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 : AUX_FILES_KEY, KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX,
26 : NON_INHERITED_RANGE, 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::{
63 : cmp::{max, min},
64 : ops::ControlFlow,
65 : };
66 : use std::{
67 : collections::btree_map::Entry,
68 : ops::{Deref, Range},
69 : };
70 :
71 : use crate::{
72 : aux_file::AuxFileSizeEstimator,
73 : tenant::{
74 : config::defaults::DEFAULT_PITR_INTERVAL,
75 : layer_map::{LayerMap, SearchResult},
76 : metadata::TimelineMetadata,
77 : storage_layer::PersistentLayerDesc,
78 : },
79 : walredo,
80 : };
81 : use crate::{
82 : context::{DownloadBehavior, RequestContext},
83 : disk_usage_eviction_task::DiskUsageEvictionInfo,
84 : pgdatadir_mapping::CollectKeySpaceError,
85 : };
86 : use crate::{
87 : disk_usage_eviction_task::finite_f32,
88 : tenant::storage_layer::{
89 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
90 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructResult,
91 : ValueReconstructState, ValuesReconstructState,
92 : },
93 : };
94 : use crate::{
95 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
96 : };
97 : use crate::{
98 : l0_flush::{self, L0FlushGlobalState},
99 : metrics::GetKind,
100 : };
101 : use crate::{
102 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
103 : };
104 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
105 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
106 : use crate::{
107 : pgdatadir_mapping::{AuxFilesDirectory, DirectoryKind},
108 : virtual_file::{MaybeFatalIo, VirtualFile},
109 : };
110 :
111 : use crate::config::PageServerConf;
112 : use crate::keyspace::{KeyPartitioning, KeySpace};
113 : use crate::metrics::TimelineMetrics;
114 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
115 : use crate::tenant::config::TenantConfOpt;
116 : use pageserver_api::reltag::RelTag;
117 : use pageserver_api::shard::ShardIndex;
118 :
119 : use postgres_connection::PgConnectionConfig;
120 : use postgres_ffi::to_pg_timestamp;
121 : use utils::{
122 : completion,
123 : generation::Generation,
124 : id::TimelineId,
125 : lsn::{AtomicLsn, Lsn, RecordLsn},
126 : seqwait::SeqWait,
127 : simple_rcu::{Rcu, RcuReadGuard},
128 : };
129 :
130 : use crate::repository::GcResult;
131 : use crate::repository::{Key, Value};
132 : use crate::task_mgr;
133 : use crate::task_mgr::TaskKind;
134 : use crate::ZERO_PAGE;
135 :
136 : use self::delete::DeleteTimelineFlow;
137 : pub(super) use self::eviction_task::EvictionTaskTenantState;
138 : use self::eviction_task::EvictionTaskTimelineState;
139 : use self::layer_manager::LayerManager;
140 : use self::logical_size::LogicalSize;
141 : use self::walreceiver::{WalReceiver, WalReceiverConf};
142 :
143 : use super::{config::TenantConf, upload_queue::NotInitialized};
144 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
145 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
146 : use super::{
147 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
148 : storage_layer::ReadableLayer,
149 : };
150 : use super::{
151 : secondary::heatmap::{HeatMapLayer, HeatMapTimeline},
152 : GcError,
153 : };
154 :
155 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
156 : pub(crate) enum FlushLoopState {
157 : NotStarted,
158 : Running {
159 : #[cfg(test)]
160 : expect_initdb_optimization: bool,
161 : #[cfg(test)]
162 : initdb_optimization_count: usize,
163 : },
164 : Exited,
165 : }
166 :
167 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
168 : pub enum ImageLayerCreationMode {
169 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
170 : Try,
171 : /// Force creating the image layers if possible. For now, no image layers will be created
172 : /// for metadata keys. Used in compaction code path with force flag enabled.
173 : Force,
174 : /// Initial ingestion of the data, and no data should be dropped in this function. This
175 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
176 : /// code path.
177 : Initial,
178 : }
179 :
180 : impl std::fmt::Display for ImageLayerCreationMode {
181 522 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
182 522 : write!(f, "{:?}", self)
183 522 : }
184 : }
185 :
186 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
187 : /// Can be removed after all refactors are done.
188 28 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
189 28 : drop(rlock)
190 28 : }
191 :
192 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
193 : /// Can be removed after all refactors are done.
194 550 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
195 550 : drop(rlock)
196 550 : }
197 :
198 : /// The outward-facing resources required to build a Timeline
199 : pub struct TimelineResources {
200 : pub remote_client: RemoteTimelineClient,
201 : pub timeline_get_throttle: Arc<
202 : crate::tenant::throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>,
203 : >,
204 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
205 : }
206 :
207 : pub(crate) struct AuxFilesState {
208 : pub(crate) dir: Option<AuxFilesDirectory>,
209 : pub(crate) n_deltas: usize,
210 : }
211 :
212 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
213 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
214 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
215 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
216 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
217 : pub(crate) struct RelSizeCache {
218 : pub(crate) complete_as_of: Lsn,
219 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
220 : }
221 :
222 : pub struct Timeline {
223 : conf: &'static PageServerConf,
224 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
225 :
226 : myself: Weak<Self>,
227 :
228 : pub(crate) tenant_shard_id: TenantShardId,
229 : pub timeline_id: TimelineId,
230 :
231 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
232 : /// Never changes for the lifetime of this [`Timeline`] object.
233 : ///
234 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
235 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
236 : pub(crate) generation: Generation,
237 :
238 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
239 : /// to shards, and is constant through the lifetime of this Timeline.
240 : shard_identity: ShardIdentity,
241 :
242 : pub pg_version: u32,
243 :
244 : /// The tuple has two elements.
245 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
246 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
247 : ///
248 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
249 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
250 : ///
251 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
252 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
253 : /// `PersistentLayerDesc`'s.
254 : ///
255 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
256 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
257 : /// runtime, e.g., during page reconstruction.
258 : ///
259 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
260 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
261 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
262 :
263 : last_freeze_at: AtomicLsn,
264 : // Atomic would be more appropriate here.
265 : last_freeze_ts: RwLock<Instant>,
266 :
267 : pub(crate) standby_horizon: AtomicLsn,
268 :
269 : // WAL redo manager. `None` only for broken tenants.
270 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
271 :
272 : /// Remote storage client.
273 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
274 : pub remote_client: Arc<RemoteTimelineClient>,
275 :
276 : // What page versions do we hold in the repository? If we get a
277 : // request > last_record_lsn, we need to wait until we receive all
278 : // the WAL up to the request. The SeqWait provides functions for
279 : // that. TODO: If we get a request for an old LSN, such that the
280 : // versions have already been garbage collected away, we should
281 : // throw an error, but we don't track that currently.
282 : //
283 : // last_record_lsn.load().last points to the end of last processed WAL record.
284 : //
285 : // We also remember the starting point of the previous record in
286 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
287 : // first WAL record when the node is started up. But here, we just
288 : // keep track of it.
289 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
290 :
291 : // All WAL records have been processed and stored durably on files on
292 : // local disk, up to this LSN. On crash and restart, we need to re-process
293 : // the WAL starting from this point.
294 : //
295 : // Some later WAL records might have been processed and also flushed to disk
296 : // already, so don't be surprised to see some, but there's no guarantee on
297 : // them yet.
298 : disk_consistent_lsn: AtomicLsn,
299 :
300 : // Parent timeline that this timeline was branched from, and the LSN
301 : // of the branch point.
302 : ancestor_timeline: Option<Arc<Timeline>>,
303 : ancestor_lsn: Lsn,
304 :
305 : pub(super) metrics: TimelineMetrics,
306 :
307 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
308 : // in `crate::page_service` writes these metrics.
309 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
310 :
311 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
312 :
313 : /// Ensures layers aren't frozen by checkpointer between
314 : /// [`Timeline::get_layer_for_write`] and layer reads.
315 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
316 : /// Must always be acquired before the layer map/individual layer lock
317 : /// to avoid deadlock.
318 : ///
319 : /// The state is cleared upon freezing.
320 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
321 :
322 : /// Used to avoid multiple `flush_loop` tasks running
323 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
324 :
325 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
326 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
327 : /// The flush cycle counter is sent back on the layer_flush_done channel when
328 : /// the flush finishes. You can use that to wait for the flush to finish.
329 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
330 : /// read by whoever sends an update
331 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
332 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
333 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
334 :
335 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
336 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
337 :
338 : // List of child timelines and their branch points. This is needed to avoid
339 : // garbage collecting data that is still needed by the child timelines.
340 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
341 :
342 : // It may change across major versions so for simplicity
343 : // keep it after running initdb for a timeline.
344 : // It is needed in checks when we want to error on some operations
345 : // when they are requested for pre-initdb lsn.
346 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
347 : // though let's keep them both for better error visibility.
348 : pub initdb_lsn: Lsn,
349 :
350 : /// When did we last calculate the partitioning? Make it pub to test cases.
351 : pub(super) partitioning: tokio::sync::Mutex<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
352 :
353 : /// Configuration: how often should the partitioning be recalculated.
354 : repartition_threshold: u64,
355 :
356 : last_image_layer_creation_check_at: AtomicLsn,
357 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
358 :
359 : /// Current logical size of the "datadir", at the last LSN.
360 : current_logical_size: LogicalSize,
361 :
362 : /// Information about the last processed message by the WAL receiver,
363 : /// or None if WAL receiver has not received anything for this timeline
364 : /// yet.
365 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
366 : pub walreceiver: Mutex<Option<WalReceiver>>,
367 :
368 : /// Relation size cache
369 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
370 :
371 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
372 :
373 : state: watch::Sender<TimelineState>,
374 :
375 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
376 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
377 : pub delete_progress: Arc<tokio::sync::Mutex<DeleteTimelineFlow>>,
378 :
379 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
380 :
381 : /// Load or creation time information about the disk_consistent_lsn and when the loading
382 : /// happened. Used for consumption metrics.
383 : pub(crate) loaded_at: (Lsn, SystemTime),
384 :
385 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
386 : pub(crate) gate: Gate,
387 :
388 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
389 : /// to the timeline should drop out when this token fires.
390 : pub(crate) cancel: CancellationToken,
391 :
392 : /// Make sure we only have one running compaction at a time in tests.
393 : ///
394 : /// Must only be taken in two places:
395 : /// - [`Timeline::compact`] (this file)
396 : /// - [`delete::delete_local_timeline_directory`]
397 : ///
398 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
399 : compaction_lock: tokio::sync::Mutex<()>,
400 :
401 : /// Make sure we only have one running gc at a time.
402 : ///
403 : /// Must only be taken in two places:
404 : /// - [`Timeline::gc`] (this file)
405 : /// - [`delete::delete_local_timeline_directory`]
406 : ///
407 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
408 : gc_lock: tokio::sync::Mutex<()>,
409 :
410 : /// Cloned from [`super::Tenant::timeline_get_throttle`] on construction.
411 : timeline_get_throttle: Arc<
412 : crate::tenant::throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>,
413 : >,
414 :
415 : /// Keep aux directory cache to avoid it's reconstruction on each update
416 : pub(crate) aux_files: tokio::sync::Mutex<AuxFilesState>,
417 :
418 : /// Size estimator for aux file v2
419 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
420 :
421 : /// Indicate whether aux file v2 storage is enabled.
422 : pub(crate) last_aux_file_policy: AtomicAuxFilePolicy,
423 :
424 : /// Some test cases directly place keys into the timeline without actually modifying the directory
425 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
426 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
427 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
428 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
429 : #[cfg(test)]
430 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
431 :
432 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
433 :
434 : pub(crate) handles: handle::PerTimelineState<crate::page_service::TenantManagerTypes>,
435 : }
436 :
437 : pub struct WalReceiverInfo {
438 : pub wal_source_connconf: PgConnectionConfig,
439 : pub last_received_msg_lsn: Lsn,
440 : pub last_received_msg_ts: u128,
441 : }
442 :
443 : /// Information about how much history needs to be retained, needed by
444 : /// Garbage Collection.
445 : #[derive(Default)]
446 : pub(crate) struct GcInfo {
447 : /// Specific LSNs that are needed.
448 : ///
449 : /// Currently, this includes all points where child branches have
450 : /// been forked off from. In the future, could also include
451 : /// explicit user-defined snapshot points.
452 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId)>,
453 :
454 : /// The cutoff coordinates, which are combined by selecting the minimum.
455 : pub(crate) cutoffs: GcCutoffs,
456 :
457 : /// Leases granted to particular LSNs.
458 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
459 :
460 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
461 : pub(crate) within_ancestor_pitr: bool,
462 : }
463 :
464 : impl GcInfo {
465 226 : pub(crate) fn min_cutoff(&self) -> Lsn {
466 226 : self.cutoffs.select_min()
467 226 : }
468 :
469 228 : pub(super) fn insert_child(&mut self, child_id: TimelineId, child_lsn: Lsn) {
470 228 : self.retain_lsns.push((child_lsn, child_id));
471 228 : self.retain_lsns.sort_by_key(|i| i.0);
472 228 : }
473 :
474 2 : pub(super) fn remove_child(&mut self, child_id: TimelineId) {
475 2 : self.retain_lsns.retain(|i| i.1 != child_id);
476 2 : }
477 : }
478 :
479 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
480 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
481 : /// between time-based and space-based retention for observability and consumption metrics purposes.
482 : #[derive(Debug, Clone)]
483 : pub(crate) struct GcCutoffs {
484 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
485 : /// history we must keep to retain a specified number of bytes of WAL.
486 : pub(crate) space: Lsn,
487 :
488 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
489 : /// history we must keep to enable reading back at least the PITR interval duration.
490 : pub(crate) time: Lsn,
491 : }
492 :
493 : impl Default for GcCutoffs {
494 398 : fn default() -> Self {
495 398 : Self {
496 398 : space: Lsn::INVALID,
497 398 : time: Lsn::INVALID,
498 398 : }
499 398 : }
500 : }
501 :
502 : impl GcCutoffs {
503 244 : fn select_min(&self) -> Lsn {
504 244 : std::cmp::min(self.space, self.time)
505 244 : }
506 : }
507 :
508 : pub(crate) struct TimelineVisitOutcome {
509 : completed_keyspace: KeySpace,
510 : image_covered_keyspace: KeySpace,
511 : }
512 :
513 : /// An error happened in a get() operation.
514 2 : #[derive(thiserror::Error, Debug)]
515 : pub(crate) enum PageReconstructError {
516 : #[error(transparent)]
517 : Other(#[from] anyhow::Error),
518 :
519 : #[error("Ancestor LSN wait error: {0}")]
520 : AncestorLsnTimeout(WaitLsnError),
521 :
522 : #[error("timeline shutting down")]
523 : Cancelled,
524 :
525 : /// An error happened replaying WAL records
526 : #[error(transparent)]
527 : WalRedo(anyhow::Error),
528 :
529 : #[error("{0}")]
530 : MissingKey(MissingKeyError),
531 : }
532 :
533 : impl GetVectoredError {
534 : #[cfg(test)]
535 6 : pub(crate) fn is_missing_key_error(&self) -> bool {
536 6 : matches!(self, Self::MissingKey(_))
537 6 : }
538 : }
539 :
540 : pub struct MissingKeyError {
541 : key: Key,
542 : shard: ShardNumber,
543 : cont_lsn: Lsn,
544 : request_lsn: Lsn,
545 : ancestor_lsn: Option<Lsn>,
546 : traversal_path: Vec<TraversalPathItem>,
547 : backtrace: Option<std::backtrace::Backtrace>,
548 : }
549 :
550 : impl std::fmt::Debug for MissingKeyError {
551 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
552 0 : write!(f, "{}", self)
553 0 : }
554 : }
555 :
556 : impl std::fmt::Display for MissingKeyError {
557 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
558 0 : write!(
559 0 : f,
560 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
561 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
562 0 : )?;
563 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
564 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
565 0 : }
566 :
567 0 : if !self.traversal_path.is_empty() {
568 0 : writeln!(f)?;
569 0 : }
570 :
571 0 : for (r, c, l) in &self.traversal_path {
572 0 : writeln!(
573 0 : f,
574 0 : "layer traversal: result {:?}, cont_lsn {}, layer: {}",
575 0 : r, c, l,
576 0 : )?;
577 : }
578 :
579 0 : if let Some(ref backtrace) = self.backtrace {
580 0 : write!(f, "\n{}", backtrace)?;
581 0 : }
582 :
583 0 : Ok(())
584 0 : }
585 : }
586 :
587 : impl PageReconstructError {
588 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
589 0 : pub(crate) fn is_stopping(&self) -> bool {
590 0 : use PageReconstructError::*;
591 0 : match self {
592 0 : Other(_) => false,
593 0 : AncestorLsnTimeout(_) => false,
594 0 : Cancelled => true,
595 0 : WalRedo(_) => false,
596 0 : MissingKey { .. } => false,
597 : }
598 0 : }
599 : }
600 :
601 0 : #[derive(thiserror::Error, Debug)]
602 : pub(crate) enum CreateImageLayersError {
603 : #[error("timeline shutting down")]
604 : Cancelled,
605 :
606 : #[error(transparent)]
607 : GetVectoredError(GetVectoredError),
608 :
609 : #[error(transparent)]
610 : PageReconstructError(PageReconstructError),
611 :
612 : #[error(transparent)]
613 : Other(#[from] anyhow::Error),
614 : }
615 :
616 0 : #[derive(thiserror::Error, Debug, Clone)]
617 : pub(crate) enum FlushLayerError {
618 : /// Timeline cancellation token was cancelled
619 : #[error("timeline shutting down")]
620 : Cancelled,
621 :
622 : /// We tried to flush a layer while the Timeline is in an unexpected state
623 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
624 : NotRunning(FlushLoopState),
625 :
626 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
627 : // loop via a watch channel, where we can only borrow it.
628 : #[error(transparent)]
629 : CreateImageLayersError(Arc<CreateImageLayersError>),
630 :
631 : #[error(transparent)]
632 : Other(#[from] Arc<anyhow::Error>),
633 : }
634 :
635 : impl FlushLayerError {
636 : // When crossing from generic anyhow errors to this error type, we explicitly check
637 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
638 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
639 0 : let cancelled = timeline.cancel.is_cancelled()
640 : // The upload queue might have been shut down before the official cancellation of the timeline.
641 0 : || err
642 0 : .downcast_ref::<NotInitialized>()
643 0 : .map(NotInitialized::is_stopping)
644 0 : .unwrap_or_default();
645 0 : if cancelled {
646 0 : Self::Cancelled
647 : } else {
648 0 : Self::Other(Arc::new(err))
649 : }
650 0 : }
651 : }
652 :
653 0 : #[derive(thiserror::Error, Debug)]
654 : pub(crate) enum GetVectoredError {
655 : #[error("timeline shutting down")]
656 : Cancelled,
657 :
658 : #[error("Requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
659 : Oversized(u64),
660 :
661 : #[error("Requested at invalid LSN: {0}")]
662 : InvalidLsn(Lsn),
663 :
664 : #[error("Requested key not found: {0}")]
665 : MissingKey(MissingKeyError),
666 :
667 : #[error(transparent)]
668 : GetReadyAncestorError(GetReadyAncestorError),
669 :
670 : #[error(transparent)]
671 : Other(#[from] anyhow::Error),
672 : }
673 :
674 2 : #[derive(thiserror::Error, Debug)]
675 : pub(crate) enum GetReadyAncestorError {
676 : #[error("Ancestor LSN wait error: {0}")]
677 : AncestorLsnTimeout(#[from] WaitLsnError),
678 :
679 : #[error("Bad state on timeline {timeline_id}: {state:?}")]
680 : BadState {
681 : timeline_id: TimelineId,
682 : state: TimelineState,
683 : },
684 :
685 : #[error("Cancelled")]
686 : Cancelled,
687 : }
688 :
689 : #[derive(Clone, Copy)]
690 : pub enum LogicalSizeCalculationCause {
691 : Initial,
692 : ConsumptionMetricsSyntheticSize,
693 : EvictionTaskImitation,
694 : TenantSizeHandler,
695 : }
696 :
697 : pub enum GetLogicalSizePriority {
698 : User,
699 : Background,
700 : }
701 :
702 0 : #[derive(enumset::EnumSetType)]
703 : pub(crate) enum CompactFlags {
704 : ForceRepartition,
705 : ForceImageLayerCreation,
706 : EnhancedGcBottomMostCompaction,
707 : }
708 :
709 : impl std::fmt::Debug for Timeline {
710 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
711 0 : write!(f, "Timeline<{}>", self.timeline_id)
712 0 : }
713 : }
714 :
715 0 : #[derive(thiserror::Error, Debug)]
716 : pub(crate) enum WaitLsnError {
717 : // Called on a timeline which is shutting down
718 : #[error("Shutdown")]
719 : Shutdown,
720 :
721 : // Called on an timeline not in active state or shutting down
722 : #[error("Bad timeline state: {0:?}")]
723 : BadState(TimelineState),
724 :
725 : // Timeout expired while waiting for LSN to catch up with goal.
726 : #[error("{0}")]
727 : Timeout(String),
728 : }
729 :
730 : // The impls below achieve cancellation mapping for errors.
731 : // Perhaps there's a way of achieving this with less cruft.
732 :
733 : impl From<CreateImageLayersError> for CompactionError {
734 0 : fn from(e: CreateImageLayersError) -> Self {
735 0 : match e {
736 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
737 0 : CreateImageLayersError::Other(e) => {
738 0 : CompactionError::Other(e.context("create image layers"))
739 : }
740 0 : _ => CompactionError::Other(e.into()),
741 : }
742 0 : }
743 : }
744 :
745 : impl From<CreateImageLayersError> for FlushLayerError {
746 0 : fn from(e: CreateImageLayersError) -> Self {
747 0 : match e {
748 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
749 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
750 : }
751 0 : }
752 : }
753 :
754 : impl From<PageReconstructError> for CreateImageLayersError {
755 0 : fn from(e: PageReconstructError) -> Self {
756 0 : match e {
757 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
758 0 : _ => CreateImageLayersError::PageReconstructError(e),
759 : }
760 0 : }
761 : }
762 :
763 : impl From<GetVectoredError> for CreateImageLayersError {
764 0 : fn from(e: GetVectoredError) -> Self {
765 0 : match e {
766 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
767 0 : _ => CreateImageLayersError::GetVectoredError(e),
768 : }
769 0 : }
770 : }
771 :
772 : impl From<GetVectoredError> for PageReconstructError {
773 6 : fn from(e: GetVectoredError) -> Self {
774 6 : match e {
775 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
776 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
777 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
778 4 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
779 2 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
780 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
781 : }
782 6 : }
783 : }
784 :
785 : impl From<GetReadyAncestorError> for PageReconstructError {
786 2 : fn from(e: GetReadyAncestorError) -> Self {
787 2 : use GetReadyAncestorError::*;
788 2 : match e {
789 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
790 2 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
791 0 : Cancelled => PageReconstructError::Cancelled,
792 : }
793 2 : }
794 : }
795 :
796 : #[derive(
797 : Eq,
798 : PartialEq,
799 : Debug,
800 : Copy,
801 : Clone,
802 202 : strum_macros::EnumString,
803 0 : strum_macros::Display,
804 0 : serde_with::DeserializeFromStr,
805 : serde_with::SerializeDisplay,
806 : )]
807 : #[strum(serialize_all = "kebab-case")]
808 : pub enum GetVectoredImpl {
809 : Sequential,
810 : Vectored,
811 : }
812 :
813 : #[derive(
814 : Eq,
815 : PartialEq,
816 : Debug,
817 : Copy,
818 : Clone,
819 202 : strum_macros::EnumString,
820 0 : strum_macros::Display,
821 0 : serde_with::DeserializeFromStr,
822 : serde_with::SerializeDisplay,
823 : )]
824 : #[strum(serialize_all = "kebab-case")]
825 : pub enum GetImpl {
826 : Legacy,
827 : Vectored,
828 : }
829 :
830 : pub(crate) enum WaitLsnWaiter<'a> {
831 : Timeline(&'a Timeline),
832 : Tenant,
833 : PageService,
834 : }
835 :
836 : /// Argument to [`Timeline::shutdown`].
