104 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1105 : /// the scan operation will not cause OOM in the future.
1106 12 : pub(crate) async fn scan(
1107 12 : &self,
1108 12 : keyspace: KeySpace,
1109 12 : lsn: Lsn,
1110 12 : ctx: &RequestContext,
1111 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1112 12 : if !lsn.is_valid() {
1113 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1114 12 : }
1115 12 :
1116 12 : trace!(
1117 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1118 0 : keyspace,
1119 0 : lsn,
1120 0 : ctx.task_kind()
1121 : );
1122 :
1123 : // We should generalize this into Keyspace::contains in the future.
1124 24 : for range in &keyspace.ranges {
1125 12 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1126 12 : || range.end.field1 > METADATA_KEY_END_PREFIX
1127 : {
1128 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1129 0 : "only metadata keyspace can be scanned"
1130 0 : )));
1131 12 : }
1132 : }
1133 :
1134 12 : let start = crate::metrics::SCAN_LATENCY
1135 12 : .for_task_kind(ctx.task_kind())
1136 12 : .map(ScanLatencyOngoingRecording::start_recording);
1137 :
1138 : // start counting after throttle so that throttle time
1139 : // is always less than observation time
1140 12 : let throttled = self
1141 12 : .timeline_get_throttle
1142 12 : // assume scan = 1 quota for now until we find a better way to process this
1143 12 : .throttle(ctx, 1)
1144 0 : .await;
1145 :
1146 12 : let vectored_res = self
1147 12 : .get_vectored_impl(
1148 12 : keyspace.clone(),
1149 12 : lsn,
1150 12 : &mut ValuesReconstructState::default(),
1151 12 : ctx,
1152 12 : )
1153 0 : .await;
1154 :
1155 12 : if let Some(recording) = start {
1156 0 : recording.observe(throttled);
1157 12 : }
1158 :
1159 12 : vectored_res
1160 12 : }
1161 :
1162 626591 : pub(super) async fn get_vectored_impl(
1163 626591 : &self,
1164 626591 : keyspace: KeySpace,
1165 626591 : lsn: Lsn,
1166 626591 : reconstruct_state: &mut ValuesReconstructState,
1167 626591 : ctx: &RequestContext,
1168 626591 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1169 626591 : let get_kind = if keyspace.total_raw_size() == 1 {
1170 626155 : GetKind::Singular
1171 : } else {
1172 436 : GetKind::Vectored
1173 : };
1174 :
1175 626591 : let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
1176 626591 : .for_get_kind(get_kind)
1177 626591 : .start_timer();
1178 626591 : self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1179 192236 : .await?;
1180 626575 : get_data_timer.stop_and_record();
1181 626575 :
1182 626575 : let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
1183 626575 : .for_get_kind(get_kind)
1184 626575 : .start_timer();
1185 626575 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
1186 626575 : let layers_visited = reconstruct_state.get_layers_visited();
1187 :
1188 666901 : for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
1189 666901 : match res {
1190 0 : Err(err) => {
1191 0 : results.insert(key, Err(err));
1192 0 : }
1193 666901 : Ok(state) => {
1194 666901 : let state = ValueReconstructState::from(state);
1195 :
1196 666901 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
1197 666901 : results.insert(key, reconstruct_res);
1198 : }
1199 : }
1200 : }
1201 626575 : reconstruct_timer.stop_and_record();
1202 626575 :
1203 626575 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1204 626575 : // when they're missing. Instead they are omitted from the resulting btree
1205 626575 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1206 626575 : // to avoid infinite results.
1207 626575 : if !results.is_empty() {
1208 626369 : let avg = layers_visited as f64 / results.len() as f64;
1209 626369 : if avg >= Self::VEC_GET_LAYERS_VISITED_WARN_THRESH {
1210 0 : use utils::rate_limit::RateLimit;
1211 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1212 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1213 0 : let mut rate_limit = LOGGED.lock().unwrap();
1214 0 : rate_limit.call(|| {
1215 0 : tracing::info!(
1216 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1217 0 : lsn = %lsn,
1218 0 : "Vectored read for {} visited {} layers on average per key and {} in total. {}/{} pages were returned",
1219 0 : keyspace, avg, layers_visited, results.len(), keyspace.total_raw_size());
1220 0 : });
1221 626369 : }
1222 :
1223 : // Note that this is an approximation. Tracking the exact number of layers visited
1224 : // per key requires virtually unbounded memory usage and is inefficient
1225 : // (i.e. segment tree tracking each range queried from a layer)
1226 626369 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED.observe(avg);
1227 206 : }
1228 :
1229 626575 : Ok(results)
1230 626591 : }
1231 :
1232 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1233 274290 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1234 274290 : self.last_record_lsn.load().last
1235 274290 : }
1236 :
1237 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1238 0 : self.last_record_lsn.load().prev
1239 0 : }
1240 :
1241 : /// Atomically get both last and prev.
1242 228 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1243 228 : self.last_record_lsn.load()
1244 228 : }
1245 :
1246 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1247 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1248 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1249 0 : self.last_record_lsn.status_receiver()
1250 0 : }
1251 :
1252 414 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1253 414 : self.disk_consistent_lsn.load()
1254 414 : }
1255 :
1256 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1257 : /// not validated with control plane yet.
1258 : /// See [`Self::get_remote_consistent_lsn_visible`].
1259 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1260 0 : self.remote_client.remote_consistent_lsn_projected()
1261 0 : }
1262 :
1263 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1264 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1265 : /// generation validation in the deletion queue.
1266 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1267 0 : self.remote_client.remote_consistent_lsn_visible()
1268 0 : }
1269 :
1270 : /// The sum of the file size of all historic layers in the layer map.
1271 : /// This method makes no distinction between local and remote layers.
1272 : /// Hence, the result **does not represent local filesystem usage**.
1273 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1274 0 : let guard = self.layers.read().await;
1275 0 : guard.layer_size_sum()
1276 0 : }
1277 :
1278 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1279 0 : self.metrics.resident_physical_size_get()
1280 0 : }
1281 :
1282 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1283 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1284 0 : }
1285 :
1286 : ///
1287 : /// Wait until WAL has been received and processed up to this LSN.
1288 : ///
1289 : /// You should call this before any of the other get_* or list_* functions. Calling
1290 : /// those functions with an LSN that has been processed yet is an error.
1291 : ///
1292 226556 : pub(crate) async fn wait_lsn(
1293 226556 : &self,
1294 226556 : lsn: Lsn,
1295 226556 : who_is_waiting: WaitLsnWaiter<'_>,
1296 226556 : ctx: &RequestContext, /* Prepare for use by cancellation */
1297 226556 : ) -> Result<(), WaitLsnError> {
1298 226556 : let state = self.current_state();
1299 226556 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1300 0 : return Err(WaitLsnError::Shutdown);
1301 226556 : } else if !matches!(state, TimelineState::Active) {
1302 0 : return Err(WaitLsnError::BadState(state));
1303 226556 : }
1304 226556 :
1305 226556 : if cfg!(debug_assertions) {
1306 226556 : match ctx.task_kind() {
1307 : TaskKind::WalReceiverManager
1308 : | TaskKind::WalReceiverConnectionHandler
1309 : | TaskKind::WalReceiverConnectionPoller => {
1310 0 : let is_myself = match who_is_waiting {
1311 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1312 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1313 : };
1314 0 : if is_myself {
1315 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1316 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1317 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1318 0 : }
1319 0 : } else {
1320 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1321 0 : // our walreceiver task can make progress independent of theirs
1322 0 : }
1323 : }
1324 226556 : _ => {}
1325 : }
1326 0 : }
1327 :
1328 226556 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1329 226556 :
1330 226556 : match self
1331 226556 : .last_record_lsn
1332 226556 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1333 0 : .await
1334 : {
1335 226556 : Ok(()) => Ok(()),
1336 0 : Err(e) => {
1337 : use utils::seqwait::SeqWaitError::*;
1338 0 : match e {
1339 0 : Shutdown => Err(WaitLsnError::Shutdown),
1340 : Timeout => {
1341 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1342 0 : drop(_timer);
1343 0 : let walreceiver_status = self.walreceiver_status();
1344 0 : Err(WaitLsnError::Timeout(format!(
1345 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1346 0 : lsn,
1347 0 : self.get_last_record_lsn(),
1348 0 : self.get_disk_consistent_lsn(),
1349 0 : walreceiver_status,
1350 0 : )))
1351 : }
1352 : }
1353 : }
1354 : }
1355 226556 : }
1356 :
1357 0 : pub(crate) fn walreceiver_status(&self) -> String {
1358 0 : match &*self.walreceiver.lock().unwrap() {
1359 0 : None => "stopping or stopped".to_string(),
1360 0 : Some(walreceiver) => match walreceiver.status() {
1361 0 : Some(status) => status.to_human_readable_string(),
1362 0 : None => "Not active".to_string(),
1363 : },
1364 : }
1365 0 : }
1366 :
1367 : /// Check that it is valid to request operations with that lsn.
1368 232 : pub(crate) fn check_lsn_is_in_scope(
1369 232 : &self,
1370 232 : lsn: Lsn,
1371 232 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1372 232 : ) -> anyhow::Result<()> {
1373 232 : ensure!(
1374 232 : lsn >= **latest_gc_cutoff_lsn,
1375 4 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1376 4 : lsn,
1377 4 : **latest_gc_cutoff_lsn,
1378 : );
1379 228 : Ok(())
1380 232 : }
1381 :
1382 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1383 10 : pub(crate) fn init_lsn_lease(
1384 10 : &self,
1385 10 : lsn: Lsn,
1386 10 : length: Duration,
1387 10 : ctx: &RequestContext,
1388 10 : ) -> anyhow::Result<LsnLease> {
1389 10 : self.make_lsn_lease(lsn, length, true, ctx)
1390 10 : }
1391 :
1392 : /// 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.
1393 4 : pub(crate) fn renew_lsn_lease(
1394 4 : &self,
1395 4 : lsn: Lsn,
1396 4 : length: Duration,
1397 4 : ctx: &RequestContext,
1398 4 : ) -> anyhow::Result<LsnLease> {
1399 4 : self.make_lsn_lease(lsn, length, false, ctx)
1400 4 : }
1401 :
1402 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1403 : ///
1404 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1405 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1406 : ///
1407 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1408 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1409 14 : fn make_lsn_lease(
1410 14 : &self,
1411 14 : lsn: Lsn,
1412 14 : length: Duration,
1413 14 : init: bool,
1414 14 : _ctx: &RequestContext,
1415 14 : ) -> anyhow::Result<LsnLease> {
1416 12 : let lease = {
1417 : // Normalize the requested LSN to be aligned, and move to the first record
1418 : // if it points to the beginning of the page (header).
1419 14 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1420 14 :
1421 14 : let mut gc_info = self.gc_info.write().unwrap();
1422 14 :
1423 14 : let valid_until = SystemTime::now() + length;
1424 14 :
1425 14 : let entry = gc_info.leases.entry(lsn);
1426 14 :
1427 14 : match entry {
1428 6 : Entry::Occupied(mut occupied) => {
1429 6 : let existing_lease = occupied.get_mut();
1430 6 : if valid_until > existing_lease.valid_until {
1431 2 : existing_lease.valid_until = valid_until;
1432 2 : let dt: DateTime<Utc> = valid_until.into();
1433 2 : info!("lease extended to {}", dt);
1434 : } else {
1435 4 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1436 4 : info!("existing lease covers greater length, valid until {}", dt);
1437 : }
1438 :
1439 6 : existing_lease.clone()
1440 : }
1441 8 : Entry::Vacant(vacant) => {
1442 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
1443 : // not blocked by the lsn lease deadline.
1444 8 : let validate = {
1445 8 : let conf = self.tenant_conf.load();
1446 8 : conf.location.attach_mode == AttachmentMode::Single
1447 8 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1448 : };
1449 :
1450 8 : if init || validate {
1451 8 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1452 8 : if lsn < *latest_gc_cutoff_lsn {
1453 2 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1454 6 : }
1455 0 : }
1456 :
1457 6 : let dt: DateTime<Utc> = valid_until.into();
1458 6 : info!("lease created, valid until {}", dt);
1459 6 : vacant.insert(LsnLease { valid_until }).clone()
1460 : }
1461 : }
1462 : };
1463 :
1464 12 : Ok(lease)
1465 14 : }
1466 :
1467 : /// Flush to disk all data that was written with the put_* functions
1468 1092 : #[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))]
1469 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1470 : self.freeze_and_flush0().await
1471 : }
1472 :
1473 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1474 : // polluting the span hierarchy.
1475 1092 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1476 1092 : let token = {
1477 : // Freeze the current open in-memory layer. It will be written to disk on next
1478 : // iteration.
1479 1092 : let mut g = self.write_lock.lock().await;
1480 :
1481 1092 : let to_lsn = self.get_last_record_lsn();
1482 1092 : self.freeze_inmem_layer_at(to_lsn, &mut g).await?
1483 : };
1484 1093 : self.wait_flush_completion(token).await
1485 1092 : }
1486 :
1487 : // Check if an open ephemeral layer should be closed: this provides
1488 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1489 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1490 : // ephemeral layer bytes has been breached.
1491 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1492 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1493 : // If the write lock is held, there is an active wal receiver: rolling open layers
1494 : // is their responsibility while they hold this lock.
1495 0 : return;
1496 : };
1497 :
1498 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1499 : // time, and this was missed.
1500 : // if write_guard.is_none() { return; }
1501 :
1502 0 : let Ok(layers_guard) = self.layers.try_read() else {
1503 : // Don't block if the layer lock is busy
1504 0 : return;
1505 : };
1506 :
1507 0 : let Ok(lm) = layers_guard.layer_map() else {
1508 0 : return;
1509 : };
1510 :
1511 0 : let Some(open_layer) = &lm.open_layer else {
1512 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1513 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1514 : // that didn't result in writes to this shard.
1515 :
1516 : // Must not hold the layers lock while waiting for a flush.
1517 0 : drop(layers_guard);
1518 0 :
1519 0 : let last_record_lsn = self.get_last_record_lsn();
1520 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1521 0 : if last_record_lsn > disk_consistent_lsn {
1522 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1523 : // we are a sharded tenant and have skipped some WAL
1524 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1525 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1526 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1527 : // without any data ingested (yet) doesn't write a remote index as soon as it
1528 : // sees its LSN advance: we only do this if we've been layer-less
1529 : // for some time.
1530 0 : tracing::debug!(
1531 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1532 : disk_consistent_lsn,
1533 : last_record_lsn
1534 : );
1535 :
1536 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1537 : // We know there is no open layer, so we can request freezing without actually
1538 : // freezing anything. This is true even if we have dropped the layers_guard, we
1539 : // still hold the write_guard.
1540 0 : let _ = async {
1541 0 : let token = self
1542 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1543 0 : .await?;
1544 0 : self.wait_flush_completion(token).await
1545 0 : }
1546 0 : .await;
1547 0 : }
1548 0 : }
1549 :
1550 0 : return;
1551 : };
1552 :
1553 0 : let Some(current_size) = open_layer.try_len() else {
1554 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1555 : // read lock to get size should always succeed.
