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 : mod init;
8 : pub mod layer_manager;
9 : pub(crate) mod logical_size;
10 : pub mod span;
11 : pub mod uninit;
12 : mod walreceiver;
13 :
14 : use anyhow::{anyhow, bail, ensure, Context, Result};
15 : use arc_swap::ArcSwap;
16 : use bytes::Bytes;
17 : use camino::Utf8Path;
18 : use chrono::{DateTime, Utc};
19 : use enumset::EnumSet;
20 : use fail::fail_point;
21 : use handle::ShardTimelineId;
22 : use once_cell::sync::Lazy;
23 : use pageserver_api::{
24 : key::{
25 : KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
26 : NON_INHERITED_SPARSE_RANGE,
27 : },
28 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
29 : models::{
30 : AtomicAuxFilePolicy, AuxFilePolicy, CompactionAlgorithm, CompactionAlgorithmSettings,
31 : DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy,
32 : InMemoryLayerInfo, LayerMapInfo, LsnLease, TimelineState,
33 : },
34 : reltag::BlockNumber,
35 : shard::{ShardIdentity, ShardNumber, TenantShardId},
36 : };
37 : use rand::Rng;
38 : use serde_with::serde_as;
39 : use storage_broker::BrokerClientChannel;
40 : use tokio::{
41 : runtime::Handle,
42 : sync::{oneshot, watch},
43 : };
44 : use tokio_util::sync::CancellationToken;
45 : use tracing::*;
46 : use utils::{
47 : bin_ser::BeSer,
48 : fs_ext, pausable_failpoint,
49 : sync::gate::{Gate, GateGuard},
50 : vec_map::VecMap,
51 : };
52 :
53 : use std::pin::pin;
54 : use std::sync::atomic::Ordering as AtomicOrdering;
55 : use std::sync::{Arc, Mutex, RwLock, Weak};
56 : use std::time::{Duration, Instant, SystemTime};
57 : use std::{
58 : array,
59 : collections::{BTreeMap, HashMap, HashSet},
60 : sync::atomic::AtomicU64,
61 : };
62 : use std::{cmp::min, ops::ControlFlow};
63 : use std::{
64 : collections::btree_map::Entry,
65 : ops::{Deref, Range},
66 : };
67 :
68 : use crate::{
69 : aux_file::AuxFileSizeEstimator,
70 : tenant::{
71 : config::defaults::DEFAULT_PITR_INTERVAL,
72 : layer_map::{LayerMap, SearchResult},
73 : metadata::TimelineMetadata,
74 : storage_layer::PersistentLayerDesc,
75 : },
76 : walredo,
77 : };
78 : use crate::{
79 : context::{DownloadBehavior, RequestContext},
80 : disk_usage_eviction_task::DiskUsageEvictionInfo,
81 : pgdatadir_mapping::CollectKeySpaceError,
82 : };
83 : use crate::{
84 : disk_usage_eviction_task::finite_f32,
85 : tenant::storage_layer::{
86 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
87 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
88 : ValuesReconstructState,
89 : },
90 : };
91 : use crate::{
92 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
93 : };
94 : use crate::{
95 : l0_flush::{self, L0FlushGlobalState},
96 : metrics::GetKind,
97 : };
98 : use crate::{
99 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
100 : };
101 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
102 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
103 : use crate::{
104 : pgdatadir_mapping::{AuxFilesDirectory, DirectoryKind},
105 : virtual_file::{MaybeFatalIo, VirtualFile},
106 : };
107 :
108 : use crate::config::PageServerConf;
109 : use crate::keyspace::{KeyPartitioning, KeySpace};
110 : use crate::metrics::TimelineMetrics;
111 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
112 : use crate::tenant::config::TenantConfOpt;
113 : use pageserver_api::reltag::RelTag;
114 : use pageserver_api::shard::ShardIndex;
115 :
116 : use postgres_connection::PgConnectionConfig;
117 : use postgres_ffi::to_pg_timestamp;
118 : use utils::{
119 : completion,
120 : generation::Generation,
121 : id::TimelineId,
122 : lsn::{AtomicLsn, Lsn, RecordLsn},
123 : seqwait::SeqWait,
124 : simple_rcu::{Rcu, RcuReadGuard},
125 : };
126 :
127 : use crate::repository::GcResult;
128 : use crate::repository::{Key, Value};
129 : use crate::task_mgr;
130 : use crate::task_mgr::TaskKind;
131 : use crate::ZERO_PAGE;
132 :
133 : use self::delete::DeleteTimelineFlow;
134 : pub(super) use self::eviction_task::EvictionTaskTenantState;
135 : use self::eviction_task::EvictionTaskTimelineState;
136 : use self::layer_manager::LayerManager;
137 : use self::logical_size::LogicalSize;
138 : use self::walreceiver::{WalReceiver, WalReceiverConf};
139 :
140 : use super::{config::TenantConf, storage_layer::LayerVisibilityHint, upload_queue::NotInitialized};
141 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
142 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
143 : use super::{
144 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
145 : storage_layer::ReadableLayer,
146 : };
147 : use super::{
148 : secondary::heatmap::{HeatMapLayer, HeatMapTimeline},
149 : GcError,
150 : };
151 :
152 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
153 : pub(crate) enum FlushLoopState {
154 : NotStarted,
155 : Running {
156 : #[cfg(test)]
157 : expect_initdb_optimization: bool,
158 : #[cfg(test)]
159 : initdb_optimization_count: usize,
160 : },
161 : Exited,
162 : }
163 :
164 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
165 : pub enum ImageLayerCreationMode {
166 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
167 : Try,
168 : /// Force creating the image layers if possible. For now, no image layers will be created
169 : /// for metadata keys. Used in compaction code path with force flag enabled.
170 : Force,
171 : /// Initial ingestion of the data, and no data should be dropped in this function. This
172 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
173 : /// code path.
174 : Initial,
175 : }
176 :
177 : impl std::fmt::Display for ImageLayerCreationMode {
178 530 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
179 530 : write!(f, "{:?}", self)
180 530 : }
181 : }
182 :
183 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
184 : /// Can be removed after all refactors are done.
185 28 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
186 28 : drop(rlock)
187 28 : }
188 :
189 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
190 : /// Can be removed after all refactors are done.
191 558 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
192 558 : drop(rlock)
193 558 : }
194 :
195 : /// The outward-facing resources required to build a Timeline
196 : pub struct TimelineResources {
197 : pub remote_client: RemoteTimelineClient,
198 : pub timeline_get_throttle: Arc<
199 : crate::tenant::throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>,
200 : >,
201 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
202 : }
203 :
204 : pub(crate) struct AuxFilesState {
205 : pub(crate) dir: Option<AuxFilesDirectory>,
206 : pub(crate) n_deltas: usize,
207 : }
208 :
209 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
210 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
211 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
212 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
213 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
214 : pub(crate) struct RelSizeCache {
215 : pub(crate) complete_as_of: Lsn,
216 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
217 : }
218 :
219 : pub struct Timeline {
220 : conf: &'static PageServerConf,
221 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
222 :
223 : myself: Weak<Self>,
224 :
225 : pub(crate) tenant_shard_id: TenantShardId,
226 : pub timeline_id: TimelineId,
227 :
228 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
229 : /// Never changes for the lifetime of this [`Timeline`] object.
230 : ///
231 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
232 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
233 : pub(crate) generation: Generation,
234 :
235 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
236 : /// to shards, and is constant through the lifetime of this Timeline.
237 : shard_identity: ShardIdentity,
238 :
239 : pub pg_version: u32,
240 :
241 : /// The tuple has two elements.
242 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
243 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
244 : ///
245 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
246 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
247 : ///
248 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
249 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
250 : /// `PersistentLayerDesc`'s.
251 : ///
252 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
253 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
254 : /// runtime, e.g., during page reconstruction.
255 : ///
256 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
257 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
258 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
259 :
260 : last_freeze_at: AtomicLsn,
261 : // Atomic would be more appropriate here.
262 : last_freeze_ts: RwLock<Instant>,
263 :
264 : pub(crate) standby_horizon: AtomicLsn,
265 :
266 : // WAL redo manager. `None` only for broken tenants.
267 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
268 :
269 : /// Remote storage client.
270 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
271 : pub remote_client: Arc<RemoteTimelineClient>,
272 :
273 : // What page versions do we hold in the repository? If we get a
274 : // request > last_record_lsn, we need to wait until we receive all
275 : // the WAL up to the request. The SeqWait provides functions for
276 : // that. TODO: If we get a request for an old LSN, such that the
277 : // versions have already been garbage collected away, we should
278 : // throw an error, but we don't track that currently.
279 : //
280 : // last_record_lsn.load().last points to the end of last processed WAL record.
281 : //
282 : // We also remember the starting point of the previous record in
283 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
284 : // first WAL record when the node is started up. But here, we just
285 : // keep track of it.
286 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
287 :
288 : // All WAL records have been processed and stored durably on files on
289 : // local disk, up to this LSN. On crash and restart, we need to re-process
290 : // the WAL starting from this point.
291 : //
292 : // Some later WAL records might have been processed and also flushed to disk
293 : // already, so don't be surprised to see some, but there's no guarantee on
294 : // them yet.
295 : disk_consistent_lsn: AtomicLsn,
296 :
297 : // Parent timeline that this timeline was branched from, and the LSN
298 : // of the branch point.
299 : ancestor_timeline: Option<Arc<Timeline>>,
300 : ancestor_lsn: Lsn,
301 :
302 : pub(super) metrics: TimelineMetrics,
303 :
304 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
305 : // in `crate::page_service` writes these metrics.
306 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
307 :
308 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
309 :
310 : /// Ensures layers aren't frozen by checkpointer between
311 : /// [`Timeline::get_layer_for_write`] and layer reads.
312 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
313 : /// Must always be acquired before the layer map/individual layer lock
314 : /// to avoid deadlock.
315 : ///
316 : /// The state is cleared upon freezing.
317 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
318 :
319 : /// Used to avoid multiple `flush_loop` tasks running
320 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
321 :
322 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
323 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
324 : /// The flush cycle counter is sent back on the layer_flush_done channel when
325 : /// the flush finishes. You can use that to wait for the flush to finish.
326 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
327 : /// read by whoever sends an update
328 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
329 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
330 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
331 :
332 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
333 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
334 :
335 : // List of child timelines and their branch points. This is needed to avoid
336 : // garbage collecting data that is still needed by the child timelines.
337 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
338 :
339 : // It may change across major versions so for simplicity
340 : // keep it after running initdb for a timeline.
341 : // It is needed in checks when we want to error on some operations
342 : // when they are requested for pre-initdb lsn.
343 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
344 : // though let's keep them both for better error visibility.
345 : pub initdb_lsn: Lsn,
346 :
347 : /// When did we last calculate the partitioning? Make it pub to test cases.
348 : pub(super) partitioning: tokio::sync::Mutex<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
349 :
350 : /// Configuration: how often should the partitioning be recalculated.
351 : repartition_threshold: u64,
352 :
353 : last_image_layer_creation_check_at: AtomicLsn,
354 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
355 :
356 : /// Current logical size of the "datadir", at the last LSN.
357 : current_logical_size: LogicalSize,
358 :
359 : /// Information about the last processed message by the WAL receiver,
360 : /// or None if WAL receiver has not received anything for this timeline
361 : /// yet.
362 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
363 : pub walreceiver: Mutex<Option<WalReceiver>>,
364 :
365 : /// Relation size cache
366 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
367 :
368 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
369 :
370 : state: watch::Sender<TimelineState>,
371 :
372 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
373 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
374 : pub delete_progress: Arc<tokio::sync::Mutex<DeleteTimelineFlow>>,
375 :
376 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
377 :
378 : /// Load or creation time information about the disk_consistent_lsn and when the loading
379 : /// happened. Used for consumption metrics.
380 : pub(crate) loaded_at: (Lsn, SystemTime),
381 :
382 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
383 : pub(crate) gate: Gate,
384 :
385 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
386 : /// to the timeline should drop out when this token fires.
387 : pub(crate) cancel: CancellationToken,
388 :
389 : /// Make sure we only have one running compaction at a time in tests.
390 : ///
391 : /// Must only be taken in two places:
392 : /// - [`Timeline::compact`] (this file)
393 : /// - [`delete::delete_local_timeline_directory`]
394 : ///
395 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
396 : compaction_lock: tokio::sync::Mutex<()>,
397 :
398 : /// Make sure we only have one running gc at a time.
399 : ///
400 : /// Must only be taken in two places:
401 : /// - [`Timeline::gc`] (this file)
402 : /// - [`delete::delete_local_timeline_directory`]
403 : ///
404 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
405 : gc_lock: tokio::sync::Mutex<()>,
406 :
407 : /// Cloned from [`super::Tenant::timeline_get_throttle`] on construction.
408 : timeline_get_throttle: Arc<
409 : crate::tenant::throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>,
410 : >,
411 :
412 : /// Keep aux directory cache to avoid it's reconstruction on each update
413 : pub(crate) aux_files: tokio::sync::Mutex<AuxFilesState>,
414 :
415 : /// Size estimator for aux file v2
416 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
417 :
418 : /// Indicate whether aux file v2 storage is enabled.
419 : pub(crate) last_aux_file_policy: AtomicAuxFilePolicy,
420 :
421 : /// Some test cases directly place keys into the timeline without actually modifying the directory
422 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
423 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
424 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
425 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
426 : #[cfg(test)]
427 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
428 :
429 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
430 :
431 : pub(crate) handles: handle::PerTimelineState<crate::page_service::TenantManagerTypes>,
432 : }
433 :
434 : pub struct WalReceiverInfo {
435 : pub wal_source_connconf: PgConnectionConfig,
436 : pub last_received_msg_lsn: Lsn,
437 : pub last_received_msg_ts: u128,
438 : }
439 :
440 : /// Information about how much history needs to be retained, needed by
441 : /// Garbage Collection.
442 : #[derive(Default)]
443 : pub(crate) struct GcInfo {
444 : /// Specific LSNs that are needed.
445 : ///
446 : /// Currently, this includes all points where child branches have
447 : /// been forked off from. In the future, could also include
448 : /// explicit user-defined snapshot points.
449 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId)>,
450 :
451 : /// The cutoff coordinates, which are combined by selecting the minimum.
452 : pub(crate) cutoffs: GcCutoffs,
453 :
454 : /// Leases granted to particular LSNs.
455 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
456 :
457 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
458 : pub(crate) within_ancestor_pitr: bool,
459 : }
460 :
461 : impl GcInfo {
462 226 : pub(crate) fn min_cutoff(&self) -> Lsn {
463 226 : self.cutoffs.select_min()
464 226 : }
465 :
466 228 : pub(super) fn insert_child(&mut self, child_id: TimelineId, child_lsn: Lsn) {
467 228 : self.retain_lsns.push((child_lsn, child_id));
468 228 : self.retain_lsns.sort_by_key(|i| i.0);
469 228 : }
470 :
471 2 : pub(super) fn remove_child(&mut self, child_id: TimelineId) {
472 2 : self.retain_lsns.retain(|i| i.1 != child_id);
473 2 : }
474 : }
475 :
476 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
477 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
478 : /// between time-based and space-based retention for observability and consumption metrics purposes.
479 : #[derive(Debug, Clone)]
480 : pub(crate) struct GcCutoffs {
481 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
482 : /// history we must keep to retain a specified number of bytes of WAL.
483 : pub(crate) space: Lsn,
484 :
485 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
486 : /// history we must keep to enable reading back at least the PITR interval duration.
487 : pub(crate) time: Lsn,
488 : }
489 :
490 : impl Default for GcCutoffs {
491 406 : fn default() -> Self {
492 406 : Self {
493 406 : space: Lsn::INVALID,
494 406 : time: Lsn::INVALID,
495 406 : }
496 406 : }
497 : }
498 :
499 : impl GcCutoffs {
500 246 : fn select_min(&self) -> Lsn {
501 246 : std::cmp::min(self.space, self.time)
502 246 : }
503 : }
504 :
505 : pub(crate) struct TimelineVisitOutcome {
506 : completed_keyspace: KeySpace,
507 : image_covered_keyspace: KeySpace,
508 : }
509 :
510 : /// An error happened in a get() operation.
511 2 : #[derive(thiserror::Error, Debug)]
512 : pub(crate) enum PageReconstructError {
513 : #[error(transparent)]
514 : Other(anyhow::Error),
515 :
516 : #[error("Ancestor LSN wait error: {0}")]
517 : AncestorLsnTimeout(WaitLsnError),
518 :
519 : #[error("timeline shutting down")]
520 : Cancelled,
521 :
522 : /// An error happened replaying WAL records
523 : #[error(transparent)]
524 : WalRedo(anyhow::Error),
525 :
526 : #[error("{0}")]
527 : MissingKey(MissingKeyError),
528 : }
529 :
530 : impl From<anyhow::Error> for PageReconstructError {
531 0 : fn from(value: anyhow::Error) -> Self {
532 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
533 0 : match value.downcast::<PageReconstructError>() {
534 0 : Ok(pre) => pre,
535 0 : Err(other) => PageReconstructError::Other(other),
536 : }
537 0 : }
538 : }
539 :
540 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
541 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
542 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
543 0 : }
544 : }
545 :
546 : impl From<layer_manager::Shutdown> for PageReconstructError {
547 0 : fn from(_: layer_manager::Shutdown) -> Self {
548 0 : PageReconstructError::Cancelled
549 0 : }
550 : }
551 :
552 : impl GetVectoredError {
553 : #[cfg(test)]
554 6 : pub(crate) fn is_missing_key_error(&self) -> bool {
555 6 : matches!(self, Self::MissingKey(_))
556 6 : }
557 : }
558 :
559 : impl From<layer_manager::Shutdown> for GetVectoredError {
560 0 : fn from(_: layer_manager::Shutdown) -> Self {
561 0 : GetVectoredError::Cancelled
562 0 : }
563 : }
564 :
565 : #[derive(thiserror::Error)]
566 : pub struct MissingKeyError {
567 : key: Key,
568 : shard: ShardNumber,
569 : cont_lsn: Lsn,
570 : request_lsn: Lsn,
571 : ancestor_lsn: Option<Lsn>,
572 : backtrace: Option<std::backtrace::Backtrace>,
573 : }
574 :
575 : impl std::fmt::Debug for MissingKeyError {
576 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
577 0 : write!(f, "{}", self)
578 0 : }
579 : }
580 :
581 : impl std::fmt::Display for MissingKeyError {
582 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
583 0 : write!(
584 0 : f,
585 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
586 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
587 0 : )?;
588 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
589 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
590 0 : }
591 :
592 0 : if let Some(ref backtrace) = self.backtrace {
593 0 : write!(f, "\n{}", backtrace)?;
594 0 : }
595 :
596 0 : Ok(())
597 0 : }
598 : }
599 :
600 : impl PageReconstructError {
601 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
602 0 : pub(crate) fn is_stopping(&self) -> bool {
603 0 : use PageReconstructError::*;
604 0 : match self {
605 0 : Cancelled => true,
606 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
607 : }
608 0 : }
609 : }
610 :
611 0 : #[derive(thiserror::Error, Debug)]
612 : pub(crate) enum CreateImageLayersError {
613 : #[error("timeline shutting down")]
614 : Cancelled,
615 :
616 : #[error("read failed")]
617 : GetVectoredError(#[source] GetVectoredError),
618 :
619 : #[error("reconstruction failed")]
620 : PageReconstructError(#[source] PageReconstructError),
621 :
622 : #[error(transparent)]
623 : Other(#[from] anyhow::Error),
624 : }
625 :
626 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
627 0 : fn from(_: layer_manager::Shutdown) -> Self {
628 0 : CreateImageLayersError::Cancelled
629 0 : }
630 : }
631 :
632 0 : #[derive(thiserror::Error, Debug, Clone)]
633 : pub(crate) enum FlushLayerError {
634 : /// Timeline cancellation token was cancelled
635 : #[error("timeline shutting down")]
636 : Cancelled,
637 :
638 : /// We tried to flush a layer while the Timeline is in an unexpected state
639 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
640 : NotRunning(FlushLoopState),
641 :
642 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
643 : // loop via a watch channel, where we can only borrow it.
644 : #[error("create image layers (shared)")]
645 : CreateImageLayersError(Arc<CreateImageLayersError>),
646 :
647 : #[error("other (shared)")]
648 : Other(#[from] Arc<anyhow::Error>),
649 : }
650 :
651 : impl FlushLayerError {
652 : // When crossing from generic anyhow errors to this error type, we explicitly check
653 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
654 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
655 0 : let cancelled = timeline.cancel.is_cancelled()
656 : // The upload queue might have been shut down before the official cancellation of the timeline.
657 0 : || err
658 0 : .downcast_ref::<NotInitialized>()
659 0 : .map(NotInitialized::is_stopping)
660 0 : .unwrap_or_default();
661 0 : if cancelled {
662 0 : Self::Cancelled
663 : } else {
664 0 : Self::Other(Arc::new(err))
665 : }
666 0 : }
667 : }
668 :
669 : impl From<layer_manager::Shutdown> for FlushLayerError {
670 0 : fn from(_: layer_manager::Shutdown) -> Self {
671 0 : FlushLayerError::Cancelled
672 0 : }
673 : }
674 :
675 0 : #[derive(thiserror::Error, Debug)]
676 : pub(crate) enum GetVectoredError {
677 : #[error("timeline shutting down")]
678 : Cancelled,
679 :
680 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
681 : Oversized(u64),
682 :
683 : #[error("requested at invalid LSN: {0}")]
684 : InvalidLsn(Lsn),
685 :
686 : #[error("requested key not found: {0}")]
687 : MissingKey(MissingKeyError),
688 :
689 : #[error("ancestry walk")]
690 : GetReadyAncestorError(#[source] GetReadyAncestorError),
691 :
692 : #[error(transparent)]
693 : Other(#[from] anyhow::Error),
694 : }
695 :
696 : impl From<GetReadyAncestorError> for GetVectoredError {
697 2 : fn from(value: GetReadyAncestorError) -> Self {
698 2 : use GetReadyAncestorError::*;
699 2 : match value {
700 0 : Cancelled => GetVectoredError::Cancelled,
701 : AncestorLsnTimeout(_) | BadState { .. } => {
702 2 : GetVectoredError::GetReadyAncestorError(value)
703 : }
704 : }
705 2 : }
706 : }
707 :
708 2 : #[derive(thiserror::Error, Debug)]
709 : pub(crate) enum GetReadyAncestorError {
710 : #[error("ancestor LSN wait error")]
711 : AncestorLsnTimeout(#[from] WaitLsnError),
712 :
713 : #[error("bad state on timeline {timeline_id}: {state:?}")]
714 : BadState {
715 : timeline_id: TimelineId,
716 : state: TimelineState,
717 : },
718 :
719 : #[error("cancelled")]
720 : Cancelled,
721 : }
722 :
723 : #[derive(Clone, Copy)]
724 : pub enum LogicalSizeCalculationCause {
725 : Initial,
726 : ConsumptionMetricsSyntheticSize,
727 : EvictionTaskImitation,
728 : TenantSizeHandler,
729 : }
730 :
731 : pub enum GetLogicalSizePriority {
732 : User,
733 : Background,
734 : }
735 :
736 0 : #[derive(enumset::EnumSetType)]
737 : pub(crate) enum CompactFlags {
738 : ForceRepartition,
739 : ForceImageLayerCreation,
740 : EnhancedGcBottomMostCompaction,
741 : DryRun,
742 : }
743 :
744 : impl std::fmt::Debug for Timeline {
745 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
746 0 : write!(f, "Timeline<{}>", self.timeline_id)
747 0 : }
748 : }
749 :
750 0 : #[derive(thiserror::Error, Debug)]
751 : pub(crate) enum WaitLsnError {
752 : // Called on a timeline which is shutting down
753 : #[error("Shutdown")]
754 : Shutdown,
755 :
756 : // Called on an timeline not in active state or shutting down
757 : #[error("Bad timeline state: {0:?}")]
758 : BadState(TimelineState),
759 :
760 : // Timeout expired while waiting for LSN to catch up with goal.
761 : #[error("{0}")]
762 : Timeout(String),
763 : }
764 :
765 : // The impls below achieve cancellation mapping for errors.
766 : // Perhaps there's a way of achieving this with less cruft.
767 :
768 : impl From<CreateImageLayersError> for CompactionError {
769 0 : fn from(e: CreateImageLayersError) -> Self {
770 0 : match e {
771 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
772 0 : CreateImageLayersError::Other(e) => {
773 0 : CompactionError::Other(e.context("create image layers"))
774 : }
775 0 : _ => CompactionError::Other(e.into()),
776 : }
777 0 : }
778 : }
779 :
780 : impl From<CreateImageLayersError> for FlushLayerError {
781 0 : fn from(e: CreateImageLayersError) -> Self {
782 0 : match e {
783 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
784 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
785 : }
786 0 : }
787 : }
788 :
789 : impl From<PageReconstructError> for CreateImageLayersError {
790 0 : fn from(e: PageReconstructError) -> Self {
791 0 : match e {
792 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
793 0 : _ => CreateImageLayersError::PageReconstructError(e),
794 : }
795 0 : }
796 : }
797 :
798 : impl From<GetVectoredError> for CreateImageLayersError {
799 0 : fn from(e: GetVectoredError) -> Self {
800 0 : match e {
801 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
802 0 : _ => CreateImageLayersError::GetVectoredError(e),
803 : }
804 0 : }
805 : }
806 :
807 : impl From<GetVectoredError> for PageReconstructError {
808 6 : fn from(e: GetVectoredError) -> Self {
809 6 : match e {
810 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
811 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
812 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
813 4 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
814 2 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
815 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
816 : }
817 6 : }
818 : }
819 :
820 : impl From<GetReadyAncestorError> for PageReconstructError {
821 2 : fn from(e: GetReadyAncestorError) -> Self {
822 2 : use GetReadyAncestorError::*;
823 2 : match e {
824 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
825 2 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
826 0 : Cancelled => PageReconstructError::Cancelled,
827 : }
828 2 : }
829 : }
830 :
831 : pub(crate) enum WaitLsnWaiter<'a> {
832 : Timeline(&'a Timeline),
833 : Tenant,
834 : PageService,
835 : }
836 :
837 : /// Argument to [`Timeline::shutdown`].