837 : #[derive(Debug, Clone, Copy)]
838 : pub(crate) enum ShutdownMode {
839 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
840 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
841 : ///
842 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
843 : /// the call to [`Timeline::shutdown`].
844 : FreezeAndFlush,
845 : /// Shut down immediately, without waiting for any open layers to flush.
846 : Hard,
847 : }
848 :
849 : struct ImageLayerCreationOutcome {
850 : image: Option<ResidentLayer>,
851 : next_start_key: Key,
852 : }
853 :
854 : /// Public interface functions
855 : impl Timeline {
856 : /// Get the LSN where this branch was created
857 102 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
858 102 : self.ancestor_lsn
859 102 : }
860 :
861 : /// Get the ancestor's timeline id
862 1508 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
863 1508 : self.ancestor_timeline
864 1508 : .as_ref()
865 1508 : .map(|ancestor| ancestor.timeline_id)
866 1508 : }
867 :
868 : /// Get the bytes written since the PITR cutoff on this branch, and
869 : /// whether this branch's ancestor_lsn is within its parent's PITR.
870 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
871 0 : let gc_info = self.gc_info.read().unwrap();
872 0 : let history = self
873 0 : .get_last_record_lsn()
874 0 : .checked_sub(gc_info.cutoffs.time)
875 0 : .unwrap_or(Lsn(0))
876 0 : .0;
877 0 : (history, gc_info.within_ancestor_pitr)
878 0 : }
879 :
880 : /// Lock and get timeline's GC cutoff
881 995 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
882 995 : self.latest_gc_cutoff_lsn.read()
883 995 : }
884 :
885 : /// Look up given page version.
886 : ///
887 : /// If a remote layer file is needed, it is downloaded as part of this
888 : /// call.
889 : ///
890 : /// This method enforces [`Self::timeline_get_throttle`] internally.
891 : ///
892 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
893 : /// abstraction above this needs to store suitable metadata to track what
894 : /// data exists with what keys, in separate metadata entries. If a
895 : /// non-existent key is requested, we may incorrectly return a value from
896 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
897 : /// non-existing key.
898 : ///
899 : /// # Cancel-Safety
900 : ///
901 : /// This method is cancellation-safe.
902 : #[inline(always)]
903 624874 : pub(crate) async fn get(
904 624874 : &self,
905 624874 : key: Key,
906 624874 : lsn: Lsn,
907 624874 : ctx: &RequestContext,
908 624874 : ) -> Result<Bytes, PageReconstructError> {
909 624874 : if !lsn.is_valid() {
910 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
911 624874 : }
912 624874 :
913 624874 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
914 624874 : // already checked the key against the shard_identity when looking up the Timeline from
915 624874 : // page_service.
916 624874 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
917 :
918 624874 : self.timeline_get_throttle.throttle(ctx, 1).await;
919 :
920 624874 : match self.conf.get_impl {
921 : GetImpl::Legacy => {
922 0 : let reconstruct_state = ValueReconstructState {
923 0 : records: Vec::new(),
924 0 : img: None,
925 0 : };
926 0 :
927 0 : self.get_impl(key, lsn, reconstruct_state, ctx).await
928 : }
929 : GetImpl::Vectored => {
930 624874 : let keyspace = KeySpace {
931 624874 : ranges: vec![key..key.next()],
932 624874 : };
933 624874 :
934 624874 : // Initialise the reconstruct state for the key with the cache
935 624874 : // entry returned above.
936 624874 : let mut reconstruct_state = ValuesReconstructState::new();
937 :
938 624874 : let vectored_res = self
939 624874 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
940 102581 : .await;
941 :
942 624874 : if self.conf.validate_vectored_get {
943 0 : self.validate_get_vectored_impl(&vectored_res, keyspace, lsn, ctx)
944 0 : .await;
945 624874 : }
946 :
947 624874 : let key_value = vectored_res?.pop_first();
948 624868 : match key_value {
949 624746 : Some((got_key, value)) => {
950 624746 : if got_key != key {
951 0 : error!(
952 0 : "Expected {}, but singular vectored get returned {}",
953 : key, got_key
954 : );
955 0 : Err(PageReconstructError::Other(anyhow!(
956 0 : "Singular vectored get returned wrong key"
957 0 : )))
958 : } else {
959 624746 : value
960 : }
961 : }
962 122 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
963 122 : key,
964 122 : shard: self.shard_identity.get_shard_number(&key),
965 122 : cont_lsn: Lsn(0),
966 122 : request_lsn: lsn,
967 122 : ancestor_lsn: None,
968 122 : traversal_path: Vec::new(),
969 122 : backtrace: None,
970 122 : })),
971 : }
972 : }
973 : }
974 624874 : }
975 :
976 : /// Not subject to [`Self::timeline_get_throttle`].
977 322 : async fn get_impl(
978 322 : &self,
979 322 : key: Key,
980 322 : lsn: Lsn,
981 322 : mut reconstruct_state: ValueReconstructState,
982 322 : ctx: &RequestContext,
983 322 : ) -> Result<Bytes, PageReconstructError> {
984 322 : // XXX: structured stats collection for layer eviction here.
985 322 : trace!(
986 0 : "get page request for {}@{} from task kind {:?}",
987 0 : key,
988 0 : lsn,
989 0 : ctx.task_kind()
990 : );
991 :
992 322 : let timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
993 322 : .for_get_kind(GetKind::Singular)
994 322 : .start_timer();
995 322 : let path = self
996 322 : .get_reconstruct_data(key, lsn, &mut reconstruct_state, ctx)
997 19 : .await?;
998 320 : timer.stop_and_record();
999 320 :
1000 320 : let start = Instant::now();
1001 320 : let res = self.reconstruct_value(key, lsn, reconstruct_state).await;
1002 320 : let elapsed = start.elapsed();
1003 320 : crate::metrics::RECONSTRUCT_TIME
1004 320 : .for_get_kind(GetKind::Singular)
1005 320 : .observe(elapsed.as_secs_f64());
1006 320 :
1007 320 : if cfg!(feature = "testing")
1008 320 : && res.is_err()
1009 0 : && !matches!(res, Err(PageReconstructError::Cancelled))
1010 : {
1011 : // it can only be walredo issue
1012 : use std::fmt::Write;
1013 :
1014 0 : let mut msg = String::new();
1015 0 :
1016 0 : path.into_iter().for_each(|(res, cont_lsn, layer)| {
1017 0 : writeln!(
1018 0 : msg,
1019 0 : "- layer traversal: result {res:?}, cont_lsn {cont_lsn}, layer: {}",
1020 0 : layer,
1021 0 : )
1022 0 : .expect("string grows")
1023 0 : });
1024 0 :
1025 0 : // this is to rule out or provide evidence that we could in some cases read a duplicate
1026 0 : // walrecord
1027 0 : tracing::info!("walredo failed, path:\n{msg}");
1028 320 : }
1029 :
1030 320 : res
1031 322 : }
1032 :
1033 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1034 : pub(crate) const VEC_GET_LAYERS_VISITED_WARN_THRESH: f64 = 512.0;
1035 :
1036 : /// Look up multiple page versions at a given LSN
1037 : ///
1038 : /// This naive implementation will be replaced with a more efficient one
1039 : /// which actually vectorizes the read path.
1040 1036 : pub(crate) async fn get_vectored(
1041 1036 : &self,
1042 1036 : keyspace: KeySpace,
1043 1036 : lsn: Lsn,
1044 1036 : ctx: &RequestContext,
1045 1036 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1046 1036 : if !lsn.is_valid() {
1047 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1048 1036 : }
1049 1036 :
1050 1036 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1051 1036 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1052 0 : return Err(GetVectoredError::Oversized(key_count));
1053 1036 : }
1054 :
1055 2072 : for range in &keyspace.ranges {
1056 1036 : let mut key = range.start;
1057 2410 : while key != range.end {
1058 1374 : assert!(!self.shard_identity.is_key_disposable(&key));
1059 1374 : key = key.next();
1060 : }
1061 : }
1062 :
1063 1036 : trace!(
1064 0 : "get vectored request for {:?}@{} from task kind {:?} will use {} implementation",
1065 0 : keyspace,
1066 0 : lsn,
1067 0 : ctx.task_kind(),
1068 : self.conf.get_vectored_impl
1069 : );
1070 :
1071 1036 : let start = crate::metrics::GET_VECTORED_LATENCY
1072 1036 : .for_task_kind(ctx.task_kind())
1073 1036 : .map(|metric| (metric, Instant::now()));
1074 :
1075 : // start counting after throttle so that throttle time
1076 : // is always less than observation time
1077 1036 : let throttled = self
1078 1036 : .timeline_get_throttle
1079 1036 : .throttle(ctx, key_count as usize)
1080 0 : .await;
1081 :
1082 1036 : let res = match self.conf.get_vectored_impl {
1083 : GetVectoredImpl::Sequential => {
1084 0 : self.get_vectored_sequential_impl(keyspace, lsn, ctx).await
1085 : }
1086 : GetVectoredImpl::Vectored => {
1087 1036 : let vectored_res = self
1088 1036 : .get_vectored_impl(
1089 1036 : keyspace.clone(),
1090 1036 : lsn,
1091 1036 : &mut ValuesReconstructState::new(),
1092 1036 : ctx,
1093 1036 : )
1094 49 : .await;
1095 :
1096 1036 : if self.conf.validate_vectored_get {
1097 0 : self.validate_get_vectored_impl(&vectored_res, keyspace, lsn, ctx)
1098 0 : .await;
1099 1036 : }
1100 :
1101 1036 : vectored_res
1102 : }
1103 : };
1104 :
1105 1036 : if let Some((metric, start)) = start {
1106 0 : let elapsed = start.elapsed();
1107 0 : let ex_throttled = if let Some(throttled) = throttled {
1108 0 : elapsed.checked_sub(throttled)
1109 : } else {
1110 0 : Some(elapsed)
1111 : };
1112 :
1113 0 : if let Some(ex_throttled) = ex_throttled {
1114 0 : metric.observe(ex_throttled.as_secs_f64());
1115 0 : } else {
1116 0 : use utils::rate_limit::RateLimit;
1117 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1118 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
1119 0 : let mut rate_limit = LOGGED.lock().unwrap();
1120 0 : rate_limit.call(|| {
1121 0 : warn!("error deducting time spent throttled; this message is logged at a global rate limit");
1122 0 : });
1123 0 : }
1124 1036 : }
1125 :
1126 1036 : res
1127 1036 : }
1128 :
1129 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1130 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1131 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1132 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1133 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1134 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1135 : /// the scan operation will not cause OOM in the future.
1136 12 : pub(crate) async fn scan(
1137 12 : &self,
1138 12 : keyspace: KeySpace,
1139 12 : lsn: Lsn,
1140 12 : ctx: &RequestContext,
1141 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1142 12 : if !lsn.is_valid() {
1143 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1144 12 : }
1145 12 :
1146 12 : trace!(
1147 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1148 0 : keyspace,
1149 0 : lsn,
1150 0 : ctx.task_kind()
1151 : );
1152 :
1153 : // We should generalize this into Keyspace::contains in the future.
1154 24 : for range in &keyspace.ranges {
1155 12 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1156 12 : || range.end.field1 > METADATA_KEY_END_PREFIX
1157 : {
1158 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1159 0 : "only metadata keyspace can be scanned"
1160 0 : )));
1161 12 : }
1162 : }
1163 :
1164 12 : let start = crate::metrics::SCAN_LATENCY
1165 12 : .for_task_kind(ctx.task_kind())
1166 12 : .map(ScanLatencyOngoingRecording::start_recording);
1167 :
1168 : // start counting after throttle so that throttle time
1169 : // is always less than observation time
1170 12 : let throttled = self
1171 12 : .timeline_get_throttle
1172 12 : // assume scan = 1 quota for now until we find a better way to process this
1173 12 : .throttle(ctx, 1)
1174 0 : .await;
1175 :
1176 12 : let vectored_res = self
1177 12 : .get_vectored_impl(
1178 12 : keyspace.clone(),
1179 12 : lsn,
1180 12 : &mut ValuesReconstructState::default(),
1181 12 : ctx,
1182 12 : )
1183 0 : .await;
1184 :
1185 12 : if let Some(recording) = start {
1186 0 : recording.observe(throttled);
1187 12 : }
1188 :
1189 12 : vectored_res
1190 12 : }
1191 :
1192 : /// Not subject to [`Self::timeline_get_throttle`].
1193 12 : pub(super) async fn get_vectored_sequential_impl(
1194 12 : &self,
1195 12 : keyspace: KeySpace,
1196 12 : lsn: Lsn,
1197 12 : ctx: &RequestContext,
1198 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1199 12 : let mut values = BTreeMap::new();
1200 :
1201 24 : for range in keyspace.ranges {
1202 12 : let mut key = range.start;
1203 334 : while key != range.end {
1204 322 : let block = self
1205 322 : .get_impl(key, lsn, ValueReconstructState::default(), ctx)
1206 19 : .await;
1207 :
1208 : use PageReconstructError::*;
1209 0 : match block {
1210 0 : Err(Cancelled) => return Err(GetVectoredError::Cancelled),
1211 : Err(MissingKey(_))
1212 2 : if NON_INHERITED_RANGE.contains(&key)
1213 2 : || NON_INHERITED_SPARSE_RANGE.contains(&key) =>
1214 2 : {
1215 2 : // Ignore missing key error for aux key range. TODO: currently, we assume non_inherited_range == aux_key_range.
1216 2 : // When we add more types of keys into the page server, we should revisit this part of code and throw errors
1217 2 : // accordingly.
1218 2 : key = key.next();
1219 2 : }
1220 0 : Err(MissingKey(err)) => {
1221 0 : return Err(GetVectoredError::MissingKey(err));
1222 : }
1223 0 : Err(Other(err))
1224 0 : if err
1225 0 : .to_string()
1226 0 : .contains("downloading evicted layer file failed") =>
1227 0 : {
1228 0 : return Err(GetVectoredError::Other(err))
1229 : }
1230 0 : Err(Other(err))
1231 0 : if err
1232 0 : .chain()
1233 0 : .any(|cause| cause.to_string().contains("layer loading failed")) =>
1234 : {
1235 : // The intent here is to achieve error parity with the vectored read path.
1236 : // When vectored read fails to load a layer it fails the whole read, hence
1237 : // we mimic this behaviour here to keep the validation happy.
1238 0 : return Err(GetVectoredError::Other(err));
1239 : }
1240 320 : _ => {
1241 320 : values.insert(key, block);
1242 320 : key = key.next();
1243 320 : }
1244 : }
1245 : }
1246 : }
1247 :
1248 12 : Ok(values)
1249 12 : }
1250 :
1251 626048 : pub(super) async fn get_vectored_impl(
1252 626048 : &self,
1253 626048 : keyspace: KeySpace,
1254 626048 : lsn: Lsn,
1255 626048 : reconstruct_state: &mut ValuesReconstructState,
1256 626048 : ctx: &RequestContext,
1257 626048 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1258 626048 : let get_kind = if keyspace.total_raw_size() == 1 {
1259 625808 : GetKind::Singular
1260 : } else {
1261 240 : GetKind::Vectored
1262 : };
1263 :
1264 626048 : let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
1265 626048 : .for_get_kind(get_kind)
1266 626048 : .start_timer();
1267 626048 : self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1268 114010 : .await?;
1269 626032 : get_data_timer.stop_and_record();
1270 626032 :
1271 626032 : let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
1272 626032 : .for_get_kind(get_kind)
1273 626032 : .start_timer();
1274 626032 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
1275 626032 : let layers_visited = reconstruct_state.get_layers_visited();
1276 :
1277 666556 : for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
1278 666556 : match res {
1279 0 : Err(err) => {
1280 0 : results.insert(key, Err(err));
1281 0 : }
1282 666556 : Ok(state) => {
1283 666556 : let state = ValueReconstructState::from(state);
1284 :
1285 666556 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
1286 666556 : results.insert(key, reconstruct_res);
1287 : }
1288 : }
1289 : }
1290 626032 : reconstruct_timer.stop_and_record();
1291 626032 :
1292 626032 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1293 626032 : // when they're missing. Instead they are omitted from the resulting btree
1294 626032 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1295 626032 : // to avoid infinite results.
1296 626032 : if !results.is_empty() {
1297 625892 : let avg = layers_visited as f64 / results.len() as f64;
1298 625892 : if avg >= Self::VEC_GET_LAYERS_VISITED_WARN_THRESH {
1299 0 : use utils::rate_limit::RateLimit;
1300 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1301 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1302 0 : let mut rate_limit = LOGGED.lock().unwrap();
1303 0 : rate_limit.call(|| {
1304 0 : tracing::info!(
1305 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1306 0 : lsn = %lsn,
1307 0 : "Vectored read for {} visited {} layers on average per key and {} in total. {}/{} pages were returned",
1308 0 : keyspace, avg, layers_visited, results.len(), keyspace.total_raw_size());
1309 0 : });
1310 625892 : }
1311 :
1312 : // Note that this is an approximation. Tracking the exact number of layers visited
1313 : // per key requires virtually unbounded memory usage and is inefficient
1314 : // (i.e. segment tree tracking each range queried from a layer)
1315 625892 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED.observe(avg);
1316 140 : }
1317 :
1318 626032 : Ok(results)
1319 626048 : }
1320 :
1321 : /// Not subject to [`Self::timeline_get_throttle`].
1322 12 : pub(super) async fn validate_get_vectored_impl(
1323 12 : &self,
1324 12 : vectored_res: &Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError>,
1325 12 : keyspace: KeySpace,
1326 12 : lsn: Lsn,
1327 12 : ctx: &RequestContext,
1328 12 : ) {
1329 12 : if keyspace.overlaps(&Key::metadata_key_range()) {
1330 : // skip validation for metadata key range
1331 0 : return;
1332 12 : }
1333 :
1334 12 : let sequential_res = self
1335 12 : .get_vectored_sequential_impl(keyspace.clone(), lsn, ctx)
1336 19 : .await;
1337 :
1338 0 : fn errors_match(lhs: &GetVectoredError, rhs: &GetVectoredError) -> bool {
1339 0 : use GetVectoredError::*;
1340 0 : match (lhs, rhs) {
1341 0 : (Oversized(l), Oversized(r)) => l == r,
1342 0 : (InvalidLsn(l), InvalidLsn(r)) => l == r,
1343 0 : (MissingKey(l), MissingKey(r)) => l.key == r.key,
1344 0 : (GetReadyAncestorError(_), GetReadyAncestorError(_)) => true,
1345 0 : (Other(_), Other(_)) => true,
1346 0 : _ => false,
1347 : }
1348 0 : }
1349 :
1350 12 : match (&sequential_res, vectored_res) {
1351 0 : (Err(GetVectoredError::Cancelled), _) => {},
1352 0 : (_, Err(GetVectoredError::Cancelled)) => {},
1353 0 : (Err(seq_err), Ok(_)) => {
1354 0 : panic!(concat!("Sequential get failed with {}, but vectored get did not",
1355 0 : " - keyspace={:?} lsn={}"),
1356 0 : seq_err, keyspace, lsn) },
1357 0 : (Ok(_), Err(GetVectoredError::GetReadyAncestorError(GetReadyAncestorError::AncestorLsnTimeout(_)))) => {
1358 0 : // Sequential get runs after vectored get, so it is possible for the later
1359 0 : // to time out while waiting for its ancestor's Lsn to become ready and for the
1360 0 : // former to succeed (it essentially has a doubled wait time).
1361 0 : },
1362 0 : (Ok(_), Err(vec_err)) => {
1363 0 : panic!(concat!("Vectored get failed with {}, but sequential get did not",
1364 0 : " - keyspace={:?} lsn={}"),
1365 0 : vec_err, keyspace, lsn) },
1366 0 : (Err(seq_err), Err(vec_err)) => {
1367 0 : assert!(errors_match(seq_err, vec_err),
1368 0 : "Mismatched errors: {seq_err} != {vec_err} - keyspace={keyspace:?} lsn={lsn}")},
1369 12 : (Ok(seq_values), Ok(vec_values)) => {
1370 320 : seq_values.iter().zip(vec_values.iter()).for_each(|((seq_key, seq_res), (vec_key, vec_res))| {
1371 320 : assert_eq!(seq_key, vec_key);
1372 320 : match (seq_res, vec_res) {
1373 320 : (Ok(seq_blob), Ok(vec_blob)) => {
1374 320 : Self::validate_key_equivalence(seq_key, &keyspace, lsn, seq_blob, vec_blob);
1375 320 : },
1376 0 : (Err(err), Ok(_)) => {
1377 0 : panic!(
1378 0 : concat!("Sequential get failed with {} for key {}, but vectored get did not",
1379 0 : " - keyspace={:?} lsn={}"),
1380 0 : err, seq_key, keyspace, lsn) },
1381 0 : (Ok(_), Err(err)) => {
1382 0 : panic!(
1383 0 : concat!("Vectored get failed with {} for key {}, but sequential get did not",
1384 0 : " - keyspace={:?} lsn={}"),
1385 0 : err, seq_key, keyspace, lsn) },
1386 0 : (Err(_), Err(_)) => {}
1387 : }
1388 320 : })
1389 : }
1390 : }
1391 12 : }
1392 :
1393 320 : fn validate_key_equivalence(
1394 320 : key: &Key,
1395 320 : keyspace: &KeySpace,
1396 320 : lsn: Lsn,
1397 320 : seq: &Bytes,
1398 320 : vec: &Bytes,
1399 320 : ) {
1400 320 : if *key == AUX_FILES_KEY {
1401 : // The value reconstruct of AUX_FILES_KEY from records is not deterministic
1402 : // since it uses a hash map under the hood. Hence, deserialise both results
1403 : // before comparing.
1404 0 : let seq_aux_dir_res = AuxFilesDirectory::des(seq);
1405 0 : let vec_aux_dir_res = AuxFilesDirectory::des(vec);
1406 0 : match (&seq_aux_dir_res, &vec_aux_dir_res) {
1407 0 : (Ok(seq_aux_dir), Ok(vec_aux_dir)) => {
1408 0 : assert_eq!(
1409 : seq_aux_dir, vec_aux_dir,
1410 0 : "Mismatch for key {} - keyspace={:?} lsn={}",
1411 : key, keyspace, lsn
1412 : );
1413 : }
1414 0 : (Err(_), Err(_)) => {}
1415 : _ => {
1416 0 : panic!("Mismatch for {key}: {seq_aux_dir_res:?} != {vec_aux_dir_res:?}");
1417 : }
1418 : }
1419 : } else {
1420 : // All other keys should reconstruct deterministically, so we simply compare the blobs.
1421 320 : assert_eq!(
1422 : seq, vec,
1423 0 : "Image mismatch for key {key} - keyspace={keyspace:?} lsn={lsn}"
1424 : );
1425 : }
1426 320 : }
1427 :
1428 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1429 276802 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1430 276802 : self.last_record_lsn.load().last
1431 276802 : }
1432 :
1433 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1434 0 : self.last_record_lsn.load().prev
1435 0 : }
1436 :
1437 : /// Atomically get both last and prev.
1438 226 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1439 226 : self.last_record_lsn.load()
1440 226 : }
1441 :
1442 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1443 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1444 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1445 0 : self.last_record_lsn.status_receiver()
1446 0 : }
1447 :
1448 1143 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1449 1143 : self.disk_consistent_lsn.load()
1450 1143 : }
1451 :
1452 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1453 : /// not validated with control plane yet.
1454 : /// See [`Self::get_remote_consistent_lsn_visible`].
1455 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1456 0 : self.remote_client.remote_consistent_lsn_projected()
1457 0 : }
1458 :
1459 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1460 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1461 : /// generation validation in the deletion queue.
1462 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1463 0 : self.remote_client.remote_consistent_lsn_visible()
1464 0 : }
1465 :
1466 : /// The sum of the file size of all historic layers in the layer map.
1467 : /// This method makes no distinction between local and remote layers.
1468 : /// Hence, the result **does not represent local filesystem usage**.
1469 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1470 0 : let guard = self.layers.read().await;
1471 0 : let layer_map = guard.layer_map();
1472 0 : let mut size = 0;
1473 0 : for l in layer_map.iter_historic_layers() {
1474 0 : size += l.file_size;
1475 0 : }
1476 0 : size
1477 0 : }
1478 :
1479 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1480 0 : self.metrics.resident_physical_size_get()
1481 0 : }
1482 :
1483 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1484 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1485 0 : }
1486 :
1487 : ///
1488 : /// Wait until WAL has been received and processed up to this LSN.