1556 0 : tracing::warn!("Lock conflict while reading size of open layer");
1557 0 : return;
1558 : };
1559 :
1560 0 : let current_lsn = self.get_last_record_lsn();
1561 :
1562 0 : let checkpoint_distance_override = open_layer.tick().await;
1563 :
1564 0 : if let Some(size_override) = checkpoint_distance_override {
1565 0 : if current_size > size_override {
1566 : // This is not harmful, but it only happens in relatively rare cases where
1567 : // time-based checkpoints are not happening fast enough to keep the amount of
1568 : // ephemeral data within configured limits. It's a sign of stress on the system.
1569 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1570 0 : }
1571 0 : }
1572 :
1573 0 : let checkpoint_distance =
1574 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1575 0 :
1576 0 : if self.should_roll(
1577 0 : current_size,
1578 0 : current_size,
1579 0 : checkpoint_distance,
1580 0 : self.get_last_record_lsn(),
1581 0 : self.last_freeze_at.load(),
1582 0 : open_layer.get_opened_at(),
1583 0 : ) {
1584 0 : match open_layer.info() {
1585 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1586 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1587 0 : // happens asynchronously in the background.
1588 0 : tracing::debug!(
1589 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1590 : );
1591 : }
1592 : InMemoryLayerInfo::Open { .. } => {
1593 : // Upgrade to a write lock and freeze the layer
1594 0 : drop(layers_guard);
1595 0 : let res = self
1596 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1597 0 : .await;
1598 :
1599 0 : if let Err(e) = res {
1600 0 : tracing::info!(
1601 0 : "failed to flush frozen layer after background freeze: {e:#}"
1602 : );
1603 0 : }
1604 : }
1605 : }
1606 0 : }
1607 0 : }
1608 :
1609 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1610 : ///
1611 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1612 : /// child timelines that are not offloaded yet.
1613 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1614 0 : if self.remote_client.is_archived() != Some(true) {
1615 0 : return (false, "the timeline is not archived");
1616 0 : }
1617 0 : if !self.remote_client.no_pending_work() {
1618 : // if the remote client is still processing some work, we can't offload
1619 0 : return (false, "the upload queue is not drained yet");
1620 0 : }
1621 0 :
1622 0 : (true, "ok")
1623 0 : }
1624 :
1625 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1626 : /// compaction tasks.
1627 364 : pub(crate) async fn compact(
1628 364 : self: &Arc<Self>,
1629 364 : cancel: &CancellationToken,
1630 364 : flags: EnumSet<CompactFlags>,
1631 364 : ctx: &RequestContext,
1632 364 : ) -> Result<bool, CompactionError> {
1633 364 : self.compact_with_options(
1634 364 : cancel,
1635 364 : CompactOptions {
1636 364 : flags,
1637 364 : compact_range: None,
1638 364 : },
1639 364 : ctx,
1640 364 : )
1641 37157 : .await
1642 364 : }
1643 :
1644 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1645 : /// compaction tasks.
1646 364 : pub(crate) async fn compact_with_options(
1647 364 : self: &Arc<Self>,
1648 364 : cancel: &CancellationToken,
1649 364 : options: CompactOptions,
1650 364 : ctx: &RequestContext,
1651 364 : ) -> Result<bool, CompactionError> {
1652 364 : // most likely the cancellation token is from background task, but in tests it could be the
1653 364 : // request task as well.
1654 364 :
1655 364 : let prepare = async move {
1656 364 : let guard = self.compaction_lock.lock().await;
1657 :
1658 364 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1659 364 : BackgroundLoopKind::Compaction,
1660 364 : ctx,
1661 364 : )
1662 0 : .await;
1663 :
1664 364 : (guard, permit)
1665 364 : };
1666 :
1667 : // this wait probably never needs any "long time spent" logging, because we already nag if
1668 : // compaction task goes over it's period (20s) which is quite often in production.
1669 364 : let (_guard, _permit) = tokio::select! {
1670 364 : tuple = prepare => { tuple },
1671 364 : _ = self.cancel.cancelled() => return Ok(false),
1672 364 : _ = cancel.cancelled() => return Ok(false),
1673 : };
1674 :
1675 364 : let last_record_lsn = self.get_last_record_lsn();
1676 364 :
1677 364 : // Last record Lsn could be zero in case the timeline was just created
1678 364 : if !last_record_lsn.is_valid() {
1679 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1680 0 : return Ok(false);
1681 364 : }
1682 364 :
1683 364 : match self.get_compaction_algorithm_settings().kind {
1684 : CompactionAlgorithm::Tiered => {
1685 0 : self.compact_tiered(cancel, ctx).await?;
1686 0 : Ok(false)
1687 : }
1688 37157 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1689 : }
1690 364 : }
1691 :
1692 : /// Mutate the timeline with a [`TimelineWriter`].
1693 5133162 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1694 5133162 : TimelineWriter {
1695 5133162 : tl: self,
1696 5133162 : write_guard: self.write_lock.lock().await,
1697 : }
1698 5133162 : }
1699 :
1700 0 : pub(crate) fn activate(
1701 0 : self: &Arc<Self>,
1702 0 : parent: Arc<crate::tenant::Tenant>,
1703 0 : broker_client: BrokerClientChannel,
1704 0 : background_jobs_can_start: Option<&completion::Barrier>,
1705 0 : ctx: &RequestContext,
1706 0 : ) {
1707 0 : if self.tenant_shard_id.is_shard_zero() {
1708 0 : // Logical size is only maintained accurately on shard zero.
1709 0 : self.spawn_initial_logical_size_computation_task(ctx);
1710 0 : }
1711 0 : self.launch_wal_receiver(ctx, broker_client);
1712 0 : self.set_state(TimelineState::Active);
1713 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1714 0 : }
1715 :
1716 : /// After this function returns, there are no timeline-scoped tasks are left running.
1717 : ///
1718 : /// The preferred pattern for is:
1719 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1720 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1721 : /// go the extra mile and keep track of JoinHandles
1722 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1723 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1724 : ///
1725 : /// For legacy reasons, we still have multiple tasks spawned using
1726 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1727 : /// We refer to these as "timeline-scoped task_mgr tasks".
1728 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1729 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1730 : /// or [`task_mgr::shutdown_watcher`].
1731 : /// We want to gradually convert the code base away from these.
1732 : ///
1733 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1734 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1735 : /// ones that aren't mentioned here):
1736 : /// - [`TaskKind::TimelineDeletionWorker`]
1737 : /// - NB: also used for tenant deletion
1738 : /// - [`TaskKind::RemoteUploadTask`]`
1739 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1740 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1741 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1742 : /// - [`TaskKind::Eviction`]
1743 : /// - [`TaskKind::LayerFlushTask`]
1744 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1745 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1746 10 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1747 10 : debug_assert_current_span_has_tenant_and_timeline_id();
1748 10 :
1749 10 : // Regardless of whether we're going to try_freeze_and_flush
1750 10 : // or not, stop ingesting any more data. Walreceiver only provides
1751 10 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1752 10 : // So, only after the self.gate.close() below will we know for sure that
1753 10 : // no walreceiver tasks are left.
1754 10 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1755 10 : // data during the call to `self.freeze_and_flush()` below.
1756 10 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1757 10 : // which is what we'd need to properly model early shutdown of the walreceiver
1758 10 : // task sub-tree before the other Timeline task sub-trees.
1759 10 : let walreceiver = self.walreceiver.lock().unwrap().take();
1760 10 : tracing::debug!(
1761 0 : is_some = walreceiver.is_some(),
1762 0 : "Waiting for WalReceiverManager..."
1763 : );
1764 10 : if let Some(walreceiver) = walreceiver {
1765 0 : walreceiver.cancel();
1766 10 : }
1767 : // ... and inform any waiters for newer LSNs that there won't be any.
1768 10 : self.last_record_lsn.shutdown();
1769 10 :
1770 10 : if let ShutdownMode::FreezeAndFlush = mode {
1771 6 : if let Some((open, frozen)) = self
1772 6 : .layers
1773 6 : .read()
1774 0 : .await
1775 6 : .layer_map()
1776 6 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1777 6 : .ok()
1778 6 : .filter(|(open, frozen)| *open || *frozen > 0)
1779 : {
1780 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1781 6 : } else {
1782 6 : // this is double-shutdown, ignore it
1783 6 : }
1784 :
1785 : // we shut down walreceiver above, so, we won't add anything more
1786 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1787 : // to reach the disk & upload queue, then shut the upload queue and
1788 : // wait for it to drain.
1789 6 : match self.freeze_and_flush().await {
1790 : Ok(_) => {
1791 : // drain the upload queue
1792 : // if we did not wait for completion here, it might be our shutdown process
1793 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1794 : // be spawned.
1795 : //
1796 : // what is problematic is the shutting down of RemoteTimelineClient, because
1797 : // obviously it does not make sense to stop while we wait for it, but what
1798 : // about corner cases like s3 suddenly hanging up?
1799 6 : self.remote_client.shutdown().await;
1800 : }
1801 : Err(FlushLayerError::Cancelled) => {
1802 : // this is likely the second shutdown, ignore silently.
1803 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1804 0 : debug_assert!(self.cancel.is_cancelled());
1805 : }
1806 0 : Err(e) => {
1807 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1808 0 : // we have some extra WAL replay to do next time the timeline starts.
1809 0 : warn!("failed to freeze and flush: {e:#}");
1810 : }
1811 : }
1812 :
1813 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
1814 : // we also do a final check here to ensure that the queue is empty.
1815 6 : if !self.remote_client.no_pending_work() {
1816 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1817 6 : }
1818 4 : }
1819 :
1820 10 : if let ShutdownMode::Flush = mode {
1821 : // drain the upload queue
1822 2 : self.remote_client.shutdown().await;
1823 2 : if !self.remote_client.no_pending_work() {
1824 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1825 2 : }
1826 8 : }
1827 :
1828 : // Signal any subscribers to our cancellation token to drop out
1829 10 : tracing::debug!("Cancelling CancellationToken");
1830 10 : self.cancel.cancel();
1831 10 :
1832 10 : // Ensure Prevent new page service requests from starting.
1833 10 : self.handles.shutdown();
1834 10 :
1835 10 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1836 10 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1837 10 : self.remote_client.stop();
1838 10 :
1839 10 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1840 10 : // to shut down the upload queue tasks.
1841 10 : // TODO: fix that, task management should be encapsulated inside remote_client.
1842 10 : task_mgr::shutdown_tasks(
1843 10 : Some(TaskKind::RemoteUploadTask),
1844 10 : Some(self.tenant_shard_id),
1845 10 : Some(self.timeline_id),
1846 10 : )
1847 0 : .await;
1848 :
1849 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
1850 10 : tracing::debug!("Waiting for tasks...");
1851 10 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1852 :
1853 : {
1854 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
1855 : // open.
1856 10 : let mut write_guard = self.write_lock.lock().await;
1857 10 : self.layers.write().await.shutdown(&mut write_guard);
1858 10 : }
1859 10 :
1860 10 : // Finally wait until any gate-holders are complete.
1861 10 : //
1862 10 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1863 10 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1864 10 : self.gate.close().await;
1865 :
1866 10 : self.metrics.shutdown();
1867 10 : }
1868 :
1869 420 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1870 420 : match (self.current_state(), new_state) {
1871 420 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1872 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1873 : }
1874 0 : (st, TimelineState::Loading) => {
1875 0 : error!("ignoring transition from {st:?} into Loading state");
1876 : }
1877 0 : (TimelineState::Broken { .. }, new_state) => {
1878 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1879 : }
1880 : (TimelineState::Stopping, TimelineState::Active) => {
1881 0 : error!("Not activating a Stopping timeline");
1882 : }
1883 418 : (_, new_state) => {
1884 418 : self.state.send_replace(new_state);
1885 418 : }
1886 : }
1887 420 : }
1888 :
1889 2 : pub(crate) fn set_broken(&self, reason: String) {
1890 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1891 2 : let broken_state = TimelineState::Broken {
1892 2 : reason,
1893 2 : backtrace: backtrace_str,
1894 2 : };
1895 2 : self.set_state(broken_state);
1896 2 :
1897 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1898 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1899 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1900 2 : self.cancel.cancel();
1901 2 : }
1902 :
1903 227566 : pub(crate) fn current_state(&self) -> TimelineState {
1904 227566 : self.state.borrow().clone()
1905 227566 : }
1906 :
1907 6 : pub(crate) fn is_broken(&self) -> bool {
1908 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1909 6 : }
1910 :
1911 222 : pub(crate) fn is_active(&self) -> bool {
1912 222 : self.current_state() == TimelineState::Active
1913 222 : }
1914 :
1915 2 : pub(crate) fn is_archived(&self) -> Option<bool> {
1916 2 : self.remote_client.is_archived()
1917 2 : }
1918 :
1919 368 : pub(crate) fn is_stopping(&self) -> bool {
1920 368 : self.current_state() == TimelineState::Stopping
1921 368 : }
1922 :
1923 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1924 0 : self.state.subscribe()
1925 0 : }
1926 :
1927 226558 : pub(crate) async fn wait_to_become_active(
1928 226558 : &self,
1929 226558 : _ctx: &RequestContext, // Prepare for use by cancellation
1930 226558 : ) -> Result<(), TimelineState> {
1931 226558 : let mut receiver = self.state.subscribe();
1932 : loop {
1933 226558 : let current_state = receiver.borrow().clone();
1934 226558 : match current_state {
1935 : TimelineState::Loading => {
1936 0 : receiver
1937 0 : .changed()
1938 0 : .await
1939 0 : .expect("holding a reference to self");
1940 : }
1941 : TimelineState::Active { .. } => {
1942 226556 : return Ok(());
1943 : }
1944 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1945 : // There's no chance the timeline can transition back into ::Active
1946 2 : return Err(current_state);
1947 : }
1948 : }
1949 : }
1950 226558 : }
1951 :
1952 0 : pub(crate) async fn layer_map_info(
1953 0 : &self,
1954 0 : reset: LayerAccessStatsReset,
1955 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
1956 0 : let guard = self.layers.read().await;
1957 0 : let layer_map = guard.layer_map()?;
1958 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
1959 0 : if let Some(open_layer) = &layer_map.open_layer {
1960 0 : in_memory_layers.push(open_layer.info());
1961 0 : }
1962 0 : for frozen_layer in &layer_map.frozen_layers {
1963 0 : in_memory_layers.push(frozen_layer.info());
1964 0 : }
1965 :
1966 0 : let historic_layers = layer_map
1967 0 : .iter_historic_layers()
1968 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
1969 0 : .collect();
1970 0 :
1971 0 : Ok(LayerMapInfo {
1972 0 : in_memory_layers,
1973 0 : historic_layers,
1974 0 : })
1975 0 : }
1976 :
1977 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))]
1978 : pub(crate) async fn download_layer(
1979 : &self,
1980 : layer_file_name: &LayerName,
1981 : ) -> anyhow::Result<Option<bool>> {
1982 : let Some(layer) = self.find_layer(layer_file_name).await? else {
1983 : return Ok(None);
1984 : };
1985 :
1986 : layer.download().await?;
1987 :
1988 : Ok(Some(true))
1989 : }
1990 :
1991 : /// Evict just one layer.