838 : #[derive(Debug, Clone, Copy)]
839 : pub(crate) enum ShutdownMode {
840 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
841 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
842 : ///
843 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
844 : /// the call to [`Timeline::shutdown`].
845 : FreezeAndFlush,
846 : /// Shut down immediately, without waiting for any open layers to flush.
847 : Hard,
848 : }
849 :
850 : struct ImageLayerCreationOutcome {
851 : image: Option<ResidentLayer>,
852 : next_start_key: Key,
853 : }
854 :
855 : /// Public interface functions
856 : impl Timeline {
857 : /// Get the LSN where this branch was created
858 2 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
859 2 : self.ancestor_lsn
860 2 : }
861 :
862 : /// Get the ancestor's timeline id
863 754 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
864 754 : self.ancestor_timeline
865 754 : .as_ref()
866 754 : .map(|ancestor| ancestor.timeline_id)
867 754 : }
868 :
869 : /// Get the bytes written since the PITR cutoff on this branch, and
870 : /// whether this branch's ancestor_lsn is within its parent's PITR.
871 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
872 0 : let gc_info = self.gc_info.read().unwrap();
873 0 : let history = self
874 0 : .get_last_record_lsn()
875 0 : .checked_sub(gc_info.cutoffs.time)
876 0 : .unwrap_or(Lsn(0))
877 0 : .0;
878 0 : (history, gc_info.within_ancestor_pitr)
879 0 : }
880 :
881 : /// Lock and get timeline's GC cutoff
882 996 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
883 996 : self.latest_gc_cutoff_lsn.read()
884 996 : }
885 :
886 : /// Look up given page version.
887 : ///
888 : /// If a remote layer file is needed, it is downloaded as part of this
889 : /// call.
890 : ///
891 : /// This method enforces [`Self::timeline_get_throttle`] internally.
892 : ///
893 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
894 : /// abstraction above this needs to store suitable metadata to track what
895 : /// data exists with what keys, in separate metadata entries. If a
896 : /// non-existent key is requested, we may incorrectly return a value from
897 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
898 : /// non-existing key.
899 : ///
900 : /// # Cancel-Safety
901 : ///
902 : /// This method is cancellation-safe.
903 : #[inline(always)]
904 624961 : pub(crate) async fn get(
905 624961 : &self,
906 624961 : key: Key,
907 624961 : lsn: Lsn,
908 624961 : ctx: &RequestContext,
909 624961 : ) -> Result<Bytes, PageReconstructError> {
910 624961 : if !lsn.is_valid() {
911 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
912 624961 : }
913 624961 :
914 624961 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
915 624961 : // already checked the key against the shard_identity when looking up the Timeline from
916 624961 : // page_service.
917 624961 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
918 :
919 624961 : self.timeline_get_throttle.throttle(ctx, 1).await;
920 :
921 624961 : let keyspace = KeySpace {
922 624961 : ranges: vec![key..key.next()],
923 624961 : };
924 624961 :
925 624961 : // Initialise the reconstruct state for the key with the cache
926 624961 : // entry returned above.
927 624961 : let mut reconstruct_state = ValuesReconstructState::new();
928 :
929 624961 : let vectored_res = self
930 624961 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
931 102499 : .await;
932 :
933 624961 : let key_value = vectored_res?.pop_first();
934 624955 : match key_value {
935 624833 : Some((got_key, value)) => {
936 624833 : if got_key != key {
937 0 : error!(
938 0 : "Expected {}, but singular vectored get returned {}",
939 : key, got_key
940 : );
941 0 : Err(PageReconstructError::Other(anyhow!(
942 0 : "Singular vectored get returned wrong key"
943 0 : )))
944 : } else {
945 624833 : value
946 : }
947 : }
948 122 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
949 122 : key,
950 122 : shard: self.shard_identity.get_shard_number(&key),
951 122 : cont_lsn: Lsn(0),
952 122 : request_lsn: lsn,
953 122 : ancestor_lsn: None,
954 122 : backtrace: None,
955 122 : })),
956 : }
957 624961 : }
958 :
959 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
960 : pub(crate) const VEC_GET_LAYERS_VISITED_WARN_THRESH: f64 = 512.0;
961 :
962 : /// Look up multiple page versions at a given LSN
963 : ///
964 : /// This naive implementation will be replaced with a more efficient one
965 : /// which actually vectorizes the read path.
966 1084 : pub(crate) async fn get_vectored(
967 1084 : &self,
968 1084 : keyspace: KeySpace,
969 1084 : lsn: Lsn,
970 1084 : ctx: &RequestContext,
971 1084 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
972 1084 : if !lsn.is_valid() {
973 0 : return Err(GetVectoredError::InvalidLsn(lsn));
974 1084 : }
975 1084 :
976 1084 : let key_count = keyspace.total_raw_size().try_into().unwrap();
977 1084 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
978 0 : return Err(GetVectoredError::Oversized(key_count));
979 1084 : }
980 :
981 2168 : for range in &keyspace.ranges {
982 1084 : let mut key = range.start;
983 2522 : while key != range.end {
984 1438 : assert!(!self.shard_identity.is_key_disposable(&key));
985 1438 : key = key.next();
986 : }
987 : }
988 :
989 1084 : trace!(
990 0 : "get vectored request for {:?}@{} from task kind {:?}",
991 0 : keyspace,
992 0 : lsn,
993 0 : ctx.task_kind(),
994 : );
995 :
996 1084 : let start = crate::metrics::GET_VECTORED_LATENCY
997 1084 : .for_task_kind(ctx.task_kind())
998 1084 : .map(|metric| (metric, Instant::now()));
999 :
1000 : // start counting after throttle so that throttle time
1001 : // is always less than observation time
1002 1084 : let throttled = self
1003 1084 : .timeline_get_throttle
1004 1084 : .throttle(ctx, key_count as usize)
1005 0 : .await;
1006 :
1007 1084 : let res = self
1008 1084 : .get_vectored_impl(
1009 1084 : keyspace.clone(),
1010 1084 : lsn,
1011 1084 : &mut ValuesReconstructState::new(),
1012 1084 : ctx,
1013 1084 : )
1014 50 : .await;
1015 :
1016 1084 : if let Some((metric, start)) = start {
1017 0 : let elapsed = start.elapsed();
1018 0 : let ex_throttled = if let Some(throttled) = throttled {
1019 0 : elapsed.checked_sub(throttled)
1020 : } else {
1021 0 : Some(elapsed)
1022 : };
1023 :
1024 0 : if let Some(ex_throttled) = ex_throttled {
1025 0 : metric.observe(ex_throttled.as_secs_f64());
1026 0 : } else {
1027 0 : use utils::rate_limit::RateLimit;
1028 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1029 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
1030 0 : let mut rate_limit = LOGGED.lock().unwrap();
1031 0 : rate_limit.call(|| {
1032 0 : warn!("error deducting time spent throttled; this message is logged at a global rate limit");
1033 0 : });
1034 0 : }
1035 1084 : }
1036 :
1037 1084 : res
1038 1084 : }
1039 :
1040 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1041 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1042 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1043 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1044 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1045 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1046 : /// the scan operation will not cause OOM in the future.
1047 12 : pub(crate) async fn scan(
1048 12 : &self,
1049 12 : keyspace: KeySpace,
1050 12 : lsn: Lsn,
1051 12 : ctx: &RequestContext,
1052 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1053 12 : if !lsn.is_valid() {
1054 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1055 12 : }
1056 12 :
1057 12 : trace!(
1058 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1059 0 : keyspace,
1060 0 : lsn,
1061 0 : ctx.task_kind()
1062 : );
1063 :
1064 : // We should generalize this into Keyspace::contains in the future.
1065 24 : for range in &keyspace.ranges {
1066 12 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1067 12 : || range.end.field1 > METADATA_KEY_END_PREFIX
1068 : {
1069 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1070 0 : "only metadata keyspace can be scanned"
1071 0 : )));
1072 12 : }
1073 : }
1074 :
1075 12 : let start = crate::metrics::SCAN_LATENCY
1076 12 : .for_task_kind(ctx.task_kind())
1077 12 : .map(ScanLatencyOngoingRecording::start_recording);
1078 :
1079 : // start counting after throttle so that throttle time
1080 : // is always less than observation time
1081 12 : let throttled = self
1082 12 : .timeline_get_throttle
1083 12 : // assume scan = 1 quota for now until we find a better way to process this
1084 12 : .throttle(ctx, 1)
1085 0 : .await;
1086 :
1087 12 : let vectored_res = self
1088 12 : .get_vectored_impl(
1089 12 : keyspace.clone(),
1090 12 : lsn,
1091 12 : &mut ValuesReconstructState::default(),
1092 12 : ctx,
1093 12 : )
1094 0 : .await;
1095 :
1096 12 : if let Some(recording) = start {
1097 0 : recording.observe(throttled);
1098 12 : }
1099 :
1100 12 : vectored_res
1101 12 : }
1102 :
1103 626183 : pub(super) async fn get_vectored_impl(
1104 626183 : &self,
1105 626183 : keyspace: KeySpace,
1106 626183 : lsn: Lsn,
1107 626183 : reconstruct_state: &mut ValuesReconstructState,
1108 626183 : ctx: &RequestContext,
1109 626183 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1110 626183 : let get_kind = if keyspace.total_raw_size() == 1 {
1111 625935 : GetKind::Singular
1112 : } else {
1113 248 : GetKind::Vectored
1114 : };
1115 :
1116 626183 : let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
1117 626183 : .for_get_kind(get_kind)
1118 626183 : .start_timer();
1119 626183 : self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1120 114069 : .await?;
1121 626167 : get_data_timer.stop_and_record();
1122 626167 :
1123 626167 : let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
1124 626167 : .for_get_kind(get_kind)
1125 626167 : .start_timer();
1126 626167 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
1127 626167 : let layers_visited = reconstruct_state.get_layers_visited();
1128 :
1129 666707 : for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
1130 666707 : match res {
1131 0 : Err(err) => {
1132 0 : results.insert(key, Err(err));
1133 0 : }
1134 666707 : Ok(state) => {
1135 666707 : let state = ValueReconstructState::from(state);
1136 :
1137 666707 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
1138 666707 : results.insert(key, reconstruct_res);
1139 : }
1140 : }
1141 : }
1142 626167 : reconstruct_timer.stop_and_record();
1143 626167 :
1144 626167 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1145 626167 : // when they're missing. Instead they are omitted from the resulting btree
1146 626167 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1147 626167 : // to avoid infinite results.
1148 626167 : if !results.is_empty() {
1149 626027 : let avg = layers_visited as f64 / results.len() as f64;
1150 626027 : if avg >= Self::VEC_GET_LAYERS_VISITED_WARN_THRESH {
1151 0 : use utils::rate_limit::RateLimit;
1152 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1153 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1154 0 : let mut rate_limit = LOGGED.lock().unwrap();
1155 0 : rate_limit.call(|| {
1156 0 : tracing::info!(
1157 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1158 0 : lsn = %lsn,
1159 0 : "Vectored read for {} visited {} layers on average per key and {} in total. {}/{} pages were returned",
1160 0 : keyspace, avg, layers_visited, results.len(), keyspace.total_raw_size());
1161 0 : });
1162 626027 : }
1163 :
1164 : // Note that this is an approximation. Tracking the exact number of layers visited
1165 : // per key requires virtually unbounded memory usage and is inefficient
1166 : // (i.e. segment tree tracking each range queried from a layer)
1167 626027 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED.observe(avg);
1168 140 : }
1169 :
1170 626167 : Ok(results)
1171 626183 : }
1172 :
1173 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1174 276844 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1175 276844 : self.last_record_lsn.load().last
1176 276844 : }
1177 :
1178 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1179 0 : self.last_record_lsn.load().prev
1180 0 : }
1181 :
1182 : /// Atomically get both last and prev.
1183 226 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1184 226 : self.last_record_lsn.load()
1185 226 : }
1186 :
1187 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1188 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1189 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1190 0 : self.last_record_lsn.status_receiver()
1191 0 : }
1192 :
1193 1152 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1194 1152 : self.disk_consistent_lsn.load()
1195 1152 : }
1196 :
1197 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1198 : /// not validated with control plane yet.
1199 : /// See [`Self::get_remote_consistent_lsn_visible`].
1200 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1201 0 : self.remote_client.remote_consistent_lsn_projected()
1202 0 : }
1203 :
1204 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1205 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1206 : /// generation validation in the deletion queue.
1207 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1208 0 : self.remote_client.remote_consistent_lsn_visible()
1209 0 : }
1210 :
1211 : /// The sum of the file size of all historic layers in the layer map.
1212 : /// This method makes no distinction between local and remote layers.
1213 : /// Hence, the result **does not represent local filesystem usage**.
1214 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1215 0 : let guard = self.layers.read().await;
1216 0 : guard.layer_size_sum()
1217 0 : }
1218 :
1219 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1220 0 : self.metrics.resident_physical_size_get()
1221 0 : }
1222 :
1223 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1224 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1225 0 : }
1226 :
1227 : ///
1228 : /// Wait until WAL has been received and processed up to this LSN.
1229 : ///
1230 : /// You should call this before any of the other get_* or list_* functions. Calling
1231 : /// those functions with an LSN that has been processed yet is an error.
1232 : ///
1233 224949 : pub(crate) async fn wait_lsn(
1234 224949 : &self,
1235 224949 : lsn: Lsn,
1236 224949 : who_is_waiting: WaitLsnWaiter<'_>,
1237 224949 : ctx: &RequestContext, /* Prepare for use by cancellation */
1238 224949 : ) -> Result<(), WaitLsnError> {
1239 224949 : let state = self.current_state();
1240 224949 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1241 0 : return Err(WaitLsnError::Shutdown);
1242 224949 : } else if !matches!(state, TimelineState::Active) {
1243 0 : return Err(WaitLsnError::BadState(state));
1244 224949 : }
1245 224949 :
1246 224949 : if cfg!(debug_assertions) {
1247 224949 : match ctx.task_kind() {
1248 : TaskKind::WalReceiverManager
1249 : | TaskKind::WalReceiverConnectionHandler
1250 : | TaskKind::WalReceiverConnectionPoller => {
1251 0 : let is_myself = match who_is_waiting {
1252 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1253 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1254 : };
1255 0 : if is_myself {
1256 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1257 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1258 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1259 0 : }
1260 0 : } else {
1261 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1262 0 : // our walreceiver task can make progress independent of theirs
1263 0 : }
1264 : }
1265 224949 : _ => {}
1266 : }
1267 0 : }
1268 :
1269 224949 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1270 224949 :
1271 224949 : match self
1272 224949 : .last_record_lsn
1273 224949 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1274 0 : .await
1275 : {
1276 224949 : Ok(()) => Ok(()),
1277 0 : Err(e) => {
1278 0 : use utils::seqwait::SeqWaitError::*;
1279 0 : match e {
1280 0 : Shutdown => Err(WaitLsnError::Shutdown),
1281 : Timeout => {
1282 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1283 0 : drop(_timer);
1284 0 : let walreceiver_status = self.walreceiver_status();
1285 0 : Err(WaitLsnError::Timeout(format!(
1286 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1287 0 : lsn,
1288 0 : self.get_last_record_lsn(),
1289 0 : self.get_disk_consistent_lsn(),
1290 0 : walreceiver_status,
1291 0 : )))
1292 : }
1293 : }
1294 : }
1295 : }
1296 224949 : }
1297 :
1298 0 : pub(crate) fn walreceiver_status(&self) -> String {
1299 0 : match &*self.walreceiver.lock().unwrap() {
1300 0 : None => "stopping or stopped".to_string(),
1301 0 : Some(walreceiver) => match walreceiver.status() {
1302 0 : Some(status) => status.to_human_readable_string(),
1303 0 : None => "Not active".to_string(),
1304 : },
1305 : }
1306 0 : }
1307 :
1308 : /// Check that it is valid to request operations with that lsn.
1309 230 : pub(crate) fn check_lsn_is_in_scope(
1310 230 : &self,
1311 230 : lsn: Lsn,
1312 230 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1313 230 : ) -> anyhow::Result<()> {
1314 230 : ensure!(
1315 230 : lsn >= **latest_gc_cutoff_lsn,
1316 4 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1317 4 : lsn,
1318 4 : **latest_gc_cutoff_lsn,
1319 : );
1320 226 : Ok(())
1321 230 : }
1322 :
1323 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1324 : ///
1325 : /// This function will error if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is also
1326 : /// no existing lease to renew. If there is an existing lease in the map, the lease will be renewed only if
1327 : /// the request extends the lease. The returned lease is therefore the maximum between the existing lease and
1328 : /// the requesting lease.
1329 14 : pub(crate) fn make_lsn_lease(
1330 14 : &self,
1331 14 : lsn: Lsn,
1332 14 : length: Duration,
1333 14 : _ctx: &RequestContext,
1334 14 : ) -> anyhow::Result<LsnLease> {
1335 12 : let lease = {
1336 14 : let mut gc_info = self.gc_info.write().unwrap();
1337 14 :
1338 14 : let valid_until = SystemTime::now() + length;
1339 14 :
1340 14 : let entry = gc_info.leases.entry(lsn);
1341 :
1342 12 : let lease = {
1343 14 : if let Entry::Occupied(mut occupied) = entry {
1344 6 : let existing_lease = occupied.get_mut();
1345 6 : if valid_until > existing_lease.valid_until {
1346 2 : existing_lease.valid_until = valid_until;
1347 2 : let dt: DateTime<Utc> = valid_until.into();
1348 2 : info!("lease extended to {}", dt);
1349 : } else {
1350 4 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1351 4 : info!("existing lease covers greater length, valid until {}", dt);
1352 : }
1353 :
1354 6 : existing_lease.clone()
1355 : } else {
1356 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff)
1357 8 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1358 8 : if lsn < *latest_gc_cutoff_lsn {
1359 2 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1360 6 : }
1361 6 :
1362 6 : let dt: DateTime<Utc> = valid_until.into();
1363 6 : info!("lease created, valid until {}", dt);
1364 6 : entry.or_insert(LsnLease { valid_until }).clone()
1365 : }
1366 : };
1367 :
1368 12 : lease
1369 12 : };
1370 12 :
1371 12 : Ok(lease)
1372 14 : }
1373 :
1374 : /// Flush to disk all data that was written with the put_* functions
1375 2164 : #[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))]
1376 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1377 : self.freeze_and_flush0().await
1378 : }
1379 :
1380 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1381 : // polluting the span hierarchy.
1382 1082 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1383 1082 : let token = {
1384 : // Freeze the current open in-memory layer. It will be written to disk on next
1385 : // iteration.
1386 1082 : let mut g = self.write_lock.lock().await;
1387 :
1388 1082 : let to_lsn = self.get_last_record_lsn();
1389 1082 : self.freeze_inmem_layer_at(to_lsn, &mut g).await?
1390 : };
1391 1082 : self.wait_flush_completion(token).await
1392 1082 : }
1393 :
1394 : // Check if an open ephemeral layer should be closed: this provides
1395 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1396 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1397 : // ephemeral layer bytes has been breached.
1398 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1399 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1400 : // If the write lock is held, there is an active wal receiver: rolling open layers
1401 : // is their responsibility while they hold this lock.
1402 0 : return;
1403 : };
1404 :
1405 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1406 : // time, and this was missed.
1407 : // if write_guard.is_none() { return; }
1408 :
1409 0 : let Ok(layers_guard) = self.layers.try_read() else {
1410 : // Don't block if the layer lock is busy
1411 0 : return;
1412 : };
1413 :
1414 0 : let Ok(lm) = layers_guard.layer_map() else {
1415 0 : return;
1416 : };
1417 :
1418 0 : let Some(open_layer) = &lm.open_layer else {
1419 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1420 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1421 : // that didn't result in writes to this shard.
1422 :
1423 : // Must not hold the layers lock while waiting for a flush.
1424 0 : drop(layers_guard);
1425 0 :
1426 0 : let last_record_lsn = self.get_last_record_lsn();
1427 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1428 0 : if last_record_lsn > disk_consistent_lsn {
1429 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1430 : // we are a sharded tenant and have skipped some WAL
1431 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1432 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1433 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1434 : // without any data ingested (yet) doesn't write a remote index as soon as it
1435 : // sees its LSN advance: we only do this if we've been layer-less
1436 : // for some time.
1437 0 : tracing::debug!(
1438 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1439 : disk_consistent_lsn,
1440 : last_record_lsn
1441 : );
1442 :
1443 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1444 : // We know there is no open layer, so we can request freezing without actually
1445 : // freezing anything. This is true even if we have dropped the layers_guard, we
1446 : // still hold the write_guard.
1447 0 : let _ = async {
1448 0 : let token = self
1449 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1450 0 : .await?;
1451 0 : self.wait_flush_completion(token).await
1452 0 : }
1453 0 : .await;
1454 0 : }
1455 0 : }
1456 :
1457 0 : return;
1458 : };
1459 :
1460 0 : let Some(current_size) = open_layer.try_len() else {
1461 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1462 : // read lock to get size should always succeed.
1463 0 : tracing::warn!("Lock conflict while reading size of open layer");
1464 0 : return;
1465 : };
1466 :
1467 0 : let current_lsn = self.get_last_record_lsn();
1468 :
1469 0 : let checkpoint_distance_override = open_layer.tick().await;
1470 :
1471 0 : if let Some(size_override) = checkpoint_distance_override {
1472 0 : if current_size > size_override {
1473 : // This is not harmful, but it only happens in relatively rare cases where
1474 : // time-based checkpoints are not happening fast enough to keep the amount of
1475 : // ephemeral data within configured limits. It's a sign of stress on the system.
1476 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1477 0 : }
1478 0 : }
1479 :
1480 0 : let checkpoint_distance =
1481 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1482 0 :
1483 0 : if self.should_roll(
1484 0 : current_size,
1485 0 : current_size,
1486 0 : checkpoint_distance,
1487 0 : self.get_last_record_lsn(),
1488 0 : self.last_freeze_at.load(),
1489 0 : open_layer.get_opened_at(),
1490 0 : ) {
1491 0 : match open_layer.info() {
1492 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1493 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1494 0 : // happens asynchronously in the background.
1495 0 : tracing::debug!(
1496 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1497 : );
1498 : }
1499 : InMemoryLayerInfo::Open { .. } => {
1500 : // Upgrade to a write lock and freeze the layer
1501 0 : drop(layers_guard);
1502 0 : let res = self
1503 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1504 0 : .await;
1505 :
1506 0 : if let Err(e) = res {
1507 0 : tracing::info!(
1508 0 : "failed to flush frozen layer after background freeze: {e:#}"
1509 : );
1510 0 : }
1511 : }
1512 : }
1513 0 : }
1514 0 : }
1515 :
1516 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1517 : /// compaction tasks.
1518 364 : pub(crate) async fn compact(
1519 364 : self: &Arc<Self>,
1520 364 : cancel: &CancellationToken,
1521 364 : flags: EnumSet<CompactFlags>,
1522 364 : ctx: &RequestContext,
1523 364 : ) -> Result<bool, CompactionError> {
1524 364 : // most likely the cancellation token is from background task, but in tests it could be the
1525 364 : // request task as well.
1526 364 :
1527 364 : let prepare = async move {
1528 364 : let guard = self.compaction_lock.lock().await;
1529 :
1530 364 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1531 364 : BackgroundLoopKind::Compaction,
1532 364 : ctx,
1533 364 : )
1534 0 : .await;
1535 :
1536 364 : (guard, permit)
1537 364 : };
1538 :
1539 : // this wait probably never needs any "long time spent" logging, because we already nag if
1540 : // compaction task goes over it's period (20s) which is quite often in production.
1541 364 : let (_guard, _permit) = tokio::select! {
1542 : tuple = prepare => { tuple },
1543 : _ = self.cancel.cancelled() => return Ok(false),
1544 : _ = cancel.cancelled() => return Ok(false),
1545 : };
1546 :
1547 364 : let last_record_lsn = self.get_last_record_lsn();
1548 364 :
1549 364 : // Last record Lsn could be zero in case the timeline was just created
1550 364 : if !last_record_lsn.is_valid() {
1551 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1552 0 : return Ok(false);
1553 364 : }
1554 364 :
1555 364 : match self.get_compaction_algorithm_settings().kind {
1556 : CompactionAlgorithm::Tiered => {
1557 0 : self.compact_tiered(cancel, ctx).await?;
1558 0 : Ok(false)
1559 : }
1560 76754 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, flags, ctx).await,
1561 : }
1562 364 : }
1563 :
1564 : /// Mutate the timeline with a [`TimelineWriter`].