1489 : ///
1490 : /// You should call this before any of the other get_* or list_* functions. Calling
1491 : /// those functions with an LSN that has been processed yet is an error.
1492 : ///
1493 225212 : pub(crate) async fn wait_lsn(
1494 225212 : &self,
1495 225212 : lsn: Lsn,
1496 225212 : who_is_waiting: WaitLsnWaiter<'_>,
1497 225212 : ctx: &RequestContext, /* Prepare for use by cancellation */
1498 225212 : ) -> Result<(), WaitLsnError> {
1499 225212 : let state = self.current_state();
1500 225212 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1501 0 : return Err(WaitLsnError::Shutdown);
1502 225212 : } else if !matches!(state, TimelineState::Active) {
1503 0 : return Err(WaitLsnError::BadState(state));
1504 225212 : }
1505 225212 :
1506 225212 : if cfg!(debug_assertions) {
1507 225212 : match ctx.task_kind() {
1508 : TaskKind::WalReceiverManager
1509 : | TaskKind::WalReceiverConnectionHandler
1510 : | TaskKind::WalReceiverConnectionPoller => {
1511 0 : let is_myself = match who_is_waiting {
1512 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1513 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1514 : };
1515 0 : if is_myself {
1516 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1517 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1518 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1519 0 : }
1520 0 : } else {
1521 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1522 0 : // our walreceiver task can make progress independent of theirs
1523 0 : }
1524 : }
1525 225212 : _ => {}
1526 : }
1527 0 : }
1528 :
1529 225212 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1530 225212 :
1531 225212 : match self
1532 225212 : .last_record_lsn
1533 225212 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1534 0 : .await
1535 : {
1536 225212 : Ok(()) => Ok(()),
1537 0 : Err(e) => {
1538 0 : use utils::seqwait::SeqWaitError::*;
1539 0 : match e {
1540 0 : Shutdown => Err(WaitLsnError::Shutdown),
1541 : Timeout => {
1542 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1543 0 : drop(_timer);
1544 0 : let walreceiver_status = self.walreceiver_status();
1545 0 : Err(WaitLsnError::Timeout(format!(
1546 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1547 0 : lsn,
1548 0 : self.get_last_record_lsn(),
1549 0 : self.get_disk_consistent_lsn(),
1550 0 : walreceiver_status,
1551 0 : )))
1552 : }
1553 : }
1554 : }
1555 : }
1556 225212 : }
1557 :
1558 0 : pub(crate) fn walreceiver_status(&self) -> String {
1559 0 : match &*self.walreceiver.lock().unwrap() {
1560 0 : None => "stopping or stopped".to_string(),
1561 0 : Some(walreceiver) => match walreceiver.status() {
1562 0 : Some(status) => status.to_human_readable_string(),
1563 0 : None => "Not active".to_string(),
1564 : },
1565 : }
1566 0 : }
1567 :
1568 : /// Check that it is valid to request operations with that lsn.
1569 230 : pub(crate) fn check_lsn_is_in_scope(
1570 230 : &self,
1571 230 : lsn: Lsn,
1572 230 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1573 230 : ) -> anyhow::Result<()> {
1574 230 : ensure!(
1575 230 : lsn >= **latest_gc_cutoff_lsn,
1576 4 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1577 4 : lsn,
1578 4 : **latest_gc_cutoff_lsn,
1579 : );
1580 226 : Ok(())
1581 230 : }
1582 :
1583 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1584 : ///
1585 : /// This function will error if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is also
1586 : /// no existing lease to renew. If there is an existing lease in the map, the lease will be renewed only if
1587 : /// the request extends the lease. The returned lease is therefore the maximum between the existing lease and
1588 : /// the requesting lease.
1589 14 : pub(crate) fn make_lsn_lease(
1590 14 : &self,
1591 14 : lsn: Lsn,
1592 14 : length: Duration,
1593 14 : _ctx: &RequestContext,
1594 14 : ) -> anyhow::Result<LsnLease> {
1595 12 : let lease = {
1596 14 : let mut gc_info = self.gc_info.write().unwrap();
1597 14 :
1598 14 : let valid_until = SystemTime::now() + length;
1599 14 :
1600 14 : let entry = gc_info.leases.entry(lsn);
1601 :
1602 12 : let lease = {
1603 14 : if let Entry::Occupied(mut occupied) = entry {
1604 6 : let existing_lease = occupied.get_mut();
1605 6 : if valid_until > existing_lease.valid_until {
1606 2 : existing_lease.valid_until = valid_until;
1607 2 : let dt: DateTime<Utc> = valid_until.into();
1608 2 : info!("lease extended to {}", dt);
1609 : } else {
1610 4 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1611 4 : info!("existing lease covers greater length, valid until {}", dt);
1612 : }
1613 :
1614 6 : existing_lease.clone()
1615 : } else {
1616 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff)
1617 8 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1618 8 : if lsn < *latest_gc_cutoff_lsn {
1619 2 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1620 6 : }
1621 6 :
1622 6 : let dt: DateTime<Utc> = valid_until.into();
1623 6 : info!("lease created, valid until {}", dt);
1624 6 : entry.or_insert(LsnLease { valid_until }).clone()
1625 : }
1626 : };
1627 :
1628 12 : lease
1629 12 : };
1630 12 :
1631 12 : Ok(lease)
1632 14 : }
1633 :
1634 : /// Flush to disk all data that was written with the put_* functions
1635 2148 : #[instrument(skip(self), fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%self.timeline_id))]
1636 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1637 : self.freeze_and_flush0().await
1638 : }
1639 :
1640 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1641 : // polluting the span hierarchy.
1642 1074 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1643 1074 : let to_lsn = {
1644 : // Freeze the current open in-memory layer. It will be written to disk on next
1645 : // iteration.
1646 1074 : let mut g = self.write_lock.lock().await;
1647 :
1648 1074 : let to_lsn = self.get_last_record_lsn();
1649 1074 : self.freeze_inmem_layer_at(to_lsn, &mut g).await;
1650 1074 : to_lsn
1651 1074 : };
1652 1077 : self.flush_frozen_layers_and_wait(to_lsn).await
1653 1074 : }
1654 :
1655 : // Check if an open ephemeral layer should be closed: this provides
1656 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1657 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1658 : // ephemeral layer bytes has been breached.
1659 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1660 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1661 : // If the write lock is held, there is an active wal receiver: rolling open layers
1662 : // is their responsibility while they hold this lock.
1663 0 : return;
1664 : };
1665 :
1666 0 : let Ok(layers_guard) = self.layers.try_read() else {
1667 : // Don't block if the layer lock is busy
1668 0 : return;
1669 : };
1670 :
1671 0 : let Some(open_layer) = &layers_guard.layer_map().open_layer else {
1672 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1673 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1674 : // that didn't result in writes to this shard.
1675 :
1676 : // Must not hold the layers lock while waiting for a flush.
1677 0 : drop(layers_guard);
1678 0 :
1679 0 : let last_record_lsn = self.get_last_record_lsn();
1680 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1681 0 : if last_record_lsn > disk_consistent_lsn {
1682 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1683 : // we are a sharded tenant and have skipped some WAL
1684 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1685 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1686 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1687 : // without any data ingested (yet) doesn't write a remote index as soon as it
1688 : // sees its LSN advance: we only do this if we've been layer-less
1689 : // for some time.
1690 0 : tracing::debug!(
1691 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1692 : disk_consistent_lsn,
1693 : last_record_lsn
1694 : );
1695 :
1696 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1697 0 : self.flush_frozen_layers_and_wait(last_record_lsn)
1698 0 : .await
1699 0 : .ok();
1700 0 : }
1701 0 : }
1702 :
1703 0 : return;
1704 : };
1705 :
1706 0 : let Some(current_size) = open_layer.try_len() else {
1707 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1708 : // read lock to get size should always succeed.
1709 0 : tracing::warn!("Lock conflict while reading size of open layer");
1710 0 : return;
1711 : };
1712 :
1713 0 : let current_lsn = self.get_last_record_lsn();
1714 :
1715 0 : let checkpoint_distance_override = open_layer.tick().await;
1716 :
1717 0 : if let Some(size_override) = checkpoint_distance_override {
1718 0 : if current_size > size_override {
1719 : // This is not harmful, but it only happens in relatively rare cases where
1720 : // time-based checkpoints are not happening fast enough to keep the amount of
1721 : // ephemeral data within configured limits. It's a sign of stress on the system.
1722 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1723 0 : }
1724 0 : }
1725 :
1726 0 : let checkpoint_distance =
1727 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1728 0 :
1729 0 : if self.should_roll(
1730 0 : current_size,
1731 0 : current_size,
1732 0 : checkpoint_distance,
1733 0 : self.get_last_record_lsn(),
1734 0 : self.last_freeze_at.load(),
1735 0 : open_layer.get_opened_at(),
1736 0 : ) {
1737 0 : let at_lsn = match open_layer.info() {
1738 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1739 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1740 0 : // happens asynchronously in the background.
1741 0 : tracing::debug!(
1742 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1743 : );
1744 0 : None
1745 : }
1746 : InMemoryLayerInfo::Open { .. } => {
1747 : // Upgrade to a write lock and freeze the layer
1748 0 : drop(layers_guard);
1749 0 : let mut layers_guard = self.layers.write().await;
1750 0 : let froze = layers_guard
1751 0 : .try_freeze_in_memory_layer(
1752 0 : current_lsn,
1753 0 : &self.last_freeze_at,
1754 0 : &mut write_guard,
1755 0 : )
1756 0 : .await;
1757 0 : Some(current_lsn).filter(|_| froze)
1758 : }
1759 : };
1760 0 : if let Some(lsn) = at_lsn {
1761 0 : let res: Result<u64, _> = self.flush_frozen_layers(lsn);
1762 0 : if let Err(e) = res {
1763 0 : tracing::info!("failed to flush frozen layer after background freeze: {e:#}");
1764 0 : }
1765 0 : }
1766 0 : }
1767 0 : }
1768 :
1769 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1770 : /// compaction tasks.
1771 364 : pub(crate) async fn compact(
1772 364 : self: &Arc<Self>,
1773 364 : cancel: &CancellationToken,
1774 364 : flags: EnumSet<CompactFlags>,
1775 364 : ctx: &RequestContext,
1776 364 : ) -> Result<bool, CompactionError> {
1777 364 : // most likely the cancellation token is from background task, but in tests it could be the
1778 364 : // request task as well.
1779 364 :
1780 364 : let prepare = async move {
1781 364 : let guard = self.compaction_lock.lock().await;
1782 :
1783 364 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1784 364 : BackgroundLoopKind::Compaction,
1785 364 : ctx,
1786 364 : )
1787 0 : .await;
1788 :
1789 364 : (guard, permit)
1790 364 : };
1791 :
1792 : // this wait probably never needs any "long time spent" logging, because we already nag if
1793 : // compaction task goes over it's period (20s) which is quite often in production.
1794 364 : let (_guard, _permit) = tokio::select! {
1795 : tuple = prepare => { tuple },
1796 : _ = self.cancel.cancelled() => return Ok(false),
1797 : _ = cancel.cancelled() => return Ok(false),
1798 : };
1799 :
1800 364 : let last_record_lsn = self.get_last_record_lsn();
1801 364 :
1802 364 : // Last record Lsn could be zero in case the timeline was just created
1803 364 : if !last_record_lsn.is_valid() {
1804 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1805 0 : return Ok(false);
1806 364 : }
1807 364 :
1808 364 : match self.get_compaction_algorithm_settings().kind {
1809 : CompactionAlgorithm::Tiered => {
1810 0 : self.compact_tiered(cancel, ctx).await?;
1811 0 : Ok(false)
1812 : }
1813 76988 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, flags, ctx).await,
1814 : }
1815 364 : }
1816 :
1817 : /// Mutate the timeline with a [`TimelineWriter`].
1818 5133160 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1819 5133160 : TimelineWriter {
1820 5133160 : tl: self,
1821 5133160 : write_guard: self.write_lock.lock().await,
1822 : }
1823 5133160 : }
1824 :
1825 0 : pub(crate) fn activate(
1826 0 : self: &Arc<Self>,
1827 0 : parent: Arc<crate::tenant::Tenant>,
1828 0 : broker_client: BrokerClientChannel,
1829 0 : background_jobs_can_start: Option<&completion::Barrier>,
1830 0 : ctx: &RequestContext,
1831 0 : ) {
1832 0 : if self.tenant_shard_id.is_shard_zero() {
1833 0 : // Logical size is only maintained accurately on shard zero.
1834 0 : self.spawn_initial_logical_size_computation_task(ctx);
1835 0 : }
1836 0 : self.launch_wal_receiver(ctx, broker_client);
1837 0 : self.set_state(TimelineState::Active);
1838 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1839 0 : }
1840 :
1841 : /// After this function returns, there are no timeline-scoped tasks are left running.
1842 : ///
1843 : /// The preferred pattern for is:
1844 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1845 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1846 : /// go the extra mile and keep track of JoinHandles
1847 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1848 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1849 : ///
1850 : /// For legacy reasons, we still have multiple tasks spawned using
1851 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1852 : /// We refer to these as "timeline-scoped task_mgr tasks".
1853 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1854 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1855 : /// or [`task_mgr::shutdown_watcher`].
1856 : /// We want to gradually convert the code base away from these.
1857 : ///
1858 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1859 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1860 : /// ones that aren't mentioned here):
1861 : /// - [`TaskKind::TimelineDeletionWorker`]
1862 : /// - NB: also used for tenant deletion
1863 : /// - [`TaskKind::RemoteUploadTask`]`
1864 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1865 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1866 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1867 : /// - [`TaskKind::Eviction`]
1868 : /// - [`TaskKind::LayerFlushTask`]
1869 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1870 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1871 8 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1872 8 : debug_assert_current_span_has_tenant_and_timeline_id();
1873 :
1874 8 : let try_freeze_and_flush = match mode {
1875 6 : ShutdownMode::FreezeAndFlush => true,
1876 2 : ShutdownMode::Hard => false,
1877 : };
1878 :
1879 : // Regardless of whether we're going to try_freeze_and_flush
1880 : // or not, stop ingesting any more data. Walreceiver only provides
1881 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1882 : // So, only after the self.gate.close() below will we know for sure that
1883 : // no walreceiver tasks are left.
1884 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1885 : // data during the call to `self.freeze_and_flush()` below.
1886 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1887 : // which is what we'd need to properly model early shutdown of the walreceiver
1888 : // task sub-tree before the other Timeline task sub-trees.
1889 8 : let walreceiver = self.walreceiver.lock().unwrap().take();
1890 8 : tracing::debug!(
1891 0 : is_some = walreceiver.is_some(),
1892 0 : "Waiting for WalReceiverManager..."
1893 : );
1894 8 : if let Some(walreceiver) = walreceiver {
1895 0 : walreceiver.cancel();
1896 8 : }
1897 : // ... and inform any waiters for newer LSNs that there won't be any.
1898 8 : self.last_record_lsn.shutdown();
1899 8 :
1900 8 : if try_freeze_and_flush {
1901 : // we shut down walreceiver above, so, we won't add anything more
1902 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1903 : // to reach the disk & upload queue, then shut the upload queue and
1904 : // wait for it to drain.
1905 6 : match self.freeze_and_flush().await {
1906 : Ok(_) => {
1907 : // drain the upload queue
1908 : // if we did not wait for completion here, it might be our shutdown process
1909 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1910 : // be spawned.
1911 : //
1912 : // what is problematic is the shutting down of RemoteTimelineClient, because
1913 : // obviously it does not make sense to stop while we wait for it, but what
1914 : // about corner cases like s3 suddenly hanging up?
1915 6 : self.remote_client.shutdown().await;
1916 : }
1917 0 : Err(e) => {
1918 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1919 0 : // we have some extra WAL replay to do next time the timeline starts.
1920 0 : warn!("failed to freeze and flush: {e:#}");
1921 : }
1922 : }
1923 2 : }
1924 :
1925 : // Signal any subscribers to our cancellation token to drop out
1926 8 : tracing::debug!("Cancelling CancellationToken");
1927 8 : self.cancel.cancel();
1928 8 :
1929 8 : // Ensure Prevent new page service requests from starting.
1930 8 : self.handles.shutdown();
1931 8 :
1932 8 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1933 8 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1934 8 : self.remote_client.stop();
1935 8 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1936 8 : // to shut down the upload queue tasks.
1937 8 : // TODO: fix that, task management should be encapsulated inside remote_client.
1938 8 : task_mgr::shutdown_tasks(
1939 8 : Some(TaskKind::RemoteUploadTask),
1940 8 : Some(self.tenant_shard_id),
1941 8 : Some(self.timeline_id),
1942 8 : )
1943 0 : .await;
1944 :
1945 : // TODO: work toward making this a no-op. See this funciton's doc comment for more context.
1946 8 : tracing::debug!("Waiting for tasks...");
1947 8 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1948 :
1949 : // Finally wait until any gate-holders are complete.
1950 : //
1951 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1952 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1953 8 : self.gate.close().await;
1954 :
1955 8 : self.metrics.shutdown();
1956 8 : }
1957 :
1958 398 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1959 398 : match (self.current_state(), new_state) {
1960 398 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1961 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1962 : }
1963 0 : (st, TimelineState::Loading) => {
1964 0 : error!("ignoring transition from {st:?} into Loading state");
1965 : }
1966 0 : (TimelineState::Broken { .. }, new_state) => {
1967 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1968 : }
1969 : (TimelineState::Stopping, TimelineState::Active) => {
1970 0 : error!("Not activating a Stopping timeline");
1971 : }
1972 396 : (_, new_state) => {
1973 396 : self.state.send_replace(new_state);
1974 396 : }
1975 : }
1976 398 : }
1977 :
1978 2 : pub(crate) fn set_broken(&self, reason: String) {
1979 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1980 2 : let broken_state = TimelineState::Broken {
1981 2 : reason,
1982 2 : backtrace: backtrace_str,
1983 2 : };
1984 2 : self.set_state(broken_state);
1985 2 :
1986 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1987 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1988 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1989 2 : self.cancel.cancel();
1990 2 : }
1991 :
1992 226947 : pub(crate) fn current_state(&self) -> TimelineState {
1993 226947 : self.state.borrow().clone()
1994 226947 : }
1995 :
1996 6 : pub(crate) fn is_broken(&self) -> bool {
1997 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1998 6 : }
1999 :
2000 220 : pub(crate) fn is_active(&self) -> bool {
2001 220 : self.current_state() == TimelineState::Active
2002 220 : }
2003 :
2004 : #[allow(unused)]
2005 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
2006 0 : self.remote_client.is_archived()
2007 0 : }
2008 :
2009 1117 : pub(crate) fn is_stopping(&self) -> bool {
2010 1117 : self.current_state() == TimelineState::Stopping
2011 1117 : }
2012 :
2013 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2014 0 : self.state.subscribe()
2015 0 : }
2016 :
2017 225214 : pub(crate) async fn wait_to_become_active(
2018 225214 : &self,
2019 225214 : _ctx: &RequestContext, // Prepare for use by cancellation
2020 225214 : ) -> Result<(), TimelineState> {
2021 225214 : let mut receiver = self.state.subscribe();
2022 225214 : loop {
2023 225214 : let current_state = receiver.borrow().clone();
2024 225214 : match current_state {
2025 : TimelineState::Loading => {
2026 0 : receiver
2027 0 : .changed()
2028 0 : .await
2029 0 : .expect("holding a reference to self");
2030 : }
2031 : TimelineState::Active { .. } => {
2032 225212 : return Ok(());
2033 : }
2034 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2035 : // There's no chance the timeline can transition back into ::Active
2036 2 : return Err(current_state);
2037 : }
2038 : }
2039 : }
2040 225214 : }
2041 :
2042 0 : pub(crate) async fn layer_map_info(&self, reset: LayerAccessStatsReset) -> LayerMapInfo {
2043 0 : let guard = self.layers.read().await;
2044 0 : let layer_map = guard.layer_map();
2045 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2046 0 : if let Some(open_layer) = &layer_map.open_layer {
2047 0 : in_memory_layers.push(open_layer.info());
2048 0 : }
2049 0 : for frozen_layer in &layer_map.frozen_layers {
2050 0 : in_memory_layers.push(frozen_layer.info());
2051 0 : }
2052 :
2053 0 : let mut historic_layers = Vec::new();
2054 0 : for historic_layer in layer_map.iter_historic_layers() {
2055 0 : let historic_layer = guard.get_from_desc(&historic_layer);
2056 0 : historic_layers.push(historic_layer.info(reset));
2057 0 : }
2058 :
2059 0 : LayerMapInfo {
2060 0 : in_memory_layers,
2061 0 : historic_layers,
2062 0 : }
2063 0 : }
2064 :
2065 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))]
2066 : pub(crate) async fn download_layer(
2067 : &self,
2068 : layer_file_name: &LayerName,
2069 : ) -> anyhow::Result<Option<bool>> {
2070 : let Some(layer) = self.find_layer(layer_file_name).await else {
2071 : return Ok(None);
2072 : };
2073 :
2074 : layer.download().await?;
2075 :
2076 : Ok(Some(true))
2077 : }
2078 :
2079 : /// Evict just one layer.
2080 : ///
2081 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2082 0 : pub(crate) async fn evict_layer(
2083 0 : &self,
2084 0 : layer_file_name: &LayerName,
2085 0 : ) -> anyhow::Result<Option<bool>> {
2086 0 : let _gate = self
2087 0 : .gate
2088 0 : .enter()
2089 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2090 :
2091 0 : let Some(local_layer) = self.find_layer(layer_file_name).await else {
2092 0 : return Ok(None);
2093 : };
2094 :
2095 : // curl has this by default
2096 0 : let timeout = std::time::Duration::from_secs(120);
2097 0 :
2098 0 : match local_layer.evict_and_wait(timeout).await {
2099 0 : Ok(()) => Ok(Some(true)),
2100 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2101 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2102 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2103 : }
2104 0 : }
2105 :
2106 4803026 : fn should_roll(
2107 4803026 : &self,
2108 4803026 : layer_size: u64,
2109 4803026 : projected_layer_size: u64,
2110 4803026 : checkpoint_distance: u64,
2111 4803026 : projected_lsn: Lsn,
2112 4803026 : last_freeze_at: Lsn,
2113 4803026 : opened_at: Instant,
2114 4803026 : ) -> bool {
2115 4803026 : let distance = projected_lsn.widening_sub(last_freeze_at);
2116 4803026 :
2117 4803026 : // Rolling the open layer can be triggered by:
2118 4803026 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2119 4803026 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2120 4803026 : // account for how writes are distributed across shards: we expect each node to consume
2121 4803026 : // 1/count of the LSN on average.
2122 4803026 : // 2. The size of the currently open layer.
2123 4803026 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2124 4803026 : // up and suspend activity.
2125 4803026 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2126 0 : info!(
2127 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2128 : projected_lsn, layer_size, distance
2129 : );
2130 :
2131 0 : true
2132 4803026 : } else if projected_layer_size >= checkpoint_distance {
2133 80 : info!(
2134 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2135 : projected_lsn, layer_size, projected_layer_size
2136 : );
2137 :
2138 80 : true
2139 4802946 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2140 0 : info!(
2141 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2142 0 : projected_lsn,
2143 0 : layer_size,
2144 0 : opened_at.elapsed()
2145 : );
2146 :
2147 0 : true
2148 : } else {
2149 4802946 : false
2150 : }
2151 4803026 : }
2152 : }
2153 :
2154 : /// Number of times we will compute partition within a checkpoint distance.