1992 : ///
1993 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
1994 0 : pub(crate) async fn evict_layer(
1995 0 : &self,
1996 0 : layer_file_name: &LayerName,
1997 0 : ) -> anyhow::Result<Option<bool>> {
1998 0 : let _gate = self
1999 0 : .gate
2000 0 : .enter()
2001 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2002 :
2003 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2004 0 : return Ok(None);
2005 : };
2006 :
2007 : // curl has this by default
2008 0 : let timeout = std::time::Duration::from_secs(120);
2009 0 :
2010 0 : match local_layer.evict_and_wait(timeout).await {
2011 0 : Ok(()) => Ok(Some(true)),
2012 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2013 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2014 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2015 : }
2016 0 : }
2017 :
2018 4803010 : fn should_roll(
2019 4803010 : &self,
2020 4803010 : layer_size: u64,
2021 4803010 : projected_layer_size: u64,
2022 4803010 : checkpoint_distance: u64,
2023 4803010 : projected_lsn: Lsn,
2024 4803010 : last_freeze_at: Lsn,
2025 4803010 : opened_at: Instant,
2026 4803010 : ) -> bool {
2027 4803010 : let distance = projected_lsn.widening_sub(last_freeze_at);
2028 4803010 :
2029 4803010 : // Rolling the open layer can be triggered by:
2030 4803010 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2031 4803010 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2032 4803010 : // account for how writes are distributed across shards: we expect each node to consume
2033 4803010 : // 1/count of the LSN on average.
2034 4803010 : // 2. The size of the currently open layer.
2035 4803010 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2036 4803010 : // up and suspend activity.
2037 4803010 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2038 0 : info!(
2039 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2040 : projected_lsn, layer_size, distance
2041 : );
2042 :
2043 0 : true
2044 4803010 : } else if projected_layer_size >= checkpoint_distance {
2045 : // NB: this check is relied upon by:
2046 80 : let _ = IndexEntry::validate_checkpoint_distance;
2047 80 : info!(
2048 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2049 : projected_lsn, layer_size, projected_layer_size
2050 : );
2051 :
2052 80 : true
2053 4802930 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2054 0 : info!(
2055 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2056 0 : projected_lsn,
2057 0 : layer_size,
2058 0 : opened_at.elapsed()
2059 : );
2060 :
2061 0 : true
2062 : } else {
2063 4802930 : false
2064 : }
2065 4803010 : }
2066 : }
2067 :
2068 : /// Number of times we will compute partition within a checkpoint distance.
2069 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2070 :
2071 : // Private functions
2072 : impl Timeline {
2073 12 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2074 12 : let tenant_conf = self.tenant_conf.load();
2075 12 : tenant_conf
2076 12 : .tenant_conf
2077 12 : .lsn_lease_length
2078 12 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2079 12 : }
2080 :
2081 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2082 0 : let tenant_conf = self.tenant_conf.load();
2083 0 : tenant_conf
2084 0 : .tenant_conf
2085 0 : .lsn_lease_length_for_ts
2086 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2087 0 : }
2088 :
2089 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2090 0 : let tenant_conf = self.tenant_conf.load();
2091 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2092 0 : }
2093 :
2094 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2095 0 : let tenant_conf = self.tenant_conf.load();
2096 0 : tenant_conf
2097 0 : .tenant_conf
2098 0 : .lazy_slru_download
2099 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2100 0 : }
2101 :
2102 4804616 : fn get_checkpoint_distance(&self) -> u64 {
2103 4804616 : let tenant_conf = self.tenant_conf.load();
2104 4804616 : tenant_conf
2105 4804616 : .tenant_conf
2106 4804616 : .checkpoint_distance
2107 4804616 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2108 4804616 : }
2109 :
2110 4802930 : fn get_checkpoint_timeout(&self) -> Duration {
2111 4802930 : let tenant_conf = self.tenant_conf.load();
2112 4802930 : tenant_conf
2113 4802930 : .tenant_conf
2114 4802930 : .checkpoint_timeout
2115 4802930 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2116 4802930 : }
2117 :
2118 618 : fn get_compaction_target_size(&self) -> u64 {
2119 618 : let tenant_conf = self.tenant_conf.load();
2120 618 : tenant_conf
2121 618 : .tenant_conf
2122 618 : .compaction_target_size
2123 618 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2124 618 : }
2125 :
2126 1536 : fn get_compaction_threshold(&self) -> usize {
2127 1536 : let tenant_conf = self.tenant_conf.load();
2128 1536 : tenant_conf
2129 1536 : .tenant_conf
2130 1536 : .compaction_threshold
2131 1536 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2132 1536 : }
2133 :
2134 14 : fn get_image_creation_threshold(&self) -> usize {
2135 14 : let tenant_conf = self.tenant_conf.load();
2136 14 : tenant_conf
2137 14 : .tenant_conf
2138 14 : .image_creation_threshold
2139 14 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2140 14 : }
2141 :
2142 364 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2143 364 : let tenant_conf = &self.tenant_conf.load();
2144 364 : tenant_conf
2145 364 : .tenant_conf
2146 364 : .compaction_algorithm
2147 364 : .as_ref()
2148 364 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2149 364 : .clone()
2150 364 : }
2151 :
2152 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2153 0 : let tenant_conf = self.tenant_conf.load();
2154 0 : tenant_conf
2155 0 : .tenant_conf
2156 0 : .eviction_policy
2157 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2158 0 : }
2159 :
2160 418 : fn get_evictions_low_residence_duration_metric_threshold(
2161 418 : tenant_conf: &TenantConfOpt,
2162 418 : default_tenant_conf: &TenantConf,
2163 418 : ) -> Duration {
2164 418 : tenant_conf
2165 418 : .evictions_low_residence_duration_metric_threshold
2166 418 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2167 418 : }
2168 :
2169 716 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2170 716 : let tenant_conf = self.tenant_conf.load();
2171 716 : tenant_conf
2172 716 : .tenant_conf
2173 716 : .image_layer_creation_check_threshold
2174 716 : .unwrap_or(
2175 716 : self.conf
2176 716 : .default_tenant_conf
2177 716 : .image_layer_creation_check_threshold,
2178 716 : )
2179 716 : }
2180 :
2181 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2182 0 : // NB: Most tenant conf options are read by background loops, so,
2183 0 : // changes will automatically be picked up.
2184 0 :
2185 0 : // The threshold is embedded in the metric. So, we need to update it.
2186 0 : {
2187 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2188 0 : &new_conf.tenant_conf,
2189 0 : &self.conf.default_tenant_conf,
2190 0 : );
2191 0 :
2192 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2193 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2194 0 :
2195 0 : let timeline_id_str = self.timeline_id.to_string();
2196 0 :
2197 0 : self.remote_client.update_config(&new_conf.location);
2198 0 :
2199 0 : self.metrics
2200 0 : .evictions_with_low_residence_duration
2201 0 : .write()
2202 0 : .unwrap()
2203 0 : .change_threshold(
2204 0 : &tenant_id_str,
2205 0 : &shard_id_str,
2206 0 : &timeline_id_str,
2207 0 : new_threshold,
2208 0 : );
2209 0 : }
2210 0 : }
2211 :
2212 : /// Open a Timeline handle.
2213 : ///
2214 : /// Loads the metadata for the timeline into memory, but not the layer map.
2215 : #[allow(clippy::too_many_arguments)]
2216 418 : pub(super) fn new(
2217 418 : conf: &'static PageServerConf,
2218 418 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2219 418 : metadata: &TimelineMetadata,
2220 418 : ancestor: Option<Arc<Timeline>>,
2221 418 : timeline_id: TimelineId,
2222 418 : tenant_shard_id: TenantShardId,
2223 418 : generation: Generation,
2224 418 : shard_identity: ShardIdentity,
2225 418 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2226 418 : resources: TimelineResources,
2227 418 : pg_version: u32,
2228 418 : state: TimelineState,
2229 418 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2230 418 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2231 418 : cancel: CancellationToken,
2232 418 : ) -> Arc<Self> {
2233 418 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2234 418 : let (state, _) = watch::channel(state);
2235 418 :
2236 418 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2237 418 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2238 418 :
2239 418 : let evictions_low_residence_duration_metric_threshold = {
2240 418 : let loaded_tenant_conf = tenant_conf.load();
2241 418 : Self::get_evictions_low_residence_duration_metric_threshold(
2242 418 : &loaded_tenant_conf.tenant_conf,
2243 418 : &conf.default_tenant_conf,
2244 418 : )
2245 : };
2246 :
2247 418 : if let Some(ancestor) = &ancestor {
2248 230 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2249 230 : // If we construct an explicit timeline object, it's obviously not offloaded
2250 230 : let is_offloaded = MaybeOffloaded::No;
2251 230 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2252 230 : }
2253 :
2254 418 : Arc::new_cyclic(|myself| {
2255 418 : let metrics = TimelineMetrics::new(
2256 418 : &tenant_shard_id,
2257 418 : &timeline_id,
2258 418 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2259 418 : "mtime",
2260 418 : evictions_low_residence_duration_metric_threshold,
2261 418 : ),
2262 418 : );
2263 418 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2264 :
2265 418 : let mut result = Timeline {
2266 418 : conf,
2267 418 : tenant_conf,
2268 418 : myself: myself.clone(),
2269 418 : timeline_id,
2270 418 : tenant_shard_id,
2271 418 : generation,
2272 418 : shard_identity,
2273 418 : pg_version,
2274 418 : layers: Default::default(),
2275 418 :
2276 418 : walredo_mgr,
2277 418 : walreceiver: Mutex::new(None),
2278 418 :
2279 418 : remote_client: Arc::new(resources.remote_client),
2280 418 :
2281 418 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2282 418 : last_record_lsn: SeqWait::new(RecordLsn {
2283 418 : last: disk_consistent_lsn,
2284 418 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2285 418 : }),
2286 418 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2287 418 :
2288 418 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2289 418 : last_freeze_ts: RwLock::new(Instant::now()),
2290 418 :
2291 418 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2292 418 :
2293 418 : ancestor_timeline: ancestor,
2294 418 : ancestor_lsn: metadata.ancestor_lsn(),
2295 418 :
2296 418 : metrics,
2297 418 :
2298 418 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2299 418 : &tenant_shard_id,
2300 418 : &timeline_id,
2301 418 : ),
2302 418 :
2303 2926 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2304 418 :
2305 418 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2306 418 :
2307 418 : layer_flush_start_tx,
2308 418 : layer_flush_done_tx,
2309 418 :
2310 418 : write_lock: tokio::sync::Mutex::new(None),
2311 418 :
2312 418 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2313 418 :
2314 418 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2315 418 : initdb_lsn: metadata.initdb_lsn(),
2316 418 :
2317 418 : current_logical_size: if disk_consistent_lsn.is_valid() {
2318 : // we're creating timeline data with some layer files existing locally,
2319 : // need to recalculate timeline's logical size based on data in the layers.
2320 234 : LogicalSize::deferred_initial(disk_consistent_lsn)
2321 : } else {
2322 : // we're creating timeline data without any layers existing locally,
2323 : // initial logical size is 0.
2324 184 : LogicalSize::empty_initial()
2325 : },
2326 418 : partitioning: tokio::sync::Mutex::new((
2327 418 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2328 418 : Lsn(0),
2329 418 : )),
2330 418 : repartition_threshold: 0,
2331 418 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2332 418 : last_image_layer_creation_check_instant: Mutex::new(None),
2333 418 :
2334 418 : last_received_wal: Mutex::new(None),
2335 418 : rel_size_cache: RwLock::new(RelSizeCache {
2336 418 : complete_as_of: disk_consistent_lsn,
2337 418 : map: HashMap::new(),
2338 418 : }),
2339 418 :
2340 418 : download_all_remote_layers_task_info: RwLock::new(None),
2341 418 :
2342 418 : state,
2343 418 :
2344 418 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2345 418 : EvictionTaskTimelineState::default(),
2346 418 : ),
2347 418 : delete_progress: TimelineDeleteProgress::default(),
2348 418 :
2349 418 : cancel,
2350 418 : gate: Gate::default(),
2351 418 :
2352 418 : compaction_lock: tokio::sync::Mutex::default(),
2353 418 : gc_lock: tokio::sync::Mutex::default(),
2354 418 :
2355 418 : standby_horizon: AtomicLsn::new(0),
2356 418 :
2357 418 : timeline_get_throttle: resources.timeline_get_throttle,
2358 418 :
2359 418 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2360 418 :
2361 418 : #[cfg(test)]
2362 418 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2363 418 :
2364 418 : l0_flush_global_state: resources.l0_flush_global_state,
2365 418 :
2366 418 : handles: Default::default(),
2367 418 :
2368 418 : attach_wal_lag_cooldown,
2369 418 :
2370 418 : create_idempotency,
2371 418 : };
2372 418 :
2373 418 : result.repartition_threshold =
2374 418 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2375 418 :
2376 418 : result
2377 418 : .metrics
2378 418 : .last_record_gauge
2379 418 : .set(disk_consistent_lsn.0 as i64);
2380 418 : result
2381 418 : })
2382 418 : }
2383 :
2384 588 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2385 588 : let Ok(guard) = self.gate.enter() else {
2386 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2387 0 : return;
2388 : };
2389 588 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2390 588 : match *flush_loop_state {
2391 412 : FlushLoopState::NotStarted => (),
2392 : FlushLoopState::Running { .. } => {
2393 176 : info!(
2394 0 : "skipping attempt to start flush_loop twice {}/{}",
2395 0 : self.tenant_shard_id, self.timeline_id
2396 : );
2397 176 : return;
2398 : }
2399 : FlushLoopState::Exited => {
2400 0 : warn!(
2401 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2402 0 : self.tenant_shard_id, self.timeline_id
2403 : );
2404 0 : return;
2405 : }
2406 : }
2407 :
2408 412 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2409 412 : let self_clone = Arc::clone(self);
2410 412 :
2411 412 : debug!("spawning flush loop");
2412 412 : *flush_loop_state = FlushLoopState::Running {
2413 412 : #[cfg(test)]
2414 412 : expect_initdb_optimization: false,
2415 412 : #[cfg(test)]
2416 412 : initdb_optimization_count: 0,
2417 412 : };
2418 412 : task_mgr::spawn(
2419 412 : task_mgr::BACKGROUND_RUNTIME.handle(),
2420 412 : task_mgr::TaskKind::LayerFlushTask,
2421 412 : self.tenant_shard_id,
2422 412 : Some(self.timeline_id),
2423 412 : "layer flush task",
2424 412 : async move {
2425 412 : let _guard = guard;
2426 412 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2427 18039 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2428 10 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2429 10 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2430 10 : *flush_loop_state = FlushLoopState::Exited;
2431 10 : Ok(())
2432 10 : }
2433 412 : .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))
2434 : );
2435 588 : }
2436 :
2437 : /// Creates and starts the wal receiver.
2438 : ///
2439 : /// This function is expected to be called at most once per Timeline's lifecycle
2440 : /// when the timeline is activated.