1565 5133168 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1566 5133168 : TimelineWriter {
1567 5133168 : tl: self,
1568 5133168 : write_guard: self.write_lock.lock().await,
1569 : }
1570 5133168 : }
1571 :
1572 0 : pub(crate) fn activate(
1573 0 : self: &Arc<Self>,
1574 0 : parent: Arc<crate::tenant::Tenant>,
1575 0 : broker_client: BrokerClientChannel,
1576 0 : background_jobs_can_start: Option<&completion::Barrier>,
1577 0 : ctx: &RequestContext,
1578 0 : ) {
1579 0 : if self.tenant_shard_id.is_shard_zero() {
1580 0 : // Logical size is only maintained accurately on shard zero.
1581 0 : self.spawn_initial_logical_size_computation_task(ctx);
1582 0 : }
1583 0 : self.launch_wal_receiver(ctx, broker_client);
1584 0 : self.set_state(TimelineState::Active);
1585 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1586 0 : }
1587 :
1588 : /// After this function returns, there are no timeline-scoped tasks are left running.
1589 : ///
1590 : /// The preferred pattern for is:
1591 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1592 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1593 : /// go the extra mile and keep track of JoinHandles
1594 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1595 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1596 : ///
1597 : /// For legacy reasons, we still have multiple tasks spawned using
1598 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1599 : /// We refer to these as "timeline-scoped task_mgr tasks".
1600 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1601 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1602 : /// or [`task_mgr::shutdown_watcher`].
1603 : /// We want to gradually convert the code base away from these.
1604 : ///
1605 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1606 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1607 : /// ones that aren't mentioned here):
1608 : /// - [`TaskKind::TimelineDeletionWorker`]
1609 : /// - NB: also used for tenant deletion
1610 : /// - [`TaskKind::RemoteUploadTask`]`
1611 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1612 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1613 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1614 : /// - [`TaskKind::Eviction`]
1615 : /// - [`TaskKind::LayerFlushTask`]
1616 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1617 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1618 8 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1619 8 : debug_assert_current_span_has_tenant_and_timeline_id();
1620 :
1621 8 : let try_freeze_and_flush = match mode {
1622 6 : ShutdownMode::FreezeAndFlush => true,
1623 2 : ShutdownMode::Hard => false,
1624 : };
1625 :
1626 : // Regardless of whether we're going to try_freeze_and_flush
1627 : // or not, stop ingesting any more data. Walreceiver only provides
1628 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1629 : // So, only after the self.gate.close() below will we know for sure that
1630 : // no walreceiver tasks are left.
1631 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1632 : // data during the call to `self.freeze_and_flush()` below.
1633 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1634 : // which is what we'd need to properly model early shutdown of the walreceiver
1635 : // task sub-tree before the other Timeline task sub-trees.
1636 8 : let walreceiver = self.walreceiver.lock().unwrap().take();
1637 8 : tracing::debug!(
1638 0 : is_some = walreceiver.is_some(),
1639 0 : "Waiting for WalReceiverManager..."
1640 : );
1641 8 : if let Some(walreceiver) = walreceiver {
1642 0 : walreceiver.cancel();
1643 8 : }
1644 : // ... and inform any waiters for newer LSNs that there won't be any.
1645 8 : self.last_record_lsn.shutdown();
1646 8 :
1647 8 : if try_freeze_and_flush {
1648 6 : if let Some((open, frozen)) = self
1649 6 : .layers
1650 6 : .read()
1651 0 : .await
1652 6 : .layer_map()
1653 6 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1654 6 : .ok()
1655 6 : .filter(|(open, frozen)| *open || *frozen > 0)
1656 : {
1657 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1658 6 : } else {
1659 6 : // this is double-shutdown, ignore it
1660 6 : }
1661 :
1662 : // we shut down walreceiver above, so, we won't add anything more
1663 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1664 : // to reach the disk & upload queue, then shut the upload queue and
1665 : // wait for it to drain.
1666 6 : match self.freeze_and_flush().await {
1667 : Ok(_) => {
1668 : // drain the upload queue
1669 : // if we did not wait for completion here, it might be our shutdown process
1670 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1671 : // be spawned.
1672 : //
1673 : // what is problematic is the shutting down of RemoteTimelineClient, because
1674 : // obviously it does not make sense to stop while we wait for it, but what
1675 : // about corner cases like s3 suddenly hanging up?
1676 6 : self.remote_client.shutdown().await;
1677 : }
1678 : Err(FlushLayerError::Cancelled) => {
1679 : // this is likely the second shutdown, ignore silently.
1680 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1681 0 : debug_assert!(self.cancel.is_cancelled());
1682 : }
1683 0 : Err(e) => {
1684 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1685 0 : // we have some extra WAL replay to do next time the timeline starts.
1686 0 : warn!("failed to freeze and flush: {e:#}");
1687 : }
1688 : }
1689 2 : }
1690 :
1691 : // Signal any subscribers to our cancellation token to drop out
1692 8 : tracing::debug!("Cancelling CancellationToken");
1693 8 : self.cancel.cancel();
1694 8 :
1695 8 : // Ensure Prevent new page service requests from starting.
1696 8 : self.handles.shutdown();
1697 8 :
1698 8 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1699 8 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1700 8 : self.remote_client.stop();
1701 8 :
1702 8 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1703 8 : // to shut down the upload queue tasks.
1704 8 : // TODO: fix that, task management should be encapsulated inside remote_client.
1705 8 : task_mgr::shutdown_tasks(
1706 8 : Some(TaskKind::RemoteUploadTask),
1707 8 : Some(self.tenant_shard_id),
1708 8 : Some(self.timeline_id),
1709 8 : )
1710 0 : .await;
1711 :
1712 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
1713 8 : tracing::debug!("Waiting for tasks...");
1714 8 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1715 :
1716 : {
1717 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
1718 : // open.
1719 8 : let mut write_guard = self.write_lock.lock().await;
1720 8 : self.layers.write().await.shutdown(&mut write_guard);
1721 8 : }
1722 8 :
1723 8 : // Finally wait until any gate-holders are complete.
1724 8 : //
1725 8 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1726 8 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1727 8 : self.gate.close().await;
1728 :
1729 8 : self.metrics.shutdown();
1730 8 : }
1731 :
1732 406 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1733 406 : match (self.current_state(), new_state) {
1734 406 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1735 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1736 : }
1737 0 : (st, TimelineState::Loading) => {
1738 0 : error!("ignoring transition from {st:?} into Loading state");
1739 : }
1740 0 : (TimelineState::Broken { .. }, new_state) => {
1741 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1742 : }
1743 : (TimelineState::Stopping, TimelineState::Active) => {
1744 0 : error!("Not activating a Stopping timeline");
1745 : }
1746 404 : (_, new_state) => {
1747 404 : self.state.send_replace(new_state);
1748 404 : }
1749 : }
1750 406 : }
1751 :
1752 2 : pub(crate) fn set_broken(&self, reason: String) {
1753 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1754 2 : let broken_state = TimelineState::Broken {
1755 2 : reason,
1756 2 : backtrace: backtrace_str,
1757 2 : };
1758 2 : self.set_state(broken_state);
1759 2 :
1760 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1761 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1762 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1763 2 : self.cancel.cancel();
1764 2 : }
1765 :
1766 226695 : pub(crate) fn current_state(&self) -> TimelineState {
1767 226695 : self.state.borrow().clone()
1768 226695 : }
1769 :
1770 6 : pub(crate) fn is_broken(&self) -> bool {
1771 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1772 6 : }
1773 :
1774 222 : pub(crate) fn is_active(&self) -> bool {
1775 222 : self.current_state() == TimelineState::Active
1776 222 : }
1777 :
1778 : #[allow(unused)]
1779 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
1780 0 : self.remote_client.is_archived()
1781 0 : }
1782 :
1783 1118 : pub(crate) fn is_stopping(&self) -> bool {
1784 1118 : self.current_state() == TimelineState::Stopping
1785 1118 : }
1786 :
1787 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1788 0 : self.state.subscribe()
1789 0 : }
1790 :
1791 224951 : pub(crate) async fn wait_to_become_active(
1792 224951 : &self,
1793 224951 : _ctx: &RequestContext, // Prepare for use by cancellation
1794 224951 : ) -> Result<(), TimelineState> {
1795 224951 : let mut receiver = self.state.subscribe();
1796 224951 : loop {
1797 224951 : let current_state = receiver.borrow().clone();
1798 224951 : match current_state {
1799 : TimelineState::Loading => {
1800 0 : receiver
1801 0 : .changed()
1802 0 : .await
1803 0 : .expect("holding a reference to self");
1804 : }
1805 : TimelineState::Active { .. } => {
1806 224949 : return Ok(());
1807 : }
1808 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1809 : // There's no chance the timeline can transition back into ::Active
1810 2 : return Err(current_state);
1811 : }
1812 : }
1813 : }
1814 224951 : }
1815 :
1816 0 : pub(crate) async fn layer_map_info(
1817 0 : &self,
1818 0 : reset: LayerAccessStatsReset,
1819 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
1820 0 : let guard = self.layers.read().await;
1821 0 : let layer_map = guard.layer_map()?;
1822 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
1823 0 : if let Some(open_layer) = &layer_map.open_layer {
1824 0 : in_memory_layers.push(open_layer.info());
1825 0 : }
1826 0 : for frozen_layer in &layer_map.frozen_layers {
1827 0 : in_memory_layers.push(frozen_layer.info());
1828 0 : }
1829 :
1830 0 : let historic_layers = layer_map
1831 0 : .iter_historic_layers()
1832 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
1833 0 : .collect();
1834 0 :
1835 0 : Ok(LayerMapInfo {
1836 0 : in_memory_layers,
1837 0 : historic_layers,
1838 0 : })
1839 0 : }
1840 :
1841 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))]
1842 : pub(crate) async fn download_layer(
1843 : &self,
1844 : layer_file_name: &LayerName,
1845 : ) -> anyhow::Result<Option<bool>> {
1846 : let Some(layer) = self.find_layer(layer_file_name).await? else {
1847 : return Ok(None);
1848 : };
1849 :
1850 : layer.download().await?;
1851 :
1852 : Ok(Some(true))
1853 : }
1854 :
1855 : /// Evict just one layer.
1856 : ///
1857 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
1858 0 : pub(crate) async fn evict_layer(
1859 0 : &self,
1860 0 : layer_file_name: &LayerName,
1861 0 : ) -> anyhow::Result<Option<bool>> {
1862 0 : let _gate = self
1863 0 : .gate
1864 0 : .enter()
1865 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
1866 :
1867 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
1868 0 : return Ok(None);
1869 : };
1870 :
1871 : // curl has this by default
1872 0 : let timeout = std::time::Duration::from_secs(120);
1873 0 :
1874 0 : match local_layer.evict_and_wait(timeout).await {
1875 0 : Ok(()) => Ok(Some(true)),
1876 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
1877 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
1878 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
1879 : }
1880 0 : }
1881 :
1882 4803026 : fn should_roll(
1883 4803026 : &self,
1884 4803026 : layer_size: u64,
1885 4803026 : projected_layer_size: u64,
1886 4803026 : checkpoint_distance: u64,
1887 4803026 : projected_lsn: Lsn,
1888 4803026 : last_freeze_at: Lsn,
1889 4803026 : opened_at: Instant,
1890 4803026 : ) -> bool {
1891 4803026 : let distance = projected_lsn.widening_sub(last_freeze_at);
1892 4803026 :
1893 4803026 : // Rolling the open layer can be triggered by:
1894 4803026 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
1895 4803026 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
1896 4803026 : // account for how writes are distributed across shards: we expect each node to consume
1897 4803026 : // 1/count of the LSN on average.
1898 4803026 : // 2. The size of the currently open layer.
1899 4803026 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
1900 4803026 : // up and suspend activity.
1901 4803026 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
1902 0 : info!(
1903 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
1904 : projected_lsn, layer_size, distance
1905 : );
1906 :
1907 0 : true
1908 4803026 : } else if projected_layer_size >= checkpoint_distance {
1909 80 : info!(
1910 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
1911 : projected_lsn, layer_size, projected_layer_size
1912 : );
1913 :
1914 80 : true
1915 4802946 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
1916 0 : info!(
1917 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
1918 0 : projected_lsn,
1919 0 : layer_size,
1920 0 : opened_at.elapsed()
1921 : );
1922 :
1923 0 : true
1924 : } else {
1925 4802946 : false
1926 : }
1927 4803026 : }
1928 : }
1929 :
1930 : /// Number of times we will compute partition within a checkpoint distance.
1931 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
1932 :
1933 : // Private functions
1934 : impl Timeline {
1935 12 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
1936 12 : let tenant_conf = self.tenant_conf.load();
1937 12 : tenant_conf
1938 12 : .tenant_conf
1939 12 : .lsn_lease_length
1940 12 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
1941 12 : }
1942 :
1943 : // TODO(yuchen): remove unused flag after implementing https://github.com/neondatabase/neon/issues/8072
1944 : #[allow(unused)]
1945 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
1946 0 : let tenant_conf = self.tenant_conf.load();
1947 0 : tenant_conf
1948 0 : .tenant_conf
1949 0 : .lsn_lease_length_for_ts
1950 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
1951 0 : }
1952 :
1953 224 : pub(crate) fn get_switch_aux_file_policy(&self) -> AuxFilePolicy {
1954 224 : let tenant_conf = self.tenant_conf.load();
1955 224 : tenant_conf
1956 224 : .tenant_conf
1957 224 : .switch_aux_file_policy
1958 224 : .unwrap_or(self.conf.default_tenant_conf.switch_aux_file_policy)
1959 224 : }
1960 :
1961 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
1962 0 : let tenant_conf = self.tenant_conf.load();
1963 0 : tenant_conf
1964 0 : .tenant_conf
1965 0 : .lazy_slru_download
1966 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
1967 0 : }
1968 :
1969 4804434 : fn get_checkpoint_distance(&self) -> u64 {
1970 4804434 : let tenant_conf = self.tenant_conf.load();
1971 4804434 : tenant_conf
1972 4804434 : .tenant_conf
1973 4804434 : .checkpoint_distance
1974 4804434 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
1975 4804434 : }
1976 :
1977 4802946 : fn get_checkpoint_timeout(&self) -> Duration {
1978 4802946 : let tenant_conf = self.tenant_conf.load();
1979 4802946 : tenant_conf
1980 4802946 : .tenant_conf
1981 4802946 : .checkpoint_timeout
1982 4802946 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
1983 4802946 : }
1984 :
1985 530 : fn get_compaction_target_size(&self) -> u64 {
1986 530 : let tenant_conf = self.tenant_conf.load();
1987 530 : tenant_conf
1988 530 : .tenant_conf
1989 530 : .compaction_target_size
1990 530 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
1991 530 : }
1992 :
1993 392 : fn get_compaction_threshold(&self) -> usize {
1994 392 : let tenant_conf = self.tenant_conf.load();
1995 392 : tenant_conf
1996 392 : .tenant_conf
1997 392 : .compaction_threshold
1998 392 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
1999 392 : }
2000 :
2001 14 : fn get_image_creation_threshold(&self) -> usize {
2002 14 : let tenant_conf = self.tenant_conf.load();
2003 14 : tenant_conf
2004 14 : .tenant_conf
2005 14 : .image_creation_threshold
2006 14 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2007 14 : }
2008 :
2009 364 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2010 364 : let tenant_conf = &self.tenant_conf.load();
2011 364 : tenant_conf
2012 364 : .tenant_conf
2013 364 : .compaction_algorithm
2014 364 : .as_ref()
2015 364 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2016 364 : .clone()
2017 364 : }
2018 :
2019 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2020 0 : let tenant_conf = self.tenant_conf.load();
2021 0 : tenant_conf
2022 0 : .tenant_conf
2023 0 : .eviction_policy
2024 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2025 0 : }
2026 :
2027 414 : fn get_evictions_low_residence_duration_metric_threshold(
2028 414 : tenant_conf: &TenantConfOpt,
2029 414 : default_tenant_conf: &TenantConf,
2030 414 : ) -> Duration {
2031 414 : tenant_conf
2032 414 : .evictions_low_residence_duration_metric_threshold
2033 414 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2034 414 : }
2035 :
2036 530 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2037 530 : let tenant_conf = self.tenant_conf.load();
2038 530 : tenant_conf
2039 530 : .tenant_conf
2040 530 : .image_layer_creation_check_threshold
2041 530 : .unwrap_or(
2042 530 : self.conf
2043 530 : .default_tenant_conf
2044 530 : .image_layer_creation_check_threshold,
2045 530 : )
2046 530 : }
2047 :
2048 8 : pub(super) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2049 8 : // NB: Most tenant conf options are read by background loops, so,
2050 8 : // changes will automatically be picked up.
2051 8 :
2052 8 : // The threshold is embedded in the metric. So, we need to update it.
2053 8 : {
2054 8 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2055 8 : new_conf,
2056 8 : &self.conf.default_tenant_conf,
2057 8 : );
2058 8 :
2059 8 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2060 8 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2061 8 :
2062 8 : let timeline_id_str = self.timeline_id.to_string();
2063 8 : self.metrics
2064 8 : .evictions_with_low_residence_duration
2065 8 : .write()
2066 8 : .unwrap()
2067 8 : .change_threshold(
2068 8 : &tenant_id_str,
2069 8 : &shard_id_str,
2070 8 : &timeline_id_str,
2071 8 : new_threshold,
2072 8 : );
2073 8 : }
2074 8 : }
2075 :
2076 : /// Open a Timeline handle.
2077 : ///
2078 : /// Loads the metadata for the timeline into memory, but not the layer map.
2079 : #[allow(clippy::too_many_arguments)]
2080 406 : pub(super) fn new(
2081 406 : conf: &'static PageServerConf,
2082 406 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2083 406 : metadata: &TimelineMetadata,
2084 406 : ancestor: Option<Arc<Timeline>>,
2085 406 : timeline_id: TimelineId,
2086 406 : tenant_shard_id: TenantShardId,
2087 406 : generation: Generation,
2088 406 : shard_identity: ShardIdentity,
2089 406 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2090 406 : resources: TimelineResources,
2091 406 : pg_version: u32,
2092 406 : state: TimelineState,
2093 406 : aux_file_policy: Option<AuxFilePolicy>,
2094 406 : cancel: CancellationToken,
2095 406 : ) -> Arc<Self> {
2096 406 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2097 406 : let (state, _) = watch::channel(state);
2098 406 :
2099 406 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2100 406 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2101 406 :
2102 406 : let evictions_low_residence_duration_metric_threshold = {
2103 406 : let loaded_tenant_conf = tenant_conf.load();
2104 406 : Self::get_evictions_low_residence_duration_metric_threshold(
2105 406 : &loaded_tenant_conf.tenant_conf,
2106 406 : &conf.default_tenant_conf,
2107 406 : )
2108 : };
2109 :
2110 406 : if let Some(ancestor) = &ancestor {
2111 228 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2112 228 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn());
2113 228 : }
2114 :
2115 406 : Arc::new_cyclic(|myself| {
2116 406 : let metrics = TimelineMetrics::new(
2117 406 : &tenant_shard_id,
2118 406 : &timeline_id,
2119 406 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2120 406 : "mtime",
2121 406 : evictions_low_residence_duration_metric_threshold,
2122 406 : ),
2123 406 : );
2124 406 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2125 :
2126 406 : let mut result = Timeline {
2127 406 : conf,
2128 406 : tenant_conf,
2129 406 : myself: myself.clone(),
2130 406 : timeline_id,
2131 406 : tenant_shard_id,
2132 406 : generation,
2133 406 : shard_identity,
2134 406 : pg_version,
2135 406 : layers: Default::default(),
2136 406 :
2137 406 : walredo_mgr,
2138 406 : walreceiver: Mutex::new(None),
2139 406 :
2140 406 : remote_client: Arc::new(resources.remote_client),
2141 406 :
2142 406 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2143 406 : last_record_lsn: SeqWait::new(RecordLsn {
2144 406 : last: disk_consistent_lsn,
2145 406 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2146 406 : }),
2147 406 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2148 406 :
2149 406 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2150 406 : last_freeze_ts: RwLock::new(Instant::now()),
2151 406 :
2152 406 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2153 406 :
2154 406 : ancestor_timeline: ancestor,
2155 406 : ancestor_lsn: metadata.ancestor_lsn(),
2156 406 :
2157 406 : metrics,
2158 406 :
2159 406 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2160 406 : &tenant_shard_id,
2161 406 : &timeline_id,
2162 406 : ),
2163 406 :
2164 2842 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2165 406 :
2166 406 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2167 406 :
2168 406 : layer_flush_start_tx,
2169 406 : layer_flush_done_tx,
2170 406 :
2171 406 : write_lock: tokio::sync::Mutex::new(None),
2172 406 :
2173 406 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2174 406 :
2175 406 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2176 406 : initdb_lsn: metadata.initdb_lsn(),
2177 406 :
2178 406 : current_logical_size: if disk_consistent_lsn.is_valid() {
2179 : // we're creating timeline data with some layer files existing locally,
2180 : // need to recalculate timeline's logical size based on data in the layers.
2181 232 : LogicalSize::deferred_initial(disk_consistent_lsn)
2182 : } else {
2183 : // we're creating timeline data without any layers existing locally,
2184 : // initial logical size is 0.
2185 174 : LogicalSize::empty_initial()
2186 : },
2187 406 : partitioning: tokio::sync::Mutex::new((
2188 406 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2189 406 : Lsn(0),
2190 406 : )),
2191 406 : repartition_threshold: 0,
2192 406 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2193 406 : last_image_layer_creation_check_instant: Mutex::new(None),
2194 406 :
2195 406 : last_received_wal: Mutex::new(None),
2196 406 : rel_size_cache: RwLock::new(RelSizeCache {
2197 406 : complete_as_of: disk_consistent_lsn,
2198 406 : map: HashMap::new(),
2199 406 : }),
2200 406 :
2201 406 : download_all_remote_layers_task_info: RwLock::new(None),
2202 406 :
2203 406 : state,
2204 406 :
2205 406 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2206 406 : EvictionTaskTimelineState::default(),
2207 406 : ),
2208 406 : delete_progress: Arc::new(tokio::sync::Mutex::new(DeleteTimelineFlow::default())),
2209 406 :
2210 406 : cancel,
2211 406 : gate: Gate::default(),
2212 406 :
2213 406 : compaction_lock: tokio::sync::Mutex::default(),
2214 406 : gc_lock: tokio::sync::Mutex::default(),
2215 406 :
2216 406 : standby_horizon: AtomicLsn::new(0),
2217 406 :
2218 406 : timeline_get_throttle: resources.timeline_get_throttle,
2219 406 :
2220 406 : aux_files: tokio::sync::Mutex::new(AuxFilesState {
2221 406 : dir: None,
2222 406 : n_deltas: 0,
2223 406 : }),
2224 406 :
2225 406 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2226 406 :
2227 406 : last_aux_file_policy: AtomicAuxFilePolicy::new(aux_file_policy),
2228 406 :
2229 406 : #[cfg(test)]
2230 406 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2231 406 :
2232 406 : l0_flush_global_state: resources.l0_flush_global_state,
2233 406 :
2234 406 : handles: Default::default(),
2235 406 : };
2236 406 : result.repartition_threshold =
2237 406 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2238 406 :
2239 406 : result
2240 406 : .metrics
2241 406 : .last_record_gauge
2242 406 : .set(disk_consistent_lsn.0 as i64);
2243 406 : result
2244 406 : })
2245 406 : }
2246 :
2247 566 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2248 566 : let Ok(guard) = self.gate.enter() else {
2249 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2250 0 : return;
2251 : };
2252 566 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2253 566 : match *flush_loop_state {
2254 400 : FlushLoopState::NotStarted => (),
2255 : FlushLoopState::Running { .. } => {
2256 166 : info!(
2257 0 : "skipping attempt to start flush_loop twice {}/{}",
2258 0 : self.tenant_shard_id, self.timeline_id
2259 : );
2260 166 : return;
2261 : }
2262 : FlushLoopState::Exited => {
2263 0 : warn!(
2264 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2265 0 : self.tenant_shard_id, self.timeline_id
2266 : );
2267 0 : return;
2268 : }
2269 : }
2270 :
2271 400 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2272 400 : let self_clone = Arc::clone(self);
2273 400 :
2274 400 : debug!("spawning flush loop");
2275 400 : *flush_loop_state = FlushLoopState::Running {
2276 400 : #[cfg(test)]
2277 400 : expect_initdb_optimization: false,
2278 400 : #[cfg(test)]
2279 400 : initdb_optimization_count: 0,
2280 400 : };
2281 400 : task_mgr::spawn(
2282 400 : task_mgr::BACKGROUND_RUNTIME.handle(),
2283 400 : task_mgr::TaskKind::LayerFlushTask,
2284 400 : self.tenant_shard_id,
2285 400 : Some(self.timeline_id),
2286 400 : "layer flush task",
2287 400 : async move {
2288 400 : let _guard = guard;
2289 400 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2290 18117 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2291 8 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2292 8 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2293 8 : *flush_loop_state = FlushLoopState::Exited;
2294 8 : Ok(())
2295 8 : }
2296 400 : .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))
2297 : );
2298 566 : }
2299 :
2300 : /// Creates and starts the wal receiver.
2301 : ///
2302 : /// This function is expected to be called at most once per Timeline's lifecycle
2303 : /// when the timeline is activated.