2155 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2156 :
2157 : // Private functions
2158 : impl Timeline {
2159 12 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2160 12 : let tenant_conf = self.tenant_conf.load();
2161 12 : tenant_conf
2162 12 : .tenant_conf
2163 12 : .lsn_lease_length
2164 12 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2165 12 : }
2166 :
2167 : // TODO(yuchen): remove unused flag after implementing https://github.com/neondatabase/neon/issues/8072
2168 : #[allow(unused)]
2169 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2170 0 : let tenant_conf = self.tenant_conf.load();
2171 0 : tenant_conf
2172 0 : .tenant_conf
2173 0 : .lsn_lease_length_for_ts
2174 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2175 0 : }
2176 :
2177 216 : pub(crate) fn get_switch_aux_file_policy(&self) -> AuxFilePolicy {
2178 216 : let tenant_conf = self.tenant_conf.load();
2179 216 : tenant_conf
2180 216 : .tenant_conf
2181 216 : .switch_aux_file_policy
2182 216 : .unwrap_or(self.conf.default_tenant_conf.switch_aux_file_policy)
2183 216 : }
2184 :
2185 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2186 0 : let tenant_conf = self.tenant_conf.load();
2187 0 : tenant_conf
2188 0 : .tenant_conf
2189 0 : .lazy_slru_download
2190 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2191 0 : }
2192 :
2193 4804418 : fn get_checkpoint_distance(&self) -> u64 {
2194 4804418 : let tenant_conf = self.tenant_conf.load();
2195 4804418 : tenant_conf
2196 4804418 : .tenant_conf
2197 4804418 : .checkpoint_distance
2198 4804418 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2199 4804418 : }
2200 :
2201 4802946 : fn get_checkpoint_timeout(&self) -> Duration {
2202 4802946 : let tenant_conf = self.tenant_conf.load();
2203 4802946 : tenant_conf
2204 4802946 : .tenant_conf
2205 4802946 : .checkpoint_timeout
2206 4802946 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2207 4802946 : }
2208 :
2209 522 : fn get_compaction_target_size(&self) -> u64 {
2210 522 : let tenant_conf = self.tenant_conf.load();
2211 522 : tenant_conf
2212 522 : .tenant_conf
2213 522 : .compaction_target_size
2214 522 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2215 522 : }
2216 :
2217 392 : fn get_compaction_threshold(&self) -> usize {
2218 392 : let tenant_conf = self.tenant_conf.load();
2219 392 : tenant_conf
2220 392 : .tenant_conf
2221 392 : .compaction_threshold
2222 392 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2223 392 : }
2224 :
2225 14 : fn get_image_creation_threshold(&self) -> usize {
2226 14 : let tenant_conf = self.tenant_conf.load();
2227 14 : tenant_conf
2228 14 : .tenant_conf
2229 14 : .image_creation_threshold
2230 14 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2231 14 : }
2232 :
2233 364 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2234 364 : let tenant_conf = &self.tenant_conf.load();
2235 364 : tenant_conf
2236 364 : .tenant_conf
2237 364 : .compaction_algorithm
2238 364 : .as_ref()
2239 364 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2240 364 : .clone()
2241 364 : }
2242 :
2243 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2244 0 : let tenant_conf = self.tenant_conf.load();
2245 0 : tenant_conf
2246 0 : .tenant_conf
2247 0 : .eviction_policy
2248 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2249 0 : }
2250 :
2251 406 : fn get_evictions_low_residence_duration_metric_threshold(
2252 406 : tenant_conf: &TenantConfOpt,
2253 406 : default_tenant_conf: &TenantConf,
2254 406 : ) -> Duration {
2255 406 : tenant_conf
2256 406 : .evictions_low_residence_duration_metric_threshold
2257 406 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2258 406 : }
2259 :
2260 522 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2261 522 : let tenant_conf = self.tenant_conf.load();
2262 522 : tenant_conf
2263 522 : .tenant_conf
2264 522 : .image_layer_creation_check_threshold
2265 522 : .unwrap_or(
2266 522 : self.conf
2267 522 : .default_tenant_conf
2268 522 : .image_layer_creation_check_threshold,
2269 522 : )
2270 522 : }
2271 :
2272 8 : pub(super) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2273 8 : // NB: Most tenant conf options are read by background loops, so,
2274 8 : // changes will automatically be picked up.
2275 8 :
2276 8 : // The threshold is embedded in the metric. So, we need to update it.
2277 8 : {
2278 8 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2279 8 : new_conf,
2280 8 : &self.conf.default_tenant_conf,
2281 8 : );
2282 8 :
2283 8 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2284 8 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2285 8 :
2286 8 : let timeline_id_str = self.timeline_id.to_string();
2287 8 : self.metrics
2288 8 : .evictions_with_low_residence_duration
2289 8 : .write()
2290 8 : .unwrap()
2291 8 : .change_threshold(
2292 8 : &tenant_id_str,
2293 8 : &shard_id_str,
2294 8 : &timeline_id_str,
2295 8 : new_threshold,
2296 8 : );
2297 8 : }
2298 8 : }
2299 :
2300 : /// Open a Timeline handle.
2301 : ///
2302 : /// Loads the metadata for the timeline into memory, but not the layer map.
2303 : #[allow(clippy::too_many_arguments)]
2304 398 : pub(super) fn new(
2305 398 : conf: &'static PageServerConf,
2306 398 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2307 398 : metadata: &TimelineMetadata,
2308 398 : ancestor: Option<Arc<Timeline>>,
2309 398 : timeline_id: TimelineId,
2310 398 : tenant_shard_id: TenantShardId,
2311 398 : generation: Generation,
2312 398 : shard_identity: ShardIdentity,
2313 398 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2314 398 : resources: TimelineResources,
2315 398 : pg_version: u32,
2316 398 : state: TimelineState,
2317 398 : aux_file_policy: Option<AuxFilePolicy>,
2318 398 : cancel: CancellationToken,
2319 398 : ) -> Arc<Self> {
2320 398 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2321 398 : let (state, _) = watch::channel(state);
2322 398 :
2323 398 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2324 398 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2325 398 :
2326 398 : let evictions_low_residence_duration_metric_threshold = {
2327 398 : let loaded_tenant_conf = tenant_conf.load();
2328 398 : Self::get_evictions_low_residence_duration_metric_threshold(
2329 398 : &loaded_tenant_conf.tenant_conf,
2330 398 : &conf.default_tenant_conf,
2331 398 : )
2332 : };
2333 :
2334 398 : if let Some(ancestor) = &ancestor {
2335 228 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2336 228 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn());
2337 228 : }
2338 :
2339 398 : Arc::new_cyclic(|myself| {
2340 398 : let metrics = TimelineMetrics::new(
2341 398 : &tenant_shard_id,
2342 398 : &timeline_id,
2343 398 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2344 398 : "mtime",
2345 398 : evictions_low_residence_duration_metric_threshold,
2346 398 : ),
2347 398 : );
2348 398 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2349 :
2350 398 : let mut result = Timeline {
2351 398 : conf,
2352 398 : tenant_conf,
2353 398 : myself: myself.clone(),
2354 398 : timeline_id,
2355 398 : tenant_shard_id,
2356 398 : generation,
2357 398 : shard_identity,
2358 398 : pg_version,
2359 398 : layers: Default::default(),
2360 398 :
2361 398 : walredo_mgr,
2362 398 : walreceiver: Mutex::new(None),
2363 398 :
2364 398 : remote_client: Arc::new(resources.remote_client),
2365 398 :
2366 398 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2367 398 : last_record_lsn: SeqWait::new(RecordLsn {
2368 398 : last: disk_consistent_lsn,
2369 398 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2370 398 : }),
2371 398 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2372 398 :
2373 398 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2374 398 : last_freeze_ts: RwLock::new(Instant::now()),
2375 398 :
2376 398 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2377 398 :
2378 398 : ancestor_timeline: ancestor,
2379 398 : ancestor_lsn: metadata.ancestor_lsn(),
2380 398 :
2381 398 : metrics,
2382 398 :
2383 398 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2384 398 : &tenant_shard_id,
2385 398 : &timeline_id,
2386 398 : ),
2387 398 :
2388 2786 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2389 398 :
2390 398 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2391 398 :
2392 398 : layer_flush_start_tx,
2393 398 : layer_flush_done_tx,
2394 398 :
2395 398 : write_lock: tokio::sync::Mutex::new(None),
2396 398 :
2397 398 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2398 398 :
2399 398 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2400 398 : initdb_lsn: metadata.initdb_lsn(),
2401 398 :
2402 398 : current_logical_size: if disk_consistent_lsn.is_valid() {
2403 : // we're creating timeline data with some layer files existing locally,
2404 : // need to recalculate timeline's logical size based on data in the layers.
2405 232 : LogicalSize::deferred_initial(disk_consistent_lsn)
2406 : } else {
2407 : // we're creating timeline data without any layers existing locally,
2408 : // initial logical size is 0.
2409 166 : LogicalSize::empty_initial()
2410 : },
2411 398 : partitioning: tokio::sync::Mutex::new((
2412 398 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2413 398 : Lsn(0),
2414 398 : )),
2415 398 : repartition_threshold: 0,
2416 398 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2417 398 : last_image_layer_creation_check_instant: Mutex::new(None),
2418 398 :
2419 398 : last_received_wal: Mutex::new(None),
2420 398 : rel_size_cache: RwLock::new(RelSizeCache {
2421 398 : complete_as_of: disk_consistent_lsn,
2422 398 : map: HashMap::new(),
2423 398 : }),
2424 398 :
2425 398 : download_all_remote_layers_task_info: RwLock::new(None),
2426 398 :
2427 398 : state,
2428 398 :
2429 398 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2430 398 : EvictionTaskTimelineState::default(),
2431 398 : ),
2432 398 : delete_progress: Arc::new(tokio::sync::Mutex::new(DeleteTimelineFlow::default())),
2433 398 :
2434 398 : cancel,
2435 398 : gate: Gate::default(),
2436 398 :
2437 398 : compaction_lock: tokio::sync::Mutex::default(),
2438 398 : gc_lock: tokio::sync::Mutex::default(),
2439 398 :
2440 398 : standby_horizon: AtomicLsn::new(0),
2441 398 :
2442 398 : timeline_get_throttle: resources.timeline_get_throttle,
2443 398 :
2444 398 : aux_files: tokio::sync::Mutex::new(AuxFilesState {
2445 398 : dir: None,
2446 398 : n_deltas: 0,
2447 398 : }),
2448 398 :
2449 398 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2450 398 :
2451 398 : last_aux_file_policy: AtomicAuxFilePolicy::new(aux_file_policy),
2452 398 :
2453 398 : #[cfg(test)]
2454 398 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2455 398 :
2456 398 : l0_flush_global_state: resources.l0_flush_global_state,
2457 398 :
2458 398 : handles: Default::default(),
2459 398 : };
2460 398 : result.repartition_threshold =
2461 398 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2462 398 :
2463 398 : result
2464 398 : .metrics
2465 398 : .last_record_gauge
2466 398 : .set(disk_consistent_lsn.0 as i64);
2467 398 : result
2468 398 : })
2469 398 : }
2470 :
2471 550 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2472 550 : let Ok(guard) = self.gate.enter() else {
2473 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2474 0 : return;
2475 : };
2476 550 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2477 550 : match *flush_loop_state {
2478 392 : FlushLoopState::NotStarted => (),
2479 : FlushLoopState::Running { .. } => {
2480 158 : info!(
2481 0 : "skipping attempt to start flush_loop twice {}/{}",
2482 0 : self.tenant_shard_id, self.timeline_id
2483 : );
2484 158 : return;
2485 : }
2486 : FlushLoopState::Exited => {
2487 0 : warn!(
2488 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2489 0 : self.tenant_shard_id, self.timeline_id
2490 : );
2491 0 : return;
2492 : }
2493 : }
2494 :
2495 392 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2496 392 : let self_clone = Arc::clone(self);
2497 392 :
2498 392 : debug!("spawning flush loop");
2499 392 : *flush_loop_state = FlushLoopState::Running {
2500 392 : #[cfg(test)]
2501 392 : expect_initdb_optimization: false,
2502 392 : #[cfg(test)]
2503 392 : initdb_optimization_count: 0,
2504 392 : };
2505 392 : task_mgr::spawn(
2506 392 : task_mgr::BACKGROUND_RUNTIME.handle(),
2507 392 : task_mgr::TaskKind::LayerFlushTask,
2508 392 : Some(self.tenant_shard_id),
2509 392 : Some(self.timeline_id),
2510 392 : "layer flush task",
2511 392 : async move {
2512 392 : let _guard = guard;
2513 392 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2514 18004 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2515 8 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2516 8 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2517 8 : *flush_loop_state = FlushLoopState::Exited;
2518 8 : Ok(())
2519 8 : }
2520 392 : .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))
2521 : );
2522 550 : }
2523 :
2524 : /// Creates and starts the wal receiver.
2525 : ///
2526 : /// This function is expected to be called at most once per Timeline's lifecycle
2527 : /// when the timeline is activated.
2528 0 : fn launch_wal_receiver(
2529 0 : self: &Arc<Self>,
2530 0 : ctx: &RequestContext,
2531 0 : broker_client: BrokerClientChannel,
2532 0 : ) {
2533 0 : info!(
2534 0 : "launching WAL receiver for timeline {} of tenant {}",
2535 0 : self.timeline_id, self.tenant_shard_id
2536 : );
2537 :
2538 0 : let tenant_conf = self.tenant_conf.load();
2539 0 : let wal_connect_timeout = tenant_conf
2540 0 : .tenant_conf
2541 0 : .walreceiver_connect_timeout
2542 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2543 0 : let lagging_wal_timeout = tenant_conf
2544 0 : .tenant_conf
2545 0 : .lagging_wal_timeout
2546 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2547 0 : let max_lsn_wal_lag = tenant_conf
2548 0 : .tenant_conf
2549 0 : .max_lsn_wal_lag
2550 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2551 0 :
2552 0 : let mut guard = self.walreceiver.lock().unwrap();
2553 0 : assert!(
2554 0 : guard.is_none(),
2555 0 : "multiple launches / re-launches of WAL receiver are not supported"
2556 : );
2557 0 : *guard = Some(WalReceiver::start(
2558 0 : Arc::clone(self),
2559 0 : WalReceiverConf {
2560 0 : wal_connect_timeout,
2561 0 : lagging_wal_timeout,
2562 0 : max_lsn_wal_lag,
2563 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2564 0 : availability_zone: self.conf.availability_zone.clone(),
2565 0 : ingest_batch_size: self.conf.ingest_batch_size,
2566 0 : },
2567 0 : broker_client,
2568 0 : ctx,
2569 0 : ));
2570 0 : }
2571 :
2572 : /// Initialize with an empty layer map. Used when creating a new timeline.
2573 392 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2574 392 : let mut layers = self.layers.try_write().expect(
2575 392 : "in the context where we call this function, no other task has access to the object",
2576 392 : );
2577 392 : layers.initialize_empty(Lsn(start_lsn.0));
2578 392 : }
2579 :
2580 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2581 : /// files.
2582 6 : pub(super) async fn load_layer_map(
2583 6 : &self,
2584 6 : disk_consistent_lsn: Lsn,
2585 6 : index_part: Option<IndexPart>,
2586 6 : ) -> anyhow::Result<()> {
2587 : use init::{Decision::*, Discovered, DismissedLayer};
2588 : use LayerName::*;
2589 :
2590 6 : let mut guard = self.layers.write().await;
2591 :
2592 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2593 6 :
2594 6 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2595 6 : // structs representing all files on disk
2596 6 : let timeline_path = self
2597 6 : .conf
2598 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2599 6 : let conf = self.conf;
2600 6 : let span = tracing::Span::current();
2601 6 :
2602 6 : // Copy to move into the task we're about to spawn
2603 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2604 :
2605 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2606 6 : move || {
2607 6 : let _g = span.entered();
2608 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2609 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2610 6 : let mut unrecognized_files = Vec::new();
2611 6 :
2612 6 : let mut path = timeline_path;
2613 :
2614 22 : for discovered in discovered {
2615 16 : let (name, kind) = match discovered {
2616 16 : Discovered::Layer(layer_file_name, local_metadata) => {
2617 16 : discovered_layers.push((layer_file_name, local_metadata));
2618 16 : continue;
2619 : }
2620 0 : Discovered::IgnoredBackup(path) => {
2621 0 : std::fs::remove_file(path)
2622 0 : .or_else(fs_ext::ignore_not_found)
2623 0 : .fatal_err("Removing .old file");
2624 0 : continue;
2625 : }
2626 0 : Discovered::Unknown(file_name) => {
2627 0 : // we will later error if there are any
2628 0 : unrecognized_files.push(file_name);
2629 0 : continue;
2630 : }
2631 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2632 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2633 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2634 : };
2635 0 : path.push(Utf8Path::new(&name));
2636 0 : init::cleanup(&path, kind)?;
2637 0 : path.pop();
2638 : }
2639 :
2640 6 : if !unrecognized_files.is_empty() {
2641 : // assume that if there are any there are many many.
2642 0 : let n = unrecognized_files.len();
2643 0 : let first = &unrecognized_files[..n.min(10)];
2644 0 : anyhow::bail!(
2645 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2646 0 : );
2647 6 : }
2648 6 :
2649 6 : let decided =
2650 6 : init::reconcile(discovered_layers, index_part.as_ref(), disk_consistent_lsn);
2651 6 :
2652 6 : let mut loaded_layers = Vec::new();
2653 6 : let mut needs_cleanup = Vec::new();
2654 6 : let mut total_physical_size = 0;
2655 :
2656 22 : for (name, decision) in decided {
2657 16 : let decision = match decision {
2658 16 : Ok(decision) => decision,
2659 0 : Err(DismissedLayer::Future { local }) => {
2660 0 : if let Some(local) = local {
2661 0 : init::cleanup_future_layer(
2662 0 : &local.local_path,
2663 0 : &name,
2664 0 : disk_consistent_lsn,
2665 0 : )?;
2666 0 : }
2667 0 : needs_cleanup.push(name);
2668 0 : continue;
2669 : }
2670 0 : Err(DismissedLayer::LocalOnly(local)) => {
2671 0 : init::cleanup_local_only_file(&name, &local)?;
2672 : // this file never existed remotely, we will have to do rework
2673 0 : continue;
2674 : }
2675 0 : Err(DismissedLayer::BadMetadata(local)) => {
2676 0 : init::cleanup_local_file_for_remote(&local)?;
2677 : // this file never existed remotely, we will have to do rework
2678 0 : continue;
2679 : }
2680 : };
2681 :
2682 16 : match &name {
2683 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2684 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2685 : }
2686 :
2687 16 : tracing::debug!(layer=%name, ?decision, "applied");
2688 :
2689 16 : let layer = match decision {
2690 16 : Resident { local, remote } => {
2691 16 : total_physical_size += local.file_size;
2692 16 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2693 16 : .drop_eviction_guard()
2694 : }
2695 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2696 : };
2697 :
2698 16 : loaded_layers.push(layer);
2699 : }
2700 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2701 6 : }
2702 6 : })
2703 4 : .await
2704 6 : .map_err(anyhow::Error::new)
2705 6 : .and_then(|x| x)?;
2706 :
2707 6 : let num_layers = loaded_layers.len();
2708 6 :
2709 6 : guard.initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2710 6 :
2711 6 : self.remote_client
2712 6 : .schedule_layer_file_deletion(&needs_cleanup)?;
2713 6 : self.remote_client
2714 6 : .schedule_index_upload_for_file_changes()?;
2715 : // This barrier orders above DELETEs before any later operations.
2716 : // This is critical because code executing after the barrier might
2717 : // create again objects with the same key that we just scheduled for deletion.
2718 : // For example, if we just scheduled deletion of an image layer "from the future",
2719 : // later compaction might run again and re-create the same image layer.
2720 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2721 : // "same" here means same key range and LSN.
2722 : //
2723 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2724 : // the upload queue may execute the PUT first, then the DELETE.
2725 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2726 : //
2727 : // 1. a future image layer is created and uploaded
2728 : // 2. ps restart
2729 : // 3. the future layer from (1) is deleted during load layer map
2730 : // 4. image layer is re-created and uploaded
2731 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2732 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2733 : //
2734 : // See https://github.com/neondatabase/neon/issues/5878
2735 : //
2736 : // NB: generation numbers naturally protect against this because they disambiguate
2737 : // (1) and (4)
2738 6 : self.remote_client.schedule_barrier()?;
2739 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2740 : // on retry.
2741 :
2742 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2743 6 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2744 6 : self.update_layer_visibility().await;
2745 :
2746 6 : info!(
2747 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2748 : num_layers, disk_consistent_lsn, total_physical_size
2749 : );
2750 :
2751 6 : timer.stop_and_record();
2752 6 : Ok(())
2753 6 : }
2754 :
2755 : /// Retrieve current logical size of the timeline.
2756 : ///
2757 : /// The size could be lagging behind the actual number, in case
2758 : /// the initial size calculation has not been run (gets triggered on the first size access).
2759 : ///
2760 : /// return size and boolean flag that shows if the size is exact
2761 0 : pub(crate) fn get_current_logical_size(
2762 0 : self: &Arc<Self>,
2763 0 : priority: GetLogicalSizePriority,
2764 0 : ctx: &RequestContext,
2765 0 : ) -> logical_size::CurrentLogicalSize {
2766 0 : if !self.tenant_shard_id.is_shard_zero() {
2767 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2768 : // when HTTP API is serving a GET for timeline zero, return zero
2769 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2770 0 : }
2771 0 :
2772 0 : let current_size = self.current_logical_size.current_size();
2773 0 : debug!("Current size: {current_size:?}");
2774 :
2775 0 : match (current_size.accuracy(), priority) {
2776 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2777 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2778 0 : // background task will eventually deliver an exact value, we're in no rush
2779 0 : }
2780 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2781 : // background task is not ready, but user is asking for it now;
2782 : // => make the background task skip the line
2783 : // (The alternative would be to calculate the size here, but,
2784 : // it can actually take a long time if the user has a lot of rels.
2785 : // And we'll inevitable need it again; So, let the background task do the work.)
2786 0 : match self
2787 0 : .current_logical_size
2788 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2789 0 : .get()
2790 : {
2791 0 : Some(cancel) => cancel.cancel(),
2792 : None => {
2793 0 : let state = self.current_state();
2794 0 : if matches!(
2795 0 : state,
2796 : TimelineState::Broken { .. } | TimelineState::Stopping
2797 0 : ) {
2798 0 :
2799 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2800 0 : // Don't make noise.
2801 0 : } else {
2802 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2803 0 : debug_assert!(false);
2804 : }
2805 : }
2806 : };
2807 : }
2808 : }
2809 :
2810 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2811 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2812 0 : let first = self
2813 0 : .current_logical_size
2814 0 : .did_return_approximate_to_walreceiver
2815 0 : .compare_exchange(
2816 0 : false,
2817 0 : true,
2818 0 : AtomicOrdering::Relaxed,
2819 0 : AtomicOrdering::Relaxed,
2820 0 : )
2821 0 : .is_ok();
2822 0 : if first {
2823 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2824 0 : }
2825 0 : }
2826 0 : }
2827 :
2828 0 : current_size
2829 0 : }
2830 :
2831 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2832 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2833 : // nothing to do for freshly created timelines;
2834 0 : assert_eq!(
2835 0 : self.current_logical_size.current_size().accuracy(),
2836 0 : logical_size::Accuracy::Exact,
2837 0 : );
2838 0 : self.current_logical_size.initialized.add_permits(1);
2839 0 : return;
2840 : };
2841 :
2842 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2843 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2844 0 : self.current_logical_size
2845 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2846 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2847 0 :
2848 0 : let self_clone = Arc::clone(self);
2849 0 : let background_ctx = ctx.detached_child(
2850 0 : TaskKind::InitialLogicalSizeCalculation,
2851 0 : DownloadBehavior::Download,
2852 0 : );
2853 0 : task_mgr::spawn(
2854 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2855 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2856 0 : Some(self.tenant_shard_id),
2857 0 : Some(self.timeline_id),
2858 0 : "initial size calculation",
2859 : // NB: don't log errors here, task_mgr will do that.