2441 0 : fn launch_wal_receiver(
2442 0 : self: &Arc<Self>,
2443 0 : ctx: &RequestContext,
2444 0 : broker_client: BrokerClientChannel,
2445 0 : ) {
2446 0 : info!(
2447 0 : "launching WAL receiver for timeline {} of tenant {}",
2448 0 : self.timeline_id, self.tenant_shard_id
2449 : );
2450 :
2451 0 : let tenant_conf = self.tenant_conf.load();
2452 0 : let wal_connect_timeout = tenant_conf
2453 0 : .tenant_conf
2454 0 : .walreceiver_connect_timeout
2455 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2456 0 : let lagging_wal_timeout = tenant_conf
2457 0 : .tenant_conf
2458 0 : .lagging_wal_timeout
2459 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2460 0 : let max_lsn_wal_lag = tenant_conf
2461 0 : .tenant_conf
2462 0 : .max_lsn_wal_lag
2463 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2464 0 :
2465 0 : let mut guard = self.walreceiver.lock().unwrap();
2466 0 : assert!(
2467 0 : guard.is_none(),
2468 0 : "multiple launches / re-launches of WAL receiver are not supported"
2469 : );
2470 0 : *guard = Some(WalReceiver::start(
2471 0 : Arc::clone(self),
2472 0 : WalReceiverConf {
2473 0 : wal_connect_timeout,
2474 0 : lagging_wal_timeout,
2475 0 : max_lsn_wal_lag,
2476 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2477 0 : availability_zone: self.conf.availability_zone.clone(),
2478 0 : ingest_batch_size: self.conf.ingest_batch_size,
2479 0 : },
2480 0 : broker_client,
2481 0 : ctx,
2482 0 : ));
2483 0 : }
2484 :
2485 : /// Initialize with an empty layer map. Used when creating a new timeline.
2486 412 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2487 412 : let mut layers = self.layers.try_write().expect(
2488 412 : "in the context where we call this function, no other task has access to the object",
2489 412 : );
2490 412 : layers
2491 412 : .open_mut()
2492 412 : .expect("in this context the LayerManager must still be open")
2493 412 : .initialize_empty(Lsn(start_lsn.0));
2494 412 : }
2495 :
2496 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2497 : /// files.
2498 6 : pub(super) async fn load_layer_map(
2499 6 : &self,
2500 6 : disk_consistent_lsn: Lsn,
2501 6 : index_part: IndexPart,
2502 6 : ) -> anyhow::Result<()> {
2503 : use init::{Decision::*, Discovered, DismissedLayer};
2504 : use LayerName::*;
2505 :
2506 6 : let mut guard = self.layers.write().await;
2507 :
2508 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2509 6 :
2510 6 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2511 6 : // structs representing all files on disk
2512 6 : let timeline_path = self
2513 6 : .conf
2514 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2515 6 : let conf = self.conf;
2516 6 : let span = tracing::Span::current();
2517 6 :
2518 6 : // Copy to move into the task we're about to spawn
2519 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2520 :
2521 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2522 6 : move || {
2523 6 : let _g = span.entered();
2524 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2525 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2526 6 : let mut unrecognized_files = Vec::new();
2527 6 :
2528 6 : let mut path = timeline_path;
2529 :
2530 22 : for discovered in discovered {
2531 16 : let (name, kind) = match discovered {
2532 16 : Discovered::Layer(layer_file_name, local_metadata) => {
2533 16 : discovered_layers.push((layer_file_name, local_metadata));
2534 16 : continue;
2535 : }
2536 0 : Discovered::IgnoredBackup(path) => {
2537 0 : std::fs::remove_file(path)
2538 0 : .or_else(fs_ext::ignore_not_found)
2539 0 : .fatal_err("Removing .old file");
2540 0 : continue;
2541 : }
2542 0 : Discovered::Unknown(file_name) => {
2543 0 : // we will later error if there are any
2544 0 : unrecognized_files.push(file_name);
2545 0 : continue;
2546 : }
2547 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2548 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2549 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2550 : };
2551 0 : path.push(Utf8Path::new(&name));
2552 0 : init::cleanup(&path, kind)?;
2553 0 : path.pop();
2554 : }
2555 :
2556 6 : if !unrecognized_files.is_empty() {
2557 : // assume that if there are any there are many many.
2558 0 : let n = unrecognized_files.len();
2559 0 : let first = &unrecognized_files[..n.min(10)];
2560 0 : anyhow::bail!(
2561 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2562 0 : );
2563 6 : }
2564 6 :
2565 6 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2566 6 :
2567 6 : let mut loaded_layers = Vec::new();
2568 6 : let mut needs_cleanup = Vec::new();
2569 6 : let mut total_physical_size = 0;
2570 :
2571 22 : for (name, decision) in decided {
2572 16 : let decision = match decision {
2573 16 : Ok(decision) => decision,
2574 0 : Err(DismissedLayer::Future { local }) => {
2575 0 : if let Some(local) = local {
2576 0 : init::cleanup_future_layer(
2577 0 : &local.local_path,
2578 0 : &name,
2579 0 : disk_consistent_lsn,
2580 0 : )?;
2581 0 : }
2582 0 : needs_cleanup.push(name);
2583 0 : continue;
2584 : }
2585 0 : Err(DismissedLayer::LocalOnly(local)) => {
2586 0 : init::cleanup_local_only_file(&name, &local)?;
2587 : // this file never existed remotely, we will have to do rework
2588 0 : continue;
2589 : }
2590 0 : Err(DismissedLayer::BadMetadata(local)) => {
2591 0 : init::cleanup_local_file_for_remote(&local)?;
2592 : // this file never existed remotely, we will have to do rework
2593 0 : continue;
2594 : }
2595 : };
2596 :
2597 16 : match &name {
2598 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2599 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2600 : }
2601 :
2602 16 : tracing::debug!(layer=%name, ?decision, "applied");
2603 :
2604 16 : let layer = match decision {
2605 16 : Resident { local, remote } => {
2606 16 : total_physical_size += local.file_size;
2607 16 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2608 16 : .drop_eviction_guard()
2609 : }
2610 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2611 : };
2612 :
2613 16 : loaded_layers.push(layer);
2614 : }
2615 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2616 6 : }
2617 6 : })
2618 4 : .await
2619 6 : .map_err(anyhow::Error::new)
2620 6 : .and_then(|x| x)?;
2621 :
2622 6 : let num_layers = loaded_layers.len();
2623 6 :
2624 6 : guard
2625 6 : .open_mut()
2626 6 : .expect("layermanager must be open during init")
2627 6 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2628 6 :
2629 6 : self.remote_client
2630 6 : .schedule_layer_file_deletion(&needs_cleanup)?;
2631 6 : self.remote_client
2632 6 : .schedule_index_upload_for_file_changes()?;
2633 : // This barrier orders above DELETEs before any later operations.
2634 : // This is critical because code executing after the barrier might
2635 : // create again objects with the same key that we just scheduled for deletion.
2636 : // For example, if we just scheduled deletion of an image layer "from the future",
2637 : // later compaction might run again and re-create the same image layer.
2638 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2639 : // "same" here means same key range and LSN.
2640 : //
2641 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2642 : // the upload queue may execute the PUT first, then the DELETE.
2643 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2644 : //
2645 : // 1. a future image layer is created and uploaded
2646 : // 2. ps restart
2647 : // 3. the future layer from (1) is deleted during load layer map
2648 : // 4. image layer is re-created and uploaded
2649 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2650 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2651 : //
2652 : // See https://github.com/neondatabase/neon/issues/5878
2653 : //
2654 : // NB: generation numbers naturally protect against this because they disambiguate
2655 : // (1) and (4)
2656 : // TODO: this is basically a no-op now, should we remove it?
2657 6 : self.remote_client.schedule_barrier()?;
2658 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2659 : // on retry.
2660 :
2661 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2662 6 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2663 6 : self.update_layer_visibility().await?;
2664 :
2665 6 : info!(
2666 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2667 : num_layers, disk_consistent_lsn, total_physical_size
2668 : );
2669 :
2670 6 : timer.stop_and_record();
2671 6 : Ok(())
2672 6 : }
2673 :
2674 : /// Retrieve current logical size of the timeline.
2675 : ///
2676 : /// The size could be lagging behind the actual number, in case
2677 : /// the initial size calculation has not been run (gets triggered on the first size access).
2678 : ///
2679 : /// return size and boolean flag that shows if the size is exact
2680 0 : pub(crate) fn get_current_logical_size(
2681 0 : self: &Arc<Self>,
2682 0 : priority: GetLogicalSizePriority,
2683 0 : ctx: &RequestContext,
2684 0 : ) -> logical_size::CurrentLogicalSize {
2685 0 : if !self.tenant_shard_id.is_shard_zero() {
2686 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2687 : // when HTTP API is serving a GET for timeline zero, return zero
2688 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2689 0 : }
2690 0 :
2691 0 : let current_size = self.current_logical_size.current_size();
2692 0 : debug!("Current size: {current_size:?}");
2693 :
2694 0 : match (current_size.accuracy(), priority) {
2695 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2696 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2697 0 : // background task will eventually deliver an exact value, we're in no rush
2698 0 : }
2699 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2700 : // background task is not ready, but user is asking for it now;
2701 : // => make the background task skip the line
2702 : // (The alternative would be to calculate the size here, but,
2703 : // it can actually take a long time if the user has a lot of rels.
2704 : // And we'll inevitable need it again; So, let the background task do the work.)
2705 0 : match self
2706 0 : .current_logical_size
2707 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2708 0 : .get()
2709 : {
2710 0 : Some(cancel) => cancel.cancel(),
2711 : None => {
2712 0 : match self.current_state() {
2713 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2714 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2715 0 : // Don't make noise.
2716 0 : }
2717 : TimelineState::Loading => {
2718 : // Import does not return an activated timeline.
2719 0 : info!("discarding priority boost for logical size calculation because timeline is not yet active");
2720 : }
2721 : TimelineState::Active => {
2722 : // activation should be setting the once cell
2723 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2724 0 : debug_assert!(false);
2725 : }
2726 : }
2727 : }
2728 : }
2729 : }
2730 : }
2731 :
2732 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2733 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2734 0 : let first = self
2735 0 : .current_logical_size
2736 0 : .did_return_approximate_to_walreceiver
2737 0 : .compare_exchange(
2738 0 : false,
2739 0 : true,
2740 0 : AtomicOrdering::Relaxed,
2741 0 : AtomicOrdering::Relaxed,
2742 0 : )
2743 0 : .is_ok();
2744 0 : if first {
2745 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2746 0 : }
2747 0 : }
2748 0 : }
2749 :
2750 0 : current_size
2751 0 : }
2752 :
2753 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2754 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2755 : // nothing to do for freshly created timelines;
2756 0 : assert_eq!(
2757 0 : self.current_logical_size.current_size().accuracy(),
2758 0 : logical_size::Accuracy::Exact,
2759 0 : );
2760 0 : self.current_logical_size.initialized.add_permits(1);
2761 0 : return;
2762 : };
2763 :
2764 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2765 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2766 0 : self.current_logical_size
2767 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2768 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2769 0 :
2770 0 : let self_clone = Arc::clone(self);
2771 0 : let background_ctx = ctx.detached_child(
2772 0 : TaskKind::InitialLogicalSizeCalculation,
2773 0 : DownloadBehavior::Download,
2774 0 : );
2775 0 : task_mgr::spawn(
2776 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2777 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2778 0 : self.tenant_shard_id,
2779 0 : Some(self.timeline_id),
2780 0 : "initial size calculation",
2781 : // NB: don't log errors here, task_mgr will do that.
2782 0 : async move {
2783 0 : let cancel = task_mgr::shutdown_token();
2784 0 : self_clone
2785 0 : .initial_logical_size_calculation_task(
2786 0 : initial_part_end,
2787 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2788 0 : cancel,
2789 0 : background_ctx,
2790 0 : )
2791 0 : .await;
2792 0 : Ok(())
2793 0 : }
2794 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)),
2795 : );
2796 0 : }
2797 :
2798 0 : async fn initial_logical_size_calculation_task(
2799 0 : self: Arc<Self>,
2800 0 : initial_part_end: Lsn,
2801 0 : skip_concurrency_limiter: CancellationToken,
2802 0 : cancel: CancellationToken,
2803 0 : background_ctx: RequestContext,
2804 0 : ) {
2805 0 : scopeguard::defer! {
2806 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2807 0 : self.current_logical_size.initialized.add_permits(1);
2808 0 : }
2809 0 :
2810 0 : let try_once = |attempt: usize| {
2811 0 : let background_ctx = &background_ctx;
2812 0 : let self_ref = &self;
2813 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2814 0 : async move {
2815 0 : let cancel = task_mgr::shutdown_token();
2816 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2817 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2818 0 : background_ctx,
2819 0 : );
2820 :
2821 : use crate::metrics::initial_logical_size::StartCircumstances;
2822 0 : let (_maybe_permit, circumstances) = tokio::select! {
2823 0 : permit = wait_for_permit => {
2824 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2825 : }
2826 0 : _ = self_ref.cancel.cancelled() => {
2827 0 : return Err(CalculateLogicalSizeError::Cancelled);
2828 : }
2829 0 : _ = cancel.cancelled() => {
2830 0 : return Err(CalculateLogicalSizeError::Cancelled);
2831 : },
2832 0 : () = skip_concurrency_limiter.cancelled() => {
2833 : // Some action that is part of a end user interaction requested logical size
2834 : // => break out of the rate limit
2835 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2836 : // but then again what happens if they cancel; also, we should just be using
2837 : // one runtime across the entire process, so, let's leave this for now.
2838 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2839 : }
2840 : };
2841 :
2842 0 : let metrics_guard = if attempt == 1 {
2843 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2844 : } else {
2845 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2846 : };
2847 :
2848 0 : let calculated_size = self_ref
2849 0 : .logical_size_calculation_task(
2850 0 : initial_part_end,
2851 0 : LogicalSizeCalculationCause::Initial,
2852 0 : background_ctx,
2853 0 : )
2854 0 : .await?;
2855 :
2856 0 : self_ref
2857 0 : .trigger_aux_file_size_computation(initial_part_end, background_ctx)
2858 0 : .await?;
2859 :
2860 : // TODO: add aux file size to logical size
2861 :
2862 0 : Ok((calculated_size, metrics_guard))
2863 0 : }
2864 0 : };
2865 :
2866 0 : let retrying = async {
2867 0 : let mut attempt = 0;
2868 : loop {
2869 0 : attempt += 1;
2870 0 :
2871 0 : match try_once(attempt).await {
2872 0 : Ok(res) => return ControlFlow::Continue(res),
2873 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
2874 : Err(
2875 0 : e @ (CalculateLogicalSizeError::Decode(_)
2876 0 : | CalculateLogicalSizeError::PageRead(_)),
2877 0 : ) => {
2878 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2879 : // exponential back-off doesn't make sense at these long intervals;
2880 : // use fixed retry interval with generous jitter instead
2881 0 : let sleep_duration = Duration::from_secs(
2882 0 : u64::try_from(
2883 0 : // 1hour base
2884 0 : (60_i64 * 60_i64)
2885 0 : // 10min jitter
2886 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2887 0 : )
2888 0 : .expect("10min < 1hour"),
2889 0 : );
2890 0 : tokio::time::sleep(sleep_duration).await;
2891 : }
2892 : }
2893 : }
2894 0 : };
2895 :
2896 0 : let (calculated_size, metrics_guard) = tokio::select! {
2897 0 : res = retrying => {
2898 0 : match res {
2899 0 : ControlFlow::Continue(calculated_size) => calculated_size,
2900 0 : ControlFlow::Break(()) => return,
2901 : }
2902 : }
2903 0 : _ = cancel.cancelled() => {
2904 0 : return;
2905 : }
2906 : };
2907 :
2908 : // we cannot query current_logical_size.current_size() to know the current
2909 : // *negative* value, only truncated to u64.