2304 0 : fn launch_wal_receiver(
2305 0 : self: &Arc<Self>,
2306 0 : ctx: &RequestContext,
2307 0 : broker_client: BrokerClientChannel,
2308 0 : ) {
2309 0 : info!(
2310 0 : "launching WAL receiver for timeline {} of tenant {}",
2311 0 : self.timeline_id, self.tenant_shard_id
2312 : );
2313 :
2314 0 : let tenant_conf = self.tenant_conf.load();
2315 0 : let wal_connect_timeout = tenant_conf
2316 0 : .tenant_conf
2317 0 : .walreceiver_connect_timeout
2318 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2319 0 : let lagging_wal_timeout = tenant_conf
2320 0 : .tenant_conf
2321 0 : .lagging_wal_timeout
2322 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2323 0 : let max_lsn_wal_lag = tenant_conf
2324 0 : .tenant_conf
2325 0 : .max_lsn_wal_lag
2326 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2327 0 :
2328 0 : let mut guard = self.walreceiver.lock().unwrap();
2329 0 : assert!(
2330 0 : guard.is_none(),
2331 0 : "multiple launches / re-launches of WAL receiver are not supported"
2332 : );
2333 0 : *guard = Some(WalReceiver::start(
2334 0 : Arc::clone(self),
2335 0 : WalReceiverConf {
2336 0 : wal_connect_timeout,
2337 0 : lagging_wal_timeout,
2338 0 : max_lsn_wal_lag,
2339 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2340 0 : availability_zone: self.conf.availability_zone.clone(),
2341 0 : ingest_batch_size: self.conf.ingest_batch_size,
2342 0 : },
2343 0 : broker_client,
2344 0 : ctx,
2345 0 : ));
2346 0 : }
2347 :
2348 : /// Initialize with an empty layer map. Used when creating a new timeline.
2349 400 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2350 400 : let mut layers = self.layers.try_write().expect(
2351 400 : "in the context where we call this function, no other task has access to the object",
2352 400 : );
2353 400 : layers
2354 400 : .open_mut()
2355 400 : .expect("in this context the LayerManager must still be open")
2356 400 : .initialize_empty(Lsn(start_lsn.0));
2357 400 : }
2358 :
2359 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2360 : /// files.
2361 6 : pub(super) async fn load_layer_map(
2362 6 : &self,
2363 6 : disk_consistent_lsn: Lsn,
2364 6 : index_part: Option<IndexPart>,
2365 6 : ) -> anyhow::Result<()> {
2366 : use init::{Decision::*, Discovered, DismissedLayer};
2367 : use LayerName::*;
2368 :
2369 6 : let mut guard = self.layers.write().await;
2370 :
2371 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2372 6 :
2373 6 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2374 6 : // structs representing all files on disk
2375 6 : let timeline_path = self
2376 6 : .conf
2377 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2378 6 : let conf = self.conf;
2379 6 : let span = tracing::Span::current();
2380 6 :
2381 6 : // Copy to move into the task we're about to spawn
2382 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2383 :
2384 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2385 6 : move || {
2386 6 : let _g = span.entered();
2387 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2388 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2389 6 : let mut unrecognized_files = Vec::new();
2390 6 :
2391 6 : let mut path = timeline_path;
2392 :
2393 22 : for discovered in discovered {
2394 16 : let (name, kind) = match discovered {
2395 16 : Discovered::Layer(layer_file_name, local_metadata) => {
2396 16 : discovered_layers.push((layer_file_name, local_metadata));
2397 16 : continue;
2398 : }
2399 0 : Discovered::IgnoredBackup(path) => {
2400 0 : std::fs::remove_file(path)
2401 0 : .or_else(fs_ext::ignore_not_found)
2402 0 : .fatal_err("Removing .old file");
2403 0 : continue;
2404 : }
2405 0 : Discovered::Unknown(file_name) => {
2406 0 : // we will later error if there are any
2407 0 : unrecognized_files.push(file_name);
2408 0 : continue;
2409 : }
2410 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2411 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2412 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2413 : };
2414 0 : path.push(Utf8Path::new(&name));
2415 0 : init::cleanup(&path, kind)?;
2416 0 : path.pop();
2417 : }
2418 :
2419 6 : if !unrecognized_files.is_empty() {
2420 : // assume that if there are any there are many many.
2421 0 : let n = unrecognized_files.len();
2422 0 : let first = &unrecognized_files[..n.min(10)];
2423 0 : anyhow::bail!(
2424 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2425 0 : );
2426 6 : }
2427 6 :
2428 6 : let decided =
2429 6 : init::reconcile(discovered_layers, index_part.as_ref(), disk_consistent_lsn);
2430 6 :
2431 6 : let mut loaded_layers = Vec::new();
2432 6 : let mut needs_cleanup = Vec::new();
2433 6 : let mut total_physical_size = 0;
2434 :
2435 22 : for (name, decision) in decided {
2436 16 : let decision = match decision {
2437 16 : Ok(decision) => decision,
2438 0 : Err(DismissedLayer::Future { local }) => {
2439 0 : if let Some(local) = local {
2440 0 : init::cleanup_future_layer(
2441 0 : &local.local_path,
2442 0 : &name,
2443 0 : disk_consistent_lsn,
2444 0 : )?;
2445 0 : }
2446 0 : needs_cleanup.push(name);
2447 0 : continue;
2448 : }
2449 0 : Err(DismissedLayer::LocalOnly(local)) => {
2450 0 : init::cleanup_local_only_file(&name, &local)?;
2451 : // this file never existed remotely, we will have to do rework
2452 0 : continue;
2453 : }
2454 0 : Err(DismissedLayer::BadMetadata(local)) => {
2455 0 : init::cleanup_local_file_for_remote(&local)?;
2456 : // this file never existed remotely, we will have to do rework
2457 0 : continue;
2458 : }
2459 : };
2460 :
2461 16 : match &name {
2462 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2463 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2464 : }
2465 :
2466 16 : tracing::debug!(layer=%name, ?decision, "applied");
2467 :
2468 16 : let layer = match decision {
2469 16 : Resident { local, remote } => {
2470 16 : total_physical_size += local.file_size;
2471 16 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2472 16 : .drop_eviction_guard()
2473 : }
2474 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2475 : };
2476 :
2477 16 : loaded_layers.push(layer);
2478 : }
2479 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2480 6 : }
2481 6 : })
2482 6 : .await
2483 6 : .map_err(anyhow::Error::new)
2484 6 : .and_then(|x| x)?;
2485 :
2486 6 : let num_layers = loaded_layers.len();
2487 6 :
2488 6 : guard
2489 6 : .open_mut()
2490 6 : .expect("layermanager must be open during init")
2491 6 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2492 6 :
2493 6 : self.remote_client
2494 6 : .schedule_layer_file_deletion(&needs_cleanup)?;
2495 6 : self.remote_client
2496 6 : .schedule_index_upload_for_file_changes()?;
2497 : // This barrier orders above DELETEs before any later operations.
2498 : // This is critical because code executing after the barrier might
2499 : // create again objects with the same key that we just scheduled for deletion.
2500 : // For example, if we just scheduled deletion of an image layer "from the future",
2501 : // later compaction might run again and re-create the same image layer.
2502 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2503 : // "same" here means same key range and LSN.
2504 : //
2505 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2506 : // the upload queue may execute the PUT first, then the DELETE.
2507 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2508 : //
2509 : // 1. a future image layer is created and uploaded
2510 : // 2. ps restart
2511 : // 3. the future layer from (1) is deleted during load layer map
2512 : // 4. image layer is re-created and uploaded
2513 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2514 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2515 : //
2516 : // See https://github.com/neondatabase/neon/issues/5878
2517 : //
2518 : // NB: generation numbers naturally protect against this because they disambiguate
2519 : // (1) and (4)
2520 6 : self.remote_client.schedule_barrier()?;
2521 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2522 : // on retry.
2523 :
2524 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2525 6 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2526 6 : self.update_layer_visibility().await?;
2527 :
2528 6 : info!(
2529 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2530 : num_layers, disk_consistent_lsn, total_physical_size
2531 : );
2532 :
2533 6 : timer.stop_and_record();
2534 6 : Ok(())
2535 6 : }
2536 :
2537 : /// Retrieve current logical size of the timeline.
2538 : ///
2539 : /// The size could be lagging behind the actual number, in case
2540 : /// the initial size calculation has not been run (gets triggered on the first size access).
2541 : ///
2542 : /// return size and boolean flag that shows if the size is exact
2543 0 : pub(crate) fn get_current_logical_size(
2544 0 : self: &Arc<Self>,
2545 0 : priority: GetLogicalSizePriority,
2546 0 : ctx: &RequestContext,
2547 0 : ) -> logical_size::CurrentLogicalSize {
2548 0 : if !self.tenant_shard_id.is_shard_zero() {
2549 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2550 : // when HTTP API is serving a GET for timeline zero, return zero
2551 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2552 0 : }
2553 0 :
2554 0 : let current_size = self.current_logical_size.current_size();
2555 0 : debug!("Current size: {current_size:?}");
2556 :
2557 0 : match (current_size.accuracy(), priority) {
2558 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2559 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2560 0 : // background task will eventually deliver an exact value, we're in no rush
2561 0 : }
2562 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2563 : // background task is not ready, but user is asking for it now;
2564 : // => make the background task skip the line
2565 : // (The alternative would be to calculate the size here, but,
2566 : // it can actually take a long time if the user has a lot of rels.
2567 : // And we'll inevitable need it again; So, let the background task do the work.)
2568 0 : match self
2569 0 : .current_logical_size
2570 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2571 0 : .get()
2572 : {
2573 0 : Some(cancel) => cancel.cancel(),
2574 : None => {
2575 0 : let state = self.current_state();
2576 0 : if matches!(
2577 0 : state,
2578 : TimelineState::Broken { .. } | TimelineState::Stopping
2579 0 : ) {
2580 0 :
2581 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2582 0 : // Don't make noise.
2583 0 : } else {
2584 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2585 0 : debug_assert!(false);
2586 : }
2587 : }
2588 : };
2589 : }
2590 : }
2591 :
2592 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2593 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2594 0 : let first = self
2595 0 : .current_logical_size
2596 0 : .did_return_approximate_to_walreceiver
2597 0 : .compare_exchange(
2598 0 : false,
2599 0 : true,
2600 0 : AtomicOrdering::Relaxed,
2601 0 : AtomicOrdering::Relaxed,
2602 0 : )
2603 0 : .is_ok();
2604 0 : if first {
2605 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2606 0 : }
2607 0 : }
2608 0 : }
2609 :
2610 0 : current_size
2611 0 : }
2612 :
2613 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2614 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2615 : // nothing to do for freshly created timelines;
2616 0 : assert_eq!(
2617 0 : self.current_logical_size.current_size().accuracy(),
2618 0 : logical_size::Accuracy::Exact,
2619 0 : );
2620 0 : self.current_logical_size.initialized.add_permits(1);
2621 0 : return;
2622 : };
2623 :
2624 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2625 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2626 0 : self.current_logical_size
2627 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2628 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2629 0 :
2630 0 : let self_clone = Arc::clone(self);
2631 0 : let background_ctx = ctx.detached_child(
2632 0 : TaskKind::InitialLogicalSizeCalculation,
2633 0 : DownloadBehavior::Download,
2634 0 : );
2635 0 : task_mgr::spawn(
2636 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2637 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2638 0 : self.tenant_shard_id,
2639 0 : Some(self.timeline_id),
2640 0 : "initial size calculation",
2641 : // NB: don't log errors here, task_mgr will do that.
2642 0 : async move {
2643 0 : let cancel = task_mgr::shutdown_token();
2644 0 : self_clone
2645 0 : .initial_logical_size_calculation_task(
2646 0 : initial_part_end,
2647 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2648 0 : cancel,
2649 0 : background_ctx,
2650 0 : )
2651 0 : .await;
2652 0 : Ok(())
2653 0 : }
2654 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)),
2655 : );
2656 0 : }
2657 :
2658 0 : async fn initial_logical_size_calculation_task(
2659 0 : self: Arc<Self>,
2660 0 : initial_part_end: Lsn,
2661 0 : skip_concurrency_limiter: CancellationToken,
2662 0 : cancel: CancellationToken,
2663 0 : background_ctx: RequestContext,
2664 0 : ) {
2665 : scopeguard::defer! {
2666 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2667 : self.current_logical_size.initialized.add_permits(1);
2668 : }
2669 :
2670 0 : let try_once = |attempt: usize| {
2671 0 : let background_ctx = &background_ctx;
2672 0 : let self_ref = &self;
2673 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2674 0 : async move {
2675 0 : let cancel = task_mgr::shutdown_token();
2676 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2677 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2678 0 : background_ctx,
2679 0 : );
2680 :
2681 : use crate::metrics::initial_logical_size::StartCircumstances;
2682 0 : let (_maybe_permit, circumstances) = tokio::select! {
2683 : permit = wait_for_permit => {
2684 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2685 : }
2686 : _ = self_ref.cancel.cancelled() => {
2687 : return Err(CalculateLogicalSizeError::Cancelled);
2688 : }
2689 : _ = cancel.cancelled() => {
2690 : return Err(CalculateLogicalSizeError::Cancelled);
2691 : },
2692 : () = skip_concurrency_limiter.cancelled() => {
2693 : // Some action that is part of a end user interaction requested logical size
2694 : // => break out of the rate limit
2695 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2696 : // but then again what happens if they cancel; also, we should just be using
2697 : // one runtime across the entire process, so, let's leave this for now.
2698 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2699 : }
2700 : };
2701 :
2702 0 : let metrics_guard = if attempt == 1 {
2703 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2704 : } else {
2705 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2706 : };
2707 :
2708 0 : let calculated_size = self_ref
2709 0 : .logical_size_calculation_task(
2710 0 : initial_part_end,
2711 0 : LogicalSizeCalculationCause::Initial,
2712 0 : background_ctx,
2713 0 : )
2714 0 : .await?;
2715 :
2716 0 : self_ref
2717 0 : .trigger_aux_file_size_computation(initial_part_end, background_ctx)
2718 0 : .await?;
2719 :
2720 : // TODO: add aux file size to logical size
2721 :
2722 0 : Ok((calculated_size, metrics_guard))
2723 0 : }
2724 0 : };
2725 :
2726 0 : let retrying = async {
2727 0 : let mut attempt = 0;
2728 0 : loop {
2729 0 : attempt += 1;
2730 0 :
2731 0 : match try_once(attempt).await {
2732 0 : Ok(res) => return ControlFlow::Continue(res),
2733 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
2734 : Err(
2735 0 : e @ (CalculateLogicalSizeError::Decode(_)
2736 0 : | CalculateLogicalSizeError::PageRead(_)),
2737 0 : ) => {
2738 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2739 : // exponential back-off doesn't make sense at these long intervals;
2740 : // use fixed retry interval with generous jitter instead
2741 0 : let sleep_duration = Duration::from_secs(
2742 0 : u64::try_from(
2743 0 : // 1hour base
2744 0 : (60_i64 * 60_i64)
2745 0 : // 10min jitter
2746 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2747 0 : )
2748 0 : .expect("10min < 1hour"),
2749 0 : );
2750 0 : tokio::time::sleep(sleep_duration).await;
2751 : }
2752 : }
2753 : }
2754 0 : };
2755 :
2756 0 : let (calculated_size, metrics_guard) = tokio::select! {
2757 : res = retrying => {
2758 : match res {
2759 : ControlFlow::Continue(calculated_size) => calculated_size,
2760 : ControlFlow::Break(()) => return,
2761 : }
2762 : }
2763 : _ = cancel.cancelled() => {
2764 : return;
2765 : }
2766 : };
2767 :
2768 : // we cannot query current_logical_size.current_size() to know the current
2769 : // *negative* value, only truncated to u64.
2770 0 : let added = self
2771 0 : .current_logical_size
2772 0 : .size_added_after_initial
2773 0 : .load(AtomicOrdering::Relaxed);
2774 0 :
2775 0 : let sum = calculated_size.saturating_add_signed(added);
2776 0 :
2777 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2778 0 : self.metrics.current_logical_size_gauge.set(sum);
2779 0 :
2780 0 : self.current_logical_size
2781 0 : .initial_logical_size
2782 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
2783 0 : .ok()
2784 0 : .expect("only this task sets it");
2785 0 : }
2786 :
2787 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
2788 0 : self: &Arc<Self>,
2789 0 : lsn: Lsn,
2790 0 : cause: LogicalSizeCalculationCause,
2791 0 : ctx: RequestContext,
2792 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
2793 0 : let (sender, receiver) = oneshot::channel();
2794 0 : let self_clone = Arc::clone(self);
2795 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
2796 0 : // we should stop the size calculation work and return an error.
2797 0 : // That would require restructuring this function's API to
2798 0 : // return the result directly, instead of a Receiver for the result.
2799 0 : let ctx = ctx.detached_child(
2800 0 : TaskKind::OndemandLogicalSizeCalculation,
2801 0 : DownloadBehavior::Download,
2802 0 : );
2803 0 : task_mgr::spawn(
2804 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2805 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
2806 0 : self.tenant_shard_id,
2807 0 : Some(self.timeline_id),
2808 0 : "ondemand logical size calculation",
2809 0 : async move {
2810 0 : let res = self_clone
2811 0 : .logical_size_calculation_task(lsn, cause, &ctx)
2812 0 : .await;
2813 0 : let _ = sender.send(res).ok();
2814 0 : Ok(()) // Receiver is responsible for handling errors
2815 0 : }
2816 0 : .in_current_span(),
2817 0 : );
2818 0 : receiver
2819 0 : }
2820 :
2821 : /// # Cancel-Safety
2822 : ///
2823 : /// This method is cancellation-safe.
2824 0 : #[instrument(skip_all)]
2825 : async fn logical_size_calculation_task(
2826 : self: &Arc<Self>,
2827 : lsn: Lsn,
2828 : cause: LogicalSizeCalculationCause,
2829 : ctx: &RequestContext,
2830 : ) -> Result<u64, CalculateLogicalSizeError> {
2831 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
2832 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
2833 : // accurate relation sizes, and they do not emit consumption metrics.
2834 : debug_assert!(self.tenant_shard_id.is_shard_zero());
2835 :
2836 : let guard = self
2837 : .gate
2838 : .enter()
2839 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
2840 :
2841 : let self_calculation = Arc::clone(self);
2842 :
2843 0 : let mut calculation = pin!(async {
2844 0 : let ctx = ctx.attached_child();
2845 0 : self_calculation
2846 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
2847 0 : .await
2848 0 : });
2849 :
2850 : tokio::select! {
2851 : res = &mut calculation => { res }
2852 : _ = self.cancel.cancelled() => {
2853 : debug!("cancelling logical size calculation for timeline shutdown");
2854 : calculation.await
2855 : }
2856 : }
2857 : }
2858 :
2859 : /// Calculate the logical size of the database at the latest LSN.
2860 : ///
2861 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
2862 : /// especially if we need to download remote layers.
2863 : ///
2864 : /// # Cancel-Safety
2865 : ///
2866 : /// This method is cancellation-safe.
2867 0 : async fn calculate_logical_size(
2868 0 : &self,
2869 0 : up_to_lsn: Lsn,
2870 0 : cause: LogicalSizeCalculationCause,
2871 0 : _guard: &GateGuard,
2872 0 : ctx: &RequestContext,
2873 0 : ) -> Result<u64, CalculateLogicalSizeError> {
2874 0 : info!(
2875 0 : "Calculating logical size for timeline {} at {}",
2876 : self.timeline_id, up_to_lsn
2877 : );
2878 :
2879 : pausable_failpoint!("timeline-calculate-logical-size-pause");
2880 :
2881 : // See if we've already done the work for initial size calculation.
2882 : // This is a short-cut for timelines that are mostly unused.
2883 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
2884 0 : return Ok(size);
2885 0 : }
2886 0 : let storage_time_metrics = match cause {
2887 : LogicalSizeCalculationCause::Initial
2888 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
2889 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
2890 : LogicalSizeCalculationCause::EvictionTaskImitation => {
2891 0 : &self.metrics.imitate_logical_size_histo
2892 : }
2893 : };
2894 0 : let timer = storage_time_metrics.start_timer();
2895 0 : let logical_size = self
2896 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
2897 0 : .await?;
2898 0 : debug!("calculated logical size: {logical_size}");
2899 0 : timer.stop_and_record();
2900 0 : Ok(logical_size)
2901 0 : }
2902 :
2903 : /// Update current logical size, adding `delta' to the old value.
2904 270570 : fn update_current_logical_size(&self, delta: i64) {
2905 270570 : let logical_size = &self.current_logical_size;
2906 270570 : logical_size.increment_size(delta);
2907 270570 :
2908 270570 : // Also set the value in the prometheus gauge. Note that
2909 270570 : // there is a race condition here: if this is is called by two
2910 270570 : // threads concurrently, the prometheus gauge might be set to
2911 270570 : // one value while current_logical_size is set to the
2912 270570 : // other.
2913 270570 : match logical_size.current_size() {
2914 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
2915 270570 : .metrics
2916 270570 : .current_logical_size_gauge
2917 270570 : .set(new_current_size.into()),
2918 0 : CurrentLogicalSize::Approximate(_) => {
2919 0 : // don't update the gauge yet, this allows us not to update the gauge back and
2920 0 : // forth between the initial size calculation task.
2921 0 : }
2922 : }
2923 270570 : }
2924 :
2925 2988 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
2926 2988 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
2927 2988 : let aux_metric =
2928 2988 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
2929 2988 :
2930 2988 : let sum_of_entries = self
2931 2988 : .directory_metrics
2932 2988 : .iter()
2933 20916 : .map(|v| v.load(AtomicOrdering::Relaxed))
2934 2988 : .sum();
2935 2988 : // Set a high general threshold and a lower threshold for the auxiliary files,
2936 2988 : // as we can have large numbers of relations in the db directory.
2937 2988 : const SUM_THRESHOLD: u64 = 5000;
2938 2988 : const AUX_THRESHOLD: u64 = 1000;
2939 2988 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
2940 0 : self.metrics
2941 0 : .directory_entries_count_gauge
2942 0 : .set(sum_of_entries);
2943 2988 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
2944 0 : metric.set(sum_of_entries);
2945 2988 : }
2946 2988 : }
2947 :
2948 0 : async fn find_layer(
2949 0 : &self,
2950 0 : layer_name: &LayerName,
2951 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
2952 0 : let guard = self.layers.read().await;
2953 0 : let layer = guard
2954 0 : .layer_map()?
2955 0 : .iter_historic_layers()
2956 0 : .find(|l| &l.layer_name() == layer_name)
2957 0 : .map(|found| guard.get_from_desc(&found));
2958 0 : Ok(layer)
2959 0 : }
2960 :
2961 : /// The timeline heatmap is a hint to secondary locations from the primary location,
2962 : /// indicating which layers are currently on-disk on the primary.
2963 : ///
2964 : /// None is returned if the Timeline is in a state where uploading a heatmap
2965 : /// doesn't make sense, such as shutting down or initializing. The caller
2966 : /// should treat this as a cue to simply skip doing any heatmap uploading
2967 : /// for this timeline.
2968 2 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
2969 2 : if !self.is_active() {
2970 0 : return None;
2971 2 : }
2972 :
2973 2 : let guard = self.layers.read().await;
2974 :
2975 10 : let resident = guard.likely_resident_layers().filter_map(|layer| {
2976 10 : match layer.visibility() {
2977 : LayerVisibilityHint::Visible => {
2978 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
2979 8 : let last_activity_ts = layer.latest_activity();
2980 8 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
2981 : }
2982 : LayerVisibilityHint::Covered => {
2983 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
2984 2 : None
2985 : }
2986 : }
2987 10 : });
2988 2 :
2989 2 : let mut layers = resident.collect::<Vec<_>>();
2990 2 :
2991 2 : // Sort layers in order of which to download first. For a large set of layers to download, we
2992 2 : // want to prioritize those layers which are most likely to still be in the resident many minutes
2993 2 : // or hours later:
2994 2 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
2995 2 : // only exist for a few minutes before being compacted into L1s.
2996 2 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
2997 2 : // the layer is likely to be covered by an image layer during compaction.
2998 20 : layers.sort_by_key(|(desc, _meta, _atime)| {
2999 20 : std::cmp::Reverse((!LayerMap::is_l0(&desc.key_range), desc.lsn_range.end))
3000 20 : });
3001 2 :
3002 2 : let layers = layers
3003 2 : .into_iter()
3004 8 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3005 2 : .collect();
3006 2 :
3007 2 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3008 2 : }
3009 :
3010 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3011 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3012 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3013 0 : // branchpoint in the value in IndexPart::lineage
3014 0 : self.ancestor_lsn == lsn
3015 0 : || (self.ancestor_lsn == Lsn::INVALID
3016 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3017 0 : }
3018 : }
3019 :
3020 : impl Timeline {
3021 : #[allow(unknown_lints)] // doc_lazy_continuation is still a new lint
3022 : #[allow(clippy::doc_lazy_continuation)]
3023 : /// Get the data needed to reconstruct all keys in the provided keyspace
3024 : ///
3025 : /// The algorithm is as follows:
3026 : /// 1. While some keys are still not done and there's a timeline to visit:
3027 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3028 : /// 2.1: Build the fringe for the current keyspace
3029 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3030 : /// intersects
3031 : /// 2.3. Pop the timeline from the fringe
3032 : /// 2.4. If the fringe is empty, go back to 1
3033 626183 : async fn get_vectored_reconstruct_data(
3034 626183 : &self,
3035 626183 : mut keyspace: KeySpace,
3036 626183 : request_lsn: Lsn,
3037 626183 : reconstruct_state: &mut ValuesReconstructState,
3038 626183 : ctx: &RequestContext,
3039 626183 : ) -> Result<(), GetVectoredError> {
3040 626183 : let mut timeline_owned: Arc<Timeline>;
3041 626183 : let mut timeline = self;
3042 626183 :
3043 626183 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3044 :
3045 626181 : let missing_keyspace = loop {
3046 851132 : if self.cancel.is_cancelled() {
3047 0 : return Err(GetVectoredError::Cancelled);
3048 851132 : }
3049 :
3050 : let TimelineVisitOutcome {
3051 851132 : completed_keyspace: completed,
3052 851132 : image_covered_keyspace,
3053 851132 : } = Self::get_vectored_reconstruct_data_timeline(
3054 851132 : timeline,
3055 851132 : keyspace.clone(),
3056 851132 : cont_lsn,
3057 851132 : reconstruct_state,
3058 851132 : &self.cancel,
3059 851132 : ctx,
3060 851132 : )
3061 114069 : .await?;
3062 :
3063 851132 : keyspace.remove_overlapping_with(&completed);
3064 851132 :
3065 851132 : // Do not descend into the ancestor timeline for aux files.