2860 0 : async move {
2861 0 : let cancel = task_mgr::shutdown_token();
2862 0 : self_clone
2863 0 : .initial_logical_size_calculation_task(
2864 0 : initial_part_end,
2865 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2866 0 : cancel,
2867 0 : background_ctx,
2868 0 : )
2869 0 : .await;
2870 0 : Ok(())
2871 0 : }
2872 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)),
2873 : );
2874 0 : }
2875 :
2876 0 : async fn initial_logical_size_calculation_task(
2877 0 : self: Arc<Self>,
2878 0 : initial_part_end: Lsn,
2879 0 : skip_concurrency_limiter: CancellationToken,
2880 0 : cancel: CancellationToken,
2881 0 : background_ctx: RequestContext,
2882 0 : ) {
2883 : scopeguard::defer! {
2884 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2885 : self.current_logical_size.initialized.add_permits(1);
2886 : }
2887 :
2888 0 : let try_once = |attempt: usize| {
2889 0 : let background_ctx = &background_ctx;
2890 0 : let self_ref = &self;
2891 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2892 0 : async move {
2893 0 : let cancel = task_mgr::shutdown_token();
2894 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2895 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2896 0 : background_ctx,
2897 0 : );
2898 :
2899 : use crate::metrics::initial_logical_size::StartCircumstances;
2900 0 : let (_maybe_permit, circumstances) = tokio::select! {
2901 : permit = wait_for_permit => {
2902 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2903 : }
2904 : _ = self_ref.cancel.cancelled() => {
2905 : return Err(CalculateLogicalSizeError::Cancelled);
2906 : }
2907 : _ = cancel.cancelled() => {
2908 : return Err(CalculateLogicalSizeError::Cancelled);
2909 : },
2910 : () = skip_concurrency_limiter.cancelled() => {
2911 : // Some action that is part of a end user interaction requested logical size
2912 : // => break out of the rate limit
2913 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2914 : // but then again what happens if they cancel; also, we should just be using
2915 : // one runtime across the entire process, so, let's leave this for now.
2916 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2917 : }
2918 : };
2919 :
2920 0 : let metrics_guard = if attempt == 1 {
2921 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2922 : } else {
2923 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2924 : };
2925 :
2926 0 : let calculated_size = self_ref
2927 0 : .logical_size_calculation_task(
2928 0 : initial_part_end,
2929 0 : LogicalSizeCalculationCause::Initial,
2930 0 : background_ctx,
2931 0 : )
2932 0 : .await?;
2933 :
2934 0 : self_ref
2935 0 : .trigger_aux_file_size_computation(initial_part_end, background_ctx)
2936 0 : .await?;
2937 :
2938 : // TODO: add aux file size to logical size
2939 :
2940 0 : Ok((calculated_size, metrics_guard))
2941 0 : }
2942 0 : };
2943 :
2944 0 : let retrying = async {
2945 0 : let mut attempt = 0;
2946 0 : loop {
2947 0 : attempt += 1;
2948 0 :
2949 0 : match try_once(attempt).await {
2950 0 : Ok(res) => return ControlFlow::Continue(res),
2951 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
2952 : Err(
2953 0 : e @ (CalculateLogicalSizeError::Decode(_)
2954 0 : | CalculateLogicalSizeError::PageRead(_)),
2955 0 : ) => {
2956 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2957 : // exponential back-off doesn't make sense at these long intervals;
2958 : // use fixed retry interval with generous jitter instead
2959 0 : let sleep_duration = Duration::from_secs(
2960 0 : u64::try_from(
2961 0 : // 1hour base
2962 0 : (60_i64 * 60_i64)
2963 0 : // 10min jitter
2964 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2965 0 : )
2966 0 : .expect("10min < 1hour"),
2967 0 : );
2968 0 : tokio::time::sleep(sleep_duration).await;
2969 : }
2970 : }
2971 : }
2972 0 : };
2973 :
2974 0 : let (calculated_size, metrics_guard) = tokio::select! {
2975 : res = retrying => {
2976 : match res {
2977 : ControlFlow::Continue(calculated_size) => calculated_size,
2978 : ControlFlow::Break(()) => return,
2979 : }
2980 : }
2981 : _ = cancel.cancelled() => {
2982 : return;
2983 : }
2984 : };
2985 :
2986 : // we cannot query current_logical_size.current_size() to know the current
2987 : // *negative* value, only truncated to u64.
2988 0 : let added = self
2989 0 : .current_logical_size
2990 0 : .size_added_after_initial
2991 0 : .load(AtomicOrdering::Relaxed);
2992 0 :
2993 0 : let sum = calculated_size.saturating_add_signed(added);
2994 0 :
2995 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2996 0 : self.metrics.current_logical_size_gauge.set(sum);
2997 0 :
2998 0 : self.current_logical_size
2999 0 : .initial_logical_size
3000 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3001 0 : .ok()
3002 0 : .expect("only this task sets it");
3003 0 : }
3004 :
3005 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3006 0 : self: &Arc<Self>,
3007 0 : lsn: Lsn,
3008 0 : cause: LogicalSizeCalculationCause,
3009 0 : ctx: RequestContext,
3010 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3011 0 : let (sender, receiver) = oneshot::channel();
3012 0 : let self_clone = Arc::clone(self);
3013 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3014 0 : // we should stop the size calculation work and return an error.
3015 0 : // That would require restructuring this function's API to
3016 0 : // return the result directly, instead of a Receiver for the result.
3017 0 : let ctx = ctx.detached_child(
3018 0 : TaskKind::OndemandLogicalSizeCalculation,
3019 0 : DownloadBehavior::Download,
3020 0 : );
3021 0 : task_mgr::spawn(
3022 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3023 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3024 0 : Some(self.tenant_shard_id),
3025 0 : Some(self.timeline_id),
3026 0 : "ondemand logical size calculation",
3027 0 : async move {
3028 0 : let res = self_clone
3029 0 : .logical_size_calculation_task(lsn, cause, &ctx)
3030 0 : .await;
3031 0 : let _ = sender.send(res).ok();
3032 0 : Ok(()) // Receiver is responsible for handling errors
3033 0 : }
3034 0 : .in_current_span(),
3035 0 : );
3036 0 : receiver
3037 0 : }
3038 :
3039 : /// # Cancel-Safety
3040 : ///
3041 : /// This method is cancellation-safe.
3042 0 : #[instrument(skip_all)]
3043 : async fn logical_size_calculation_task(
3044 : self: &Arc<Self>,
3045 : lsn: Lsn,
3046 : cause: LogicalSizeCalculationCause,
3047 : ctx: &RequestContext,
3048 : ) -> Result<u64, CalculateLogicalSizeError> {
3049 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3050 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3051 : // accurate relation sizes, and they do not emit consumption metrics.
3052 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3053 :
3054 : let guard = self
3055 : .gate
3056 : .enter()
3057 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3058 :
3059 : let self_calculation = Arc::clone(self);
3060 :
3061 0 : let mut calculation = pin!(async {
3062 0 : let ctx = ctx.attached_child();
3063 0 : self_calculation
3064 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
3065 0 : .await
3066 0 : });
3067 :
3068 : tokio::select! {
3069 : res = &mut calculation => { res }
3070 : _ = self.cancel.cancelled() => {
3071 : debug!("cancelling logical size calculation for timeline shutdown");
3072 : calculation.await
3073 : }
3074 : }
3075 : }
3076 :
3077 : /// Calculate the logical size of the database at the latest LSN.
3078 : ///
3079 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3080 : /// especially if we need to download remote layers.
3081 : ///
3082 : /// # Cancel-Safety
3083 : ///
3084 : /// This method is cancellation-safe.
3085 0 : async fn calculate_logical_size(
3086 0 : &self,
3087 0 : up_to_lsn: Lsn,
3088 0 : cause: LogicalSizeCalculationCause,
3089 0 : _guard: &GateGuard,
3090 0 : ctx: &RequestContext,
3091 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3092 0 : info!(
3093 0 : "Calculating logical size for timeline {} at {}",
3094 : self.timeline_id, up_to_lsn
3095 : );
3096 :
3097 : pausable_failpoint!("timeline-calculate-logical-size-pause");
3098 :
3099 : // See if we've already done the work for initial size calculation.
3100 : // This is a short-cut for timelines that are mostly unused.
3101 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3102 0 : return Ok(size);
3103 0 : }
3104 0 : let storage_time_metrics = match cause {
3105 : LogicalSizeCalculationCause::Initial
3106 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3107 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3108 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3109 0 : &self.metrics.imitate_logical_size_histo
3110 : }
3111 : };
3112 0 : let timer = storage_time_metrics.start_timer();
3113 0 : let logical_size = self
3114 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3115 0 : .await?;
3116 0 : debug!("calculated logical size: {logical_size}");
3117 0 : timer.stop_and_record();
3118 0 : Ok(logical_size)
3119 0 : }
3120 :
3121 : /// Update current logical size, adding `delta' to the old value.
3122 270570 : fn update_current_logical_size(&self, delta: i64) {
3123 270570 : let logical_size = &self.current_logical_size;
3124 270570 : logical_size.increment_size(delta);
3125 270570 :
3126 270570 : // Also set the value in the prometheus gauge. Note that
3127 270570 : // there is a race condition here: if this is is called by two
3128 270570 : // threads concurrently, the prometheus gauge might be set to
3129 270570 : // one value while current_logical_size is set to the
3130 270570 : // other.
3131 270570 : match logical_size.current_size() {
3132 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
3133 270570 : .metrics
3134 270570 : .current_logical_size_gauge
3135 270570 : .set(new_current_size.into()),
3136 0 : CurrentLogicalSize::Approximate(_) => {
3137 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3138 0 : // forth between the initial size calculation task.
3139 0 : }
3140 : }
3141 270570 : }
3142 :
3143 2940 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3144 2940 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3145 2940 : let aux_metric =
3146 2940 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3147 2940 :
3148 2940 : let sum_of_entries = self
3149 2940 : .directory_metrics
3150 2940 : .iter()
3151 20580 : .map(|v| v.load(AtomicOrdering::Relaxed))
3152 2940 : .sum();
3153 2940 : // Set a high general threshold and a lower threshold for the auxiliary files,
3154 2940 : // as we can have large numbers of relations in the db directory.
3155 2940 : const SUM_THRESHOLD: u64 = 5000;
3156 2940 : const AUX_THRESHOLD: u64 = 1000;
3157 2940 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3158 0 : self.metrics
3159 0 : .directory_entries_count_gauge
3160 0 : .set(sum_of_entries);
3161 2940 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3162 0 : metric.set(sum_of_entries);
3163 2940 : }
3164 2940 : }
3165 :
3166 0 : async fn find_layer(&self, layer_name: &LayerName) -> Option<Layer> {
3167 0 : let guard = self.layers.read().await;
3168 0 : for historic_layer in guard.layer_map().iter_historic_layers() {
3169 0 : let historic_layer_name = historic_layer.layer_name();
3170 0 : if layer_name == &historic_layer_name {
3171 0 : return Some(guard.get_from_desc(&historic_layer));
3172 0 : }
3173 : }
3174 :
3175 0 : None
3176 0 : }
3177 :
3178 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3179 : /// indicating which layers are currently on-disk on the primary.
3180 : ///
3181 : /// None is returned if the Timeline is in a state where uploading a heatmap
3182 : /// doesn't make sense, such as shutting down or initializing. The caller
3183 : /// should treat this as a cue to simply skip doing any heatmap uploading
3184 : /// for this timeline.
3185 0 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3186 0 : if !self.is_active() {
3187 0 : return None;
3188 0 : }
3189 :
3190 0 : let guard = self.layers.read().await;
3191 :
3192 0 : let resident = guard.likely_resident_layers().map(|layer| {
3193 0 : let last_activity_ts = layer.access_stats().latest_activity();
3194 0 :
3195 0 : HeatMapLayer::new(
3196 0 : layer.layer_desc().layer_name(),
3197 0 : layer.metadata(),
3198 0 : last_activity_ts,
3199 0 : )
3200 0 : });
3201 0 :
3202 0 : let layers = resident.collect();
3203 0 :
3204 0 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3205 0 : }
3206 :
3207 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3208 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3209 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3210 0 : // branchpoint in the value in IndexPart::lineage
3211 0 : self.ancestor_lsn == lsn
3212 0 : || (self.ancestor_lsn == Lsn::INVALID
3213 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3214 0 : }
3215 : }
3216 :
3217 : type TraversalId = Arc<str>;
3218 :
3219 : trait TraversalLayerExt {
3220 : fn traversal_id(&self) -> TraversalId;
3221 : }
3222 :
3223 : impl TraversalLayerExt for Layer {
3224 320 : fn traversal_id(&self) -> TraversalId {
3225 320 : Arc::clone(self.debug_str())
3226 320 : }
3227 : }
3228 :
3229 : impl TraversalLayerExt for Arc<InMemoryLayer> {
3230 0 : fn traversal_id(&self) -> TraversalId {
3231 0 : Arc::clone(self.local_path_str())
3232 0 : }
3233 : }
3234 :
3235 : impl Timeline {
3236 : ///
3237 : /// Get a handle to a Layer for reading.
3238 : ///
3239 : /// The returned Layer might be from an ancestor timeline, if the
3240 : /// segment hasn't been updated on this timeline yet.
3241 : ///
3242 : /// This function takes the current timeline's locked LayerMap as an argument,
3243 : /// so callers can avoid potential race conditions.
3244 : ///
3245 : /// # Cancel-Safety
3246 : ///
3247 : /// This method is cancellation-safe.
3248 322 : async fn get_reconstruct_data(
3249 322 : &self,
3250 322 : key: Key,
3251 322 : request_lsn: Lsn,
3252 322 : reconstruct_state: &mut ValueReconstructState,
3253 322 : ctx: &RequestContext,
3254 322 : ) -> Result<Vec<TraversalPathItem>, PageReconstructError> {
3255 322 : // Start from the current timeline.
3256 322 : let mut timeline_owned;
3257 322 : let mut timeline = self;
3258 322 :
3259 322 : let mut read_count = scopeguard::guard(0, |cnt| {
3260 322 : crate::metrics::READ_NUM_LAYERS_VISITED.observe(cnt as f64)
3261 322 : });
3262 322 :
3263 322 : // For debugging purposes, collect the path of layers that we traversed
3264 322 : // through. It's included in the error message if we fail to find the key.
3265 322 : let mut traversal_path = Vec::<TraversalPathItem>::new();
3266 :
3267 322 : let cached_lsn = if let Some((cached_lsn, _)) = &reconstruct_state.img {
3268 0 : *cached_lsn
3269 : } else {
3270 322 : Lsn(0)
3271 : };
3272 :
3273 : // 'prev_lsn' tracks the last LSN that we were at in our search. It's used
3274 : // to check that each iteration make some progress, to break infinite
3275 : // looping if something goes wrong.
3276 322 : let mut prev_lsn = None;
3277 322 :
3278 322 : let mut result = ValueReconstructResult::Continue;
3279 322 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3280 :
3281 644 : 'outer: loop {
3282 644 : if self.cancel.is_cancelled() {
3283 0 : return Err(PageReconstructError::Cancelled);
3284 644 : }
3285 644 :
3286 644 : // The function should have updated 'state'
3287 644 : //info!("CALLED for {} at {}: {:?} with {} records, cached {}", key, cont_lsn, result, reconstruct_state.records.len(), cached_lsn);
3288 644 : match result {
3289 320 : ValueReconstructResult::Complete => return Ok(traversal_path),
3290 : ValueReconstructResult::Continue => {
3291 : // If we reached an earlier cached page image, we're done.
3292 324 : if cont_lsn == cached_lsn + 1 {
3293 0 : return Ok(traversal_path);
3294 324 : }
3295 324 : if let Some(prev) = prev_lsn {
3296 2 : if prev <= cont_lsn {
3297 : // Didn't make any progress in last iteration. Error out to avoid
3298 : // getting stuck in the loop.
3299 2 : return Err(PageReconstructError::MissingKey(MissingKeyError {
3300 2 : key,
3301 2 : shard: self.shard_identity.get_shard_number(&key),
3302 2 : cont_lsn: Lsn(cont_lsn.0 - 1),
3303 2 : request_lsn,
3304 2 : ancestor_lsn: Some(timeline.ancestor_lsn),
3305 2 : traversal_path,
3306 2 : backtrace: None,
3307 2 : }));
3308 0 : }
3309 322 : }
3310 322 : prev_lsn = Some(cont_lsn);
3311 : }
3312 : ValueReconstructResult::Missing => {
3313 : return Err(PageReconstructError::MissingKey(MissingKeyError {
3314 0 : key,
3315 0 : shard: self.shard_identity.get_shard_number(&key),
3316 0 : cont_lsn,
3317 0 : request_lsn,
3318 0 : ancestor_lsn: None,
3319 0 : traversal_path,
3320 0 : backtrace: if cfg!(test) {
3321 0 : Some(std::backtrace::Backtrace::force_capture())
3322 : } else {
3323 0 : None
3324 : },
3325 : }));
3326 : }
3327 : }
3328 :
3329 : // Recurse into ancestor if needed
3330 322 : if let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() {
3331 2 : if key.is_inherited_key() && Lsn(cont_lsn.0 - 1) <= timeline.ancestor_lsn {
3332 0 : trace!(
3333 0 : "going into ancestor {}, cont_lsn is {}",
3334 : timeline.ancestor_lsn,
3335 : cont_lsn
3336 : );
3337 :
3338 0 : timeline_owned = timeline
3339 0 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3340 0 : .await?;
3341 0 : timeline = &*timeline_owned;
3342 0 : prev_lsn = None;
3343 0 : continue 'outer;
3344 2 : }
3345 320 : }
3346 :
3347 322 : let guard = timeline.layers.read().await;
3348 322 : let layers = guard.layer_map();
3349 :
3350 : // Check the open and frozen in-memory layers first, in order from newest
3351 : // to oldest.
3352 322 : if let Some(open_layer) = &layers.open_layer {
3353 0 : let start_lsn = open_layer.get_lsn_range().start;
3354 0 : if cont_lsn > start_lsn {
3355 : //info!("CHECKING for {} at {} on open layer {}", key, cont_lsn, open_layer.layer_name().display());
3356 : // Get all the data needed to reconstruct the page version from this layer.
3357 : // But if we have an older cached page image, no need to go past that.
3358 0 : let lsn_floor = max(cached_lsn + 1, start_lsn);
3359 0 :
3360 0 : let open_layer = open_layer.clone();
3361 0 : drop(guard);
3362 0 :
3363 0 : result = match open_layer
3364 0 : .get_value_reconstruct_data(
3365 0 : key,
3366 0 : lsn_floor..cont_lsn,
3367 0 : reconstruct_state,
3368 0 : ctx,
3369 0 : )
3370 0 : .await
3371 : {
3372 0 : Ok(result) => result,
3373 0 : Err(e) => return Err(PageReconstructError::from(e)),
3374 : };
3375 0 : cont_lsn = lsn_floor;
3376 0 : *read_count += 1;
3377 0 : traversal_path.push((result, cont_lsn, open_layer.traversal_id()));
3378 0 : continue 'outer;
3379 0 : }
3380 322 : }
3381 322 : for frozen_layer in layers.frozen_layers.iter().rev() {
3382 0 : let start_lsn = frozen_layer.get_lsn_range().start;
3383 0 : if cont_lsn > start_lsn {
3384 : //info!("CHECKING for {} at {} on frozen layer {}", key, cont_lsn, frozen_layer.layer_name().display());
3385 0 : let lsn_floor = max(cached_lsn + 1, start_lsn);
3386 0 :
3387 0 : let frozen_layer = frozen_layer.clone();
3388 0 : drop(guard);
3389 0 :
3390 0 : result = match frozen_layer
3391 0 : .get_value_reconstruct_data(
3392 0 : key,
3393 0 : lsn_floor..cont_lsn,
3394 0 : reconstruct_state,
3395 0 : ctx,
3396 0 : )
3397 0 : .await
3398 : {
3399 0 : Ok(result) => result,
3400 0 : Err(e) => return Err(PageReconstructError::from(e)),
3401 : };
3402 0 : cont_lsn = lsn_floor;
3403 0 : *read_count += 1;
3404 0 : traversal_path.push((result, cont_lsn, frozen_layer.traversal_id()));
3405 0 : continue 'outer;
3406 0 : }
3407 : }
3408 :
3409 322 : if let Some(SearchResult { lsn_floor, layer }) = layers.search(key, cont_lsn) {
3410 320 : let layer = guard.get_from_desc(&layer);
3411 320 : drop(guard);
3412 320 : // Get all the data needed to reconstruct the page version from this layer.
3413 320 : // But if we have an older cached page image, no need to go past that.
3414 320 : let lsn_floor = max(cached_lsn + 1, lsn_floor);
3415 320 : result = match layer
3416 320 : .get_value_reconstruct_data(key, lsn_floor..cont_lsn, reconstruct_state, ctx)
3417 17 : .await
3418 : {
3419 320 : Ok(result) => result,
3420 0 : Err(e) => return Err(PageReconstructError::from(e)),
3421 : };
3422 320 : cont_lsn = lsn_floor;
3423 320 : *read_count += 1;
3424 320 : traversal_path.push((result, cont_lsn, layer.traversal_id()));
3425 320 : continue 'outer;
3426 2 : } else if timeline.ancestor_timeline.is_some() {
3427 : // Nothing on this timeline. Traverse to parent
3428 2 : result = ValueReconstructResult::Continue;
3429 2 : cont_lsn = Lsn(timeline.ancestor_lsn.0 + 1);
3430 2 : continue 'outer;
3431 : } else {
3432 : // Nothing found
3433 0 : result = ValueReconstructResult::Missing;
3434 0 : continue 'outer;
3435 : }
3436 : }
3437 322 : }
3438 :
3439 : #[allow(clippy::doc_lazy_continuation)]
3440 : /// Get the data needed to reconstruct all keys in the provided keyspace
3441 : ///
3442 : /// The algorithm is as follows:
3443 : /// 1. While some keys are still not done and there's a timeline to visit:
3444 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3445 : /// 2.1: Build the fringe for the current keyspace
3446 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3447 : /// intersects
3448 : /// 2.3. Pop the timeline from the fringe
3449 : /// 2.4. If the fringe is empty, go back to 1
3450 626048 : async fn get_vectored_reconstruct_data(
3451 626048 : &self,
3452 626048 : mut keyspace: KeySpace,
3453 626048 : request_lsn: Lsn,
3454 626048 : reconstruct_state: &mut ValuesReconstructState,
3455 626048 : ctx: &RequestContext,
3456 626048 : ) -> Result<(), GetVectoredError> {
3457 626048 : let mut timeline_owned: Arc<Timeline>;
3458 626048 : let mut timeline = self;
3459 626048 :
3460 626048 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3461 :
3462 626046 : let missing_keyspace = loop {
3463 851260 : if self.cancel.is_cancelled() {
3464 0 : return Err(GetVectoredError::Cancelled);
3465 851260 : }
3466 :
3467 : let TimelineVisitOutcome {
3468 851260 : completed_keyspace: completed,
3469 851260 : image_covered_keyspace,
3470 851260 : } = Self::get_vectored_reconstruct_data_timeline(
3471 851260 : timeline,
3472 851260 : keyspace.clone(),
3473 851260 : cont_lsn,
3474 851260 : reconstruct_state,
3475 851260 : &self.cancel,
3476 851260 : ctx,
3477 851260 : )
3478 114010 : .await?;
3479 :
3480 851260 : keyspace.remove_overlapping_with(&completed);
3481 851260 :
3482 851260 : // Do not descend into the ancestor timeline for aux files.
3483 851260 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3484 851260 : // stalling compaction.
3485 851260 : keyspace.remove_overlapping_with(&KeySpace {
3486 851260 : ranges: vec![NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE],
3487 851260 : });
3488 851260 :
3489 851260 : // Keyspace is fully retrieved
3490 851260 : if keyspace.is_empty() {
3491 626032 : break None;
3492 225228 : }
3493 :
3494 225228 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3495 : // Not fully retrieved but no ancestor timeline.
3496 14 : break Some(keyspace);
3497 : };
3498 :
3499 : // Now we see if there are keys covered by the image layer but does not exist in the
3500 : // image layer, which means that the key does not exist.
3501 :
3502 : // The block below will stop the vectored search if any of the keys encountered an image layer
3503 : // which did not contain a snapshot for said key. Since we have already removed all completed
3504 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3505 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3506 : // and stop the search as a result of that.
3507 225214 : let removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3508 225214 : if !removed.is_empty() {
3509 0 : break Some(removed);
3510 225214 : }
3511 225214 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3512 225214 : // keyspace.