2910 0 : let added = self
2911 0 : .current_logical_size
2912 0 : .size_added_after_initial
2913 0 : .load(AtomicOrdering::Relaxed);
2914 0 :
2915 0 : let sum = calculated_size.saturating_add_signed(added);
2916 0 :
2917 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2918 0 : self.metrics.current_logical_size_gauge.set(sum);
2919 0 :
2920 0 : self.current_logical_size
2921 0 : .initial_logical_size
2922 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
2923 0 : .ok()
2924 0 : .expect("only this task sets it");
2925 0 : }
2926 :
2927 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
2928 0 : self: &Arc<Self>,
2929 0 : lsn: Lsn,
2930 0 : cause: LogicalSizeCalculationCause,
2931 0 : ctx: RequestContext,
2932 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
2933 0 : let (sender, receiver) = oneshot::channel();
2934 0 : let self_clone = Arc::clone(self);
2935 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
2936 0 : // we should stop the size calculation work and return an error.
2937 0 : // That would require restructuring this function's API to
2938 0 : // return the result directly, instead of a Receiver for the result.
2939 0 : let ctx = ctx.detached_child(
2940 0 : TaskKind::OndemandLogicalSizeCalculation,
2941 0 : DownloadBehavior::Download,
2942 0 : );
2943 0 : task_mgr::spawn(
2944 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2945 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
2946 0 : self.tenant_shard_id,
2947 0 : Some(self.timeline_id),
2948 0 : "ondemand logical size calculation",
2949 0 : async move {
2950 0 : let res = self_clone
2951 0 : .logical_size_calculation_task(lsn, cause, &ctx)
2952 0 : .await;
2953 0 : let _ = sender.send(res).ok();
2954 0 : Ok(()) // Receiver is responsible for handling errors
2955 0 : }
2956 0 : .in_current_span(),
2957 0 : );
2958 0 : receiver
2959 0 : }
2960 :
2961 : /// # Cancel-Safety
2962 : ///
2963 : /// This method is cancellation-safe.
2964 0 : #[instrument(skip_all)]
2965 : async fn logical_size_calculation_task(
2966 : self: &Arc<Self>,
2967 : lsn: Lsn,
2968 : cause: LogicalSizeCalculationCause,
2969 : ctx: &RequestContext,
2970 : ) -> Result<u64, CalculateLogicalSizeError> {
2971 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
2972 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
2973 : // accurate relation sizes, and they do not emit consumption metrics.
2974 : debug_assert!(self.tenant_shard_id.is_shard_zero());
2975 :
2976 : let guard = self
2977 : .gate
2978 : .enter()
2979 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
2980 :
2981 : let self_calculation = Arc::clone(self);
2982 :
2983 0 : let mut calculation = pin!(async {
2984 0 : let ctx = ctx.attached_child();
2985 0 : self_calculation
2986 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
2987 0 : .await
2988 0 : });
2989 :
2990 : tokio::select! {
2991 : res = &mut calculation => { res }
2992 : _ = self.cancel.cancelled() => {
2993 : debug!("cancelling logical size calculation for timeline shutdown");
2994 : calculation.await
2995 : }
2996 : }
2997 : }
2998 :
2999 : /// Calculate the logical size of the database at the latest LSN.
3000 : ///
3001 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3002 : /// especially if we need to download remote layers.
3003 : ///
3004 : /// # Cancel-Safety
3005 : ///
3006 : /// This method is cancellation-safe.
3007 0 : async fn calculate_logical_size(
3008 0 : &self,
3009 0 : up_to_lsn: Lsn,
3010 0 : cause: LogicalSizeCalculationCause,
3011 0 : _guard: &GateGuard,
3012 0 : ctx: &RequestContext,
3013 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3014 0 : info!(
3015 0 : "Calculating logical size for timeline {} at {}",
3016 : self.timeline_id, up_to_lsn
3017 : );
3018 :
3019 0 : pausable_failpoint!("timeline-calculate-logical-size-pause");
3020 :
3021 : // See if we've already done the work for initial size calculation.
3022 : // This is a short-cut for timelines that are mostly unused.
3023 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3024 0 : return Ok(size);
3025 0 : }
3026 0 : let storage_time_metrics = match cause {
3027 : LogicalSizeCalculationCause::Initial
3028 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3029 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3030 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3031 0 : &self.metrics.imitate_logical_size_histo
3032 : }
3033 : };
3034 0 : let timer = storage_time_metrics.start_timer();
3035 0 : let logical_size = self
3036 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3037 0 : .await?;
3038 0 : debug!("calculated logical size: {logical_size}");
3039 0 : timer.stop_and_record();
3040 0 : Ok(logical_size)
3041 0 : }
3042 :
3043 : /// Update current logical size, adding `delta' to the old value.
3044 270570 : fn update_current_logical_size(&self, delta: i64) {
3045 270570 : let logical_size = &self.current_logical_size;
3046 270570 : logical_size.increment_size(delta);
3047 270570 :
3048 270570 : // Also set the value in the prometheus gauge. Note that
3049 270570 : // there is a race condition here: if this is is called by two
3050 270570 : // threads concurrently, the prometheus gauge might be set to
3051 270570 : // one value while current_logical_size is set to the
3052 270570 : // other.
3053 270570 : match logical_size.current_size() {
3054 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
3055 270570 : .metrics
3056 270570 : .current_logical_size_gauge
3057 270570 : .set(new_current_size.into()),
3058 0 : CurrentLogicalSize::Approximate(_) => {
3059 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3060 0 : // forth between the initial size calculation task.
3061 0 : }
3062 : }
3063 270570 : }
3064 :
3065 2834 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3066 2834 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3067 2834 : let aux_metric =
3068 2834 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3069 2834 :
3070 2834 : let sum_of_entries = self
3071 2834 : .directory_metrics
3072 2834 : .iter()
3073 19838 : .map(|v| v.load(AtomicOrdering::Relaxed))
3074 2834 : .sum();
3075 : // Set a high general threshold and a lower threshold for the auxiliary files,
3076 : // as we can have large numbers of relations in the db directory.
3077 : const SUM_THRESHOLD: u64 = 5000;
3078 : const AUX_THRESHOLD: u64 = 1000;
3079 2834 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3080 0 : self.metrics
3081 0 : .directory_entries_count_gauge
3082 0 : .set(sum_of_entries);
3083 2834 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3084 0 : metric.set(sum_of_entries);
3085 2834 : }
3086 2834 : }
3087 :
3088 0 : async fn find_layer(
3089 0 : &self,
3090 0 : layer_name: &LayerName,
3091 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3092 0 : let guard = self.layers.read().await;
3093 0 : let layer = guard
3094 0 : .layer_map()?
3095 0 : .iter_historic_layers()
3096 0 : .find(|l| &l.layer_name() == layer_name)
3097 0 : .map(|found| guard.get_from_desc(&found));
3098 0 : Ok(layer)
3099 0 : }
3100 :
3101 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3102 : /// indicating which layers are currently on-disk on the primary.
3103 : ///
3104 : /// None is returned if the Timeline is in a state where uploading a heatmap
3105 : /// doesn't make sense, such as shutting down or initializing. The caller
3106 : /// should treat this as a cue to simply skip doing any heatmap uploading
3107 : /// for this timeline.
3108 2 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3109 2 : if !self.is_active() {
3110 0 : return None;
3111 2 : }
3112 :
3113 2 : let guard = self.layers.read().await;
3114 :
3115 10 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3116 10 : match layer.visibility() {
3117 : LayerVisibilityHint::Visible => {
3118 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3119 8 : let last_activity_ts = layer.latest_activity();
3120 8 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
3121 : }
3122 : LayerVisibilityHint::Covered => {
3123 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3124 2 : None
3125 : }
3126 : }
3127 10 : });
3128 2 :
3129 2 : let mut layers = resident.collect::<Vec<_>>();
3130 2 :
3131 2 : // Sort layers in order of which to download first. For a large set of layers to download, we
3132 2 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3133 2 : // or hours later:
3134 2 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3135 2 : // only exist for a few minutes before being compacted into L1s.
3136 2 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3137 2 : // the layer is likely to be covered by an image layer during compaction.
3138 20 : layers.sort_by_key(|(desc, _meta, _atime)| {
3139 20 : std::cmp::Reverse((
3140 20 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3141 20 : desc.lsn_range.end,
3142 20 : ))
3143 20 : });
3144 2 :
3145 2 : let layers = layers
3146 2 : .into_iter()
3147 8 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3148 2 : .collect();
3149 2 :
3150 2 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3151 2 : }
3152 :
3153 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3154 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3155 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3156 0 : // branchpoint in the value in IndexPart::lineage
3157 0 : self.ancestor_lsn == lsn
3158 0 : || (self.ancestor_lsn == Lsn::INVALID
3159 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3160 0 : }
3161 : }
3162 :
3163 : impl Timeline {
3164 : #[allow(clippy::doc_lazy_continuation)]
3165 : /// Get the data needed to reconstruct all keys in the provided keyspace
3166 : ///
3167 : /// The algorithm is as follows:
3168 : /// 1. While some keys are still not done and there's a timeline to visit:
3169 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3170 : /// 2.1: Build the fringe for the current keyspace
3171 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3172 : /// intersects
3173 : /// 2.3. Pop the timeline from the fringe
3174 : /// 2.4. If the fringe is empty, go back to 1
3175 626591 : async fn get_vectored_reconstruct_data(
3176 626591 : &self,
3177 626591 : mut keyspace: KeySpace,
3178 626591 : request_lsn: Lsn,
3179 626591 : reconstruct_state: &mut ValuesReconstructState,
3180 626591 : ctx: &RequestContext,
3181 626591 : ) -> Result<(), GetVectoredError> {
3182 626591 : let mut timeline_owned: Arc<Timeline>;
3183 626591 : let mut timeline = self;
3184 626591 :
3185 626591 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3186 :
3187 626589 : let missing_keyspace = loop {
3188 853147 : if self.cancel.is_cancelled() {
3189 0 : return Err(GetVectoredError::Cancelled);
3190 853147 : }
3191 :
3192 : let TimelineVisitOutcome {
3193 853147 : completed_keyspace: completed,
3194 853147 : image_covered_keyspace,
3195 853147 : } = Self::get_vectored_reconstruct_data_timeline(
3196 853147 : timeline,
3197 853147 : keyspace.clone(),
3198 853147 : cont_lsn,
3199 853147 : reconstruct_state,
3200 853147 : &self.cancel,
3201 853147 : ctx,
3202 853147 : )
3203 192236 : .await?;
3204 :
3205 853147 : keyspace.remove_overlapping_with(&completed);
3206 853147 :
3207 853147 : // Do not descend into the ancestor timeline for aux files.
3208 853147 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3209 853147 : // stalling compaction.
3210 853147 : keyspace.remove_overlapping_with(&KeySpace {
3211 853147 : ranges: vec![NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE],
3212 853147 : });
3213 853147 :
3214 853147 : // Keyspace is fully retrieved
3215 853147 : if keyspace.is_empty() {
3216 626575 : break None;
3217 226572 : }
3218 :
3219 226572 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3220 : // Not fully retrieved but no ancestor timeline.
3221 14 : break Some(keyspace);
3222 : };
3223 :
3224 : // Now we see if there are keys covered by the image layer but does not exist in the
3225 : // image layer, which means that the key does not exist.
3226 :
3227 : // The block below will stop the vectored search if any of the keys encountered an image layer
3228 : // which did not contain a snapshot for said key. Since we have already removed all completed
3229 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3230 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3231 : // and stop the search as a result of that.
3232 226558 : let removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3233 226558 : if !removed.is_empty() {
3234 0 : break Some(removed);
3235 226558 : }
3236 226558 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3237 226558 : // keyspace.
3238 226558 :
3239 226558 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3240 226558 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3241 226558 : timeline_owned = timeline
3242 226558 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3243 2 : .await?;
3244 226556 : timeline = &*timeline_owned;
3245 : };
3246 :
3247 626589 : if let Some(missing_keyspace) = missing_keyspace {
3248 14 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3249 14 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3250 14 : shard: self
3251 14 : .shard_identity
3252 14 : .get_shard_number(&missing_keyspace.start().unwrap()),
3253 14 : cont_lsn,
3254 14 : request_lsn,
3255 14 : ancestor_lsn: Some(timeline.ancestor_lsn),
3256 14 : backtrace: None,
3257 14 : }));
3258 626575 : }
3259 626575 :
3260 626575 : Ok(())
3261 626591 : }
3262 :
3263 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3264 : ///
3265 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3266 : /// the current keyspace. The current keyspace of the search at any given timeline
3267 : /// is the original keyspace minus all the keys that have been completed minus
3268 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3269 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3270 : ///
3271 : /// This is basically a depth-first search visitor implementation where a vertex
3272 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3273 : ///
3274 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3275 : /// and get all the required reconstruct data from the layer in one go.
3276 : ///
3277 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3278 : /// decides how to deal with these two keyspaces.
3279 853147 : async fn get_vectored_reconstruct_data_timeline(
3280 853147 : timeline: &Timeline,
3281 853147 : keyspace: KeySpace,
3282 853147 : mut cont_lsn: Lsn,
3283 853147 : reconstruct_state: &mut ValuesReconstructState,
3284 853147 : cancel: &CancellationToken,
3285 853147 : ctx: &RequestContext,
3286 853147 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3287 853147 : let mut unmapped_keyspace = keyspace.clone();
3288 853147 : let mut fringe = LayerFringe::new();
3289 853147 :
3290 853147 : let mut completed_keyspace = KeySpace::default();
3291 853147 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3292 :
3293 : loop {
3294 1698575 : if cancel.is_cancelled() {
3295 0 : return Err(GetVectoredError::Cancelled);
3296 1698575 : }
3297 1698575 :
3298 1698575 : let (keys_done_last_step, keys_with_image_coverage) =
3299 1698575 : reconstruct_state.consume_done_keys();
3300 1698575 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3301 1698575 : completed_keyspace.merge(&keys_done_last_step);
3302 1698575 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3303 21808 : unmapped_keyspace
3304 21808 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3305 21808 : image_covered_keyspace.add_range(keys_with_image_coverage);
3306 1676767 : }
3307 :
3308 : // Do not descent any further if the last layer we visited
3309 : // completed all keys in the keyspace it inspected. This is not
3310 : // required for correctness, but avoids visiting extra layers
3311 : // which turns out to be a perf bottleneck in some cases.