3066 851132 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3067 851132 : // stalling compaction.
3068 851132 : keyspace.remove_overlapping_with(&KeySpace {
3069 851132 : ranges: vec![NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE],
3070 851132 : });
3071 851132 :
3072 851132 : // Keyspace is fully retrieved
3073 851132 : if keyspace.is_empty() {
3074 626167 : break None;
3075 224965 : }
3076 :
3077 224965 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3078 : // Not fully retrieved but no ancestor timeline.
3079 14 : break Some(keyspace);
3080 : };
3081 :
3082 : // Now we see if there are keys covered by the image layer but does not exist in the
3083 : // image layer, which means that the key does not exist.
3084 :
3085 : // The block below will stop the vectored search if any of the keys encountered an image layer
3086 : // which did not contain a snapshot for said key. Since we have already removed all completed
3087 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3088 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3089 : // and stop the search as a result of that.
3090 224951 : let removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3091 224951 : if !removed.is_empty() {
3092 0 : break Some(removed);
3093 224951 : }
3094 224951 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3095 224951 : // keyspace.
3096 224951 :
3097 224951 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3098 224951 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3099 224951 : timeline_owned = timeline
3100 224951 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3101 2 : .await?;
3102 224949 : timeline = &*timeline_owned;
3103 : };
3104 :
3105 626181 : if let Some(missing_keyspace) = missing_keyspace {
3106 14 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3107 14 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3108 14 : shard: self
3109 14 : .shard_identity
3110 14 : .get_shard_number(&missing_keyspace.start().unwrap()),
3111 14 : cont_lsn,
3112 14 : request_lsn,
3113 14 : ancestor_lsn: Some(timeline.ancestor_lsn),
3114 14 : backtrace: None,
3115 14 : }));
3116 626167 : }
3117 626167 :
3118 626167 : Ok(())
3119 626183 : }
3120 :
3121 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3122 : ///
3123 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3124 : /// the current keyspace. The current keyspace of the search at any given timeline
3125 : /// is the original keyspace minus all the keys that have been completed minus
3126 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3127 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3128 : ///
3129 : /// This is basically a depth-first search visitor implementation where a vertex
3130 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3131 : ///
3132 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3133 : /// and get all the required reconstruct data from the layer in one go.
3134 : ///
3135 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3136 : /// decides how to deal with these two keyspaces.
3137 851132 : async fn get_vectored_reconstruct_data_timeline(
3138 851132 : timeline: &Timeline,
3139 851132 : keyspace: KeySpace,
3140 851132 : mut cont_lsn: Lsn,
3141 851132 : reconstruct_state: &mut ValuesReconstructState,
3142 851132 : cancel: &CancellationToken,
3143 851132 : ctx: &RequestContext,
3144 851132 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3145 851132 : let mut unmapped_keyspace = keyspace.clone();
3146 851132 : let mut fringe = LayerFringe::new();
3147 851132 :
3148 851132 : let mut completed_keyspace = KeySpace::default();
3149 851132 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3150 :
3151 1669453 : loop {
3152 1669453 : if cancel.is_cancelled() {
3153 0 : return Err(GetVectoredError::Cancelled);
3154 1669453 : }
3155 1669453 :
3156 1669453 : let (keys_done_last_step, keys_with_image_coverage) =
3157 1669453 : reconstruct_state.consume_done_keys();
3158 1669453 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3159 1669453 : completed_keyspace.merge(&keys_done_last_step);
3160 1669453 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3161 7726 : unmapped_keyspace
3162 7726 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3163 7726 : image_covered_keyspace.add_range(keys_with_image_coverage);
3164 1661727 : }
3165 :
3166 : // Do not descent any further if the last layer we visited
3167 : // completed all keys in the keyspace it inspected. This is not
3168 : // required for correctness, but avoids visiting extra layers
3169 : // which turns out to be a perf bottleneck in some cases.
3170 1669453 : if !unmapped_keyspace.is_empty() {
3171 1043412 : let guard = timeline.layers.read().await;
3172 1043412 : let layers = guard.layer_map()?;
3173 :
3174 1043412 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3175 911192 : let start_lsn = l.get_lsn_range().start;
3176 911192 : cont_lsn > start_lsn
3177 1043412 : });
3178 1043412 :
3179 1043412 : match in_memory_layer {
3180 606143 : Some(l) => {
3181 606143 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3182 606143 : fringe.update(
3183 606143 : ReadableLayer::InMemoryLayer(l),
3184 606143 : unmapped_keyspace.clone(),
3185 606143 : lsn_range,
3186 606143 : );
3187 606143 : }
3188 : None => {
3189 508921 : for range in unmapped_keyspace.ranges.iter() {
3190 508921 : let results = layers.range_search(range.clone(), cont_lsn);
3191 508921 :
3192 508921 : results
3193 508921 : .found
3194 508921 : .into_iter()
3195 508921 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3196 275808 : (
3197 275808 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3198 275808 : keyspace_accum.to_keyspace(),
3199 275808 : lsn_floor..cont_lsn,
3200 275808 : )
3201 508921 : })
3202 508921 : .for_each(|(layer, keyspace, lsn_range)| {
3203 275808 : fringe.update(layer, keyspace, lsn_range)
3204 508921 : });
3205 508921 : }
3206 : }
3207 : }
3208 :
3209 : // It's safe to drop the layer map lock after planning the next round of reads.
3210 : // The fringe keeps readable handles for the layers which are safe to read even
3211 : // if layers were compacted or flushed.
3212 : //
3213 : // The more interesting consideration is: "Why is the read algorithm still correct
3214 : // if the layer map changes while it is operating?". Doing a vectored read on a
3215 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3216 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3217 : // range at *a particular point in time*. It is fine for the answer to be different
3218 : // at two different time points.
3219 1043412 : drop(guard);
3220 626041 : }
3221 :
3222 1669453 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3223 818321 : let next_cont_lsn = lsn_range.start;
3224 818321 : layer_to_read
3225 818321 : .get_values_reconstruct_data(
3226 818321 : keyspace_to_read.clone(),
3227 818321 : lsn_range,
3228 818321 : reconstruct_state,
3229 818321 : ctx,
3230 818321 : )
3231 107835 : .await?;
3232 :
3233 818321 : unmapped_keyspace = keyspace_to_read;
3234 818321 : cont_lsn = next_cont_lsn;
3235 818321 :
3236 818321 : reconstruct_state.on_layer_visited(&layer_to_read);
3237 : } else {
3238 851132 : break;
3239 851132 : }
3240 851132 : }
3241 851132 :
3242 851132 : Ok(TimelineVisitOutcome {
3243 851132 : completed_keyspace,
3244 851132 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3245 851132 : })
3246 851132 : }
3247 :
3248 224951 : async fn get_ready_ancestor_timeline(
3249 224951 : &self,
3250 224951 : ancestor: &Arc<Timeline>,
3251 224951 : ctx: &RequestContext,
3252 224951 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3253 224951 : // It's possible that the ancestor timeline isn't active yet, or
3254 224951 : // is active but hasn't yet caught up to the branch point. Wait
3255 224951 : // for it.
3256 224951 : //
3257 224951 : // This cannot happen while the pageserver is running normally,
3258 224951 : // because you cannot create a branch from a point that isn't
3259 224951 : // present in the pageserver yet. However, we don't wait for the
3260 224951 : // branch point to be uploaded to cloud storage before creating
3261 224951 : // a branch. I.e., the branch LSN need not be remote consistent
3262 224951 : // for the branching operation to succeed.
3263 224951 : //
3264 224951 : // Hence, if we try to load a tenant in such a state where
3265 224951 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3266 224951 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3267 224951 : // then we will need to wait for the ancestor timeline to
3268 224951 : // re-stream WAL up to branch_lsn before we access it.
3269 224951 : //
3270 224951 : // How can a tenant get in such a state?
3271 224951 : // - ungraceful pageserver process exit
3272 224951 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3273 224951 : //
3274 224951 : // NB: this could be avoided by requiring
3275 224951 : // branch_lsn >= remote_consistent_lsn
3276 224951 : // during branch creation.
3277 224951 : match ancestor.wait_to_become_active(ctx).await {
3278 224949 : Ok(()) => {}
3279 : Err(TimelineState::Stopping) => {
3280 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3281 0 : return Err(GetReadyAncestorError::Cancelled);
3282 : }
3283 2 : Err(state) => {
3284 2 : return Err(GetReadyAncestorError::BadState {
3285 2 : timeline_id: ancestor.timeline_id,
3286 2 : state,
3287 2 : });
3288 : }
3289 : }
3290 224949 : ancestor
3291 224949 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3292 0 : .await
3293 224949 : .map_err(|e| match e {
3294 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3295 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3296 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3297 0 : timeline_id: ancestor.timeline_id,
3298 0 : state,
3299 0 : },
3300 224949 : })?;
3301 :
3302 224949 : Ok(ancestor.clone())
3303 224951 : }
3304 :
3305 5452 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3306 5452 : &self.shard_identity
3307 5452 : }
3308 :
3309 : #[inline(always)]
3310 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3311 0 : ShardTimelineId {
3312 0 : shard_index: ShardIndex {
3313 0 : shard_number: self.shard_identity.number,
3314 0 : shard_count: self.shard_identity.count,
3315 0 : },
3316 0 : timeline_id: self.timeline_id,
3317 0 : }
3318 0 : }
3319 :
3320 : /// Returns a non-frozen open in-memory layer for ingestion.
3321 : ///
3322 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3323 : /// this function without holding the mutex.
3324 1264 : async fn get_layer_for_write(
3325 1264 : &self,
3326 1264 : lsn: Lsn,
3327 1264 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3328 1264 : ctx: &RequestContext,
3329 1264 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3330 1264 : let mut guard = self.layers.write().await;
3331 1264 : let gate_guard = self.gate.enter().context("enter gate for inmem layer")?;
3332 :
3333 1264 : let last_record_lsn = self.get_last_record_lsn();
3334 1264 : ensure!(
3335 1264 : lsn > last_record_lsn,
3336 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3337 : lsn,
3338 : last_record_lsn,
3339 : );
3340 :
3341 1264 : let layer = guard
3342 1264 : .open_mut()?
3343 1264 : .get_layer_for_write(
3344 1264 : lsn,
3345 1264 : self.conf,
3346 1264 : self.timeline_id,
3347 1264 : self.tenant_shard_id,
3348 1264 : gate_guard,
3349 1264 : ctx,
3350 1264 : )
3351 712 : .await?;
3352 1264 : Ok(layer)
3353 1264 : }
3354 :
3355 5279070 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3356 5279070 : assert!(new_lsn.is_aligned());
3357 :
3358 5279070 : self.metrics.last_record_gauge.set(new_lsn.0 as i64);
3359 5279070 : self.last_record_lsn.advance(new_lsn);
3360 5279070 : }
3361 :
3362 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3363 : ///
3364 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3365 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3366 1162 : async fn freeze_inmem_layer_at(
3367 1162 : &self,
3368 1162 : at: Lsn,
3369 1162 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3370 1162 : ) -> Result<u64, FlushLayerError> {
3371 1162 : let frozen = {
3372 1162 : let mut guard = self.layers.write().await;
3373 1162 : guard
3374 1162 : .open_mut()?
3375 1162 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock)
3376 0 : .await
3377 : };
3378 :
3379 1162 : if frozen {
3380 1134 : let now = Instant::now();
3381 1134 : *(self.last_freeze_ts.write().unwrap()) = now;
3382 1134 : }
3383 :
3384 : // Increment the flush cycle counter and wake up the flush task.
3385 : // Remember the new value, so that when we listen for the flush
3386 : // to finish, we know when the flush that we initiated has
3387 : // finished, instead of some other flush that was started earlier.
3388 1162 : let mut my_flush_request = 0;
3389 1162 :
3390 1162 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3391 1162 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3392 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3393 1162 : }
3394 1162 :
3395 1162 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3396 1162 : my_flush_request = *counter + 1;
3397 1162 : *counter = my_flush_request;
3398 1162 : *lsn = std::cmp::max(at, *lsn);
3399 1162 : });
3400 1162 :
3401 1162 : assert_ne!(my_flush_request, 0);
3402 :
3403 1162 : Ok(my_flush_request)
3404 1162 : }
3405 :
3406 : /// Layer flusher task's main loop.
3407 400 : async fn flush_loop(
3408 400 : self: &Arc<Self>,
3409 400 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3410 400 : ctx: &RequestContext,
3411 400 : ) {
3412 400 : info!("started flush loop");
3413 1122 : loop {
3414 1122 : tokio::select! {
3415 : _ = self.cancel.cancelled() => {
3416 : info!("shutting down layer flush task due to Timeline::cancel");
3417 : break;
3418 : },
3419 : _ = layer_flush_start_rx.changed() => {}
3420 : }
3421 1122 : trace!("waking up");
3422 1122 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3423 1122 :
3424 1122 : // The highest LSN to which we flushed in the loop over frozen layers
3425 1122 : let mut flushed_to_lsn = Lsn(0);
3426 :
3427 1122 : let result = loop {
3428 2256 : if self.cancel.is_cancelled() {
3429 0 : info!("dropping out of flush loop for timeline shutdown");
3430 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3431 : // anyone waiting on that will respect self.cancel as well: they will stop
3432 : // waiting at the same time we as drop out of this loop.
3433 0 : return;
3434 2256 : }
3435 2256 :
3436 2256 : let timer = self.metrics.flush_time_histo.start_timer();
3437 :
3438 2256 : let layer_to_flush = {
3439 2256 : let guard = self.layers.read().await;
3440 2256 : let Ok(lm) = guard.layer_map() else {
3441 0 : info!("dropping out of flush loop for timeline shutdown");
3442 0 : return;
3443 : };
3444 2256 : lm.frozen_layers.front().cloned()
3445 : // drop 'layers' lock to allow concurrent reads and writes
3446 : };
3447 2256 : let Some(layer_to_flush) = layer_to_flush else {
3448 1122 : break Ok(());
3449 : };
3450 17158 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3451 1134 : Ok(this_layer_to_lsn) => {
3452 1134 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3453 1134 : }
3454 : Err(FlushLayerError::Cancelled) => {
3455 0 : info!("dropping out of flush loop for timeline shutdown");
3456 0 : return;
3457 : }
3458 0 : err @ Err(
3459 0 : FlushLayerError::NotRunning(_)
3460 0 : | FlushLayerError::Other(_)
3461 0 : | FlushLayerError::CreateImageLayersError(_),
3462 0 : ) => {
3463 0 : error!("could not flush frozen layer: {err:?}");
3464 0 : break err.map(|_| ());
3465 : }
3466 : }
3467 1134 : timer.stop_and_record();
3468 : };
3469 :
3470 : // Unsharded tenants should never advance their LSN beyond the end of the
3471 : // highest layer they write: such gaps between layer data and the frozen LSN
3472 : // are only legal on sharded tenants.
3473 1122 : debug_assert!(
3474 1122 : self.shard_identity.count.count() > 1
3475 1122 : || flushed_to_lsn >= frozen_to_lsn
3476 66 : || !flushed_to_lsn.is_valid()
3477 : );
3478 :
3479 1122 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3480 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3481 : // to us via layer_flush_start_rx, then advance it here.
3482 : //
3483 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3484 : // never encounter a gap in the wal.
3485 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3486 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3487 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3488 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3489 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3490 0 : }
3491 0 : }
3492 1122 : }
3493 :
3494 : // Notify any listeners that we're done
3495 1122 : let _ = self
3496 1122 : .layer_flush_done_tx
3497 1122 : .send_replace((flush_counter, result));
3498 : }
3499 8 : }
3500 :
3501 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
3502 1082 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3503 1082 : let mut rx = self.layer_flush_done_tx.subscribe();
3504 2159 : loop {
3505 2159 : {
3506 2159 : let (last_result_counter, last_result) = &*rx.borrow();
3507 2159 : if *last_result_counter >= request {
3508 1082 : if let Err(err) = last_result {
3509 : // We already logged the original error in
3510 : // flush_loop. We cannot propagate it to the caller
3511 : // here, because it might not be Cloneable
3512 0 : return Err(err.clone());
3513 : } else {
3514 1082 : return Ok(());
3515 : }
3516 1077 : }
3517 1077 : }
3518 1077 : trace!("waiting for flush to complete");
3519 : tokio::select! {
3520 : rx_e = rx.changed() => {
3521 0 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3522 : },
3523 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3524 : // the notification from [`flush_loop`] that it completed.
3525 : _ = self.cancel.cancelled() => {
3526 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3527 : return Ok(())
3528 : }
3529 : };
3530 1077 : trace!("done")
3531 : }
3532 1082 : }
3533 :
3534 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3535 : ///
3536 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3537 2268 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3538 : async fn flush_frozen_layer(
3539 : self: &Arc<Self>,
3540 : frozen_layer: Arc<InMemoryLayer>,
3541 : ctx: &RequestContext,
3542 : ) -> Result<Lsn, FlushLayerError> {
3543 : debug_assert_current_span_has_tenant_and_timeline_id();
3544 :
3545 : // As a special case, when we have just imported an image into the repository,
3546 : // instead of writing out a L0 delta layer, we directly write out image layer
3547 : // files instead. This is possible as long as *all* the data imported into the
3548 : // repository have the same LSN.
3549 : let lsn_range = frozen_layer.get_lsn_range();
3550 :
3551 : // Whether to directly create image layers for this flush, or flush them as delta layers
3552 : let create_image_layer =
3553 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3554 :
3555 : #[cfg(test)]
3556 : {
3557 : match &mut *self.flush_loop_state.lock().unwrap() {
3558 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3559 : panic!("flush loop not running")
3560 : }
3561 : FlushLoopState::Running {
3562 : expect_initdb_optimization,
3563 : initdb_optimization_count,
3564 : ..
3565 : } => {
3566 : if create_image_layer {
3567 : *initdb_optimization_count += 1;
3568 : } else {
3569 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3570 : }
3571 : }
3572 : }
3573 : }
3574 :
3575 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3576 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3577 : // require downloading anything during initial import.
3578 : let ((rel_partition, metadata_partition), _lsn) = self
3579 : .repartition(
3580 : self.initdb_lsn,
3581 : self.get_compaction_target_size(),
3582 : EnumSet::empty(),
3583 : ctx,
3584 : )
3585 : .await
3586 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
3587 :
3588 : if self.cancel.is_cancelled() {
3589 : return Err(FlushLayerError::Cancelled);
3590 : }
3591 :
3592 : // FIXME(auxfilesv2): support multiple metadata key partitions might need initdb support as well?
3593 : // This code path will not be hit during regression tests. After #7099 we have a single partition
3594 : // with two key ranges. If someone wants to fix initdb optimization in the future, this might need
3595 : // to be fixed.
3596 :
3597 : // For metadata, always create delta layers.
3598 : let delta_layer = if !metadata_partition.parts.is_empty() {
3599 : assert_eq!(
3600 : metadata_partition.parts.len(),
3601 : 1,
3602 : "currently sparse keyspace should only contain a single metadata keyspace"
3603 : );
3604 : let metadata_keyspace = &metadata_partition.parts[0];
3605 : self.create_delta_layer(
3606 : &frozen_layer,
3607 : Some(
3608 : metadata_keyspace.0.ranges.first().unwrap().start
3609 : ..metadata_keyspace.0.ranges.last().unwrap().end,
3610 : ),
3611 : ctx,
3612 : )
3613 : .await
3614 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
3615 : } else {
3616 : None
3617 : };
3618 :
3619 : // For image layers, we add them immediately into the layer map.
3620 : let mut layers_to_upload = Vec::new();
3621 : layers_to_upload.extend(
3622 : self.create_image_layers(
3623 : &rel_partition,
3624 : self.initdb_lsn,
3625 : ImageLayerCreationMode::Initial,
3626 : ctx,
3627 : )
3628 : .await?,
3629 : );
3630 :
3631 : if let Some(delta_layer) = delta_layer {
3632 : layers_to_upload.push(delta_layer.clone());
3633 : (layers_to_upload, Some(delta_layer))
3634 : } else {
3635 : (layers_to_upload, None)
3636 : }
3637 : } else {
3638 : // Normal case, write out a L0 delta layer file.
3639 : // `create_delta_layer` will not modify the layer map.
3640 : // We will remove frozen layer and add delta layer in one atomic operation later.
3641 : let Some(layer) = self
3642 : .create_delta_layer(&frozen_layer, None, ctx)
3643 : .await
3644 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
3645 : else {
3646 : panic!("delta layer cannot be empty if no filter is applied");
3647 : };
3648 : (
3649 : // FIXME: even though we have a single image and single delta layer assumption
3650 : // we push them to vec
3651 : vec![layer.clone()],
3652 : Some(layer),
3653 : )
3654 : };
3655 :
3656 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
3657 :
3658 : if self.cancel.is_cancelled() {
3659 : return Err(FlushLayerError::Cancelled);
3660 : }
3661 :
3662 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
3663 :
3664 : // The new on-disk layers are now in the layer map. We can remove the
3665 : // in-memory layer from the map now. The flushed layer is stored in
3666 : // the mapping in `create_delta_layer`.
3667 : {
3668 : let mut guard = self.layers.write().await;
3669 :
3670 : guard.open_mut()?.finish_flush_l0_layer(
3671 : delta_layer_to_add.as_ref(),
3672 : &frozen_layer,
3673 : &self.metrics,
3674 : );
3675 :
3676 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
3677 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
3678 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
3679 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
3680 : }
3681 : // release lock on 'layers'
3682 : };
3683 :
3684 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
3685 : // This makes us refuse ingest until the new layers have been persisted to the remote.
3686 : self.remote_client
3687 : .wait_completion()
3688 : .await
3689 0 : .map_err(|e| match e {
3690 : WaitCompletionError::UploadQueueShutDownOrStopped
3691 : | WaitCompletionError::NotInitialized(
3692 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
3693 0 : ) => FlushLayerError::Cancelled,
3694 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
3695 0 : FlushLayerError::Other(anyhow!(e).into())
3696 : }
3697 0 : })?;
3698 :
3699 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
3700 : // a compaction can delete the file and then it won't be available for uploads any more.
3701 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
3702 : // race situation.
3703 : // See https://github.com/neondatabase/neon/issues/4526
3704 : pausable_failpoint!("flush-frozen-pausable");
3705 :
3706 : // This failpoint is used by another test case `test_pageserver_recovery`.
3707 : fail_point!("flush-frozen-exit");
3708 :
3709 : Ok(Lsn(lsn_range.end.0 - 1))
3710 : }
3711 :
3712 : /// Return true if the value changed
3713 : ///
3714 : /// This function must only be used from the layer flush task.
3715 1134 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
3716 1134 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
3717 1134 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
3718 1134 : new_value != old_value
3719 1134 : }
3720 :
3721 : /// Update metadata file
3722 1140 : fn schedule_uploads(
3723 1140 : &self,
3724 1140 : disk_consistent_lsn: Lsn,
3725 1140 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
3726 1140 : ) -> anyhow::Result<()> {
3727 1140 : // We can only save a valid 'prev_record_lsn' value on disk if we
3728 1140 : // flushed *all* in-memory changes to disk. We only track
3729 1140 : // 'prev_record_lsn' in memory for the latest processed record, so we
3730 1140 : // don't remember what the correct value that corresponds to some old
3731 1140 : // LSN is. But if we flush everything, then the value corresponding
3732 1140 : // current 'last_record_lsn' is correct and we can store it on disk.
3733 1140 : let RecordLsn {
3734 1140 : last: last_record_lsn,
3735 1140 : prev: prev_record_lsn,
3736 1140 : } = self.last_record_lsn.load();
3737 1140 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
3738 1061 : Some(prev_record_lsn)
3739 : } else {
3740 79 : None
3741 : };
3742 :
3743 1140 : let update = crate::tenant::metadata::MetadataUpdate::new(
3744 1140 : disk_consistent_lsn,
3745 1140 : ondisk_prev_record_lsn,
3746 1140 : *self.latest_gc_cutoff_lsn.read(),
3747 1140 : );
3748 1140 :
3749 1140 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
3750 0 : "{}",
3751 0 : x.unwrap()
3752 1140 : ));
3753 :
3754 2288 : for layer in layers_to_upload {
3755 1148 : self.remote_client.schedule_layer_file_upload(layer)?;
3756 : }
3757 1140 : self.remote_client
3758 1140 : .schedule_index_upload_for_metadata_update(&update)?;
3759 :
3760 1140 : Ok(())
3761 1140 : }
3762 :
3763 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
3764 0 : self.remote_client
3765 0 : .preserve_initdb_archive(
3766 0 : &self.tenant_shard_id.tenant_id,
3767 0 : &self.timeline_id,
3768 0 : &self.cancel,
3769 0 : )
3770 0 : .await
3771 0 : }
3772 :
3773 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
3774 : // in layer map immediately. The caller is responsible to put it into the layer map.