3513 225214 :
3514 225214 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3515 225214 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3516 225214 : timeline_owned = timeline
3517 225214 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3518 0 : .await
3519 225214 : .map_err(GetVectoredError::GetReadyAncestorError)?;
3520 225212 : timeline = &*timeline_owned;
3521 : };
3522 :
3523 626046 : if let Some(missing_keyspace) = missing_keyspace {
3524 14 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3525 14 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3526 14 : shard: self
3527 14 : .shard_identity
3528 14 : .get_shard_number(&missing_keyspace.start().unwrap()),
3529 14 : cont_lsn,
3530 14 : request_lsn,
3531 14 : ancestor_lsn: Some(timeline.ancestor_lsn),
3532 14 : traversal_path: vec![],
3533 14 : backtrace: None,
3534 14 : }));
3535 626032 : }
3536 626032 :
3537 626032 : Ok(())
3538 626048 : }
3539 :
3540 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3541 : ///
3542 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3543 : /// the current keyspace. The current keyspace of the search at any given timeline
3544 : /// is the original keyspace minus all the keys that have been completed minus
3545 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3546 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3547 : ///
3548 : /// This is basically a depth-first search visitor implementation where a vertex
3549 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3550 : ///
3551 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3552 : /// and get all the required reconstruct data from the layer in one go.
3553 : ///
3554 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3555 : /// decides how to deal with these two keyspaces.
3556 851260 : async fn get_vectored_reconstruct_data_timeline(
3557 851260 : timeline: &Timeline,
3558 851260 : keyspace: KeySpace,
3559 851260 : mut cont_lsn: Lsn,
3560 851260 : reconstruct_state: &mut ValuesReconstructState,
3561 851260 : cancel: &CancellationToken,
3562 851260 : ctx: &RequestContext,
3563 851260 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3564 851260 : let mut unmapped_keyspace = keyspace.clone();
3565 851260 : let mut fringe = LayerFringe::new();
3566 851260 :
3567 851260 : let mut completed_keyspace = KeySpace::default();
3568 851260 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3569 :
3570 1669354 : loop {
3571 1669354 : if cancel.is_cancelled() {
3572 0 : return Err(GetVectoredError::Cancelled);
3573 1669354 : }
3574 1669354 :
3575 1669354 : let (keys_done_last_step, keys_with_image_coverage) =
3576 1669354 : reconstruct_state.consume_done_keys();
3577 1669354 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3578 1669354 : completed_keyspace.merge(&keys_done_last_step);
3579 1669354 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3580 7598 : unmapped_keyspace
3581 7598 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3582 7598 : image_covered_keyspace.add_range(keys_with_image_coverage);
3583 1661756 : }
3584 :
3585 : // Do not descent any further if the last layer we visited
3586 : // completed all keys in the keyspace it inspected. This is not
3587 : // required for correctness, but avoids visiting extra layers
3588 : // which turns out to be a perf bottleneck in some cases.
3589 1669354 : if !unmapped_keyspace.is_empty() {
3590 1043448 : let guard = timeline.layers.read().await;
3591 1043448 : let layers = guard.layer_map();
3592 1043448 :
3593 1043448 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3594 910896 : let start_lsn = l.get_lsn_range().start;
3595 910896 : cont_lsn > start_lsn
3596 1043448 : });
3597 1043448 :
3598 1043448 : match in_memory_layer {
3599 606197 : Some(l) => {
3600 606197 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3601 606197 : fringe.update(
3602 606197 : ReadableLayer::InMemoryLayer(l),
3603 606197 : unmapped_keyspace.clone(),
3604 606197 : lsn_range,
3605 606197 : );
3606 606197 : }
3607 : None => {
3608 509090 : for range in unmapped_keyspace.ranges.iter() {
3609 509090 : let results = layers.range_search(range.clone(), cont_lsn);
3610 509090 :
3611 509090 : results
3612 509090 : .found
3613 509090 : .into_iter()
3614 509090 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3615 275713 : (
3616 275713 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3617 275713 : keyspace_accum.to_keyspace(),
3618 275713 : lsn_floor..cont_lsn,
3619 275713 : )
3620 509090 : })
3621 509090 : .for_each(|(layer, keyspace, lsn_range)| {
3622 275713 : fringe.update(layer, keyspace, lsn_range)
3623 509090 : });
3624 509090 : }
3625 : }
3626 : }
3627 :
3628 : // It's safe to drop the layer map lock after planning the next round of reads.
3629 : // The fringe keeps readable handles for the layers which are safe to read even
3630 : // if layers were compacted or flushed.
3631 : //
3632 : // The more interesting consideration is: "Why is the read algorithm still correct
3633 : // if the layer map changes while it is operating?". Doing a vectored read on a
3634 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3635 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3636 : // range at *a particular point in time*. It is fine for the answer to be different
3637 : // at two different time points.
3638 1043448 : drop(guard);
3639 625906 : }
3640 :
3641 1669354 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3642 818094 : let next_cont_lsn = lsn_range.start;
3643 818094 : layer_to_read
3644 818094 : .get_values_reconstruct_data(
3645 818094 : keyspace_to_read.clone(),
3646 818094 : lsn_range,
3647 818094 : reconstruct_state,
3648 818094 : ctx,
3649 818094 : )
3650 107815 : .await?;
3651 :
3652 818094 : unmapped_keyspace = keyspace_to_read;
3653 818094 : cont_lsn = next_cont_lsn;
3654 818094 :
3655 818094 : reconstruct_state.on_layer_visited(&layer_to_read);
3656 : } else {
3657 851260 : break;
3658 851260 : }
3659 851260 : }
3660 851260 :
3661 851260 : Ok(TimelineVisitOutcome {
3662 851260 : completed_keyspace,
3663 851260 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3664 851260 : })
3665 851260 : }
3666 :
3667 225214 : async fn get_ready_ancestor_timeline(
3668 225214 : &self,
3669 225214 : ancestor: &Arc<Timeline>,
3670 225214 : ctx: &RequestContext,
3671 225214 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3672 225214 : // It's possible that the ancestor timeline isn't active yet, or
3673 225214 : // is active but hasn't yet caught up to the branch point. Wait
3674 225214 : // for it.
3675 225214 : //
3676 225214 : // This cannot happen while the pageserver is running normally,
3677 225214 : // because you cannot create a branch from a point that isn't
3678 225214 : // present in the pageserver yet. However, we don't wait for the
3679 225214 : // branch point to be uploaded to cloud storage before creating
3680 225214 : // a branch. I.e., the branch LSN need not be remote consistent
3681 225214 : // for the branching operation to succeed.
3682 225214 : //
3683 225214 : // Hence, if we try to load a tenant in such a state where
3684 225214 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3685 225214 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3686 225214 : // then we will need to wait for the ancestor timeline to
3687 225214 : // re-stream WAL up to branch_lsn before we access it.
3688 225214 : //
3689 225214 : // How can a tenant get in such a state?
3690 225214 : // - ungraceful pageserver process exit
3691 225214 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3692 225214 : //
3693 225214 : // NB: this could be avoided by requiring
3694 225214 : // branch_lsn >= remote_consistent_lsn
3695 225214 : // during branch creation.
3696 225214 : match ancestor.wait_to_become_active(ctx).await {
3697 225212 : Ok(()) => {}
3698 : Err(TimelineState::Stopping) => {
3699 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3700 0 : return Err(GetReadyAncestorError::Cancelled);
3701 : }
3702 2 : Err(state) => {
3703 2 : return Err(GetReadyAncestorError::BadState {
3704 2 : timeline_id: ancestor.timeline_id,
3705 2 : state,
3706 2 : });
3707 : }
3708 : }
3709 225212 : ancestor
3710 225212 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3711 0 : .await
3712 225212 : .map_err(|e| match e {
3713 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3714 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3715 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3716 0 : timeline_id: ancestor.timeline_id,
3717 0 : state,
3718 0 : },
3719 225212 : })?;
3720 :
3721 225212 : Ok(ancestor.clone())
3722 225214 : }
3723 :
3724 0 : pub(crate) fn get_ancestor_timeline(&self) -> Option<Arc<Timeline>> {
3725 0 : self.ancestor_timeline.clone()
3726 0 : }
3727 :
3728 5452 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3729 5452 : &self.shard_identity
3730 5452 : }
3731 :
3732 : #[inline(always)]
3733 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3734 0 : ShardTimelineId {
3735 0 : shard_index: ShardIndex {
3736 0 : shard_number: self.shard_identity.number,
3737 0 : shard_count: self.shard_identity.count,
3738 0 : },
3739 0 : timeline_id: self.timeline_id,
3740 0 : }
3741 0 : }
3742 :
3743 : ///
3744 : /// Get a handle to the latest layer for appending.
3745 : ///
3746 1256 : async fn get_layer_for_write(
3747 1256 : &self,
3748 1256 : lsn: Lsn,
3749 1256 : ctx: &RequestContext,
3750 1256 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3751 1256 : let mut guard = self.layers.write().await;
3752 1256 : let layer = guard
3753 1256 : .get_layer_for_write(
3754 1256 : lsn,
3755 1256 : self.get_last_record_lsn(),
3756 1256 : self.conf,
3757 1256 : self.timeline_id,
3758 1256 : self.tenant_shard_id,
3759 1256 : ctx,
3760 1256 : )
3761 704 : .await?;
3762 1256 : Ok(layer)
3763 1256 : }
3764 :
3765 5279062 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3766 5279062 : assert!(new_lsn.is_aligned());
3767 :
3768 5279062 : self.metrics.last_record_gauge.set(new_lsn.0 as i64);
3769 5279062 : self.last_record_lsn.advance(new_lsn);
3770 5279062 : }
3771 :
3772 1154 : async fn freeze_inmem_layer_at(
3773 1154 : &self,
3774 1154 : at: Lsn,
3775 1154 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3776 1154 : ) {
3777 1154 : let frozen = {
3778 1154 : let mut guard = self.layers.write().await;
3779 1154 : guard
3780 1154 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock)
3781 1 : .await
3782 : };
3783 1154 : if frozen {
3784 1126 : let now = Instant::now();
3785 1126 : *(self.last_freeze_ts.write().unwrap()) = now;
3786 1126 : }
3787 1154 : }
3788 :
3789 : /// Layer flusher task's main loop.
3790 392 : async fn flush_loop(
3791 392 : self: &Arc<Self>,
3792 392 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3793 392 : ctx: &RequestContext,
3794 392 : ) {
3795 392 : info!("started flush loop");
3796 1114 : loop {
3797 1114 : tokio::select! {
3798 : _ = self.cancel.cancelled() => {
3799 : info!("shutting down layer flush task due to Timeline::cancel");
3800 : break;
3801 : },
3802 : _ = layer_flush_start_rx.changed() => {}
3803 : }
3804 1114 : trace!("waking up");
3805 1114 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3806 1114 :
3807 1114 : // The highest LSN to which we flushed in the loop over frozen layers
3808 1114 : let mut flushed_to_lsn = Lsn(0);
3809 :
3810 1114 : let result = loop {
3811 2240 : if self.cancel.is_cancelled() {
3812 0 : info!("dropping out of flush loop for timeline shutdown");
3813 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3814 : // anyone waiting on that will respect self.cancel as well: they will stop
3815 : // waiting at the same time we as drop out of this loop.
3816 0 : return;
3817 2240 : }
3818 2240 :
3819 2240 : let timer = self.metrics.flush_time_histo.start_timer();
3820 :
3821 2240 : let layer_to_flush = {
3822 2240 : let guard = self.layers.read().await;
3823 2240 : guard.layer_map().frozen_layers.front().cloned()
3824 : // drop 'layers' lock to allow concurrent reads and writes
3825 : };
3826 2240 : let Some(layer_to_flush) = layer_to_flush else {
3827 1114 : break Ok(());
3828 : };
3829 17039 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3830 1126 : Ok(this_layer_to_lsn) => {
3831 1126 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3832 1126 : }
3833 : Err(FlushLayerError::Cancelled) => {
3834 0 : info!("dropping out of flush loop for timeline shutdown");
3835 0 : return;
3836 : }
3837 0 : err @ Err(
3838 0 : FlushLayerError::NotRunning(_)
3839 0 : | FlushLayerError::Other(_)
3840 0 : | FlushLayerError::CreateImageLayersError(_),
3841 0 : ) => {
3842 0 : error!("could not flush frozen layer: {err:?}");
3843 0 : break err.map(|_| ());
3844 : }
3845 : }
3846 1126 : timer.stop_and_record();
3847 : };
3848 :
3849 : // Unsharded tenants should never advance their LSN beyond the end of the
3850 : // highest layer they write: such gaps between layer data and the frozen LSN
3851 : // are only legal on sharded tenants.
3852 1114 : debug_assert!(
3853 1114 : self.shard_identity.count.count() > 1
3854 1114 : || flushed_to_lsn >= frozen_to_lsn
3855 68 : || !flushed_to_lsn.is_valid()
3856 : );
3857 :
3858 1114 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3859 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3860 : // to us via layer_flush_start_rx, then advance it here.
3861 : //
3862 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3863 : // never encounter a gap in the wal.
3864 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3865 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3866 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3867 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3868 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3869 0 : }
3870 0 : }
3871 1114 : }
3872 :
3873 : // Notify any listeners that we're done
3874 1114 : let _ = self
3875 1114 : .layer_flush_done_tx
3876 1114 : .send_replace((flush_counter, result));
3877 : }
3878 8 : }
3879 :
3880 : /// Request the flush loop to write out all frozen layers up to `at_lsn` as Delta L0 files to disk.
3881 : /// The caller is responsible for the freezing, e.g., [`Self::freeze_inmem_layer_at`].
3882 : ///
3883 : /// `at_lsn` may be higher than the highest LSN of a frozen layer: if this is the
3884 : /// case, it means no data will be written between the top of the highest frozen layer and
3885 : /// to_lsn, e.g. because this tenant shard has ingested up to to_lsn and not written any data
3886 : /// locally for that part of the WAL.
3887 1154 : fn flush_frozen_layers(&self, at_lsn: Lsn) -> Result<u64, FlushLayerError> {
3888 1154 : // Increment the flush cycle counter and wake up the flush task.
3889 1154 : // Remember the new value, so that when we listen for the flush
3890 1154 : // to finish, we know when the flush that we initiated has
3891 1154 : // finished, instead of some other flush that was started earlier.
3892 1154 : let mut my_flush_request = 0;
3893 1154 :
3894 1154 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3895 1154 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3896 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3897 1154 : }
3898 1154 :
3899 1154 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3900 1154 : my_flush_request = *counter + 1;
3901 1154 : *counter = my_flush_request;
3902 1154 : *lsn = std::cmp::max(at_lsn, *lsn);
3903 1154 : });
3904 1154 :
3905 1154 : Ok(my_flush_request)
3906 1154 : }
3907 :
3908 1074 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3909 1074 : let mut rx = self.layer_flush_done_tx.subscribe();
3910 2148 : loop {
3911 2148 : {
3912 2148 : let (last_result_counter, last_result) = &*rx.borrow();
3913 2148 : if *last_result_counter >= request {
3914 1074 : if let Err(err) = last_result {
3915 : // We already logged the original error in
3916 : // flush_loop. We cannot propagate it to the caller
3917 : // here, because it might not be Cloneable
3918 0 : return Err(err.clone());
3919 : } else {
3920 1074 : return Ok(());
3921 : }
3922 1074 : }
3923 1074 : }
3924 1074 : trace!("waiting for flush to complete");
3925 : tokio::select! {
3926 : rx_e = rx.changed() => {
3927 0 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3928 : },
3929 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3930 : // the notification from [`flush_loop`] that it completed.
3931 : _ = self.cancel.cancelled() => {
3932 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3933 : return Ok(())
3934 : }
3935 : };
3936 1074 : trace!("done")
3937 : }
3938 1074 : }
3939 :
3940 1074 : async fn flush_frozen_layers_and_wait(&self, at_lsn: Lsn) -> Result<(), FlushLayerError> {
3941 1074 : let token = self.flush_frozen_layers(at_lsn)?;
3942 1077 : self.wait_flush_completion(token).await
3943 1074 : }
3944 :
3945 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3946 : ///
3947 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3948 2252 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3949 : async fn flush_frozen_layer(
3950 : self: &Arc<Self>,
3951 : frozen_layer: Arc<InMemoryLayer>,
3952 : ctx: &RequestContext,
3953 : ) -> Result<Lsn, FlushLayerError> {
3954 : debug_assert_current_span_has_tenant_and_timeline_id();
3955 :
3956 : // As a special case, when we have just imported an image into the repository,
3957 : // instead of writing out a L0 delta layer, we directly write out image layer
3958 : // files instead. This is possible as long as *all* the data imported into the
3959 : // repository have the same LSN.
3960 : let lsn_range = frozen_layer.get_lsn_range();
3961 :
3962 : // Whether to directly create image layers for this flush, or flush them as delta layers
3963 : let create_image_layer =
3964 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3965 :
3966 : #[cfg(test)]
3967 : {
3968 : match &mut *self.flush_loop_state.lock().unwrap() {
3969 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3970 : panic!("flush loop not running")
3971 : }
3972 : FlushLoopState::Running {
3973 : expect_initdb_optimization,
3974 : initdb_optimization_count,
3975 : ..
3976 : } => {
3977 : if create_image_layer {
3978 : *initdb_optimization_count += 1;
3979 : } else {
3980 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3981 : }
3982 : }
3983 : }
3984 : }
3985 :
3986 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3987 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3988 : // require downloading anything during initial import.
3989 : let ((rel_partition, metadata_partition), _lsn) = self
3990 : .repartition(
3991 : self.initdb_lsn,
3992 : self.get_compaction_target_size(),
3993 : EnumSet::empty(),
3994 : ctx,
3995 : )
3996 : .await
3997 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
3998 :
3999 : if self.cancel.is_cancelled() {
4000 : return Err(FlushLayerError::Cancelled);
4001 : }
4002 :
4003 : // FIXME(auxfilesv2): support multiple metadata key partitions might need initdb support as well?
4004 : // This code path will not be hit during regression tests. After #7099 we have a single partition
4005 : // with two key ranges. If someone wants to fix initdb optimization in the future, this might need
4006 : // to be fixed.
4007 :
4008 : // For metadata, always create delta layers.
4009 : let delta_layer = if !metadata_partition.parts.is_empty() {
4010 : assert_eq!(
4011 : metadata_partition.parts.len(),
4012 : 1,
4013 : "currently sparse keyspace should only contain a single metadata keyspace"
4014 : );
4015 : let metadata_keyspace = &metadata_partition.parts[0];
4016 : self.create_delta_layer(
4017 : &frozen_layer,
4018 : Some(
4019 : metadata_keyspace.0.ranges.first().unwrap().start
4020 : ..metadata_keyspace.0.ranges.last().unwrap().end,
4021 : ),
4022 : ctx,
4023 : )
4024 : .await
4025 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
4026 : } else {
4027 : None
4028 : };
4029 :
4030 : // For image layers, we add them immediately into the layer map.
4031 : let mut layers_to_upload = Vec::new();
4032 : layers_to_upload.extend(
4033 : self.create_image_layers(
4034 : &rel_partition,
4035 : self.initdb_lsn,
4036 : ImageLayerCreationMode::Initial,
4037 : ctx,
4038 : )
4039 : .await?,
4040 : );
4041 :
4042 : if let Some(delta_layer) = delta_layer {
4043 : layers_to_upload.push(delta_layer.clone());
4044 : (layers_to_upload, Some(delta_layer))
4045 : } else {
4046 : (layers_to_upload, None)
4047 : }
4048 : } else {
4049 : // Normal case, write out a L0 delta layer file.
4050 : // `create_delta_layer` will not modify the layer map.
4051 : // We will remove frozen layer and add delta layer in one atomic operation later.
4052 : let Some(layer) = self
4053 : .create_delta_layer(&frozen_layer, None, ctx)
4054 : .await
4055 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
4056 : else {
4057 : panic!("delta layer cannot be empty if no filter is applied");
4058 : };
4059 : (
4060 : // FIXME: even though we have a single image and single delta layer assumption
4061 : // we push them to vec
4062 : vec![layer.clone()],
4063 : Some(layer),
4064 : )
4065 : };
4066 :
4067 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
4068 :
4069 : if self.cancel.is_cancelled() {
4070 : return Err(FlushLayerError::Cancelled);
4071 : }
4072 :
4073 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
4074 :
4075 : // The new on-disk layers are now in the layer map. We can remove the
4076 : // in-memory layer from the map now. The flushed layer is stored in
4077 : // the mapping in `create_delta_layer`.
4078 : {
4079 : let mut guard = self.layers.write().await;
4080 :
4081 : if self.cancel.is_cancelled() {
4082 : return Err(FlushLayerError::Cancelled);
4083 : }
4084 :
4085 : guard.finish_flush_l0_layer(delta_layer_to_add.as_ref(), &frozen_layer, &self.metrics);
4086 :
4087 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
4088 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
4089 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
4090 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
4091 : }
4092 : // release lock on 'layers'
4093 : };
4094 :
4095 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
4096 : // This makes us refuse ingest until the new layers have been persisted to the remote.
4097 : self.remote_client
4098 : .wait_completion()
4099 : .await
4100 0 : .map_err(|e| match e {
4101 : WaitCompletionError::UploadQueueShutDownOrStopped
4102 : | WaitCompletionError::NotInitialized(
4103 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
4104 0 : ) => FlushLayerError::Cancelled,
4105 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
4106 0 : FlushLayerError::Other(anyhow!(e).into())
4107 : }
4108 0 : })?;
4109 :
4110 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
4111 : // a compaction can delete the file and then it won't be available for uploads any more.
4112 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
4113 : // race situation.
4114 : // See https://github.com/neondatabase/neon/issues/4526
4115 : pausable_failpoint!("flush-frozen-pausable");
4116 :
4117 : // This failpoint is used by another test case `test_pageserver_recovery`.
4118 : fail_point!("flush-frozen-exit");
4119 :
4120 : Ok(Lsn(lsn_range.end.0 - 1))
4121 : }
4122 :
4123 : /// Return true if the value changed
4124 : ///
4125 : /// This function must only be used from the layer flush task, and may not be called concurrently.
4126 1126 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
4127 1126 : // We do a simple load/store cycle: that's why this function isn't safe for concurrent use.
4128 1126 : let old_value = self.disk_consistent_lsn.load();
4129 1126 : if new_value != old_value {
4130 1126 : assert!(new_value >= old_value);
4131 1126 : self.disk_consistent_lsn.store(new_value);
4132 1126 : true
4133 : } else {
4134 0 : false
4135 : }
4136 1126 : }
4137 :
4138 : /// Update metadata file
4139 1132 : fn schedule_uploads(
4140 1132 : &self,
4141 1132 : disk_consistent_lsn: Lsn,
4142 1132 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
4143 1132 : ) -> anyhow::Result<()> {
4144 1132 : // We can only save a valid 'prev_record_lsn' value on disk if we
4145 1132 : // flushed *all* in-memory changes to disk. We only track
4146 1132 : // 'prev_record_lsn' in memory for the latest processed record, so we
4147 1132 : // don't remember what the correct value that corresponds to some old
4148 1132 : // LSN is. But if we flush everything, then the value corresponding
4149 1132 : // current 'last_record_lsn' is correct and we can store it on disk.
4150 1132 : let RecordLsn {
4151 1132 : last: last_record_lsn,
4152 1132 : prev: prev_record_lsn,
4153 1132 : } = self.last_record_lsn.load();
4154 1132 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
4155 1052 : Some(prev_record_lsn)
4156 : } else {
4157 80 : None
4158 : };
4159 :
4160 1132 : let update = crate::tenant::metadata::MetadataUpdate::new(
4161 1132 : disk_consistent_lsn,
4162 1132 : ondisk_prev_record_lsn,
4163 1132 : *self.latest_gc_cutoff_lsn.read(),
4164 1132 : );
4165 1132 :
4166 1132 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
4167 0 : "{}",
4168 0 : x.unwrap()
4169 1132 : ));
4170 :
4171 2272 : for layer in layers_to_upload {
4172 1140 : self.remote_client.schedule_layer_file_upload(layer)?;
4173 : }
4174 1132 : self.remote_client
4175 1132 : .schedule_index_upload_for_metadata_update(&update)?;
4176 :
4177 1132 : Ok(())
4178 1132 : }
4179 :
4180 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
4181 0 : self.remote_client
4182 0 : .preserve_initdb_archive(
4183 0 : &self.tenant_shard_id.tenant_id,
4184 0 : &self.timeline_id,
4185 0 : &self.cancel,
4186 0 : )
4187 0 : .await
4188 0 : }
4189 :
4190 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
4191 : // in layer map immediately. The caller is responsible to put it into the layer map.