3312 1698575 : if !unmapped_keyspace.is_empty() {
3313 1074244 : let guard = timeline.layers.read().await;
3314 1074244 : let layers = guard.layer_map()?;
3315 :
3316 1074244 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3317 913928 : let start_lsn = l.get_lsn_range().start;
3318 913928 : cont_lsn > start_lsn
3319 1074244 : });
3320 1074244 :
3321 1074244 : match in_memory_layer {
3322 606331 : Some(l) => {
3323 606331 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3324 606331 : fringe.update(
3325 606331 : ReadableLayer::InMemoryLayer(l),
3326 606331 : unmapped_keyspace.clone(),
3327 606331 : lsn_range,
3328 606331 : );
3329 606331 : }
3330 : None => {
3331 467935 : for range in unmapped_keyspace.ranges.iter() {
3332 467935 : let results = layers.range_search(range.clone(), cont_lsn);
3333 467935 :
3334 467935 : results
3335 467935 : .found
3336 467935 : .into_iter()
3337 467935 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3338 239111 : (
3339 239111 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3340 239111 : keyspace_accum.to_keyspace(),
3341 239111 : lsn_floor..cont_lsn,
3342 239111 : )
3343 467935 : })
3344 467935 : .for_each(|(layer, keyspace, lsn_range)| {
3345 239111 : fringe.update(layer, keyspace, lsn_range)
3346 467935 : });
3347 467935 : }
3348 : }
3349 : }
3350 :
3351 : // It's safe to drop the layer map lock after planning the next round of reads.
3352 : // The fringe keeps readable handles for the layers which are safe to read even
3353 : // if layers were compacted or flushed.
3354 : //
3355 : // The more interesting consideration is: "Why is the read algorithm still correct
3356 : // if the layer map changes while it is operating?". Doing a vectored read on a
3357 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3358 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3359 : // range at *a particular point in time*. It is fine for the answer to be different
3360 : // at two different time points.
3361 1074244 : drop(guard);
3362 624331 : }
3363 :
3364 1698575 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3365 845428 : let next_cont_lsn = lsn_range.start;
3366 845428 : layer_to_read
3367 845428 : .get_values_reconstruct_data(
3368 845428 : keyspace_to_read.clone(),
3369 845428 : lsn_range,
3370 845428 : reconstruct_state,
3371 845428 : ctx,
3372 845428 : )
3373 183078 : .await?;
3374 :
3375 845428 : unmapped_keyspace = keyspace_to_read;
3376 845428 : cont_lsn = next_cont_lsn;
3377 845428 :
3378 845428 : reconstruct_state.on_layer_visited(&layer_to_read);
3379 : } else {
3380 853147 : break;
3381 853147 : }
3382 853147 : }
3383 853147 :
3384 853147 : Ok(TimelineVisitOutcome {
3385 853147 : completed_keyspace,
3386 853147 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3387 853147 : })
3388 853147 : }
3389 :
3390 226558 : async fn get_ready_ancestor_timeline(
3391 226558 : &self,
3392 226558 : ancestor: &Arc<Timeline>,
3393 226558 : ctx: &RequestContext,
3394 226558 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3395 226558 : // It's possible that the ancestor timeline isn't active yet, or
3396 226558 : // is active but hasn't yet caught up to the branch point. Wait
3397 226558 : // for it.
3398 226558 : //
3399 226558 : // This cannot happen while the pageserver is running normally,
3400 226558 : // because you cannot create a branch from a point that isn't
3401 226558 : // present in the pageserver yet. However, we don't wait for the
3402 226558 : // branch point to be uploaded to cloud storage before creating
3403 226558 : // a branch. I.e., the branch LSN need not be remote consistent
3404 226558 : // for the branching operation to succeed.
3405 226558 : //
3406 226558 : // Hence, if we try to load a tenant in such a state where
3407 226558 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3408 226558 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3409 226558 : // then we will need to wait for the ancestor timeline to
3410 226558 : // re-stream WAL up to branch_lsn before we access it.
3411 226558 : //
3412 226558 : // How can a tenant get in such a state?
3413 226558 : // - ungraceful pageserver process exit
3414 226558 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3415 226558 : //
3416 226558 : // NB: this could be avoided by requiring
3417 226558 : // branch_lsn >= remote_consistent_lsn
3418 226558 : // during branch creation.
3419 226558 : match ancestor.wait_to_become_active(ctx).await {
3420 226556 : Ok(()) => {}
3421 : Err(TimelineState::Stopping) => {
3422 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3423 0 : return Err(GetReadyAncestorError::Cancelled);
3424 : }
3425 2 : Err(state) => {
3426 2 : return Err(GetReadyAncestorError::BadState {
3427 2 : timeline_id: ancestor.timeline_id,
3428 2 : state,
3429 2 : });
3430 : }
3431 : }
3432 226556 : ancestor
3433 226556 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3434 0 : .await
3435 226556 : .map_err(|e| match e {
3436 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3437 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3438 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3439 0 : timeline_id: ancestor.timeline_id,
3440 0 : state,
3441 0 : },
3442 226556 : })?;
3443 :
3444 226556 : Ok(ancestor.clone())
3445 226558 : }
3446 :
3447 151304 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3448 151304 : &self.shard_identity
3449 151304 : }
3450 :
3451 : #[inline(always)]
3452 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3453 0 : ShardTimelineId {
3454 0 : shard_index: ShardIndex {
3455 0 : shard_number: self.shard_identity.number,
3456 0 : shard_count: self.shard_identity.count,
3457 0 : },
3458 0 : timeline_id: self.timeline_id,
3459 0 : }
3460 0 : }
3461 :
3462 : /// Returns a non-frozen open in-memory layer for ingestion.
3463 : ///
3464 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3465 : /// this function without holding the mutex.
3466 1268 : async fn get_layer_for_write(
3467 1268 : &self,
3468 1268 : lsn: Lsn,
3469 1268 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3470 1268 : ctx: &RequestContext,
3471 1268 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3472 1268 : let mut guard = self.layers.write().await;
3473 1268 : let gate_guard = self.gate.enter().context("enter gate for inmem layer")?;
3474 :
3475 1268 : let last_record_lsn = self.get_last_record_lsn();
3476 1268 : ensure!(
3477 1268 : lsn > last_record_lsn,
3478 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3479 : lsn,
3480 : last_record_lsn,
3481 : );
3482 :
3483 1268 : let layer = guard
3484 1268 : .open_mut()?
3485 1268 : .get_layer_for_write(
3486 1268 : lsn,
3487 1268 : self.conf,
3488 1268 : self.timeline_id,
3489 1268 : self.tenant_shard_id,
3490 1268 : gate_guard,
3491 1268 : ctx,
3492 1268 : )
3493 719 : .await?;
3494 1268 : Ok(layer)
3495 1268 : }
3496 :
3497 5279064 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3498 5279064 : assert!(new_lsn.is_aligned());
3499 :
3500 5279064 : self.metrics.last_record_gauge.set(new_lsn.0 as i64);
3501 5279064 : self.last_record_lsn.advance(new_lsn);
3502 5279064 : }
3503 :
3504 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3505 : ///
3506 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3507 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3508 1172 : async fn freeze_inmem_layer_at(
3509 1172 : &self,
3510 1172 : at: Lsn,
3511 1172 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3512 1172 : ) -> Result<u64, FlushLayerError> {
3513 1172 : let frozen = {
3514 1172 : let mut guard = self.layers.write().await;
3515 1172 : guard
3516 1172 : .open_mut()?
3517 1172 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock)
3518 4 : .await
3519 : };
3520 :
3521 1172 : if frozen {
3522 1144 : let now = Instant::now();
3523 1144 : *(self.last_freeze_ts.write().unwrap()) = now;
3524 1144 : }
3525 :
3526 : // Increment the flush cycle counter and wake up the flush task.
3527 : // Remember the new value, so that when we listen for the flush
3528 : // to finish, we know when the flush that we initiated has
3529 : // finished, instead of some other flush that was started earlier.
3530 1172 : let mut my_flush_request = 0;
3531 1172 :
3532 1172 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3533 1172 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3534 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3535 1172 : }
3536 1172 :
3537 1172 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3538 1172 : my_flush_request = *counter + 1;
3539 1172 : *counter = my_flush_request;
3540 1172 : *lsn = std::cmp::max(at, *lsn);
3541 1172 : });
3542 1172 :
3543 1172 : assert_ne!(my_flush_request, 0);
3544 :
3545 1172 : Ok(my_flush_request)
3546 1172 : }
3547 :
3548 : /// Layer flusher task's main loop.
3549 412 : async fn flush_loop(
3550 412 : self: &Arc<Self>,
3551 412 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3552 412 : ctx: &RequestContext,
3553 412 : ) {
3554 412 : info!("started flush loop");
3555 : loop {
3556 1545 : tokio::select! {
3557 1545 : _ = self.cancel.cancelled() => {
3558 10 : info!("shutting down layer flush task due to Timeline::cancel");
3559 10 : break;
3560 : },
3561 1545 : _ = layer_flush_start_rx.changed() => {}
3562 1133 : }
3563 1133 : trace!("waking up");
3564 1133 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3565 1133 :
3566 1133 : // The highest LSN to which we flushed in the loop over frozen layers
3567 1133 : let mut flushed_to_lsn = Lsn(0);
3568 :
3569 1133 : let result = loop {
3570 2277 : if self.cancel.is_cancelled() {
3571 0 : info!("dropping out of flush loop for timeline shutdown");
3572 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3573 : // anyone waiting on that will respect self.cancel as well: they will stop
3574 : // waiting at the same time we as drop out of this loop.
3575 0 : return;
3576 2277 : }
3577 2277 :
3578 2277 : let timer = self.metrics.flush_time_histo.start_timer();
3579 :
3580 : let num_frozen_layers;
3581 : let frozen_layer_total_size;
3582 2277 : let layer_to_flush = {
3583 2277 : let guard = self.layers.read().await;
3584 2277 : let Ok(lm) = guard.layer_map() else {
3585 0 : info!("dropping out of flush loop for timeline shutdown");
3586 0 : return;
3587 : };
3588 2277 : num_frozen_layers = lm.frozen_layers.len();
3589 2277 : frozen_layer_total_size = lm
3590 2277 : .frozen_layers
3591 2277 : .iter()
3592 2277 : .map(|l| l.estimated_in_mem_size())
3593 2277 : .sum::<u64>();
3594 2277 : lm.frozen_layers.front().cloned()
3595 : // drop 'layers' lock to allow concurrent reads and writes
3596 : };
3597 2277 : let Some(layer_to_flush) = layer_to_flush else {
3598 1133 : break Ok(());
3599 : };
3600 1144 : if num_frozen_layers
3601 1144 : > std::cmp::max(
3602 1144 : self.get_compaction_threshold(),
3603 1144 : DEFAULT_COMPACTION_THRESHOLD,
3604 1144 : )
3605 0 : && frozen_layer_total_size >= /* 128 MB */ 128000000
3606 : {
3607 0 : tracing::warn!(
3608 0 : "too many frozen layers: {num_frozen_layers} layers with estimated in-mem size of {frozen_layer_total_size} bytes",
3609 : );
3610 1144 : }
3611 17085 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3612 1144 : Ok(this_layer_to_lsn) => {
3613 1144 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3614 1144 : }
3615 : Err(FlushLayerError::Cancelled) => {
3616 0 : info!("dropping out of flush loop for timeline shutdown");
3617 0 : return;
3618 : }
3619 0 : err @ Err(
3620 0 : FlushLayerError::NotRunning(_)
3621 0 : | FlushLayerError::Other(_)
3622 0 : | FlushLayerError::CreateImageLayersError(_),
3623 0 : ) => {
3624 0 : error!("could not flush frozen layer: {err:?}");
3625 0 : break err.map(|_| ());
3626 : }
3627 : }
3628 1144 : timer.stop_and_record();
3629 : };
3630 :
3631 : // Unsharded tenants should never advance their LSN beyond the end of the
3632 : // highest layer they write: such gaps between layer data and the frozen LSN
3633 : // are only legal on sharded tenants.
3634 1133 : debug_assert!(
3635 1133 : self.shard_identity.count.count() > 1
3636 1133 : || flushed_to_lsn >= frozen_to_lsn
3637 67 : || !flushed_to_lsn.is_valid()
3638 : );
3639 :
3640 1133 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3641 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3642 : // to us via layer_flush_start_rx, then advance it here.
3643 : //
3644 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3645 : // never encounter a gap in the wal.
3646 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3647 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3648 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3649 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3650 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3651 0 : }
3652 0 : }
3653 1133 : }
3654 :
3655 : // Notify any listeners that we're done
3656 1133 : let _ = self
3657 1133 : .layer_flush_done_tx
3658 1133 : .send_replace((flush_counter, result));
3659 : }
3660 10 : }
3661 :
3662 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
3663 1092 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3664 1092 : let mut rx = self.layer_flush_done_tx.subscribe();
3665 : loop {
3666 : {
3667 2183 : let (last_result_counter, last_result) = &*rx.borrow();
3668 2183 : if *last_result_counter >= request {
3669 1092 : if let Err(err) = last_result {
3670 : // We already logged the original error in
3671 : // flush_loop. We cannot propagate it to the caller
3672 : // here, because it might not be Cloneable
3673 0 : return Err(err.clone());
3674 : } else {
3675 1092 : return Ok(());
3676 : }
3677 1091 : }
3678 1091 : }
3679 1091 : trace!("waiting for flush to complete");
3680 1091 : tokio::select! {
3681 1091 : rx_e = rx.changed() => {
3682 1091 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3683 : },
3684 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3685 : // the notification from [`flush_loop`] that it completed.
3686 1091 : _ = self.cancel.cancelled() => {
3687 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3688 0 : return Ok(())
3689 : }
3690 : };
3691 1091 : trace!("done")
3692 : }
3693 1092 : }
3694 :
3695 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3696 : ///
3697 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3698 1144 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3699 : async fn flush_frozen_layer(
3700 : self: &Arc<Self>,
3701 : frozen_layer: Arc<InMemoryLayer>,
3702 : ctx: &RequestContext,
3703 : ) -> Result<Lsn, FlushLayerError> {
3704 : debug_assert_current_span_has_tenant_and_timeline_id();
3705 :
3706 : // As a special case, when we have just imported an image into the repository,
3707 : // instead of writing out a L0 delta layer, we directly write out image layer
3708 : // files instead. This is possible as long as *all* the data imported into the
3709 : // repository have the same LSN.
3710 : let lsn_range = frozen_layer.get_lsn_range();
3711 :
3712 : // Whether to directly create image layers for this flush, or flush them as delta layers
3713 : let create_image_layer =
3714 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3715 :
3716 : #[cfg(test)]
3717 : {
3718 : match &mut *self.flush_loop_state.lock().unwrap() {
3719 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3720 : panic!("flush loop not running")
3721 : }
3722 : FlushLoopState::Running {
3723 : expect_initdb_optimization,
3724 : initdb_optimization_count,
3725 : ..
3726 : } => {
3727 : if create_image_layer {
3728 : *initdb_optimization_count += 1;
3729 : } else {
3730 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3731 : }
3732 : }
3733 : }
3734 : }
3735 :
3736 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3737 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3738 : // require downloading anything during initial import.
3739 : let ((rel_partition, metadata_partition), _lsn) = self
3740 : .repartition(
3741 : self.initdb_lsn,
3742 : self.get_compaction_target_size(),
3743 : EnumSet::empty(),
3744 : ctx,
3745 : )
3746 : .await
3747 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
3748 :
3749 : if self.cancel.is_cancelled() {
3750 : return Err(FlushLayerError::Cancelled);
3751 : }
3752 :
3753 : let mut layers_to_upload = Vec::new();
3754 : layers_to_upload.extend(
3755 : self.create_image_layers(
3756 : &rel_partition,
3757 : self.initdb_lsn,
3758 : ImageLayerCreationMode::Initial,
3759 : ctx,
3760 : )
3761 : .await?,
3762 : );
3763 : if !metadata_partition.parts.is_empty() {
3764 : assert_eq!(
3765 : metadata_partition.parts.len(),
3766 : 1,
3767 : "currently sparse keyspace should only contain a single metadata keyspace"
3768 : );
3769 : layers_to_upload.extend(
3770 : self.create_image_layers(
3771 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
3772 : // every single key within the keyspace, and therefore, it's safe to force converting it
3773 : // into a dense keyspace before calling this function.