3775 1134 : async fn create_delta_layer(
3776 1134 : self: &Arc<Self>,
3777 1134 : frozen_layer: &Arc<InMemoryLayer>,
3778 1134 : key_range: Option<Range<Key>>,
3779 1134 : ctx: &RequestContext,
3780 1134 : ) -> anyhow::Result<Option<ResidentLayer>> {
3781 1134 : let self_clone = Arc::clone(self);
3782 1134 : let frozen_layer = Arc::clone(frozen_layer);
3783 1134 : let ctx = ctx.attached_child();
3784 1134 : let work = async move {
3785 1134 : let Some((desc, path)) = frozen_layer
3786 1134 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
3787 10264 : .await?
3788 : else {
3789 166 : return Ok(None);
3790 : };
3791 968 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
3792 :
3793 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
3794 : // We just need to fsync the directory in which these inodes are linked,
3795 : // which we know to be the timeline directory.
3796 : //
3797 : // We use fatal_err() below because the after write_to_disk returns with success,
3798 : // the in-memory state of the filesystem already has the layer file in its final place,
3799 : // and subsequent pageserver code could think it's durable while it really isn't.
3800 968 : let timeline_dir = VirtualFile::open(
3801 968 : &self_clone
3802 968 : .conf
3803 968 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
3804 968 : &ctx,
3805 968 : )
3806 486 : .await
3807 968 : .fatal_err("VirtualFile::open for timeline dir fsync");
3808 968 : timeline_dir
3809 968 : .sync_all()
3810 484 : .await
3811 968 : .fatal_err("VirtualFile::sync_all timeline dir");
3812 968 : anyhow::Ok(Some(new_delta))
3813 1134 : };
3814 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
3815 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
3816 : use crate::virtual_file::io_engine::IoEngine;
3817 1134 : match crate::virtual_file::io_engine::get() {
3818 0 : IoEngine::NotSet => panic!("io engine not set"),
3819 : IoEngine::StdFs => {
3820 567 : let span = tracing::info_span!("blocking");
3821 567 : tokio::task::spawn_blocking({
3822 567 : move || Handle::current().block_on(work.instrument(span))
3823 567 : })
3824 567 : .await
3825 567 : .context("spawn_blocking")
3826 567 : .and_then(|x| x)
3827 : }
3828 : #[cfg(target_os = "linux")]
3829 11229 : IoEngine::TokioEpollUring => work.await,
3830 : }
3831 1134 : }
3832 :
3833 530 : async fn repartition(
3834 530 : &self,
3835 530 : lsn: Lsn,
3836 530 : partition_size: u64,
3837 530 : flags: EnumSet<CompactFlags>,
3838 530 : ctx: &RequestContext,
3839 530 : ) -> anyhow::Result<((KeyPartitioning, SparseKeyPartitioning), Lsn)> {
3840 530 : let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
3841 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
3842 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
3843 : // and hence before the compaction task starts.
3844 0 : anyhow::bail!("repartition() called concurrently, this should not happen");
3845 : };
3846 530 : let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
3847 530 : if lsn < *partition_lsn {
3848 0 : anyhow::bail!("repartition() called with LSN going backwards, this should not happen");
3849 530 : }
3850 530 :
3851 530 : let distance = lsn.0 - partition_lsn.0;
3852 530 : if *partition_lsn != Lsn(0)
3853 262 : && distance <= self.repartition_threshold
3854 262 : && !flags.contains(CompactFlags::ForceRepartition)
3855 : {
3856 248 : debug!(
3857 : distance,
3858 : threshold = self.repartition_threshold,
3859 0 : "no repartitioning needed"
3860 : );
3861 248 : return Ok((
3862 248 : (dense_partition.clone(), sparse_partition.clone()),
3863 248 : *partition_lsn,
3864 248 : ));
3865 282 : }
3866 :
3867 16406 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
3868 282 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
3869 282 : let sparse_partitioning = SparseKeyPartitioning {
3870 282 : parts: vec![sparse_ks],
3871 282 : }; // no partitioning for metadata keys for now
3872 282 : *partitioning_guard = ((dense_partitioning, sparse_partitioning), lsn);
3873 282 :
3874 282 : Ok((partitioning_guard.0.clone(), partitioning_guard.1))
3875 530 : }
3876 :
3877 : // Is it time to create a new image layer for the given partition?
3878 14 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
3879 14 : let threshold = self.get_image_creation_threshold();
3880 :
3881 14 : let guard = self.layers.read().await;
3882 14 : let Ok(layers) = guard.layer_map() else {
3883 0 : return false;
3884 : };
3885 :
3886 14 : let mut max_deltas = 0;
3887 28 : for part_range in &partition.ranges {
3888 14 : let image_coverage = layers.image_coverage(part_range, lsn);
3889 28 : for (img_range, last_img) in image_coverage {
3890 14 : let img_lsn = if let Some(last_img) = last_img {
3891 0 : last_img.get_lsn_range().end
3892 : } else {
3893 14 : Lsn(0)
3894 : };
3895 : // Let's consider an example:
3896 : //
3897 : // delta layer with LSN range 71-81
3898 : // delta layer with LSN range 81-91
3899 : // delta layer with LSN range 91-101
3900 : // image layer at LSN 100
3901 : //
3902 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
3903 : // there's no need to create a new one. We check this case explicitly, to avoid passing
3904 : // a bogus range to count_deltas below, with start > end. It's even possible that there
3905 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
3906 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
3907 14 : if img_lsn < lsn {
3908 14 : let num_deltas =
3909 14 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
3910 14 :
3911 14 : max_deltas = max_deltas.max(num_deltas);
3912 14 : if num_deltas >= threshold {
3913 0 : debug!(
3914 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
3915 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
3916 : );
3917 0 : return true;
3918 14 : }
3919 0 : }
3920 : }
3921 : }
3922 :
3923 14 : debug!(
3924 : max_deltas,
3925 0 : "none of the partitioned ranges had >= {threshold} deltas"
3926 : );
3927 14 : false
3928 14 : }
3929 :
3930 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
3931 : /// so that at most one image layer will be produced from this function.
3932 194 : async fn create_image_layer_for_rel_blocks(
3933 194 : self: &Arc<Self>,
3934 194 : partition: &KeySpace,
3935 194 : mut image_layer_writer: ImageLayerWriter,
3936 194 : lsn: Lsn,
3937 194 : ctx: &RequestContext,
3938 194 : img_range: Range<Key>,
3939 194 : start: Key,
3940 194 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
3941 194 : let mut wrote_keys = false;
3942 194 :
3943 194 : let mut key_request_accum = KeySpaceAccum::new();
3944 1278 : for range in &partition.ranges {
3945 1084 : let mut key = range.start;
3946 2522 : while key < range.end {
3947 : // Decide whether to retain this key: usually we do, but sharded tenants may
3948 : // need to drop keys that don't belong to them. If we retain the key, add it
3949 : // to `key_request_accum` for later issuing a vectored get
3950 1438 : if self.shard_identity.is_key_disposable(&key) {
3951 0 : debug!(
3952 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
3953 0 : key,
3954 0 : self.shard_identity.get_shard_number(&key)
3955 : );
3956 1438 : } else {
3957 1438 : key_request_accum.add_key(key);
3958 1438 : }
3959 :
3960 1438 : let last_key_in_range = key.next() == range.end;
3961 1438 : key = key.next();
3962 1438 :
3963 1438 : // Maybe flush `key_rest_accum`
3964 1438 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
3965 1438 : || (last_key_in_range && key_request_accum.raw_size() > 0)
3966 : {
3967 1084 : let results = self
3968 1084 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
3969 50 : .await?;
3970 :
3971 1084 : if self.cancel.is_cancelled() {
3972 0 : return Err(CreateImageLayersError::Cancelled);
3973 1084 : }
3974 :
3975 2522 : for (img_key, img) in results {
3976 1438 : let img = match img {
3977 1438 : Ok(img) => img,
3978 0 : Err(err) => {
3979 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
3980 0 : // page without losing any actual user data. That seems better
3981 0 : // than failing repeatedly and getting stuck.
3982 0 : //
3983 0 : // We had a bug at one point, where we truncated the FSM and VM
3984 0 : // in the pageserver, but the Postgres didn't know about that
3985 0 : // and continued to generate incremental WAL records for pages
3986 0 : // that didn't exist in the pageserver. Trying to replay those
3987 0 : // WAL records failed to find the previous image of the page.
3988 0 : // This special case allows us to recover from that situation.
3989 0 : // See https://github.com/neondatabase/neon/issues/2601.
3990 0 : //
3991 0 : // Unfortunately we cannot do this for the main fork, or for
3992 0 : // any metadata keys, keys, as that would lead to actual data
3993 0 : // loss.
3994 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
3995 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
3996 0 : ZERO_PAGE.clone()
3997 : } else {
3998 0 : return Err(CreateImageLayersError::from(err));
3999 : }
4000 : }
4001 : };
4002 :
4003 : // Write all the keys we just read into our new image layer.
4004 1579 : image_layer_writer.put_image(img_key, img, ctx).await?;
4005 1438 : wrote_keys = true;
4006 : }
4007 354 : }
4008 : }
4009 : }
4010 :
4011 194 : if wrote_keys {
4012 : // Normal path: we have written some data into the new image layer for this
4013 : // partition, so flush it to disk.
4014 401 : let image_layer = image_layer_writer.finish(self, ctx).await?;
4015 194 : Ok(ImageLayerCreationOutcome {
4016 194 : image: Some(image_layer),
4017 194 : next_start_key: img_range.end,
4018 194 : })
4019 : } else {
4020 : // Special case: the image layer may be empty if this is a sharded tenant and the
4021 : // partition does not cover any keys owned by this shard. In this case, to ensure
4022 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4023 : // layer we write will cover the key range that we just scanned.
4024 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4025 0 : Ok(ImageLayerCreationOutcome {
4026 0 : image: None,
4027 0 : next_start_key: start,
4028 0 : })
4029 : }
4030 194 : }
4031 :
4032 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4033 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4034 : /// would not be too large to fit in a single image layer.
4035 : #[allow(clippy::too_many_arguments)]
4036 16 : async fn create_image_layer_for_metadata_keys(
4037 16 : self: &Arc<Self>,
4038 16 : partition: &KeySpace,
4039 16 : mut image_layer_writer: ImageLayerWriter,
4040 16 : lsn: Lsn,
4041 16 : ctx: &RequestContext,
4042 16 : img_range: Range<Key>,
4043 16 : mode: ImageLayerCreationMode,
4044 16 : start: Key,
4045 16 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4046 16 : assert!(!matches!(mode, ImageLayerCreationMode::Initial));
4047 :
4048 : // Metadata keys image layer creation.
4049 16 : let mut reconstruct_state = ValuesReconstructState::default();
4050 16 : let data = self
4051 16 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4052 4125 : .await?;
4053 16 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4054 16 : let mut new_data = BTreeMap::new();
4055 16 : let mut total_kb_retrieved = 0;
4056 16 : let mut total_keys_retrieved = 0;
4057 10028 : for (k, v) in data {
4058 10012 : let v = v?;
4059 10012 : total_kb_retrieved += KEY_SIZE + v.len();
4060 10012 : total_keys_retrieved += 1;
4061 10012 : new_data.insert(k, v);
4062 : }
4063 16 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4064 16 : };
4065 16 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4066 16 :
4067 16 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4068 16 : debug!(
4069 : trigger_generation,
4070 : delta_files_accessed,
4071 : total_kb_retrieved,
4072 : total_keys_retrieved,
4073 0 : "generate metadata images"
4074 : );
4075 :
4076 16 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4077 2 : return Ok(ImageLayerCreationOutcome {
4078 2 : image: None,
4079 2 : next_start_key: img_range.end,
4080 2 : });
4081 14 : }
4082 14 : if self.cancel.is_cancelled() {
4083 0 : return Err(CreateImageLayersError::Cancelled);
4084 14 : }
4085 14 : let mut wrote_any_image = false;
4086 10026 : for (k, v) in data {
4087 10012 : if v.is_empty() {
4088 : // the key has been deleted, it does not need an image
4089 : // in metadata keyspace, an empty image == tombstone
4090 8 : continue;
4091 10004 : }
4092 10004 : wrote_any_image = true;
4093 10004 :
4094 10004 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4095 10004 :
4096 10004 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4097 10004 : // on the normal data path either.
4098 10160 : image_layer_writer.put_image(k, v, ctx).await?;
4099 : }
4100 :
4101 14 : if wrote_any_image {
4102 : // Normal path: we have written some data into the new image layer for this
4103 : // partition, so flush it to disk.
4104 24 : let image_layer = image_layer_writer.finish(self, ctx).await?;
4105 12 : Ok(ImageLayerCreationOutcome {
4106 12 : image: Some(image_layer),
4107 12 : next_start_key: img_range.end,
4108 12 : })
4109 : } else {
4110 : // Special case: the image layer may be empty if this is a sharded tenant and the
4111 : // partition does not cover any keys owned by this shard. In this case, to ensure
4112 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4113 : // layer we write will cover the key range that we just scanned.
4114 2 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4115 2 : Ok(ImageLayerCreationOutcome {
4116 2 : image: None,
4117 2 : next_start_key: start,
4118 2 : })
4119 : }
4120 16 : }
4121 :
4122 : /// Predicate function which indicates whether we should check if new image layers
4123 : /// are required. Since checking if new image layers are required is expensive in
4124 : /// terms of CPU, we only do it in the following cases:
4125 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4126 : /// 2. If enough time has passed since the last check:
4127 : /// 1. For large tenants, we wish to perform the check more often since they
4128 : /// suffer from the lack of image layers
4129 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4130 530 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4131 530 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4132 530 :
4133 530 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4134 530 : let distance = lsn
4135 530 : .checked_sub(last_checks_at)
4136 530 : .expect("Attempt to compact with LSN going backwards");
4137 530 : let min_distance =
4138 530 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4139 530 :
4140 530 : let distance_based_decision = distance.0 >= min_distance;
4141 530 :
4142 530 : let mut time_based_decision = false;
4143 530 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4144 530 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4145 428 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4146 : {
4147 0 : self.get_checkpoint_timeout()
4148 : } else {
4149 428 : Duration::from_secs(3600 * 48)
4150 : };
4151 :
4152 428 : time_based_decision = match *last_check_instant {
4153 262 : Some(last_check) => {
4154 262 : let elapsed = last_check.elapsed();
4155 262 : elapsed >= check_required_after
4156 : }
4157 166 : None => true,
4158 : };
4159 102 : }
4160 :
4161 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4162 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4163 : // check.
4164 530 : let decision = distance_based_decision || time_based_decision;
4165 :
4166 530 : if decision {
4167 168 : self.last_image_layer_creation_check_at.store(lsn);
4168 168 : *last_check_instant = Some(Instant::now());
4169 362 : }
4170 :
4171 530 : decision
4172 530 : }
4173 :
4174 1060 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4175 : async fn create_image_layers(
4176 : self: &Arc<Timeline>,
4177 : partitioning: &KeyPartitioning,
4178 : lsn: Lsn,
4179 : mode: ImageLayerCreationMode,
4180 : ctx: &RequestContext,
4181 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4182 : let timer = self.metrics.create_images_time_histo.start_timer();
4183 : let mut image_layers = Vec::new();
4184 :
4185 : // We need to avoid holes between generated image layers.
4186 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4187 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4188 : //
4189 : // How such hole between partitions can appear?
4190 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4191 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4192 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4193 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4194 : let mut start = Key::MIN;
4195 :
4196 : let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
4197 :
4198 : for partition in partitioning.parts.iter() {
4199 : if self.cancel.is_cancelled() {
4200 : return Err(CreateImageLayersError::Cancelled);
4201 : }
4202 :
4203 : let img_range = start..partition.ranges.last().unwrap().end;
4204 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4205 : if compact_metadata {
4206 : for range in &partition.ranges {
4207 : assert!(
4208 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4209 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4210 : "metadata keys must be partitioned separately"
4211 : );
4212 : }
4213 : if mode == ImageLayerCreationMode::Initial {
4214 : return Err(CreateImageLayersError::Other(anyhow::anyhow!("no image layer should be created for metadata keys when flushing frozen layers")));
4215 : }
4216 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4217 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4218 : // might mess up with evictions.
4219 : start = img_range.end;
4220 : continue;
4221 : }
4222 : } else if let ImageLayerCreationMode::Try = mode {
4223 : // check_for_image_layers = false -> skip
4224 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4225 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4226 : start = img_range.end;
4227 : continue;
4228 : }
4229 : } else if let ImageLayerCreationMode::Force = mode {
4230 : // When forced to create image layers, we might try and create them where they already
4231 : // exist. This mode is only used in tests/debug.
4232 : let layers = self.layers.read().await;
4233 : if layers.contains_key(&PersistentLayerKey {
4234 : key_range: img_range.clone(),
4235 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4236 : is_delta: false,
4237 : }) {
4238 : tracing::info!(
4239 : "Skipping image layer at {lsn} {}..{}, already exists",
4240 : img_range.start,
4241 : img_range.end
4242 : );
4243 : continue;
4244 : }
4245 : }
4246 :
4247 : let image_layer_writer = ImageLayerWriter::new(
4248 : self.conf,
4249 : self.timeline_id,
4250 : self.tenant_shard_id,
4251 : &img_range,
4252 : lsn,
4253 : ctx,
4254 : )
4255 : .await?;
4256 :
4257 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4258 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4259 0 : "failpoint image-layer-writer-fail-before-finish"
4260 0 : )))
4261 0 : });
4262 :
4263 : if !compact_metadata {
4264 : let ImageLayerCreationOutcome {
4265 : image,
4266 : next_start_key,
4267 : } = self
4268 : .create_image_layer_for_rel_blocks(
4269 : partition,
4270 : image_layer_writer,
4271 : lsn,
4272 : ctx,
4273 : img_range,
4274 : start,
4275 : )
4276 : .await?;
4277 :
4278 : start = next_start_key;
4279 : image_layers.extend(image);
4280 : } else {
4281 : let ImageLayerCreationOutcome {
4282 : image,
4283 : next_start_key,
4284 : } = self
4285 : .create_image_layer_for_metadata_keys(
4286 : partition,
4287 : image_layer_writer,
4288 : lsn,
4289 : ctx,
4290 : img_range,
4291 : mode,
4292 : start,
4293 : )
4294 : .await?;
4295 : start = next_start_key;
4296 : image_layers.extend(image);
4297 : }
4298 : }
4299 :
4300 : let mut guard = self.layers.write().await;
4301 :
4302 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4303 : // now they are being scheduled outside of write lock; current way is inconsistent with
4304 : // compaction lock order.
4305 : guard
4306 : .open_mut()?
4307 : .track_new_image_layers(&image_layers, &self.metrics);
4308 : drop_wlock(guard);
4309 : timer.stop_and_record();
4310 :
4311 : // Creating image layers may have caused some previously visible layers to be covered
4312 : self.update_layer_visibility().await?;
4313 :
4314 : Ok(image_layers)
4315 : }
4316 :
4317 : /// Wait until the background initial logical size calculation is complete, or
4318 : /// this Timeline is shut down. Calling this function will cause the initial
4319 : /// logical size calculation to skip waiting for the background jobs barrier.
4320 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4321 0 : if !self.shard_identity.is_shard_zero() {
4322 : // We don't populate logical size on shard >0: skip waiting for it.
4323 0 : return;
4324 0 : }
4325 0 :
4326 0 : if self.remote_client.is_deleting() {
4327 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4328 0 : return;
4329 0 : }
4330 0 :
4331 0 : if self.current_logical_size.current_size().is_exact() {
4332 : // root timelines are initialized with exact count, but never start the background
4333 : // calculation
4334 0 : return;
4335 0 : }
4336 :
4337 0 : if let Some(await_bg_cancel) = self
4338 0 : .current_logical_size
4339 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4340 0 : .get()
4341 0 : {
4342 0 : await_bg_cancel.cancel();
4343 0 : } else {
4344 : // We should not wait if we were not able to explicitly instruct
4345 : // the logical size cancellation to skip the concurrency limit semaphore.
4346 : // TODO: this is an unexpected case. We should restructure so that it
4347 : // can't happen.
4348 0 : tracing::warn!(
4349 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4350 : );
4351 0 : debug_assert!(false);
4352 : }
4353 :
4354 : tokio::select!(
4355 : _ = self.current_logical_size.initialized.acquire() => {},
4356 : _ = self.cancel.cancelled() => {}
4357 : )
4358 0 : }
4359 :
4360 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4361 : /// Timelines layers up to the ancestor_lsn.
4362 : ///
4363 : /// Requires a timeline that:
4364 : /// - has an ancestor to detach from
4365 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4366 : /// a technical requirement
4367 : ///
4368 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4369 : /// polled again until completion.
4370 : ///
4371 : /// During the operation all timelines sharing the data with this timeline will be reparented
4372 : /// from our ancestor to be branches of this timeline.
4373 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4374 0 : self: &Arc<Timeline>,
4375 0 : tenant: &crate::tenant::Tenant,
4376 0 : options: detach_ancestor::Options,
4377 0 : ctx: &RequestContext,
4378 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
4379 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4380 0 : }
4381 :
4382 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
4383 : /// reparents any reparentable children of previous ancestor.
4384 : ///
4385 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
4386 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
4387 : /// successfully, tenant must be reloaded.
4388 : ///
4389 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
4390 : /// resetting the tenant.
4391 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
4392 0 : self: &Arc<Timeline>,
4393 0 : tenant: &crate::tenant::Tenant,
4394 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4395 0 : ctx: &RequestContext,
4396 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
4397 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
4398 0 : }
4399 :
4400 : /// Final step which unblocks the GC.
4401 : ///
4402 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
4403 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4404 0 : self: &Arc<Timeline>,
4405 0 : tenant: &crate::tenant::Tenant,
4406 0 : attempt: detach_ancestor::Attempt,
4407 0 : ctx: &RequestContext,
4408 0 : ) -> Result<(), detach_ancestor::Error> {
4409 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
4410 0 : }
4411 :
4412 : /// Switch aux file policy and schedule upload to the index part.
4413 16 : pub(crate) fn do_switch_aux_policy(&self, policy: AuxFilePolicy) -> anyhow::Result<()> {
4414 16 : self.last_aux_file_policy.store(Some(policy));
4415 16 : self.remote_client
4416 16 : .schedule_index_upload_for_aux_file_policy_update(Some(policy))?;
4417 16 : Ok(())
4418 16 : }
4419 : }
4420 :
4421 : impl Drop for Timeline {
4422 8 : fn drop(&mut self) {
4423 8 : if let Some(ancestor) = &self.ancestor_timeline {
4424 : // This lock should never be poisoned, but in case it is we do a .map() instead of
4425 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
4426 2 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
4427 2 : gc_info.remove_child(self.timeline_id)
4428 0 : }
4429 6 : }
4430 8 : }
4431 : }
4432 :
4433 : /// Top-level failure to compact.
4434 0 : #[derive(Debug, thiserror::Error)]
4435 : pub(crate) enum CompactionError {
4436 : #[error("The timeline or pageserver is shutting down")]
4437 : ShuttingDown,
4438 : /// Compaction cannot be done right now; page reconstruction and so on.
4439 : #[error(transparent)]
4440 : Other(anyhow::Error),
4441 : }
4442 :
4443 : impl From<CollectKeySpaceError> for CompactionError {
4444 0 : fn from(err: CollectKeySpaceError) -> Self {
4445 0 : match err {
4446 : CollectKeySpaceError::Cancelled
4447 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4448 0 : CompactionError::ShuttingDown
4449 : }
4450 0 : e => CompactionError::Other(e.into()),
4451 : }
4452 0 : }
4453 : }
4454 :
4455 : impl From<super::upload_queue::NotInitialized> for CompactionError {
4456 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
4457 0 : match value {
4458 : super::upload_queue::NotInitialized::Uninitialized => {
4459 0 : CompactionError::Other(anyhow::anyhow!(value))
4460 : }
4461 : super::upload_queue::NotInitialized::ShuttingDown
4462 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
4463 : }
4464 0 : }
4465 : }
4466 :
4467 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
4468 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
4469 0 : match e {
4470 : super::storage_layer::layer::DownloadError::TimelineShutdown
4471 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
4472 0 : CompactionError::ShuttingDown
4473 : }
4474 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
4475 : | super::storage_layer::layer::DownloadError::DownloadRequired
4476 : | super::storage_layer::layer::DownloadError::NotFile(_)
4477 : | super::storage_layer::layer::DownloadError::DownloadFailed
4478 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
4479 0 : CompactionError::Other(anyhow::anyhow!(e))
4480 : }
4481 : #[cfg(test)]
4482 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
4483 0 : CompactionError::Other(anyhow::anyhow!(e))
4484 : }
4485 : }
4486 0 : }
4487 : }
4488 :
4489 : impl From<layer_manager::Shutdown> for CompactionError {
4490 0 : fn from(_: layer_manager::Shutdown) -> Self {
4491 0 : CompactionError::ShuttingDown
4492 0 : }
4493 : }
4494 :
4495 : #[serde_as]
4496 196 : #[derive(serde::Serialize)]
4497 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4498 :
4499 : #[derive(Default)]
4500 : enum DurationRecorder {
4501 : #[default]
4502 : NotStarted,
4503 : Recorded(RecordedDuration, tokio::time::Instant),
4504 : }
4505 :
4506 : impl DurationRecorder {
4507 504 : fn till_now(&self) -> DurationRecorder {
4508 504 : match self {
4509 : DurationRecorder::NotStarted => {
4510 0 : panic!("must only call on recorded measurements")
4511 : }
4512 504 : DurationRecorder::Recorded(_, ended) => {
4513 504 : let now = tokio::time::Instant::now();
4514 504 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4515 504 : }
4516 504 : }
4517 504 : }
4518 196 : fn into_recorded(self) -> Option<RecordedDuration> {
4519 196 : match self {
4520 0 : DurationRecorder::NotStarted => None,
4521 196 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4522 : }
4523 196 : }
4524 : }
4525 :
4526 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
4527 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
4528 : /// the layer descriptor requires the user to provide the ranges, which should cover all
4529 : /// keys specified in the `data` field.