4192 1126 : async fn create_delta_layer(
4193 1126 : self: &Arc<Self>,
4194 1126 : frozen_layer: &Arc<InMemoryLayer>,
4195 1126 : key_range: Option<Range<Key>>,
4196 1126 : ctx: &RequestContext,
4197 1126 : ) -> anyhow::Result<Option<ResidentLayer>> {
4198 1126 : let self_clone = Arc::clone(self);
4199 1126 : let frozen_layer = Arc::clone(frozen_layer);
4200 1126 : let ctx = ctx.attached_child();
4201 1126 : let work = async move {
4202 1126 : let Some(new_delta) = frozen_layer
4203 1126 : .write_to_disk(&self_clone, &ctx, key_range)
4204 10269 : .await?
4205 : else {
4206 158 : return Ok(None);
4207 : };
4208 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
4209 : // We just need to fsync the directory in which these inodes are linked,
4210 : // which we know to be the timeline directory.
4211 : //
4212 : // We use fatal_err() below because the after write_to_disk returns with success,
4213 : // the in-memory state of the filesystem already has the layer file in its final place,
4214 : // and subsequent pageserver code could think it's durable while it really isn't.
4215 968 : let timeline_dir = VirtualFile::open(
4216 968 : &self_clone
4217 968 : .conf
4218 968 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
4219 968 : &ctx,
4220 968 : )
4221 487 : .await
4222 968 : .fatal_err("VirtualFile::open for timeline dir fsync");
4223 968 : timeline_dir
4224 968 : .sync_all()
4225 484 : .await
4226 968 : .fatal_err("VirtualFile::sync_all timeline dir");
4227 968 : anyhow::Ok(Some(new_delta))
4228 1126 : };
4229 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
4230 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
4231 : use crate::virtual_file::io_engine::IoEngine;
4232 1126 : match crate::virtual_file::io_engine::get() {
4233 0 : IoEngine::NotSet => panic!("io engine not set"),
4234 : IoEngine::StdFs => {
4235 563 : let span = tracing::info_span!("blocking");
4236 563 : tokio::task::spawn_blocking({
4237 563 : move || Handle::current().block_on(work.instrument(span))
4238 563 : })
4239 563 : .await
4240 563 : .context("spawn_blocking")
4241 563 : .and_then(|x| x)
4242 : }
4243 : #[cfg(target_os = "linux")]
4244 11235 : IoEngine::TokioEpollUring => work.await,
4245 : }
4246 1126 : }
4247 :
4248 522 : async fn repartition(
4249 522 : &self,
4250 522 : lsn: Lsn,
4251 522 : partition_size: u64,
4252 522 : flags: EnumSet<CompactFlags>,
4253 522 : ctx: &RequestContext,
4254 522 : ) -> anyhow::Result<((KeyPartitioning, SparseKeyPartitioning), Lsn)> {
4255 522 : let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
4256 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
4257 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
4258 : // and hence before the compaction task starts.
4259 0 : anyhow::bail!("repartition() called concurrently, this should not happen");
4260 : };
4261 522 : let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
4262 522 : if lsn < *partition_lsn {
4263 0 : anyhow::bail!("repartition() called with LSN going backwards, this should not happen");
4264 522 : }
4265 522 :
4266 522 : let distance = lsn.0 - partition_lsn.0;
4267 522 : if *partition_lsn != Lsn(0)
4268 262 : && distance <= self.repartition_threshold
4269 262 : && !flags.contains(CompactFlags::ForceRepartition)
4270 : {
4271 248 : debug!(
4272 : distance,
4273 : threshold = self.repartition_threshold,
4274 0 : "no repartitioning needed"
4275 : );
4276 248 : return Ok((
4277 248 : (dense_partition.clone(), sparse_partition.clone()),
4278 248 : *partition_lsn,
4279 248 : ));
4280 274 : }
4281 :
4282 16708 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4283 274 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4284 274 : let sparse_partitioning = SparseKeyPartitioning {
4285 274 : parts: vec![sparse_ks],
4286 274 : }; // no partitioning for metadata keys for now
4287 274 : *partitioning_guard = ((dense_partitioning, sparse_partitioning), lsn);
4288 274 :
4289 274 : Ok((partitioning_guard.0.clone(), partitioning_guard.1))
4290 522 : }
4291 :
4292 : // Is it time to create a new image layer for the given partition?
4293 14 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4294 14 : let threshold = self.get_image_creation_threshold();
4295 :
4296 14 : let guard = self.layers.read().await;
4297 14 : let layers = guard.layer_map();
4298 14 :
4299 14 : let mut max_deltas = 0;
4300 28 : for part_range in &partition.ranges {
4301 14 : let image_coverage = layers.image_coverage(part_range, lsn);
4302 28 : for (img_range, last_img) in image_coverage {
4303 14 : let img_lsn = if let Some(last_img) = last_img {
4304 0 : last_img.get_lsn_range().end
4305 : } else {
4306 14 : Lsn(0)
4307 : };
4308 : // Let's consider an example:
4309 : //
4310 : // delta layer with LSN range 71-81
4311 : // delta layer with LSN range 81-91
4312 : // delta layer with LSN range 91-101
4313 : // image layer at LSN 100
4314 : //
4315 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4316 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4317 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4318 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4319 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4320 14 : if img_lsn < lsn {
4321 14 : let num_deltas =
4322 14 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4323 14 :
4324 14 : max_deltas = max_deltas.max(num_deltas);
4325 14 : if num_deltas >= threshold {
4326 0 : debug!(
4327 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4328 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4329 : );
4330 0 : return true;
4331 14 : }
4332 0 : }
4333 : }
4334 : }
4335 :
4336 14 : debug!(
4337 : max_deltas,
4338 0 : "none of the partitioned ranges had >= {threshold} deltas"
4339 : );
4340 14 : false
4341 14 : }
4342 :
4343 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4344 : /// so that at most one image layer will be produced from this function.
4345 186 : async fn create_image_layer_for_rel_blocks(
4346 186 : self: &Arc<Self>,
4347 186 : partition: &KeySpace,
4348 186 : mut image_layer_writer: ImageLayerWriter,
4349 186 : lsn: Lsn,
4350 186 : ctx: &RequestContext,
4351 186 : img_range: Range<Key>,
4352 186 : start: Key,
4353 186 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4354 186 : let mut wrote_keys = false;
4355 186 :
4356 186 : let mut key_request_accum = KeySpaceAccum::new();
4357 1222 : for range in &partition.ranges {
4358 1036 : let mut key = range.start;
4359 2410 : while key < range.end {
4360 : // Decide whether to retain this key: usually we do, but sharded tenants may
4361 : // need to drop keys that don't belong to them. If we retain the key, add it
4362 : // to `key_request_accum` for later issuing a vectored get
4363 1374 : if self.shard_identity.is_key_disposable(&key) {
4364 0 : debug!(
4365 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4366 0 : key,
4367 0 : self.shard_identity.get_shard_number(&key)
4368 : );
4369 1374 : } else {
4370 1374 : key_request_accum.add_key(key);
4371 1374 : }
4372 :
4373 1374 : let last_key_in_range = key.next() == range.end;
4374 1374 : key = key.next();
4375 1374 :
4376 1374 : // Maybe flush `key_rest_accum`
4377 1374 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4378 1374 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4379 : {
4380 1036 : let results = self
4381 1036 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
4382 49 : .await?;
4383 :
4384 2410 : for (img_key, img) in results {
4385 1374 : let img = match img {
4386 1374 : Ok(img) => img,
4387 0 : Err(err) => {
4388 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4389 0 : // page without losing any actual user data. That seems better
4390 0 : // than failing repeatedly and getting stuck.
4391 0 : //
4392 0 : // We had a bug at one point, where we truncated the FSM and VM
4393 0 : // in the pageserver, but the Postgres didn't know about that
4394 0 : // and continued to generate incremental WAL records for pages
4395 0 : // that didn't exist in the pageserver. Trying to replay those
4396 0 : // WAL records failed to find the previous image of the page.
4397 0 : // This special case allows us to recover from that situation.
4398 0 : // See https://github.com/neondatabase/neon/issues/2601.
4399 0 : //
4400 0 : // Unfortunately we cannot do this for the main fork, or for
4401 0 : // any metadata keys, keys, as that would lead to actual data
4402 0 : // loss.
4403 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4404 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4405 0 : ZERO_PAGE.clone()
4406 : } else {
4407 0 : return Err(CreateImageLayersError::PageReconstructError(err));
4408 : }
4409 : }
4410 : };
4411 :
4412 : // Write all the keys we just read into our new image layer.
4413 1484 : image_layer_writer.put_image(img_key, img, ctx).await?;
4414 1374 : wrote_keys = true;
4415 : }
4416 338 : }
4417 : }
4418 : }
4419 :
4420 186 : if wrote_keys {
4421 : // Normal path: we have written some data into the new image layer for this
4422 : // partition, so flush it to disk.
4423 375 : let image_layer = image_layer_writer.finish(self, ctx).await?;
4424 186 : Ok(ImageLayerCreationOutcome {
4425 186 : image: Some(image_layer),
4426 186 : next_start_key: img_range.end,
4427 186 : })
4428 : } else {
4429 : // Special case: the image layer may be empty if this is a sharded tenant and the
4430 : // partition does not cover any keys owned by this shard. In this case, to ensure
4431 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4432 : // layer we write will cover the key range that we just scanned.
4433 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4434 0 : Ok(ImageLayerCreationOutcome {
4435 0 : image: None,
4436 0 : next_start_key: start,
4437 0 : })
4438 : }
4439 186 : }
4440 :
4441 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4442 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4443 : /// would not be too large to fit in a single image layer.
4444 : #[allow(clippy::too_many_arguments)]
4445 16 : async fn create_image_layer_for_metadata_keys(
4446 16 : self: &Arc<Self>,
4447 16 : partition: &KeySpace,
4448 16 : mut image_layer_writer: ImageLayerWriter,
4449 16 : lsn: Lsn,
4450 16 : ctx: &RequestContext,
4451 16 : img_range: Range<Key>,
4452 16 : mode: ImageLayerCreationMode,
4453 16 : start: Key,
4454 16 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4455 16 : assert!(!matches!(mode, ImageLayerCreationMode::Initial));
4456 :
4457 : // Metadata keys image layer creation.
4458 16 : let mut reconstruct_state = ValuesReconstructState::default();
4459 16 : let data = self
4460 16 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4461 4052 : .await?;
4462 16 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4463 16 : let mut new_data = BTreeMap::new();
4464 16 : let mut total_kb_retrieved = 0;
4465 16 : let mut total_keys_retrieved = 0;
4466 10028 : for (k, v) in data {
4467 10012 : let v = v.map_err(CreateImageLayersError::PageReconstructError)?;
4468 10012 : total_kb_retrieved += KEY_SIZE + v.len();
4469 10012 : total_keys_retrieved += 1;
4470 10012 : new_data.insert(k, v);
4471 : }
4472 16 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4473 16 : };
4474 16 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4475 16 :
4476 16 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4477 16 : debug!(
4478 : trigger_generation,
4479 : delta_files_accessed,
4480 : total_kb_retrieved,
4481 : total_keys_retrieved,
4482 0 : "generate metadata images"
4483 : );
4484 :
4485 16 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4486 2 : return Ok(ImageLayerCreationOutcome {
4487 2 : image: None,
4488 2 : next_start_key: img_range.end,
4489 2 : });
4490 14 : }
4491 14 : let mut wrote_any_image = false;
4492 10026 : for (k, v) in data {
4493 10012 : if v.is_empty() {
4494 : // the key has been deleted, it does not need an image
4495 : // in metadata keyspace, an empty image == tombstone
4496 8 : continue;
4497 10004 : }
4498 10004 : wrote_any_image = true;
4499 10004 :
4500 10004 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4501 10004 :
4502 10004 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4503 10004 : // on the normal data path either.
4504 10162 : image_layer_writer.put_image(k, v, ctx).await?;
4505 : }
4506 :
4507 14 : if wrote_any_image {
4508 : // Normal path: we have written some data into the new image layer for this
4509 : // partition, so flush it to disk.
4510 25 : let image_layer = image_layer_writer.finish(self, ctx).await?;
4511 12 : Ok(ImageLayerCreationOutcome {
4512 12 : image: Some(image_layer),
4513 12 : next_start_key: img_range.end,
4514 12 : })
4515 : } else {
4516 : // Special case: the image layer may be empty if this is a sharded tenant and the
4517 : // partition does not cover any keys owned by this shard. In this case, to ensure
4518 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4519 : // layer we write will cover the key range that we just scanned.
4520 2 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4521 2 : Ok(ImageLayerCreationOutcome {
4522 2 : image: None,
4523 2 : next_start_key: start,
4524 2 : })
4525 : }
4526 16 : }
4527 :
4528 : /// Predicate function which indicates whether we should check if new image layers
4529 : /// are required. Since checking if new image layers are required is expensive in
4530 : /// terms of CPU, we only do it in the following cases:
4531 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4532 : /// 2. If enough time has passed since the last check:
4533 : /// 1. For large tenants, we wish to perform the check more often since they
4534 : /// suffer from the lack of image layers
4535 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4536 522 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4537 522 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4538 522 :
4539 522 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4540 522 : let distance = lsn
4541 522 : .checked_sub(last_checks_at)
4542 522 : .expect("Attempt to compact with LSN going backwards");
4543 522 : let min_distance =
4544 522 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4545 522 :
4546 522 : let distance_based_decision = distance.0 >= min_distance;
4547 522 :
4548 522 : let mut time_based_decision = false;
4549 522 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4550 522 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4551 420 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4552 : {
4553 0 : self.get_checkpoint_timeout()
4554 : } else {
4555 420 : Duration::from_secs(3600 * 48)
4556 : };
4557 :
4558 420 : time_based_decision = match *last_check_instant {
4559 262 : Some(last_check) => {
4560 262 : let elapsed = last_check.elapsed();
4561 262 : elapsed >= check_required_after
4562 : }
4563 158 : None => true,
4564 : };
4565 102 : }
4566 :
4567 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4568 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4569 : // check.
4570 522 : let decision = distance_based_decision || time_based_decision;
4571 :
4572 522 : if decision {
4573 160 : self.last_image_layer_creation_check_at.store(lsn);
4574 160 : *last_check_instant = Some(Instant::now());
4575 362 : }
4576 :
4577 522 : decision
4578 522 : }
4579 :
4580 1044 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4581 : async fn create_image_layers(
4582 : self: &Arc<Timeline>,
4583 : partitioning: &KeyPartitioning,
4584 : lsn: Lsn,
4585 : mode: ImageLayerCreationMode,
4586 : ctx: &RequestContext,
4587 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4588 : let timer = self.metrics.create_images_time_histo.start_timer();
4589 : let mut image_layers = Vec::new();
4590 :
4591 : // We need to avoid holes between generated image layers.
4592 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4593 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4594 : //
4595 : // How such hole between partitions can appear?
4596 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4597 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4598 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4599 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4600 : let mut start = Key::MIN;
4601 :
4602 : let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
4603 :
4604 : for partition in partitioning.parts.iter() {
4605 : let img_range = start..partition.ranges.last().unwrap().end;
4606 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4607 : if compact_metadata {
4608 : for range in &partition.ranges {
4609 : assert!(
4610 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4611 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4612 : "metadata keys must be partitioned separately"
4613 : );
4614 : }
4615 : if mode == ImageLayerCreationMode::Initial {
4616 : return Err(CreateImageLayersError::Other(anyhow::anyhow!("no image layer should be created for metadata keys when flushing frozen layers")));
4617 : }
4618 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4619 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4620 : // might mess up with evictions.
4621 : start = img_range.end;
4622 : continue;
4623 : }
4624 : } else if let ImageLayerCreationMode::Try = mode {
4625 : // check_for_image_layers = false -> skip
4626 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4627 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4628 : start = img_range.end;
4629 : continue;
4630 : }
4631 : } else if let ImageLayerCreationMode::Force = mode {
4632 : // When forced to create image layers, we might try and create them where they already
4633 : // exist. This mode is only used in tests/debug.
4634 : let layers = self.layers.read().await;
4635 : if layers.contains_key(&PersistentLayerKey {
4636 : key_range: img_range.clone(),
4637 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4638 : is_delta: false,
4639 : }) {
4640 : tracing::info!(
4641 : "Skipping image layer at {lsn} {}..{}, already exists",
4642 : img_range.start,
4643 : img_range.end
4644 : );
4645 : continue;
4646 : }
4647 : }
4648 :
4649 : let image_layer_writer = ImageLayerWriter::new(
4650 : self.conf,
4651 : self.timeline_id,
4652 : self.tenant_shard_id,
4653 : &img_range,
4654 : lsn,
4655 : ctx,
4656 : )
4657 : .await?;
4658 :
4659 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4660 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4661 0 : "failpoint image-layer-writer-fail-before-finish"
4662 0 : )))
4663 0 : });
4664 :
4665 : if !compact_metadata {
4666 : let ImageLayerCreationOutcome {
4667 : image,
4668 : next_start_key,
4669 : } = self
4670 : .create_image_layer_for_rel_blocks(
4671 : partition,
4672 : image_layer_writer,
4673 : lsn,
4674 : ctx,
4675 : img_range,
4676 : start,
4677 : )
4678 : .await?;
4679 :
4680 : start = next_start_key;
4681 : image_layers.extend(image);
4682 : } else {
4683 : let ImageLayerCreationOutcome {
4684 : image,
4685 : next_start_key,
4686 : } = self
4687 : .create_image_layer_for_metadata_keys(
4688 : partition,
4689 : image_layer_writer,
4690 : lsn,
4691 : ctx,
4692 : img_range,
4693 : mode,
4694 : start,
4695 : )
4696 : .await?;
4697 : start = next_start_key;
4698 : image_layers.extend(image);
4699 : }
4700 : }
4701 :
4702 : let mut guard = self.layers.write().await;
4703 :
4704 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4705 : // now they are being scheduled outside of write lock
4706 : guard.track_new_image_layers(&image_layers, &self.metrics);
4707 : drop_wlock(guard);
4708 : timer.stop_and_record();
4709 :
4710 : // Creating image layers may have caused some previously visible layers to be covered
4711 : self.update_layer_visibility().await;
4712 :
4713 : Ok(image_layers)
4714 : }
4715 :
4716 : /// Wait until the background initial logical size calculation is complete, or
4717 : /// this Timeline is shut down. Calling this function will cause the initial
4718 : /// logical size calculation to skip waiting for the background jobs barrier.
4719 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4720 0 : if !self.shard_identity.is_shard_zero() {
4721 : // We don't populate logical size on shard >0: skip waiting for it.
4722 0 : return;
4723 0 : }
4724 0 :
4725 0 : if self.remote_client.is_deleting() {
4726 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4727 0 : return;
4728 0 : }
4729 :
4730 0 : if let Some(await_bg_cancel) = self
4731 0 : .current_logical_size
4732 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4733 0 : .get()
4734 0 : {
4735 0 : await_bg_cancel.cancel();
4736 0 : } else {
4737 : // We should not wait if we were not able to explicitly instruct
4738 : // the logical size cancellation to skip the concurrency limit semaphore.
4739 : // TODO: this is an unexpected case. We should restructure so that it
4740 : // can't happen.
4741 0 : tracing::warn!(
4742 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4743 : );
4744 0 : debug_assert!(false);
4745 : }
4746 :
4747 : tokio::select!(
4748 : _ = self.current_logical_size.initialized.acquire() => {},
4749 : _ = self.cancel.cancelled() => {}
4750 : )
4751 0 : }
4752 :
4753 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4754 : /// Timelines layers up to the ancestor_lsn.
4755 : ///
4756 : /// Requires a timeline that:
4757 : /// - has an ancestor to detach from
4758 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4759 : /// a technical requirement
4760 : ///
4761 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4762 : /// polled again until completion.
4763 : ///
4764 : /// During the operation all timelines sharing the data with this timeline will be reparented
4765 : /// from our ancestor to be branches of this timeline.
4766 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4767 0 : self: &Arc<Timeline>,
4768 0 : tenant: &crate::tenant::Tenant,
4769 0 : options: detach_ancestor::Options,
4770 0 : ctx: &RequestContext,
4771 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
4772 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4773 0 : }
4774 :
4775 : /// Completes the ancestor detach. This method is to be called while holding the
4776 : /// TenantManager's tenant slot, so during this method we cannot be deleted nor can any
4777 : /// timeline be deleted. After this method returns successfully, tenant must be reloaded.
4778 : ///
4779 : /// Pageserver receiving a SIGKILL during this operation is not supported (yet).
4780 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4781 0 : self: &Arc<Timeline>,
4782 0 : tenant: &crate::tenant::Tenant,
4783 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4784 0 : ctx: &RequestContext,
4785 0 : ) -> Result<Vec<TimelineId>, anyhow::Error> {
4786 0 : detach_ancestor::complete(self, tenant, prepared, ctx).await
4787 0 : }
4788 :
4789 : /// Switch aux file policy and schedule upload to the index part.
4790 16 : pub(crate) fn do_switch_aux_policy(&self, policy: AuxFilePolicy) -> anyhow::Result<()> {
4791 16 : self.last_aux_file_policy.store(Some(policy));
4792 16 : self.remote_client
4793 16 : .schedule_index_upload_for_aux_file_policy_update(Some(policy))?;
4794 16 : Ok(())
4795 16 : }
4796 : }
4797 :
4798 : impl Drop for Timeline {
4799 8 : fn drop(&mut self) {
4800 8 : if let Some(ancestor) = &self.ancestor_timeline {
4801 : // This lock should never be poisoned, but in case it is we do a .map() instead of
4802 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
4803 2 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
4804 2 : gc_info.remove_child(self.timeline_id)
4805 0 : }
4806 6 : }
4807 8 : }
4808 : }
4809 :
4810 : /// Top-level failure to compact.
4811 0 : #[derive(Debug, thiserror::Error)]
4812 : pub(crate) enum CompactionError {
4813 : #[error("The timeline or pageserver is shutting down")]
4814 : ShuttingDown,
4815 : /// Compaction cannot be done right now; page reconstruction and so on.
4816 : #[error(transparent)]
4817 : Other(anyhow::Error),
4818 : }
4819 :
4820 : impl From<CollectKeySpaceError> for CompactionError {
4821 0 : fn from(err: CollectKeySpaceError) -> Self {
4822 0 : match err {
4823 : CollectKeySpaceError::Cancelled
4824 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4825 0 : CompactionError::ShuttingDown
4826 : }
4827 0 : e => CompactionError::Other(e.into()),
4828 : }
4829 0 : }
4830 : }
4831 :
4832 : impl From<super::upload_queue::NotInitialized> for CompactionError {
4833 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
4834 0 : match value {
4835 : super::upload_queue::NotInitialized::Uninitialized
4836 : | super::upload_queue::NotInitialized::Stopped => {
4837 0 : CompactionError::Other(anyhow::anyhow!(value))
4838 : }
4839 0 : super::upload_queue::NotInitialized::ShuttingDown => CompactionError::ShuttingDown,
4840 : }
4841 0 : }
4842 : }
4843 :
4844 : impl CompactionError {
4845 : /// We cannot do compaction because we could not download a layer that is input to the compaction.