3774 : &metadata_partition.into_dense(),
3775 : self.initdb_lsn,
3776 : ImageLayerCreationMode::Initial,
3777 : ctx,
3778 : )
3779 : .await?,
3780 : );
3781 : }
3782 :
3783 : (layers_to_upload, None)
3784 : } else {
3785 : // Normal case, write out a L0 delta layer file.
3786 : // `create_delta_layer` will not modify the layer map.
3787 : // We will remove frozen layer and add delta layer in one atomic operation later.
3788 : let Some(layer) = self
3789 : .create_delta_layer(&frozen_layer, None, ctx)
3790 : .await
3791 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
3792 : else {
3793 : panic!("delta layer cannot be empty if no filter is applied");
3794 : };
3795 : (
3796 : // FIXME: even though we have a single image and single delta layer assumption
3797 : // we push them to vec
3798 : vec![layer.clone()],
3799 : Some(layer),
3800 : )
3801 : };
3802 :
3803 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
3804 :
3805 : if self.cancel.is_cancelled() {
3806 : return Err(FlushLayerError::Cancelled);
3807 : }
3808 :
3809 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
3810 :
3811 : // The new on-disk layers are now in the layer map. We can remove the
3812 : // in-memory layer from the map now. The flushed layer is stored in
3813 : // the mapping in `create_delta_layer`.
3814 : {
3815 : let mut guard = self.layers.write().await;
3816 :
3817 : guard.open_mut()?.finish_flush_l0_layer(
3818 : delta_layer_to_add.as_ref(),
3819 : &frozen_layer,
3820 : &self.metrics,
3821 : );
3822 :
3823 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
3824 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
3825 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
3826 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
3827 : }
3828 : // release lock on 'layers'
3829 : };
3830 :
3831 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
3832 : // This makes us refuse ingest until the new layers have been persisted to the remote.
3833 : self.remote_client
3834 : .wait_completion()
3835 : .await
3836 0 : .map_err(|e| match e {
3837 : WaitCompletionError::UploadQueueShutDownOrStopped
3838 : | WaitCompletionError::NotInitialized(
3839 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
3840 0 : ) => FlushLayerError::Cancelled,
3841 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
3842 0 : FlushLayerError::Other(anyhow!(e).into())
3843 : }
3844 0 : })?;
3845 :
3846 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
3847 : // a compaction can delete the file and then it won't be available for uploads any more.
3848 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
3849 : // race situation.
3850 : // See https://github.com/neondatabase/neon/issues/4526
3851 : pausable_failpoint!("flush-frozen-pausable");
3852 :
3853 : // This failpoint is used by another test case `test_pageserver_recovery`.
3854 : fail_point!("flush-frozen-exit");
3855 :
3856 : Ok(Lsn(lsn_range.end.0 - 1))
3857 : }
3858 :
3859 : /// Return true if the value changed
3860 : ///
3861 : /// This function must only be used from the layer flush task.
3862 1144 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
3863 1144 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
3864 1144 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
3865 1144 : new_value != old_value
3866 1144 : }
3867 :
3868 : /// Update metadata file
3869 1146 : fn schedule_uploads(
3870 1146 : &self,
3871 1146 : disk_consistent_lsn: Lsn,
3872 1146 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
3873 1146 : ) -> anyhow::Result<()> {
3874 1146 : // We can only save a valid 'prev_record_lsn' value on disk if we
3875 1146 : // flushed *all* in-memory changes to disk. We only track
3876 1146 : // 'prev_record_lsn' in memory for the latest processed record, so we
3877 1146 : // don't remember what the correct value that corresponds to some old
3878 1146 : // LSN is. But if we flush everything, then the value corresponding
3879 1146 : // current 'last_record_lsn' is correct and we can store it on disk.
3880 1146 : let RecordLsn {
3881 1146 : last: last_record_lsn,
3882 1146 : prev: prev_record_lsn,
3883 1146 : } = self.last_record_lsn.load();
3884 1146 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
3885 1066 : Some(prev_record_lsn)
3886 : } else {
3887 80 : None
3888 : };
3889 :
3890 1146 : let update = crate::tenant::metadata::MetadataUpdate::new(
3891 1146 : disk_consistent_lsn,
3892 1146 : ondisk_prev_record_lsn,
3893 1146 : *self.latest_gc_cutoff_lsn.read(),
3894 1146 : );
3895 1146 :
3896 1146 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
3897 0 : "{}",
3898 0 : x.unwrap()
3899 1146 : ));
3900 :
3901 2302 : for layer in layers_to_upload {
3902 1156 : self.remote_client.schedule_layer_file_upload(layer)?;
3903 : }
3904 1146 : self.remote_client
3905 1146 : .schedule_index_upload_for_metadata_update(&update)?;
3906 :
3907 1146 : Ok(())
3908 1146 : }
3909 :
3910 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
3911 0 : self.remote_client
3912 0 : .preserve_initdb_archive(
3913 0 : &self.tenant_shard_id.tenant_id,
3914 0 : &self.timeline_id,
3915 0 : &self.cancel,
3916 0 : )
3917 0 : .await
3918 0 : }
3919 :
3920 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
3921 : // in layer map immediately. The caller is responsible to put it into the layer map.
3922 968 : async fn create_delta_layer(
3923 968 : self: &Arc<Self>,
3924 968 : frozen_layer: &Arc<InMemoryLayer>,
3925 968 : key_range: Option<Range<Key>>,
3926 968 : ctx: &RequestContext,
3927 968 : ) -> anyhow::Result<Option<ResidentLayer>> {
3928 968 : let self_clone = Arc::clone(self);
3929 968 : let frozen_layer = Arc::clone(frozen_layer);
3930 968 : let ctx = ctx.attached_child();
3931 968 : let work = async move {
3932 968 : let Some((desc, path)) = frozen_layer
3933 968 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
3934 10244 : .await?
3935 : else {
3936 0 : return Ok(None);
3937 : };
3938 968 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
3939 :
3940 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
3941 : // We just need to fsync the directory in which these inodes are linked,
3942 : // which we know to be the timeline directory.
3943 : //
3944 : // We use fatal_err() below because the after write_to_disk returns with success,
3945 : // the in-memory state of the filesystem already has the layer file in its final place,
3946 : // and subsequent pageserver code could think it's durable while it really isn't.
3947 968 : let timeline_dir = VirtualFile::open(
3948 968 : &self_clone
3949 968 : .conf
3950 968 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
3951 968 : &ctx,
3952 968 : )
3953 487 : .await
3954 968 : .fatal_err("VirtualFile::open for timeline dir fsync");
3955 968 : timeline_dir
3956 968 : .sync_all()
3957 484 : .await
3958 968 : .fatal_err("VirtualFile::sync_all timeline dir");
3959 968 : anyhow::Ok(Some(new_delta))
3960 968 : };
3961 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
3962 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
3963 : use crate::virtual_file::io_engine::IoEngine;
3964 968 : match crate::virtual_file::io_engine::get() {
3965 0 : IoEngine::NotSet => panic!("io engine not set"),
3966 : IoEngine::StdFs => {
3967 484 : let span = tracing::info_span!("blocking");
3968 484 : tokio::task::spawn_blocking({
3969 484 : move || Handle::current().block_on(work.instrument(span))
3970 484 : })
3971 484 : .await
3972 484 : .context("spawn_blocking")
3973 484 : .and_then(|x| x)
3974 : }
3975 : #[cfg(target_os = "linux")]
3976 11210 : IoEngine::TokioEpollUring => work.await,
3977 : }
3978 968 : }
3979 :
3980 540 : async fn repartition(
3981 540 : &self,
3982 540 : lsn: Lsn,
3983 540 : partition_size: u64,
3984 540 : flags: EnumSet<CompactFlags>,
3985 540 : ctx: &RequestContext,
3986 540 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
3987 540 : let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
3988 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
3989 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
3990 : // and hence before the compaction task starts.
3991 0 : return Err(CompactionError::Other(anyhow!(
3992 0 : "repartition() called concurrently, this should not happen"
3993 0 : )));
3994 : };
3995 540 : let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
3996 540 : 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 540 : }
4001 540 :
4002 540 : let distance = lsn.0 - partition_lsn.0;
4003 540 : 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 292 : }
4017 :
4018 15815 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4019 292 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4020 292 : let sparse_partitioning = SparseKeyPartitioning {
4021 292 : parts: vec![sparse_ks],
4022 292 : }; // no partitioning for metadata keys for now
4023 292 : *partitioning_guard = ((dense_partitioning, sparse_partitioning), lsn);
4024 292 :
4025 292 : Ok((partitioning_guard.0.clone(), partitioning_guard.1))
4026 540 : }
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 202 : async fn create_image_layer_for_rel_blocks(
4084 202 : self: &Arc<Self>,
4085 202 : partition: &KeySpace,
4086 202 : mut image_layer_writer: ImageLayerWriter,
4087 202 : lsn: Lsn,
4088 202 : ctx: &RequestContext,
4089 202 : img_range: Range<Key>,
4090 202 : start: Key,
4091 202 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4092 202 : let mut wrote_keys = false;
4093 202 :
4094 202 : let mut key_request_accum = KeySpaceAccum::new();
4095 1344 : for range in &partition.ranges {
4096 1142 : let mut key = range.start;
4097 2472 : 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 1330 : 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 1330 : } else {
4108 1330 : key_request_accum.add_key(key);
4109 1330 : }
4110 :
4111 1330 : let last_key_in_range = key.next() == range.end;
4112 1330 : key = key.next();
4113 1330 :
4114 1330 : // Maybe flush `key_rest_accum`
4115 1330 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4116 1330 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4117 : {
4118 1142 : let results = self
4119 1142 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
4120 49 : .await?;
4121 :
4122 1142 : if self.cancel.is_cancelled() {
4123 0 : return Err(CreateImageLayersError::Cancelled);
4124 1142 : }
4125 :
4126 2472 : for (img_key, img) in results {
4127 1330 : let img = match img {
4128 1330 : 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 1469 : image_layer_writer.put_image(img_key, img, ctx).await?;
4156 1330 : wrote_keys = true;
4157 : }
4158 188 : }
4159 : }
4160 : }
4161 :
4162 202 : 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 407 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4166 202 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4167 202 : info!("created image layer for rel {}", image_layer.local_path());
4168 202 : Ok(ImageLayerCreationOutcome {
4169 202 : image: Some(image_layer),
4170 202 : next_start_key: img_range.end,
4171 202 : })
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 202 : }
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 192 : async fn create_image_layer_for_metadata_keys(
4190 192 : self: &Arc<Self>,
4191 192 : partition: &KeySpace,
4192 192 : mut image_layer_writer: ImageLayerWriter,
4193 192 : lsn: Lsn,
4194 192 : ctx: &RequestContext,
4195 192 : img_range: Range<Key>,
4196 192 : mode: ImageLayerCreationMode,
4197 192 : start: Key,
4198 192 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4199 192 : // Metadata keys image layer creation.
4200 192 : let mut reconstruct_state = ValuesReconstructState::default();
4201 192 : let begin = Instant::now();
4202 192 : let data = self
4203 192 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4204 1066 : .await?;
4205 192 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4206 192 : let mut new_data = BTreeMap::new();
4207 192 : let mut total_kb_retrieved = 0;
4208 192 : let mut total_keys_retrieved = 0;
4209 10204 : 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 192 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4216 192 : };
4217 192 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4218 192 : let elapsed = begin.elapsed();
4219 192 :
4220 192 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4221 192 : 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 192 : 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 190 : }
4231 190 : if self.cancel.is_cancelled() {
4232 0 : return Err(CreateImageLayersError::Cancelled);
4233 190 : }
4234 190 : let mut wrote_any_image = false;
4235 10202 : 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 10161 : image_layer_writer.put_image(k, v, ctx).await?;
4248 : }
4249 :
4250 190 : 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 24 : 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 178 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4269 178 : Ok(ImageLayerCreationOutcome {
4270 178 : image: None,
4271 178 : next_start_key: start,
4272 178 : })
4273 : }
4274 192 : }
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 716 : 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 716 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4288 716 : let distance = lsn
4289 716 : .checked_sub(last_checks_at)
4290 716 : .expect("Attempt to compact with LSN going backwards");
4291 716 : let min_distance =
4292 716 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4293 716 :
4294 716 : let distance_based_decision = distance.0 >= min_distance;
4295 716 :
4296 716 : let mut time_based_decision = false;
4297 716 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4298 716 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4299 614 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4300 : {
4301 0 : self.get_checkpoint_timeout()
4302 : } else {
4303 614 : Duration::from_secs(3600 * 48)
4304 : };
4305 :
4306 614 : time_based_decision = match *last_check_instant {
4307 438 : Some(last_check) => {
4308 438 : let elapsed = last_check.elapsed();
4309 438 : elapsed >= check_required_after
4310 : }
4311 176 : 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 716 : let decision = distance_based_decision || time_based_decision;
4319 :
4320 716 : if decision {
4321 178 : self.last_image_layer_creation_check_at.store(lsn);
4322 178 : *last_check_instant = Some(Instant::now());
4323 538 : }
4324 :
4325 716 : decision
4326 716 : }
4327 :
4328 716 : #[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 0 : #[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 72 : pub fn new_with_inferred_key_range(
4718 72 : lsn_range: Range<Lsn>,
4719 72 : data: Vec<(Key, Lsn, Value)>,
4720 72 : ) -> Self {
4721 196 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
4722 196 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
4723 72 : Self {
4724 72 : key_range: (*key_min)..(key_max.next()),
4725 72 : lsn_range,
4726 72 : data,
4727 72 : }
4728 72 : }
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 0 : .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 667427 : async fn reconstruct_value(
5287 667427 : &self,
5288 667427 : key: Key,
5289 667427 : request_lsn: Lsn,
5290 667427 : mut data: ValueReconstructState,
5291 667427 : ) -> Result<Bytes, PageReconstructError> {
5292 667427 : // Perform WAL redo if needed
5293 667427 : data.records.reverse();
5294 667427 :
5295 667427 : // If we have a page image, and no WAL, we're all set
5296 667427 : if data.records.is_empty() {
5297 667017 : if let Some((img_lsn, img)) = &data.img {
5298 667017 : 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 667017 : 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 410 : 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 410 : if data.img.is_some() {
5324 344 : 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 410 : let res = self
5334 410 : .walredo_mgr
5335 410 : .as_ref()
5336 410 : .context("timeline has no walredo manager")
5337 410 : .map_err(PageReconstructError::WalRedo)?