4530 : #[cfg(test)]
4531 : #[derive(Clone)]
4532 : pub struct DeltaLayerTestDesc {
4533 : pub lsn_range: Range<Lsn>,
4534 : pub key_range: Range<Key>,
4535 : pub data: Vec<(Key, Lsn, Value)>,
4536 : }
4537 :
4538 : #[cfg(test)]
4539 : impl DeltaLayerTestDesc {
4540 : #[allow(dead_code)]
4541 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
4542 2 : Self {
4543 2 : lsn_range,
4544 2 : key_range,
4545 2 : data,
4546 2 : }
4547 2 : }
4548 :
4549 50 : pub fn new_with_inferred_key_range(
4550 50 : lsn_range: Range<Lsn>,
4551 50 : data: Vec<(Key, Lsn, Value)>,
4552 50 : ) -> Self {
4553 112 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
4554 112 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
4555 50 : Self {
4556 50 : key_range: (*key_min)..(key_max.next()),
4557 50 : lsn_range,
4558 50 : data,
4559 50 : }
4560 50 : }
4561 :
4562 10 : pub(crate) fn layer_name(&self) -> LayerName {
4563 10 : LayerName::Delta(super::storage_layer::DeltaLayerName {
4564 10 : key_range: self.key_range.clone(),
4565 10 : lsn_range: self.lsn_range.clone(),
4566 10 : })
4567 10 : }
4568 : }
4569 :
4570 : impl Timeline {
4571 28 : async fn finish_compact_batch(
4572 28 : self: &Arc<Self>,
4573 28 : new_deltas: &[ResidentLayer],
4574 28 : new_images: &[ResidentLayer],
4575 28 : layers_to_remove: &[Layer],
4576 28 : ) -> Result<(), CompactionError> {
4577 28 : let mut guard = tokio::select! {
4578 : guard = self.layers.write() => guard,
4579 : _ = self.cancel.cancelled() => {
4580 : return Err(CompactionError::ShuttingDown);
4581 : }
4582 : };
4583 :
4584 28 : let mut duplicated_layers = HashSet::new();
4585 28 :
4586 28 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
4587 :
4588 336 : for l in new_deltas {
4589 308 : if guard.contains(l.as_ref()) {
4590 : // expected in tests
4591 0 : tracing::error!(layer=%l, "duplicated L1 layer");
4592 :
4593 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
4594 : // `guard` on self.layers. as of writing this, there are no error returns except
4595 : // for compact_level0_phase1 creating an L0, which does not happen in practice
4596 : // because we have not implemented L0 => L0 compaction.
4597 0 : duplicated_layers.insert(l.layer_desc().key());
4598 308 : } else if LayerMap::is_l0(&l.layer_desc().key_range) {
4599 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
4600 308 : } else {
4601 308 : insert_layers.push(l.clone());
4602 308 : }
4603 : }
4604 :
4605 : // only remove those inputs which were not outputs
4606 28 : let remove_layers: Vec<Layer> = layers_to_remove
4607 28 : .iter()
4608 402 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
4609 28 : .cloned()
4610 28 : .collect();
4611 28 :
4612 28 : if !new_images.is_empty() {
4613 0 : guard
4614 0 : .open_mut()?
4615 0 : .track_new_image_layers(new_images, &self.metrics);
4616 28 : }
4617 :
4618 28 : guard
4619 28 : .open_mut()?
4620 28 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
4621 28 :
4622 28 : self.remote_client
4623 28 : .schedule_compaction_update(&remove_layers, new_deltas)?;
4624 :
4625 28 : drop_wlock(guard);
4626 28 :
4627 28 : Ok(())
4628 28 : }
4629 :
4630 0 : async fn rewrite_layers(
4631 0 : self: &Arc<Self>,
4632 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
4633 0 : mut drop_layers: Vec<Layer>,
4634 0 : ) -> Result<(), CompactionError> {
4635 0 : let mut guard = self.layers.write().await;
4636 :
4637 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
4638 : // to avoid double-removing, and avoid rewriting something that was removed.
4639 0 : replace_layers.retain(|(l, _)| guard.contains(l));
4640 0 : drop_layers.retain(|l| guard.contains(l));
4641 0 :
4642 0 : guard
4643 0 : .open_mut()?
4644 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
4645 0 :
4646 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
4647 0 :
4648 0 : self.remote_client
4649 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
4650 :
4651 0 : Ok(())
4652 0 : }
4653 :
4654 : /// Schedules the uploads of the given image layers
4655 364 : fn upload_new_image_layers(
4656 364 : self: &Arc<Self>,
4657 364 : new_images: impl IntoIterator<Item = ResidentLayer>,
4658 364 : ) -> Result<(), super::upload_queue::NotInitialized> {
4659 390 : for layer in new_images {
4660 26 : self.remote_client.schedule_layer_file_upload(layer)?;
4661 : }
4662 : // should any new image layer been created, not uploading index_part will
4663 : // result in a mismatch between remote_physical_size and layermap calculated
4664 : // size, which will fail some tests, but should not be an issue otherwise.
4665 364 : self.remote_client
4666 364 : .schedule_index_upload_for_file_changes()?;
4667 364 : Ok(())
4668 364 : }
4669 :
4670 : /// Find the Lsns above which layer files need to be retained on
4671 : /// garbage collection.
4672 : ///
4673 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
4674 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
4675 : /// the space-based retention.
4676 : ///
4677 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
4678 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
4679 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
4680 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
4681 : /// in the response as the GC cutoff point for the timeline.
4682 1508 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
4683 : pub(super) async fn find_gc_cutoffs(
4684 : &self,
4685 : space_cutoff: Lsn,
4686 : pitr: Duration,
4687 : cancel: &CancellationToken,
4688 : ctx: &RequestContext,
4689 : ) -> Result<GcCutoffs, PageReconstructError> {
4690 : let _timer = self
4691 : .metrics
4692 : .find_gc_cutoffs_histo
4693 : .start_timer()
4694 : .record_on_drop();
4695 :
4696 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
4697 :
4698 : if cfg!(test) {
4699 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
4700 : if pitr == Duration::ZERO {
4701 : return Ok(GcCutoffs {
4702 : time: self.get_last_record_lsn(),
4703 : space: space_cutoff,
4704 : });
4705 : }
4706 : }
4707 :
4708 : // Calculate a time-based limit on how much to retain:
4709 : // - if PITR interval is set, then this is our cutoff.
4710 : // - if PITR interval is not set, then we do a lookup
4711 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
4712 : let time_cutoff = {
4713 : let now = SystemTime::now();
4714 : let time_range = if pitr == Duration::ZERO {
4715 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
4716 : } else {
4717 : pitr
4718 : };
4719 :
4720 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
4721 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
4722 : let timestamp = to_pg_timestamp(time_cutoff);
4723 :
4724 : match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
4725 : LsnForTimestamp::Present(lsn) => Some(lsn),
4726 : LsnForTimestamp::Future(lsn) => {
4727 : // The timestamp is in the future. That sounds impossible,
4728 : // but what it really means is that there hasn't been
4729 : // any commits since the cutoff timestamp.
4730 : //
4731 : // In this case we should use the LSN of the most recent commit,
4732 : // which is implicitly the last LSN in the log.
4733 : debug!("future({})", lsn);
4734 : Some(self.get_last_record_lsn())
4735 : }
4736 : LsnForTimestamp::Past(lsn) => {
4737 : debug!("past({})", lsn);
4738 : None
4739 : }
4740 : LsnForTimestamp::NoData(lsn) => {
4741 : debug!("nodata({})", lsn);
4742 : None
4743 : }
4744 : }
4745 : };
4746 :
4747 : Ok(match (pitr, time_cutoff) {
4748 : (Duration::ZERO, Some(time_cutoff)) => {
4749 : // PITR is not set. Retain the size-based limit, or the default time retention,
4750 : // whichever requires less data.
4751 : GcCutoffs {
4752 : time: self.get_last_record_lsn(),
4753 : space: std::cmp::max(time_cutoff, space_cutoff),
4754 : }
4755 : }
4756 : (Duration::ZERO, None) => {
4757 : // PITR is not set, and time lookup failed
4758 : GcCutoffs {
4759 : time: self.get_last_record_lsn(),
4760 : space: space_cutoff,
4761 : }
4762 : }
4763 : (_, None) => {
4764 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
4765 : // cannot advance beyond what was already GC'd, and respect space-based retention
4766 : GcCutoffs {
4767 : time: *self.get_latest_gc_cutoff_lsn(),
4768 : space: space_cutoff,
4769 : }
4770 : }
4771 : (_, Some(time_cutoff)) => {
4772 : // PITR interval is set and we looked up timestamp successfully. Ignore
4773 : // size based retention and make time cutoff authoritative
4774 : GcCutoffs {
4775 : time: time_cutoff,
4776 : space: time_cutoff,
4777 : }
4778 : }
4779 : })
4780 : }
4781 :
4782 : /// Garbage collect layer files on a timeline that are no longer needed.
4783 : ///
4784 : /// Currently, we don't make any attempt at removing unneeded page versions
4785 : /// within a layer file. We can only remove the whole file if it's fully
4786 : /// obsolete.
4787 754 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
4788 : // this is most likely the background tasks, but it might be the spawned task from
4789 : // immediate_gc
4790 754 : let _g = tokio::select! {
4791 : guard = self.gc_lock.lock() => guard,
4792 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
4793 : };
4794 754 : let timer = self.metrics.garbage_collect_histo.start_timer();
4795 :
4796 : fail_point!("before-timeline-gc");
4797 :
4798 : // Is the timeline being deleted?
4799 754 : if self.is_stopping() {
4800 0 : return Err(GcError::TimelineCancelled);
4801 754 : }
4802 754 :
4803 754 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
4804 754 : let gc_info = self.gc_info.read().unwrap();
4805 754 :
4806 754 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
4807 754 : let time_cutoff = gc_info.cutoffs.time;
4808 754 : let retain_lsns = gc_info
4809 754 : .retain_lsns
4810 754 : .iter()
4811 754 : .map(|(lsn, _child_id)| *lsn)
4812 754 : .collect();
4813 754 :
4814 754 : // Gets the maximum LSN that holds the valid lease.
4815 754 : //
4816 754 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
4817 754 : // Here, we do not check for stale leases again.
4818 754 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
4819 754 :
4820 754 : (
4821 754 : space_cutoff,
4822 754 : time_cutoff,
4823 754 : retain_lsns,
4824 754 : max_lsn_with_valid_lease,
4825 754 : )
4826 754 : };
4827 754 :
4828 754 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
4829 754 : let standby_horizon = self.standby_horizon.load();
4830 754 : // Hold GC for the standby, but as a safety guard do it only within some
4831 754 : // reasonable lag.
4832 754 : if standby_horizon != Lsn::INVALID {
4833 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
4834 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
4835 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
4836 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
4837 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
4838 : } else {
4839 0 : warn!(
4840 0 : "standby is lagging for more than {}MB, not holding gc for it",
4841 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
4842 : )
4843 : }
4844 0 : }
4845 754 : }
4846 :
4847 : // Reset standby horizon to ignore it if it is not updated till next GC.
4848 : // It is an easy way to unset it when standby disappears without adding
4849 : // more conf options.
4850 754 : self.standby_horizon.store(Lsn::INVALID);
4851 754 : self.metrics
4852 754 : .standby_horizon_gauge
4853 754 : .set(Lsn::INVALID.0 as i64);
4854 :
4855 754 : let res = self
4856 754 : .gc_timeline(
4857 754 : space_cutoff,
4858 754 : time_cutoff,
4859 754 : retain_lsns,
4860 754 : max_lsn_with_valid_lease,
4861 754 : new_gc_cutoff,
4862 754 : )
4863 754 : .instrument(
4864 754 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
4865 : )
4866 0 : .await?;
4867 :
4868 : // only record successes
4869 754 : timer.stop_and_record();
4870 754 :
4871 754 : Ok(res)
4872 754 : }
4873 :
4874 754 : async fn gc_timeline(
4875 754 : &self,
4876 754 : space_cutoff: Lsn,
4877 754 : time_cutoff: Lsn,
4878 754 : retain_lsns: Vec<Lsn>,
4879 754 : max_lsn_with_valid_lease: Option<Lsn>,
4880 754 : new_gc_cutoff: Lsn,
4881 754 : ) -> Result<GcResult, GcError> {
4882 754 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
4883 754 :
4884 754 : let now = SystemTime::now();
4885 754 : let mut result: GcResult = GcResult::default();
4886 754 :
4887 754 : // Nothing to GC. Return early.
4888 754 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
4889 754 : if latest_gc_cutoff >= new_gc_cutoff {
4890 22 : info!(
4891 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
4892 : );
4893 22 : return Ok(result);
4894 732 : }
4895 :
4896 : // We need to ensure that no one tries to read page versions or create
4897 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
4898 : // for details. This will block until the old value is no longer in use.
4899 : //
4900 : // The GC cutoff should only ever move forwards.
4901 732 : let waitlist = {
4902 732 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
4903 732 : if *write_guard > new_gc_cutoff {
4904 0 : return Err(GcError::BadLsn {
4905 0 : why: format!(
4906 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
4907 0 : *write_guard, new_gc_cutoff
4908 0 : ),
4909 0 : });
4910 732 : }
4911 732 :
4912 732 : write_guard.store_and_unlock(new_gc_cutoff)
4913 732 : };
4914 732 : waitlist.wait().await;
4915 :
4916 732 : info!("GC starting");
4917 :
4918 732 : debug!("retain_lsns: {:?}", retain_lsns);
4919 :
4920 732 : let mut layers_to_remove = Vec::new();
4921 :
4922 : // Scan all layers in the timeline (remote or on-disk).
4923 : //
4924 : // Garbage collect the layer if all conditions are satisfied:
4925 : // 1. it is older than cutoff LSN;
4926 : // 2. it is older than PITR interval;
4927 : // 3. it doesn't need to be retained for 'retain_lsns';
4928 : // 4. it does not need to be kept for LSNs holding valid leases.
4929 : // 5. newer on-disk image layers cover the layer's whole key range
4930 : //
4931 : // TODO holding a write lock is too agressive and avoidable
4932 732 : let mut guard = self.layers.write().await;
4933 732 : let layers = guard.layer_map()?;
4934 12418 : 'outer: for l in layers.iter_historic_layers() {
4935 12418 : result.layers_total += 1;
4936 12418 :
4937 12418 : // 1. Is it newer than GC horizon cutoff point?
4938 12418 : if l.get_lsn_range().end > space_cutoff {
4939 742 : debug!(
4940 0 : "keeping {} because it's newer than space_cutoff {}",
4941 0 : l.layer_name(),
4942 : space_cutoff,
4943 : );
4944 742 : result.layers_needed_by_cutoff += 1;
4945 742 : continue 'outer;
4946 11676 : }
4947 11676 :
4948 11676 : // 2. It is newer than PiTR cutoff point?
4949 11676 : if l.get_lsn_range().end > time_cutoff {
4950 0 : debug!(
4951 0 : "keeping {} because it's newer than time_cutoff {}",
4952 0 : l.layer_name(),
4953 : time_cutoff,
4954 : );
4955 0 : result.layers_needed_by_pitr += 1;
4956 0 : continue 'outer;
4957 11676 : }
4958 :
4959 : // 3. Is it needed by a child branch?
4960 : // NOTE With that we would keep data that
4961 : // might be referenced by child branches forever.
4962 : // We can track this in child timeline GC and delete parent layers when
4963 : // they are no longer needed. This might be complicated with long inheritance chains.
4964 : //
4965 : // TODO Vec is not a great choice for `retain_lsns`
4966 11676 : for retain_lsn in &retain_lsns {
4967 : // start_lsn is inclusive
4968 12 : if &l.get_lsn_range().start <= retain_lsn {
4969 12 : debug!(
4970 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
4971 0 : l.layer_name(),
4972 0 : retain_lsn,
4973 0 : l.is_incremental(),
4974 : );
4975 12 : result.layers_needed_by_branches += 1;
4976 12 : continue 'outer;
4977 0 : }
4978 : }
4979 :
4980 : // 4. Is there a valid lease that requires us to keep this layer?
4981 11664 : if let Some(lsn) = &max_lsn_with_valid_lease {
4982 : // keep if layer start <= any of the lease
4983 18 : if &l.get_lsn_range().start <= lsn {
4984 14 : debug!(
4985 0 : "keeping {} because there is a valid lease preventing GC at {}",
4986 0 : l.layer_name(),
4987 : lsn,
4988 : );
4989 14 : result.layers_needed_by_leases += 1;
4990 14 : continue 'outer;
4991 4 : }
4992 11646 : }
4993 :
4994 : // 5. Is there a later on-disk layer for this relation?
4995 : //
4996 : // The end-LSN is exclusive, while disk_consistent_lsn is
4997 : // inclusive. For example, if disk_consistent_lsn is 100, it is
4998 : // OK for a delta layer to have end LSN 101, but if the end LSN
4999 : // is 102, then it might not have been fully flushed to disk
5000 : // before crash.
5001 : //
5002 : // For example, imagine that the following layers exist:
5003 : //
5004 : // 1000 - image (A)
5005 : // 1000-2000 - delta (B)
5006 : // 2000 - image (C)
5007 : // 2000-3000 - delta (D)
5008 : // 3000 - image (E)
5009 : //
5010 : // If GC horizon is at 2500, we can remove layers A and B, but
5011 : // we cannot remove C, even though it's older than 2500, because
5012 : // the delta layer 2000-3000 depends on it.
5013 11650 : if !layers
5014 11650 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5015 : {
5016 11642 : debug!("keeping {} because it is the latest layer", l.layer_name());
5017 11642 : result.layers_not_updated += 1;
5018 11642 : continue 'outer;
5019 8 : }
5020 8 :
5021 8 : // We didn't find any reason to keep this file, so remove it.
5022 8 : debug!(
5023 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5024 0 : l.layer_name(),
5025 0 : l.is_incremental(),
5026 : );
5027 8 : layers_to_remove.push(l);
5028 : }
5029 :
5030 732 : if !layers_to_remove.is_empty() {
5031 : // Persist the new GC cutoff value before we actually remove anything.
5032 : // This unconditionally schedules also an index_part.json update, even though, we will
5033 : // be doing one a bit later with the unlinked gc'd layers.
5034 6 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5035 6 : self.schedule_uploads(disk_consistent_lsn, None)
5036 6 : .map_err(|e| {
5037 0 : if self.cancel.is_cancelled() {
5038 0 : GcError::TimelineCancelled
5039 : } else {
5040 0 : GcError::Remote(e)
5041 : }
5042 6 : })?;
5043 :
5044 6 : let gc_layers = layers_to_remove
5045 6 : .iter()
5046 8 : .map(|x| guard.get_from_desc(x))
5047 6 : .collect::<Vec<Layer>>();
5048 6 :
5049 6 : result.layers_removed = gc_layers.len() as u64;
5050 6 :
5051 6 : self.remote_client.schedule_gc_update(&gc_layers)?;
5052 :
5053 6 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5054 6 :
5055 6 : #[cfg(feature = "testing")]
5056 6 : {
5057 6 : result.doomed_layers = gc_layers;
5058 6 : }
5059 726 : }
5060 :
5061 732 : info!(
5062 0 : "GC completed removing {} layers, cutoff {}",
5063 : result.layers_removed, new_gc_cutoff
5064 : );
5065 :
5066 732 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5067 732 : Ok(result)
5068 754 : }
5069 :
5070 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5071 667041 : async fn reconstruct_value(
5072 667041 : &self,
5073 667041 : key: Key,
5074 667041 : request_lsn: Lsn,
5075 667041 : mut data: ValueReconstructState,
5076 667041 : ) -> Result<Bytes, PageReconstructError> {
5077 667041 : // Perform WAL redo if needed
5078 667041 : data.records.reverse();
5079 667041 :
5080 667041 : // If we have a page image, and no WAL, we're all set
5081 667041 : if data.records.is_empty() {
5082 666745 : if let Some((img_lsn, img)) = &data.img {
5083 666745 : trace!(
5084 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5085 : key,
5086 : img_lsn,
5087 : request_lsn,
5088 : );
5089 666745 : Ok(img.clone())
5090 : } else {
5091 0 : Err(PageReconstructError::from(anyhow!(
5092 0 : "base image for {key} at {request_lsn} not found"
5093 0 : )))
5094 : }
5095 : } else {
5096 : // We need to do WAL redo.
5097 : //
5098 : // If we don't have a base image, then the oldest WAL record better initialize
5099 : // the page
5100 296 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5101 0 : Err(PageReconstructError::from(anyhow!(
5102 0 : "Base image for {} at {} not found, but got {} WAL records",
5103 0 : key,
5104 0 : request_lsn,
5105 0 : data.records.len()
5106 0 : )))
5107 : } else {
5108 296 : if data.img.is_some() {
5109 296 : trace!(
5110 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5111 0 : data.records.len(),
5112 : key,
5113 : request_lsn
5114 : );
5115 : } else {
5116 0 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5117 : };
5118 296 : let res = self
5119 296 : .walredo_mgr
5120 296 : .as_ref()
5121 296 : .context("timeline has no walredo manager")
5122 296 : .map_err(PageReconstructError::WalRedo)?
5123 296 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5124 0 : .await;
5125 296 : let img = match res {
5126 296 : Ok(img) => img,
5127 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5128 0 : Err(walredo::Error::Other(e)) => {
5129 0 : return Err(PageReconstructError::WalRedo(
5130 0 : e.context("reconstruct a page image"),
5131 0 : ))
5132 : }
5133 : };
5134 296 : Ok(img)
5135 : }
5136 : }
5137 667041 : }
5138 :
5139 0 : pub(crate) async fn spawn_download_all_remote_layers(
5140 0 : self: Arc<Self>,
5141 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5142 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5143 0 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5144 0 :
5145 0 : // this is not really needed anymore; it has tests which really check the return value from
5146 0 : // http api. it would be better not to maintain this anymore.
5147 0 :
5148 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5149 0 : if let Some(st) = &*status_guard {
5150 0 : match &st.state {
5151 : DownloadRemoteLayersTaskState::Running => {
5152 0 : return Err(st.clone());
5153 : }
5154 : DownloadRemoteLayersTaskState::ShutDown
5155 0 : | DownloadRemoteLayersTaskState::Completed => {
5156 0 : *status_guard = None;
5157 0 : }
5158 : }
5159 0 : }
5160 :
5161 0 : let self_clone = Arc::clone(&self);
5162 0 : let task_id = task_mgr::spawn(
5163 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5164 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5165 0 : self.tenant_shard_id,
5166 0 : Some(self.timeline_id),
5167 0 : "download all remote layers task",
5168 0 : async move {
5169 0 : self_clone.download_all_remote_layers(request).await;
5170 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5171 0 : match &mut *status_guard {
5172 : None => {
5173 0 : warn!("tasks status is supposed to be Some(), since we are running");
5174 : }
5175 0 : Some(st) => {
5176 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5177 0 : if st.task_id != exp_task_id {
5178 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5179 0 : } else {
5180 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5181 0 : }
5182 : }
5183 : };
5184 0 : Ok(())
5185 0 : }
5186 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))
5187 : );
5188 :
5189 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5190 0 : task_id: format!("{task_id}"),
5191 0 : state: DownloadRemoteLayersTaskState::Running,
5192 0 : total_layer_count: 0,
5193 0 : successful_download_count: 0,
5194 0 : failed_download_count: 0,
5195 0 : };
5196 0 : *status_guard = Some(initial_info.clone());
5197 0 :
5198 0 : Ok(initial_info)
5199 0 : }
5200 :
5201 0 : async fn download_all_remote_layers(
5202 0 : self: &Arc<Self>,
5203 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5204 0 : ) {
5205 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5206 :
5207 0 : let remaining = {
5208 0 : let guard = self.layers.read().await;
5209 0 : let Ok(lm) = guard.layer_map() else {
5210 : // technically here we could look into iterating accessible layers, but downloading
5211 : // all layers of a shutdown timeline makes no sense regardless.
5212 0 : tracing::info!("attempted to download all layers of shutdown timeline");
5213 0 : return;
5214 : };
5215 0 : lm.iter_historic_layers()
5216 0 : .map(|desc| guard.get_from_desc(&desc))
5217 0 : .collect::<Vec<_>>()
5218 0 : };
5219 0 : let total_layer_count = remaining.len();
5220 0 :
5221 0 : macro_rules! lock_status {
5222 0 : ($st:ident) => {
5223 0 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5224 0 : let st = st
5225 0 : .as_mut()
5226 0 : .expect("this function is only called after the task has been spawned");
5227 0 : assert_eq!(
5228 0 : st.task_id,
5229 0 : format!(
5230 0 : "{}",
5231 0 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5232 0 : )
5233 0 : );
5234 0 : let $st = st;
5235 0 : };
5236 0 : }
5237 0 :
5238 0 : {
5239 0 : lock_status!(st);
5240 0 : st.total_layer_count = total_layer_count as u64;
5241 0 : }
5242 0 :
5243 0 : let mut remaining = remaining.into_iter();
5244 0 : let mut have_remaining = true;
5245 0 : let mut js = tokio::task::JoinSet::new();
5246 0 :
5247 0 : let cancel = task_mgr::shutdown_token();
5248 0 :
5249 0 : let limit = request.max_concurrent_downloads;
5250 :
5251 : loop {
5252 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5253 0 : let Some(next) = remaining.next() else {
5254 0 : have_remaining = false;
5255 0 : break;
5256 : };
5257 :
5258 0 : let span = tracing::info_span!("download", layer = %next);
5259 :
5260 0 : js.spawn(
5261 0 : async move {
5262 0 : let res = next.download().await;
5263 0 : (next, res)
5264 0 : }
5265 0 : .instrument(span),
5266 0 : );
5267 0 : }
5268 :
5269 0 : while let Some(res) = js.join_next().await {
5270 0 : match res {
5271 : Ok((_, Ok(_))) => {
5272 0 : lock_status!(st);
5273 0 : st.successful_download_count += 1;
5274 : }
5275 0 : Ok((layer, Err(e))) => {
5276 0 : tracing::error!(%layer, "download failed: {e:#}");
5277 0 : lock_status!(st);
5278 0 : st.failed_download_count += 1;
5279 : }
5280 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5281 0 : Err(je) if je.is_panic() => {
5282 0 : lock_status!(st);
5283 0 : st.failed_download_count += 1;
5284 : }
5285 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5286 : }
5287 : }
5288 :
5289 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5290 0 : break;
5291 0 : }
5292 : }
5293 :
5294 : {
5295 0 : lock_status!(st);
5296 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5297 : }
5298 0 : }
5299 :
5300 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5301 0 : &self,
5302 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5303 0 : self.download_all_remote_layers_task_info
5304 0 : .read()
5305 0 : .unwrap()
5306 0 : .clone()
5307 0 : }
5308 : }
5309 :
5310 : impl Timeline {
5311 : /// Returns non-remote layers for eviction.
5312 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5313 0 : let guard = self.layers.read().await;
5314 0 : let mut max_layer_size: Option<u64> = None;
5315 0 :
5316 0 : let resident_layers = guard
5317 0 : .likely_resident_layers()
5318 0 : .map(|layer| {
5319 0 : let file_size = layer.layer_desc().file_size;
5320 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5321 0 :
5322 0 : let last_activity_ts = layer.latest_activity();
5323 0 :
5324 0 : EvictionCandidate {
5325 0 : layer: layer.to_owned().into(),
5326 0 : last_activity_ts,
5327 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5328 0 : visibility: layer.visibility(),
5329 0 : }
5330 0 : })
5331 0 : .collect();
5332 0 :
5333 0 : DiskUsageEvictionInfo {
5334 0 : max_layer_size,
5335 0 : resident_layers,
5336 0 : }
5337 0 : }
5338 :
5339 1682 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5340 1682 : ShardIndex {
5341 1682 : shard_number: self.tenant_shard_id.shard_number,
5342 1682 : shard_count: self.tenant_shard_id.shard_count,
5343 1682 : }
5344 1682 : }
5345 :
5346 : /// Persistently blocks gc for `Manual` reason.
5347 : ///
5348 : /// Returns true if no such block existed before, false otherwise.
5349 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
5350 0 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5351 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5352 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
5353 0 : }
5354 :
5355 : /// Persistently unblocks gc for `Manual` reason.
5356 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
5357 0 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5358 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5359 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
5360 0 : }
5361 :
5362 : #[cfg(test)]
5363 32 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5364 32 : self.last_record_lsn.advance(new_lsn);
5365 32 : }
5366 :
5367 : #[cfg(test)]
5368 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
5369 2 : self.disk_consistent_lsn.store(new_value);
5370 2 : }
5371 :
5372 : /// Force create an image layer and place it into the layer map.
5373 : ///
5374 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5375 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are placed into the layer map in one run.
5376 : #[cfg(test)]
5377 44 : pub(super) async fn force_create_image_layer(
5378 44 : self: &Arc<Timeline>,
5379 44 : lsn: Lsn,
5380 44 : mut images: Vec<(Key, Bytes)>,
5381 44 : check_start_lsn: Option<Lsn>,
5382 44 : ctx: &RequestContext,
5383 44 : ) -> anyhow::Result<()> {
5384 44 : let last_record_lsn = self.get_last_record_lsn();
5385 44 : assert!(
5386 44 : lsn <= last_record_lsn,
5387 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5388 : );
5389 44 : if let Some(check_start_lsn) = check_start_lsn {
5390 44 : assert!(lsn >= check_start_lsn);
5391 0 : }
5392 74 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5393 44 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5394 44 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5395 44 : let mut image_layer_writer = ImageLayerWriter::new(
5396 44 : self.conf,
5397 44 : self.timeline_id,
5398 44 : self.tenant_shard_id,
5399 44 : &(min_key..end_key),
5400 44 : lsn,
5401 44 : ctx,
5402 44 : )
5403 22 : .await?;
5404 162 : for (key, img) in images {
5405 118 : image_layer_writer.put_image(key, img, ctx).await?;
5406 : }
5407 88 : let image_layer = image_layer_writer.finish(self, ctx).await?;
5408 :
5409 44 : {
5410 44 : let mut guard = self.layers.write().await;
5411 44 : guard.open_mut().unwrap().force_insert_layer(image_layer);
5412 44 : }
5413 44 :
5414 44 : Ok(())
5415 44 : }
5416 :
5417 : /// Force create a delta layer and place it into the layer map.
5418 : ///
5419 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5420 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are placed into the layer map in one run.
5421 : #[cfg(test)]
5422 52 : pub(super) async fn force_create_delta_layer(
5423 52 : self: &Arc<Timeline>,
5424 52 : mut deltas: DeltaLayerTestDesc,
5425 52 : check_start_lsn: Option<Lsn>,
5426 52 : ctx: &RequestContext,
5427 52 : ) -> anyhow::Result<()> {
5428 52 : let last_record_lsn = self.get_last_record_lsn();
5429 52 : deltas
5430 52 : .data
5431 62 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5432 52 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5433 52 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5434 166 : for (_, lsn, _) in &deltas.data {
5435 114 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5436 : }
5437 52 : assert!(
5438 52 : deltas.lsn_range.end <= last_record_lsn,
5439 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
5440 : deltas.lsn_range.end,
5441 : last_record_lsn
5442 : );
5443 52 : if let Some(check_start_lsn) = check_start_lsn {
5444 52 : assert!(deltas.lsn_range.start >= check_start_lsn);
5445 0 : }
5446 : // check if the delta layer does not violate the LSN invariant, the legacy compaction should always produce a batch of
5447 : // layers of the same start/end LSN, and so should the force inserted layer
5448 : {
5449 : /// Checks if a overlaps with b, assume a/b = [start, end).
5450 54 : pub fn overlaps_with<T: Ord>(a: &Range<T>, b: &Range<T>) -> bool {
5451 54 : !(a.end <= b.start || b.end <= a.start)
5452 54 : }
5453 :
5454 52 : let guard = self.layers.read().await;
5455 100 : for layer in guard.layer_map()?.iter_historic_layers() {
5456 100 : if layer.is_delta()
5457 54 : && overlaps_with(&layer.lsn_range, &deltas.lsn_range)
5458 16 : && layer.lsn_range != deltas.lsn_range
5459 : {
5460 : // If a delta layer overlaps with another delta layer AND their LSN range is not the same, panic
5461 0 : panic!(
5462 0 : "inserted layer violates delta layer LSN invariant: current_lsn_range={}..{}, conflict_lsn_range={}..{}",
5463 0 : deltas.lsn_range.start, deltas.lsn_range.end, layer.lsn_range.start, layer.lsn_range.end
5464 0 : );
5465 100 : }
5466 : }
5467 : }
5468 52 : let mut delta_layer_writer = DeltaLayerWriter::new(
5469 52 : self.conf,
5470 52 : self.timeline_id,
5471 52 : self.tenant_shard_id,
5472 52 : deltas.key_range.start,
5473 52 : deltas.lsn_range,
5474 52 : ctx,
5475 52 : )
5476 26 : .await?;
5477 166 : for (key, lsn, val) in deltas.data {
5478 114 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5479 : }
5480 130 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
5481 52 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5482 :
5483 52 : {
5484 52 : let mut guard = self.layers.write().await;
5485 52 : guard.open_mut().unwrap().force_insert_layer(delta_layer);
5486 52 : }
5487 52 :
5488 52 : Ok(())
5489 52 : }
5490 :
5491 : /// Return all keys at the LSN in the image layers
5492 : #[cfg(test)]
5493 6 : pub(crate) async fn inspect_image_layers(
5494 6 : self: &Arc<Timeline>,
5495 6 : lsn: Lsn,
5496 6 : ctx: &RequestContext,
5497 6 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
5498 6 : let mut all_data = Vec::new();
5499 6 : let guard = self.layers.read().await;
5500 34 : for layer in guard.layer_map()?.iter_historic_layers() {
5501 34 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
5502 8 : let layer = guard.get_from_desc(&layer);
5503 8 : let mut reconstruct_data = ValuesReconstructState::default();
5504 8 : layer
5505 8 : .get_values_reconstruct_data(
5506 8 : KeySpace::single(Key::MIN..Key::MAX),
5507 8 : lsn..Lsn(lsn.0 + 1),
5508 8 : &mut reconstruct_data,
5509 8 : ctx,
5510 8 : )
5511 13 : .await?;
5512 80 : for (k, v) in reconstruct_data.keys {
5513 72 : all_data.push((k, v?.img.unwrap().1));
5514 : }
5515 26 : }
5516 : }
5517 6 : all_data.sort();
5518 6 : Ok(all_data)
5519 6 : }
5520 :
5521 : /// Get all historic layer descriptors in the layer map
5522 : #[cfg(test)]
5523 2 : pub(crate) async fn inspect_historic_layers(
5524 2 : self: &Arc<Timeline>,
5525 2 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
5526 2 : let mut layers = Vec::new();
5527 2 : let guard = self.layers.read().await;
5528 6 : for layer in guard.layer_map()?.iter_historic_layers() {
5529 6 : layers.push(layer.key());
5530 6 : }
5531 2 : Ok(layers)
5532 2 : }
5533 :
5534 : #[cfg(test)]
5535 10 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
5536 10 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
5537 10 : keyspace.merge(&ks);
5538 10 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
5539 10 : }
5540 : }
5541 :
5542 : /// Tracking writes ingestion does to a particular in-memory layer.
5543 : ///
5544 : /// Cleared upon freezing a layer.
5545 : pub(crate) struct TimelineWriterState {
5546 : open_layer: Arc<InMemoryLayer>,
5547 : current_size: u64,
5548 : // Previous Lsn which passed through
5549 : prev_lsn: Option<Lsn>,
5550 : // Largest Lsn which passed through the current writer
5551 : max_lsn: Option<Lsn>,
5552 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
5553 : cached_last_freeze_at: Lsn,
5554 : }
5555 :
5556 : impl TimelineWriterState {
5557 1264 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
5558 1264 : Self {
5559 1264 : open_layer,
5560 1264 : current_size,
5561 1264 : prev_lsn: None,
5562 1264 : max_lsn: None,
5563 1264 : cached_last_freeze_at: last_freeze_at,
5564 1264 : }
5565 1264 : }
5566 : }
5567 :
5568 : /// Various functions to mutate the timeline.
5569 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
5570 : // This is probably considered a bad practice in Rust and should be fixed eventually,
5571 : // but will cause large code changes.
5572 : pub(crate) struct TimelineWriter<'a> {
5573 : tl: &'a Timeline,
5574 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
5575 : }
5576 :
5577 : impl Deref for TimelineWriter<'_> {
5578 : type Target = Timeline;
5579 :
5580 4807358 : fn deref(&self) -> &Self::Target {
5581 4807358 : self.tl
5582 4807358 : }
5583 : }
5584 :
5585 : #[derive(PartialEq)]
5586 : enum OpenLayerAction {
5587 : Roll,
5588 : Open,
5589 : None,
5590 : }
5591 :
5592 : impl<'a> TimelineWriter<'a> {
5593 : /// Put a new page version that can be constructed from a WAL record
5594 : ///
5595 : /// This will implicitly extend the relation, if the page is beyond the
5596 : /// current end-of-file.
5597 5090610 : pub(crate) async fn put(
5598 5090610 : &mut self,
5599 5090610 : key: Key,
5600 5090610 : lsn: Lsn,
5601 5090610 : value: &Value,
5602 5090610 : ctx: &RequestContext,
5603 5090610 : ) -> anyhow::Result<()> {
5604 5090610 : // Avoid doing allocations for "small" values.
5605 5090610 : // In the regression test suite, the limit of 256 avoided allocations in 95% of cases:
5606 5090610 : // https://github.com/neondatabase/neon/pull/5056#discussion_r1301975061
5607 5090610 : let mut buf = smallvec::SmallVec::<[u8; 256]>::new();
5608 5090610 : value.ser_into(&mut buf)?;
5609 5090610 : let buf_size: u64 = buf.len().try_into().expect("oversized value buf");
5610 5090610 :
5611 5090610 : let action = self.get_open_layer_action(lsn, buf_size);
5612 5090610 : let layer = self.handle_open_layer_action(lsn, action, ctx).await?;
5613 5090610 : let res = layer.put_value(key.to_compact(), lsn, &buf, ctx).await;
5614 :
5615 5090610 : if res.is_ok() {
5616 5090610 : // Update the current size only when the entire write was ok.
5617 5090610 : // In case of failures, we may have had partial writes which
5618 5090610 : // render the size tracking out of sync. That's ok because
5619 5090610 : // the checkpoint distance should be significantly smaller
5620 5090610 : // than the S3 single shot upload limit of 5GiB.
5621 5090610 : let state = self.write_guard.as_mut().unwrap();
5622 5090610 :
5623 5090610 : state.current_size += buf_size;
5624 5090610 : state.prev_lsn = Some(lsn);
5625 5090610 : state.max_lsn = std::cmp::max(state.max_lsn, Some(lsn));
5626 5090610 : }
5627 :
5628 5090610 : res
5629 5090610 : }
5630 :
5631 5090612 : async fn handle_open_layer_action(
5632 5090612 : &mut self,
5633 5090612 : at: Lsn,
5634 5090612 : action: OpenLayerAction,
5635 5090612 : ctx: &RequestContext,
5636 5090612 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
5637 5090612 : match action {
5638 : OpenLayerAction::Roll => {
5639 80 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
5640 80 : self.roll_layer(freeze_at).await?;
5641 80 : self.open_layer(at, ctx).await?;
5642 : }
5643 1184 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
5644 : OpenLayerAction::None => {
5645 5089348 : assert!(self.write_guard.is_some());
5646 : }
5647 : }
5648 :
5649 5090612 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
5650 5090612 : }
5651 :
5652 1264 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
5653 1264 : let layer = self
5654 1264 : .tl
5655 1264 : .get_layer_for_write(at, &self.write_guard, ctx)
5656 713 : .await?;
5657 1264 : let initial_size = layer.size().await?;
5658 :
5659 1264 : let last_freeze_at = self.last_freeze_at.load();
5660 1264 : self.write_guard.replace(TimelineWriterState::new(
5661 1264 : layer,
5662 1264 : initial_size,
5663 1264 : last_freeze_at,
5664 1264 : ));
5665 1264 :
5666 1264 : Ok(())
5667 1264 : }
5668 :
5669 80 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
5670 80 : let current_size = self.write_guard.as_ref().unwrap().current_size;
5671 80 :
5672 80 : // self.write_guard will be taken by the freezing
5673 80 : self.tl
5674 80 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
5675 2 : .await?;
5676 :
5677 80 : assert!(self.write_guard.is_none());
5678 :
5679 80 : if current_size >= self.get_checkpoint_distance() * 2 {
5680 0 : warn!("Flushed oversized open layer with size {}", current_size)
5681 80 : }
5682 :
5683 80 : Ok(())
5684 80 : }
5685 :
5686 5090612 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
5687 5090612 : let state = &*self.write_guard;
5688 5090612 : let Some(state) = &state else {
5689 1184 : return OpenLayerAction::Open;
5690 : };
5691 :
5692 : #[cfg(feature = "testing")]
5693 5089428 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
5694 : // this check and assertion are not really needed because
5695 : // LayerManager::try_freeze_in_memory_layer will always clear out the
5696 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
5697 : // is no TimelineWriterState.
5698 0 : assert!(
5699 0 : state.open_layer.end_lsn.get().is_some(),
5700 0 : "our open_layer must be outdated"
5701 : );
5702 :
5703 : // this would be a memory leak waiting to happen because the in-memory layer always has
5704 : // an index
5705 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
5706 5089428 : }
5707 5089428 :
5708 5089428 : if state.prev_lsn == Some(lsn) {
5709 : // Rolling mid LSN is not supported by [downstream code].
5710 : // Hence, only roll at LSN boundaries.
5711 : //
5712 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
5713 286402 : return OpenLayerAction::None;
5714 4803026 : }
5715 4803026 :
5716 4803026 : if state.current_size == 0 {
5717 : // Don't roll empty layers
5718 0 : return OpenLayerAction::None;
5719 4803026 : }
5720 4803026 :
5721 4803026 : if self.tl.should_roll(
5722 4803026 : state.current_size,
5723 4803026 : state.current_size + new_value_size,
5724 4803026 : self.get_checkpoint_distance(),
5725 4803026 : lsn,
5726 4803026 : state.cached_last_freeze_at,
5727 4803026 : state.open_layer.get_opened_at(),
5728 4803026 : ) {
5729 80 : OpenLayerAction::Roll
5730 : } else {
5731 4802946 : OpenLayerAction::None
5732 : }
5733 5090612 : }
5734 :
5735 : /// Put a batch of keys at the specified Lsns.
5736 : ///
5737 : /// The batch is sorted by Lsn (enforced by usage of [`utils::vec_map::VecMap`].
5738 414054 : pub(crate) async fn put_batch(
5739 414054 : &mut self,
5740 414054 : batch: VecMap<Lsn, (Key, Value)>,
5741 414054 : ctx: &RequestContext,
5742 414054 : ) -> anyhow::Result<()> {
5743 1114510 : for (lsn, (key, val)) in batch {
5744 700456 : self.put(key, lsn, &val, ctx).await?
5745 : }
5746 :
5747 414054 : Ok(())
5748 414054 : }
5749 :
5750 2 : pub(crate) async fn delete_batch(
5751 2 : &mut self,
5752 2 : batch: &[(Range<Key>, Lsn)],
5753 2 : ctx: &RequestContext,
5754 2 : ) -> anyhow::Result<()> {
5755 2 : if let Some((_, lsn)) = batch.first() {
5756 2 : let action = self.get_open_layer_action(*lsn, 0);
5757 2 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
5758 2 : layer.put_tombstones(batch).await?;
5759 0 : }
5760 :
5761 2 : Ok(())
5762 2 : }
5763 :
5764 : /// Track the end of the latest digested WAL record.
5765 : /// Remember the (end of) last valid WAL record remembered in the timeline.
5766 : ///
5767 : /// Call this after you have finished writing all the WAL up to 'lsn'.
5768 : ///
5769 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
5770 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
5771 : /// the latest and can be read.
5772 5279070 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
5773 5279070 : self.tl.finish_write(new_lsn);
5774 5279070 : }
5775 :
5776 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
5777 270570 : self.tl.update_current_logical_size(delta)
5778 270570 : }
5779 : }
5780 :
5781 : // We need TimelineWriter to be send in upcoming conversion of
5782 : // Timeline::layers to tokio::sync::RwLock.
5783 : #[test]
5784 2 : fn is_send() {
5785 2 : fn _assert_send<T: Send>() {}
5786 2 : _assert_send::<TimelineWriter<'_>>();
5787 2 : }
5788 :
5789 : #[cfg(test)]
5790 : mod tests {
5791 : use pageserver_api::key::Key;
5792 : use utils::{id::TimelineId, lsn::Lsn};
5793 :
5794 : use crate::{
5795 : repository::Value,
5796 : tenant::{
5797 : harness::{test_img, TenantHarness},
5798 : layer_map::LayerMap,
5799 : storage_layer::{Layer, LayerName},
5800 : timeline::{DeltaLayerTestDesc, EvictionError},
5801 : Timeline,
5802 : },
5803 : };
5804 :
5805 : #[tokio::test]
5806 2 : async fn test_heatmap_generation() {
5807 2 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
5808 2 :
5809 2 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
5810 2 : Lsn(0x10)..Lsn(0x20),
5811 2 : vec![(
5812 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
5813 2 : Lsn(0x11),
5814 2 : Value::Image(test_img("foo")),
5815 2 : )],
5816 2 : );
5817 2 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
5818 2 : Lsn(0x10)..Lsn(0x20),
5819 2 : vec![(
5820 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
5821 2 : Lsn(0x11),
5822 2 : Value::Image(test_img("foo")),
5823 2 : )],
5824 2 : );
5825 2 : let l0_delta = DeltaLayerTestDesc::new(
5826 2 : Lsn(0x20)..Lsn(0x30),
5827 2 : Key::from_hex("000000000000000000000000000000000000").unwrap()
5828 2 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
5829 2 : vec![(
5830 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
5831 2 : Lsn(0x25),
5832 2 : Value::Image(test_img("foo")),
5833 2 : )],
5834 2 : );
5835 2 : let delta_layers = vec![
5836 2 : covered_delta.clone(),
5837 2 : visible_delta.clone(),
5838 2 : l0_delta.clone(),
5839 2 : ];
5840 2 :
5841 2 : let image_layer = (
5842 2 : Lsn(0x40),
5843 2 : vec![(
5844 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
5845 2 : test_img("bar"),
5846 2 : )],
5847 2 : );
5848 2 : let image_layers = vec![image_layer];
5849 2 :
5850 8 : let (tenant, ctx) = harness.load().await;
5851 2 : let timeline = tenant
5852 2 : .create_test_timeline_with_layers(
5853 2 : TimelineId::generate(),
5854 2 : Lsn(0x10),
5855 2 : 14,
5856 2 : &ctx,
5857 2 : delta_layers,
5858 2 : image_layers,
5859 2 : Lsn(0x100),
5860 2 : )
5861 29 : .await
5862 2 : .unwrap();
5863 2 :
5864 2 : // Layer visibility is an input to heatmap generation, so refresh it first
5865 2 : timeline.update_layer_visibility().await.unwrap();
5866 2 :
5867 2 : let heatmap = timeline
5868 2 : .generate_heatmap()
5869 2 : .await
5870 2 : .expect("Infallible while timeline is not shut down");
5871 2 :
5872 2 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
5873 2 :
5874 2 : // L0 should come last
5875 2 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
5876 2 :
5877 2 : let mut last_lsn = Lsn::MAX;
5878 10 : for layer in heatmap.layers {
5879 2 : // Covered layer should be omitted
5880 8 : assert!(layer.name != covered_delta.layer_name());
5881 2 :
5882 8 : let layer_lsn = match &layer.name {
5883 4 : LayerName::Delta(d) => d.lsn_range.end,
5884 4 : LayerName::Image(i) => i.lsn,
5885 2 : };
5886 2 :
5887 2 : // Apart from L0s, newest Layers should come first
5888 8 : if !LayerMap::is_l0(layer.name.key_range()) {
5889 6 : assert!(layer_lsn <= last_lsn);
5890 6 : last_lsn = layer_lsn;
5891 2 : }
5892 2 : }
5893 2 : }
5894 :
5895 : #[tokio::test]
5896 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
5897 2 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
5898 2 : .await
5899 2 : .unwrap();
5900 2 :
5901 8 : let (tenant, ctx) = harness.load().await;
5902 2 : let timeline = tenant
5903 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
5904 4 : .await
5905 2 : .unwrap();
5906 2 :
5907 2 : let layer = find_some_layer(&timeline).await;
5908 2 : let layer = layer
5909 2 : .keep_resident()
5910 2 : .await
5911 2 : .expect("no download => no downloading errors")
5912 2 : .drop_eviction_guard();
5913 2 :
5914 2 : let forever = std::time::Duration::from_secs(120);
5915 2 :
5916 2 : let first = layer.evict_and_wait(forever);
5917 2 : let second = layer.evict_and_wait(forever);
5918 2 :
5919 2 : let (first, second) = tokio::join!(first, second);
5920 2 :
5921 2 : let res = layer.keep_resident().await;
5922 2 : assert!(res.is_none(), "{res:?}");
5923 2 :
5924 2 : match (first, second) {
5925 2 : (Ok(()), Ok(())) => {
5926 2 : // because there are no more timeline locks being taken on eviction path, we can
5927 2 : // witness all three outcomes here.
5928 2 : }
5929 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
5930 0 : // if one completes before the other, this is fine just as well.
5931 0 : }
5932 2 : other => unreachable!("unexpected {:?}", other),
5933 2 : }
5934 2 : }
5935 :
5936 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
5937 2 : let layers = timeline.layers.read().await;
5938 2 : let desc = layers
5939 2 : .layer_map()
5940 2 : .unwrap()
5941 2 : .iter_historic_layers()
5942 2 : .next()
5943 2 : .expect("must find one layer to evict");
5944 2 :
5945 2 : layers.get_from_desc(&desc)
5946 2 : }
5947 : }
|