5338 410 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5339 0 : .await;
5340 410 : let img = match res {
5341 410 : 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 410 : Ok(img)
5350 : }
5351 : }
5352 667427 : }
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 1730 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5555 1730 : ShardIndex {
5556 1730 : shard_number: self.tenant_shard_id.shard_number,
5557 1730 : shard_count: self.tenant_shard_id.shard_count,
5558 1730 : }
5559 1730 : }
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 40 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5579 40 : self.last_record_lsn.advance(new_lsn);
5580 40 : }
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 52 : pub(super) async fn force_create_image_layer(
5594 52 : self: &Arc<Timeline>,
5595 52 : lsn: Lsn,
5596 52 : mut images: Vec<(Key, Bytes)>,
5597 52 : check_start_lsn: Option<Lsn>,
5598 52 : ctx: &RequestContext,
5599 52 : ) -> anyhow::Result<()> {
5600 52 : let last_record_lsn = self.get_last_record_lsn();
5601 52 : assert!(
5602 52 : lsn <= last_record_lsn,
5603 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5604 : );
5605 52 : if let Some(check_start_lsn) = check_start_lsn {
5606 52 : assert!(lsn >= check_start_lsn);
5607 0 : }
5608 126 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5609 52 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5610 52 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5611 52 : let mut image_layer_writer = ImageLayerWriter::new(
5612 52 : self.conf,
5613 52 : self.timeline_id,
5614 52 : self.tenant_shard_id,
5615 52 : &(min_key..end_key),
5616 52 : lsn,
5617 52 : ctx,
5618 52 : )
5619 26 : .await?;
5620 230 : for (key, img) in images {
5621 178 : image_layer_writer.put_image(key, img, ctx).await?;
5622 : }
5623 104 : let (desc, path) = image_layer_writer.finish(ctx).await?;
5624 52 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5625 52 : info!("force created image layer {}", image_layer.local_path());
5626 : {
5627 52 : let mut guard = self.layers.write().await;
5628 52 : guard.open_mut().unwrap().force_insert_layer(image_layer);
5629 52 : }
5630 52 :
5631 52 : Ok(())
5632 52 : }
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 74 : pub(super) async fn force_create_delta_layer(
5641 74 : self: &Arc<Timeline>,
5642 74 : mut deltas: DeltaLayerTestDesc,
5643 74 : check_start_lsn: Option<Lsn>,
5644 74 : ctx: &RequestContext,
5645 74 : ) -> anyhow::Result<()> {
5646 74 : let last_record_lsn = self.get_last_record_lsn();
5647 74 : deltas
5648 74 : .data
5649 124 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5650 74 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5651 74 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5652 272 : for (_, lsn, _) in &deltas.data {
5653 198 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5654 : }
5655 74 : assert!(
5656 74 : 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 74 : if let Some(check_start_lsn) = check_start_lsn {
5662 74 : assert!(deltas.lsn_range.start >= check_start_lsn);
5663 0 : }
5664 74 : let mut delta_layer_writer = DeltaLayerWriter::new(
5665 74 : self.conf,
5666 74 : self.timeline_id,
5667 74 : self.tenant_shard_id,
5668 74 : deltas.key_range.start,
5669 74 : deltas.lsn_range,
5670 74 : ctx,
5671 74 : )
5672 37 : .await?;
5673 272 : for (key, lsn, val) in deltas.data {
5674 198 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5675 : }
5676 185 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
5677 74 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5678 74 : info!("force created delta layer {}", delta_layer.local_path());
5679 : {
5680 74 : let mut guard = self.layers.write().await;
5681 74 : guard.open_mut().unwrap().force_insert_layer(delta_layer);
5682 74 : }
5683 74 :
5684 74 : Ok(())
5685 74 : }
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 13 : .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 12 : pub(crate) async fn inspect_historic_layers(
5720 12 : self: &Arc<Timeline>,
5721 12 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
5722 12 : let mut layers = Vec::new();
5723 12 : let guard = self.layers.read().await;
5724 70 : for layer in guard.layer_map()?.iter_historic_layers() {
5725 70 : layers.push(layer.key());
5726 70 : }
5727 12 : Ok(layers)
5728 12 : }
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 1268 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
5754 1268 : Self {
5755 1268 : open_layer,
5756 1268 : current_size,
5757 1268 : prev_lsn: None,
5758 1268 : max_lsn: None,
5759 1268 : cached_last_freeze_at: last_freeze_at,
5760 1268 : }
5761 1268 : }
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 4807192 : fn deref(&self) -> &Self::Target {
5777 4807192 : self.tl
5778 4807192 : }
5779 : }
5780 :
5781 : #[derive(PartialEq)]
5782 : enum OpenLayerAction {
5783 : Roll,
5784 : Open,
5785 : None,
5786 : }
5787 :
5788 : impl<'a> TimelineWriter<'a> {
5789 4804204 : async fn handle_open_layer_action(
5790 4804204 : &mut self,
5791 4804204 : at: Lsn,
5792 4804204 : action: OpenLayerAction,
5793 4804204 : ctx: &RequestContext,
5794 4804204 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
5795 4804204 : 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 1188 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
5802 : OpenLayerAction::None => {
5803 4802936 : assert!(self.write_guard.is_some());
5804 : }
5805 : }
5806 :
5807 4804204 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
5808 4804204 : }
5809 :
5810 1268 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
5811 1268 : let layer = self
5812 1268 : .tl
5813 1268 : .get_layer_for_write(at, &self.write_guard, ctx)
5814 719 : .await?;
5815 1268 : let initial_size = layer.size().await?;
5816 :
5817 1268 : let last_freeze_at = self.last_freeze_at.load();
5818 1268 : self.write_guard.replace(TimelineWriterState::new(
5819 1268 : layer,
5820 1268 : initial_size,
5821 1268 : last_freeze_at,
5822 1268 : ));
5823 1268 :
5824 1268 : Ok(())
5825 1268 : }
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 7 : .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 4804204 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
5845 4804204 : let state = &*self.write_guard;
5846 4804204 : let Some(state) = &state else {
5847 1188 : 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 4804204 : }
5892 :
5893 : /// Put a batch of keys at the specified Lsns.
5894 4804202 : pub(crate) async fn put_batch(
5895 4804202 : &mut self,
5896 4804202 : batch: SerializedValueBatch,
5897 4804202 : ctx: &RequestContext,
5898 4804202 : ) -> anyhow::Result<()> {
5899 4804202 : if batch.is_empty() {
5900 0 : return Ok(());
5901 4804202 : }
5902 4804202 :
5903 4804202 : let batch_max_lsn = batch.max_lsn;
5904 4804202 : let buf_size: u64 = batch.buffer_size() as u64;
5905 4804202 :
5906 4804202 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
5907 4804202 : let layer = self
5908 4804202 : .handle_open_layer_action(batch_max_lsn, action, ctx)
5909 726 : .await?;
5910 :
5911 4804202 : let res = layer.put_batch(batch, ctx).await;
5912 :
5913 4804202 : if res.is_ok() {
5914 4804202 : // Update the current size only when the entire write was ok.
5915 4804202 : // In case of failures, we may have had partial writes which
5916 4804202 : // render the size tracking out of sync. That's ok because
5917 4804202 : // the checkpoint distance should be significantly smaller
5918 4804202 : // than the S3 single shot upload limit of 5GiB.
5919 4804202 : let state = self.write_guard.as_mut().unwrap();
5920 4804202 :
5921 4804202 : state.current_size += buf_size;
5922 4804202 : state.prev_lsn = Some(batch_max_lsn);
5923 4804202 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
5924 4804202 : }
5925 :
5926 4804202 : res
5927 4804202 : }
5928 :
5929 : #[cfg(test)]
5930 : /// Test helper, for tests that would like to poke individual values without composing a batch
5931 4390154 : pub(crate) async fn put(
5932 4390154 : &mut self,
5933 4390154 : key: Key,
5934 4390154 : lsn: Lsn,
5935 4390154 : value: &Value,
5936 4390154 : ctx: &RequestContext,
5937 4390154 : ) -> anyhow::Result<()> {
5938 : use utils::bin_ser::BeSer;
5939 4390154 : if !key.is_valid_key_on_write_path() {
5940 0 : bail!(
5941 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
5942 0 : key
5943 0 : );
5944 4390154 : }
5945 4390154 : let val_ser_size = value.serialized_size().unwrap() as usize;
5946 4390154 : let batch = SerializedValueBatch::from_values(vec![(
5947 4390154 : key.to_compact(),
5948 4390154 : lsn,
5949 4390154 : val_ser_size,
5950 4390154 : value.clone(),
5951 4390154 : )]);
5952 4390154 :
5953 4390154 : self.put_batch(batch, ctx).await
5954 4390154 : }
5955 :
5956 2 : pub(crate) async fn delete_batch(
5957 2 : &mut self,
5958 2 : batch: &[(Range<Key>, Lsn)],
5959 2 : ctx: &RequestContext,
5960 2 : ) -> anyhow::Result<()> {
5961 2 : if let Some((_, lsn)) = batch.first() {
5962 2 : let action = self.get_open_layer_action(*lsn, 0);
5963 2 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
5964 2 : layer.put_tombstones(batch).await?;
5965 0 : }
5966 :
5967 2 : Ok(())
5968 2 : }
5969 :
5970 : /// Track the end of the latest digested WAL record.
5971 : /// Remember the (end of) last valid WAL record remembered in the timeline.
5972 : ///
5973 : /// Call this after you have finished writing all the WAL up to 'lsn'.
5974 : ///
5975 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
5976 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
5977 : /// the latest and can be read.
5978 5279064 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
5979 5279064 : self.tl.finish_write(new_lsn);
5980 5279064 : }
5981 :
5982 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
5983 270570 : self.tl.update_current_logical_size(delta)
5984 270570 : }
5985 : }
5986 :
5987 : // We need TimelineWriter to be send in upcoming conversion of
5988 : // Timeline::layers to tokio::sync::RwLock.
5989 : #[test]
5990 2 : fn is_send() {
5991 2 : fn _assert_send<T: Send>() {}
5992 2 : _assert_send::<TimelineWriter<'_>>();
5993 2 : }
5994 :
5995 : #[cfg(test)]
5996 : mod tests {
5997 : use pageserver_api::key::Key;
5998 : use pageserver_api::value::Value;
5999 : use utils::{id::TimelineId, lsn::Lsn};
6000 :
6001 : use crate::tenant::{
6002 : harness::{test_img, TenantHarness},
6003 : layer_map::LayerMap,
6004 : storage_layer::{Layer, LayerName},
6005 : timeline::{DeltaLayerTestDesc, EvictionError},
6006 : Timeline,
6007 : };
6008 :
6009 : #[tokio::test]
6010 2 : async fn test_heatmap_generation() {
6011 2 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
6012 2 :
6013 2 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6014 2 : Lsn(0x10)..Lsn(0x20),
6015 2 : vec![(
6016 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6017 2 : Lsn(0x11),
6018 2 : Value::Image(test_img("foo")),
6019 2 : )],
6020 2 : );
6021 2 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6022 2 : Lsn(0x10)..Lsn(0x20),
6023 2 : vec![(
6024 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6025 2 : Lsn(0x11),
6026 2 : Value::Image(test_img("foo")),
6027 2 : )],
6028 2 : );
6029 2 : let l0_delta = DeltaLayerTestDesc::new(
6030 2 : Lsn(0x20)..Lsn(0x30),
6031 2 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6032 2 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6033 2 : vec![(
6034 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6035 2 : Lsn(0x25),
6036 2 : Value::Image(test_img("foo")),
6037 2 : )],
6038 2 : );
6039 2 : let delta_layers = vec![
6040 2 : covered_delta.clone(),
6041 2 : visible_delta.clone(),
6042 2 : l0_delta.clone(),
6043 2 : ];
6044 2 :
6045 2 : let image_layer = (
6046 2 : Lsn(0x40),
6047 2 : vec![(
6048 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6049 2 : test_img("bar"),
6050 2 : )],
6051 2 : );
6052 2 : let image_layers = vec![image_layer];
6053 2 :
6054 20 : let (tenant, ctx) = harness.load().await;
6055 2 : let timeline = tenant
6056 2 : .create_test_timeline_with_layers(
6057 2 : TimelineId::generate(),
6058 2 : Lsn(0x10),
6059 2 : 14,
6060 2 : &ctx,
6061 2 : delta_layers,
6062 2 : image_layers,
6063 2 : Lsn(0x100),
6064 2 : )
6065 31 : .await
6066 2 : .unwrap();
6067 2 :
6068 2 : // Layer visibility is an input to heatmap generation, so refresh it first
6069 2 : timeline.update_layer_visibility().await.unwrap();
6070 2 :
6071 2 : let heatmap = timeline
6072 2 : .generate_heatmap()
6073 2 : .await
6074 2 : .expect("Infallible while timeline is not shut down");
6075 2 :
6076 2 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
6077 2 :
6078 2 : // L0 should come last
6079 2 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
6080 2 :
6081 2 : let mut last_lsn = Lsn::MAX;
6082 10 : for layer in heatmap.layers {
6083 2 : // Covered layer should be omitted
6084 8 : assert!(layer.name != covered_delta.layer_name());
6085 2 :
6086 8 : let layer_lsn = match &layer.name {
6087 4 : LayerName::Delta(d) => d.lsn_range.end,
6088 4 : LayerName::Image(i) => i.lsn,
6089 2 : };
6090 2 :
6091 2 : // Apart from L0s, newest Layers should come first
6092 8 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
6093 6 : assert!(layer_lsn <= last_lsn);
6094 6 : last_lsn = layer_lsn;
6095 2 : }
6096 2 : }
6097 2 : }
6098 :
6099 : #[tokio::test]
6100 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
6101 2 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6102 2 : .await
6103 2 : .unwrap();
6104 2 :
6105 20 : let (tenant, ctx) = harness.load().await;
6106 2 : let timeline = tenant
6107 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6108 6 : .await
6109 2 : .unwrap();
6110 2 :
6111 2 : let layer = find_some_layer(&timeline).await;
6112 2 : let layer = layer
6113 2 : .keep_resident()
6114 2 : .await
6115 2 : .expect("no download => no downloading errors")
6116 2 : .drop_eviction_guard();
6117 2 :
6118 2 : let forever = std::time::Duration::from_secs(120);
6119 2 :
6120 2 : let first = layer.evict_and_wait(forever);
6121 2 : let second = layer.evict_and_wait(forever);
6122 2 :
6123 2 : let (first, second) = tokio::join!(first, second);
6124 2 :
6125 2 : let res = layer.keep_resident().await;
6126 2 : assert!(res.is_none(), "{res:?}");
6127 2 :
6128 2 : match (first, second) {
6129 2 : (Ok(()), Ok(())) => {
6130 2 : // because there are no more timeline locks being taken on eviction path, we can
6131 2 : // witness all three outcomes here.
6132 2 : }
6133 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
6134 0 : // if one completes before the other, this is fine just as well.
6135 0 : }
6136 2 : other => unreachable!("unexpected {:?}", other),
6137 2 : }
6138 2 : }
6139 :
6140 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
6141 2 : let layers = timeline.layers.read().await;
6142 2 : let desc = layers
6143 2 : .layer_map()
6144 2 : .unwrap()
6145 2 : .iter_historic_layers()
6146 2 : .next()
6147 2 : .expect("must find one layer to evict");
6148 2 :
6149 2 : layers.get_from_desc(&desc)
6150 2 : }
6151 : }
|