4846 0 : pub(crate) fn input_layer_download_failed(
4847 0 : e: super::storage_layer::layer::DownloadError,
4848 0 : ) -> Self {
4849 0 : match e {
4850 : super::storage_layer::layer::DownloadError::TimelineShutdown |
4851 : /* TODO DownloadCancelled correct here? */
4852 0 : super::storage_layer::layer::DownloadError::DownloadCancelled => CompactionError::ShuttingDown,
4853 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads |
4854 : super::storage_layer::layer::DownloadError::DownloadRequired |
4855 : super::storage_layer::layer::DownloadError::NotFile(_) |
4856 : super::storage_layer::layer::DownloadError::DownloadFailed |
4857 0 : super::storage_layer::layer::DownloadError::PreStatFailed(_)=>CompactionError::Other(anyhow::anyhow!(e)),
4858 : #[cfg(test)]
4859 0 : super::storage_layer::layer::DownloadError::Failpoint(_) => CompactionError::Other(anyhow::anyhow!(e)),
4860 : }
4861 0 : }
4862 : }
4863 :
4864 : #[serde_as]
4865 196 : #[derive(serde::Serialize)]
4866 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4867 :
4868 : #[derive(Default)]
4869 : enum DurationRecorder {
4870 : #[default]
4871 : NotStarted,
4872 : Recorded(RecordedDuration, tokio::time::Instant),
4873 : }
4874 :
4875 : impl DurationRecorder {
4876 504 : fn till_now(&self) -> DurationRecorder {
4877 504 : match self {
4878 : DurationRecorder::NotStarted => {
4879 0 : panic!("must only call on recorded measurements")
4880 : }
4881 504 : DurationRecorder::Recorded(_, ended) => {
4882 504 : let now = tokio::time::Instant::now();
4883 504 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4884 504 : }
4885 504 : }
4886 504 : }
4887 196 : fn into_recorded(self) -> Option<RecordedDuration> {
4888 196 : match self {
4889 0 : DurationRecorder::NotStarted => None,
4890 196 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4891 : }
4892 196 : }
4893 : }
4894 :
4895 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
4896 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
4897 : /// the layer descriptor requires the user to provide the ranges, which should cover all
4898 : /// keys specified in the `data` field.
4899 : #[cfg(test)]
4900 : pub struct DeltaLayerTestDesc {
4901 : pub lsn_range: Range<Lsn>,
4902 : pub key_range: Range<Key>,
4903 : pub data: Vec<(Key, Lsn, Value)>,
4904 : }
4905 :
4906 : #[cfg(test)]
4907 : impl DeltaLayerTestDesc {
4908 : #[allow(dead_code)]
4909 0 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
4910 0 : Self {
4911 0 : lsn_range,
4912 0 : key_range,
4913 0 : data,
4914 0 : }
4915 0 : }
4916 :
4917 46 : pub fn new_with_inferred_key_range(
4918 46 : lsn_range: Range<Lsn>,
4919 46 : data: Vec<(Key, Lsn, Value)>,
4920 46 : ) -> Self {
4921 108 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
4922 108 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
4923 46 : Self {
4924 46 : key_range: (*key_min)..(key_max.next()),
4925 46 : lsn_range,
4926 46 : data,
4927 46 : }
4928 46 : }
4929 : }
4930 :
4931 : impl Timeline {
4932 28 : async fn finish_compact_batch(
4933 28 : self: &Arc<Self>,
4934 28 : new_deltas: &[ResidentLayer],
4935 28 : new_images: &[ResidentLayer],
4936 28 : layers_to_remove: &[Layer],
4937 28 : ) -> Result<(), CompactionError> {
4938 28 : let mut guard = self.layers.write().await;
4939 :
4940 28 : let mut duplicated_layers = HashSet::new();
4941 28 :
4942 28 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
4943 :
4944 336 : for l in new_deltas {
4945 308 : if guard.contains(l.as_ref()) {
4946 : // expected in tests
4947 0 : tracing::error!(layer=%l, "duplicated L1 layer");
4948 :
4949 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
4950 : // `guard` on self.layers. as of writing this, there are no error returns except
4951 : // for compact_level0_phase1 creating an L0, which does not happen in practice
4952 : // because we have not implemented L0 => L0 compaction.
4953 0 : duplicated_layers.insert(l.layer_desc().key());
4954 308 : } else if LayerMap::is_l0(&l.layer_desc().key_range) {
4955 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
4956 308 : } else {
4957 308 : insert_layers.push(l.clone());
4958 308 : }
4959 : }
4960 :
4961 : // only remove those inputs which were not outputs
4962 28 : let remove_layers: Vec<Layer> = layers_to_remove
4963 28 : .iter()
4964 402 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
4965 28 : .cloned()
4966 28 : .collect();
4967 28 :
4968 28 : if !new_images.is_empty() {
4969 0 : guard.track_new_image_layers(new_images, &self.metrics);
4970 28 : }
4971 :
4972 : // deletion will happen later, the layer file manager calls garbage_collect_on_drop
4973 28 : guard.finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
4974 28 :
4975 28 : self.remote_client
4976 28 : .schedule_compaction_update(&remove_layers, new_deltas)?;
4977 :
4978 28 : drop_wlock(guard);
4979 28 :
4980 28 : Ok(())
4981 28 : }
4982 :
4983 0 : async fn rewrite_layers(
4984 0 : self: &Arc<Self>,
4985 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
4986 0 : mut drop_layers: Vec<Layer>,
4987 0 : ) -> Result<(), super::upload_queue::NotInitialized> {
4988 0 : let mut guard = self.layers.write().await;
4989 :
4990 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
4991 : // to avoid double-removing, and avoid rewriting something that was removed.
4992 0 : replace_layers.retain(|(l, _)| guard.contains(l));
4993 0 : drop_layers.retain(|l| guard.contains(l));
4994 0 :
4995 0 : guard.rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
4996 0 :
4997 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
4998 0 :
4999 0 : self.remote_client
5000 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
5001 :
5002 0 : Ok(())
5003 0 : }
5004 :
5005 : /// Schedules the uploads of the given image layers
5006 364 : fn upload_new_image_layers(
5007 364 : self: &Arc<Self>,
5008 364 : new_images: impl IntoIterator<Item = ResidentLayer>,
5009 364 : ) -> Result<(), super::upload_queue::NotInitialized> {
5010 390 : for layer in new_images {
5011 26 : self.remote_client.schedule_layer_file_upload(layer)?;
5012 : }
5013 : // should any new image layer been created, not uploading index_part will
5014 : // result in a mismatch between remote_physical_size and layermap calculated
5015 : // size, which will fail some tests, but should not be an issue otherwise.
5016 364 : self.remote_client
5017 364 : .schedule_index_upload_for_file_changes()?;
5018 364 : Ok(())
5019 364 : }
5020 :
5021 : /// Find the Lsns above which layer files need to be retained on
5022 : /// garbage collection.
5023 : ///
5024 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
5025 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
5026 : /// the space-based retention.
5027 : ///
5028 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
5029 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
5030 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
5031 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
5032 : /// in the response as the GC cutoff point for the timeline.
5033 1508 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
5034 : pub(super) async fn find_gc_cutoffs(
5035 : &self,
5036 : space_cutoff: Lsn,
5037 : pitr: Duration,
5038 : cancel: &CancellationToken,
5039 : ctx: &RequestContext,
5040 : ) -> Result<GcCutoffs, PageReconstructError> {
5041 : let _timer = self
5042 : .metrics
5043 : .find_gc_cutoffs_histo
5044 : .start_timer()
5045 : .record_on_drop();
5046 :
5047 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
5048 :
5049 : if cfg!(test) {
5050 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
5051 : if pitr == Duration::ZERO {
5052 : return Ok(GcCutoffs {
5053 : time: self.get_last_record_lsn(),
5054 : space: space_cutoff,
5055 : });
5056 : }
5057 : }
5058 :
5059 : // Calculate a time-based limit on how much to retain:
5060 : // - if PITR interval is set, then this is our cutoff.
5061 : // - if PITR interval is not set, then we do a lookup
5062 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
5063 : let time_cutoff = {
5064 : let now = SystemTime::now();
5065 : let time_range = if pitr == Duration::ZERO {
5066 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
5067 : } else {
5068 : pitr
5069 : };
5070 :
5071 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
5072 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
5073 : let timestamp = to_pg_timestamp(time_cutoff);
5074 :
5075 : match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
5076 : LsnForTimestamp::Present(lsn) => Some(lsn),
5077 : LsnForTimestamp::Future(lsn) => {
5078 : // The timestamp is in the future. That sounds impossible,
5079 : // but what it really means is that there hasn't been
5080 : // any commits since the cutoff timestamp.
5081 : //
5082 : // In this case we should use the LSN of the most recent commit,
5083 : // which is implicitly the last LSN in the log.
5084 : debug!("future({})", lsn);
5085 : Some(self.get_last_record_lsn())
5086 : }
5087 : LsnForTimestamp::Past(lsn) => {
5088 : debug!("past({})", lsn);
5089 : None
5090 : }
5091 : LsnForTimestamp::NoData(lsn) => {
5092 : debug!("nodata({})", lsn);
5093 : None
5094 : }
5095 : }
5096 : };
5097 :
5098 : Ok(match (pitr, time_cutoff) {
5099 : (Duration::ZERO, Some(time_cutoff)) => {
5100 : // PITR is not set. Retain the size-based limit, or the default time retention,
5101 : // whichever requires less data.
5102 : GcCutoffs {
5103 : time: self.get_last_record_lsn(),
5104 : space: std::cmp::max(time_cutoff, space_cutoff),
5105 : }
5106 : }
5107 : (Duration::ZERO, None) => {
5108 : // PITR is not set, and time lookup failed
5109 : GcCutoffs {
5110 : time: self.get_last_record_lsn(),
5111 : space: space_cutoff,
5112 : }
5113 : }
5114 : (_, None) => {
5115 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
5116 : // cannot advance beyond what was already GC'd, and respect space-based retention
5117 : GcCutoffs {
5118 : time: *self.get_latest_gc_cutoff_lsn(),
5119 : space: space_cutoff,
5120 : }
5121 : }
5122 : (_, Some(time_cutoff)) => {
5123 : // PITR interval is set and we looked up timestamp successfully. Ignore
5124 : // size based retention and make time cutoff authoritative
5125 : GcCutoffs {
5126 : time: time_cutoff,
5127 : space: time_cutoff,
5128 : }
5129 : }
5130 : })
5131 : }
5132 :
5133 : /// Garbage collect layer files on a timeline that are no longer needed.
5134 : ///
5135 : /// Currently, we don't make any attempt at removing unneeded page versions
5136 : /// within a layer file. We can only remove the whole file if it's fully
5137 : /// obsolete.
5138 754 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
5139 : // this is most likely the background tasks, but it might be the spawned task from
5140 : // immediate_gc
5141 753 : let _g = tokio::select! {
5142 : guard = self.gc_lock.lock() => guard,
5143 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
5144 : };
5145 753 : let timer = self.metrics.garbage_collect_histo.start_timer();
5146 :
5147 : fail_point!("before-timeline-gc");
5148 :
5149 : // Is the timeline being deleted?
5150 753 : if self.is_stopping() {
5151 0 : return Err(GcError::TimelineCancelled);
5152 753 : }
5153 753 :
5154 753 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
5155 753 : let gc_info = self.gc_info.read().unwrap();
5156 753 :
5157 753 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
5158 753 : let time_cutoff = gc_info.cutoffs.time;
5159 753 : let retain_lsns = gc_info
5160 753 : .retain_lsns
5161 753 : .iter()
5162 753 : .map(|(lsn, _child_id)| *lsn)
5163 753 : .collect();
5164 753 :
5165 753 : // Gets the maximum LSN that holds the valid lease.
5166 753 : //
5167 753 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
5168 753 : // Here, we do not check for stale leases again.
5169 753 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
5170 753 :
5171 753 : (
5172 753 : space_cutoff,
5173 753 : time_cutoff,
5174 753 : retain_lsns,
5175 753 : max_lsn_with_valid_lease,
5176 753 : )
5177 753 : };
5178 753 :
5179 753 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
5180 753 : let standby_horizon = self.standby_horizon.load();
5181 753 : // Hold GC for the standby, but as a safety guard do it only within some
5182 753 : // reasonable lag.
5183 753 : if standby_horizon != Lsn::INVALID {
5184 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
5185 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
5186 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
5187 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
5188 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
5189 : } else {
5190 0 : warn!(
5191 0 : "standby is lagging for more than {}MB, not holding gc for it",
5192 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
5193 : )
5194 : }
5195 0 : }
5196 753 : }
5197 :
5198 : // Reset standby horizon to ignore it if it is not updated till next GC.
5199 : // It is an easy way to unset it when standby disappears without adding
5200 : // more conf options.
5201 753 : self.standby_horizon.store(Lsn::INVALID);
5202 753 : self.metrics
5203 753 : .standby_horizon_gauge
5204 753 : .set(Lsn::INVALID.0 as i64);
5205 :
5206 753 : let res = self
5207 753 : .gc_timeline(
5208 753 : space_cutoff,
5209 753 : time_cutoff,
5210 753 : retain_lsns,
5211 753 : max_lsn_with_valid_lease,
5212 753 : new_gc_cutoff,
5213 753 : )
5214 753 : .instrument(
5215 753 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5216 : )
5217 0 : .await?;
5218 :
5219 : // only record successes
5220 753 : timer.stop_and_record();
5221 753 :
5222 753 : Ok(res)
5223 754 : }
5224 :
5225 753 : async fn gc_timeline(
5226 753 : &self,
5227 753 : space_cutoff: Lsn,
5228 753 : time_cutoff: Lsn,
5229 753 : retain_lsns: Vec<Lsn>,
5230 753 : max_lsn_with_valid_lease: Option<Lsn>,
5231 753 : new_gc_cutoff: Lsn,
5232 753 : ) -> Result<GcResult, GcError> {
5233 753 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5234 753 :
5235 753 : let now = SystemTime::now();
5236 753 : let mut result: GcResult = GcResult::default();
5237 753 :
5238 753 : // Nothing to GC. Return early.
5239 753 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
5240 753 : if latest_gc_cutoff >= new_gc_cutoff {
5241 22 : info!(
5242 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5243 : );
5244 22 : return Ok(result);
5245 731 : }
5246 :
5247 : // We need to ensure that no one tries to read page versions or create
5248 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5249 : // for details. This will block until the old value is no longer in use.
5250 : //
5251 : // The GC cutoff should only ever move forwards.
5252 731 : let waitlist = {
5253 731 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
5254 731 : if *write_guard > new_gc_cutoff {
5255 0 : return Err(GcError::BadLsn {
5256 0 : why: format!(
5257 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5258 0 : *write_guard, new_gc_cutoff
5259 0 : ),
5260 0 : });
5261 731 : }
5262 731 :
5263 731 : write_guard.store_and_unlock(new_gc_cutoff)
5264 731 : };
5265 731 : waitlist.wait().await;
5266 :
5267 731 : info!("GC starting");
5268 :
5269 731 : debug!("retain_lsns: {:?}", retain_lsns);
5270 :
5271 731 : let mut layers_to_remove = Vec::new();
5272 :
5273 : // Scan all layers in the timeline (remote or on-disk).
5274 : //
5275 : // Garbage collect the layer if all conditions are satisfied:
5276 : // 1. it is older than cutoff LSN;
5277 : // 2. it is older than PITR interval;
5278 : // 3. it doesn't need to be retained for 'retain_lsns';
5279 : // 4. it does not need to be kept for LSNs holding valid leases.
5280 : // 5. newer on-disk image layers cover the layer's whole key range
5281 : //
5282 : // TODO holding a write lock is too agressive and avoidable
5283 731 : let mut guard = self.layers.write().await;
5284 731 : let layers = guard.layer_map();
5285 12415 : 'outer: for l in layers.iter_historic_layers() {
5286 12415 : result.layers_total += 1;
5287 12415 :
5288 12415 : // 1. Is it newer than GC horizon cutoff point?
5289 12415 : if l.get_lsn_range().end > space_cutoff {
5290 741 : debug!(
5291 0 : "keeping {} because it's newer than space_cutoff {}",
5292 0 : l.layer_name(),
5293 : space_cutoff,
5294 : );
5295 741 : result.layers_needed_by_cutoff += 1;
5296 741 : continue 'outer;
5297 11674 : }
5298 11674 :
5299 11674 : // 2. It is newer than PiTR cutoff point?
5300 11674 : if l.get_lsn_range().end > time_cutoff {
5301 0 : debug!(
5302 0 : "keeping {} because it's newer than time_cutoff {}",
5303 0 : l.layer_name(),
5304 : time_cutoff,
5305 : );
5306 0 : result.layers_needed_by_pitr += 1;
5307 0 : continue 'outer;
5308 11674 : }
5309 :
5310 : // 3. Is it needed by a child branch?
5311 : // NOTE With that we would keep data that
5312 : // might be referenced by child branches forever.
5313 : // We can track this in child timeline GC and delete parent layers when
5314 : // they are no longer needed. This might be complicated with long inheritance chains.
5315 : //
5316 : // TODO Vec is not a great choice for `retain_lsns`
5317 11674 : for retain_lsn in &retain_lsns {
5318 : // start_lsn is inclusive
5319 10 : if &l.get_lsn_range().start <= retain_lsn {
5320 10 : debug!(
5321 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5322 0 : l.layer_name(),
5323 0 : retain_lsn,
5324 0 : l.is_incremental(),
5325 : );
5326 10 : result.layers_needed_by_branches += 1;
5327 10 : continue 'outer;
5328 0 : }
5329 : }
5330 :
5331 : // 4. Is there a valid lease that requires us to keep this layer?
5332 11664 : if let Some(lsn) = &max_lsn_with_valid_lease {
5333 : // keep if layer start <= any of the lease
5334 18 : if &l.get_lsn_range().start <= lsn {
5335 14 : debug!(
5336 0 : "keeping {} because there is a valid lease preventing GC at {}",
5337 0 : l.layer_name(),
5338 : lsn,
5339 : );
5340 14 : result.layers_needed_by_leases += 1;
5341 14 : continue 'outer;
5342 4 : }
5343 11646 : }
5344 :
5345 : // 5. Is there a later on-disk layer for this relation?
5346 : //
5347 : // The end-LSN is exclusive, while disk_consistent_lsn is
5348 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5349 : // OK for a delta layer to have end LSN 101, but if the end LSN
5350 : // is 102, then it might not have been fully flushed to disk
5351 : // before crash.
5352 : //
5353 : // For example, imagine that the following layers exist:
5354 : //
5355 : // 1000 - image (A)
5356 : // 1000-2000 - delta (B)
5357 : // 2000 - image (C)
5358 : // 2000-3000 - delta (D)
5359 : // 3000 - image (E)
5360 : //
5361 : // If GC horizon is at 2500, we can remove layers A and B, but
5362 : // we cannot remove C, even though it's older than 2500, because
5363 : // the delta layer 2000-3000 depends on it.
5364 11650 : if !layers
5365 11650 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5366 : {
5367 11642 : debug!("keeping {} because it is the latest layer", l.layer_name());
5368 11642 : result.layers_not_updated += 1;
5369 11642 : continue 'outer;
5370 8 : }
5371 8 :
5372 8 : // We didn't find any reason to keep this file, so remove it.
5373 8 : debug!(
5374 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5375 0 : l.layer_name(),
5376 0 : l.is_incremental(),
5377 : );
5378 8 : layers_to_remove.push(l);
5379 : }
5380 :
5381 731 : if !layers_to_remove.is_empty() {
5382 : // Persist the new GC cutoff value before we actually remove anything.
5383 : // This unconditionally schedules also an index_part.json update, even though, we will
5384 : // be doing one a bit later with the unlinked gc'd layers.
5385 6 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5386 6 : self.schedule_uploads(disk_consistent_lsn, None)
5387 6 : .map_err(|e| {
5388 0 : if self.cancel.is_cancelled() {
5389 0 : GcError::TimelineCancelled
5390 : } else {
5391 0 : GcError::Remote(e)
5392 : }
5393 6 : })?;
5394 :
5395 6 : let gc_layers = layers_to_remove
5396 6 : .iter()
5397 8 : .map(|x| guard.get_from_desc(x))
5398 6 : .collect::<Vec<Layer>>();
5399 6 :
5400 6 : result.layers_removed = gc_layers.len() as u64;
5401 6 :
5402 6 : self.remote_client
5403 6 : .schedule_gc_update(&gc_layers)
5404 6 : .map_err(|e| {
5405 0 : if self.cancel.is_cancelled() {
5406 0 : GcError::TimelineCancelled
5407 : } else {
5408 0 : GcError::Remote(e)
5409 : }
5410 6 : })?;
5411 :
5412 6 : guard.finish_gc_timeline(&gc_layers);
5413 6 :
5414 6 : #[cfg(feature = "testing")]
5415 6 : {
5416 6 : result.doomed_layers = gc_layers;
5417 6 : }
5418 725 : }
5419 :
5420 731 : info!(
5421 0 : "GC completed removing {} layers, cutoff {}",
5422 : result.layers_removed, new_gc_cutoff
5423 : );
5424 :
5425 731 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5426 731 : Ok(result)
5427 753 : }
5428 :
5429 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5430 667174 : async fn reconstruct_value(
5431 667174 : &self,
5432 667174 : key: Key,
5433 667174 : request_lsn: Lsn,
5434 667174 : mut data: ValueReconstructState,
5435 667174 : ) -> Result<Bytes, PageReconstructError> {
5436 667174 : // Perform WAL redo if needed
5437 667174 : data.records.reverse();
5438 667174 :
5439 667174 : // If we have a page image, and no WAL, we're all set
5440 667174 : if data.records.is_empty() {
5441 666908 : if let Some((img_lsn, img)) = &data.img {
5442 666908 : trace!(
5443 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5444 : key,
5445 : img_lsn,
5446 : request_lsn,
5447 : );
5448 666908 : Ok(img.clone())
5449 : } else {
5450 0 : Err(PageReconstructError::from(anyhow!(
5451 0 : "base image for {key} at {request_lsn} not found"
5452 0 : )))
5453 : }
5454 : } else {
5455 : // We need to do WAL redo.
5456 : //
5457 : // If we don't have a base image, then the oldest WAL record better initialize
5458 : // the page
5459 266 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5460 0 : Err(PageReconstructError::from(anyhow!(
5461 0 : "Base image for {} at {} not found, but got {} WAL records",
5462 0 : key,
5463 0 : request_lsn,
5464 0 : data.records.len()
5465 0 : )))
5466 : } else {
5467 266 : if data.img.is_some() {
5468 266 : trace!(
5469 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5470 0 : data.records.len(),
5471 : key,
5472 : request_lsn
5473 : );
5474 : } else {
5475 0 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5476 : };
5477 266 : let res = self
5478 266 : .walredo_mgr
5479 266 : .as_ref()
5480 266 : .context("timeline has no walredo manager")
5481 266 : .map_err(PageReconstructError::WalRedo)?
5482 266 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5483 0 : .await;
5484 266 : let img = match res {
5485 266 : Ok(img) => img,
5486 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5487 0 : Err(walredo::Error::Other(e)) => {
5488 0 : return Err(PageReconstructError::WalRedo(
5489 0 : e.context("reconstruct a page image"),
5490 0 : ))
5491 : }
5492 : };
5493 266 : Ok(img)
5494 : }
5495 : }
5496 667174 : }
5497 :
5498 0 : pub(crate) async fn spawn_download_all_remote_layers(
5499 0 : self: Arc<Self>,
5500 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5501 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5502 0 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5503 0 :
5504 0 : // this is not really needed anymore; it has tests which really check the return value from
5505 0 : // http api. it would be better not to maintain this anymore.
5506 0 :
5507 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5508 0 : if let Some(st) = &*status_guard {
5509 0 : match &st.state {
5510 : DownloadRemoteLayersTaskState::Running => {
5511 0 : return Err(st.clone());
5512 : }
5513 : DownloadRemoteLayersTaskState::ShutDown
5514 0 : | DownloadRemoteLayersTaskState::Completed => {
5515 0 : *status_guard = None;
5516 0 : }
5517 : }
5518 0 : }
5519 :
5520 0 : let self_clone = Arc::clone(&self);
5521 0 : let task_id = task_mgr::spawn(
5522 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5523 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5524 0 : Some(self.tenant_shard_id),
5525 0 : Some(self.timeline_id),
5526 0 : "download all remote layers task",
5527 0 : async move {
5528 0 : self_clone.download_all_remote_layers(request).await;
5529 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5530 0 : match &mut *status_guard {
5531 : None => {
5532 0 : warn!("tasks status is supposed to be Some(), since we are running");
5533 : }
5534 0 : Some(st) => {
5535 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5536 0 : if st.task_id != exp_task_id {
5537 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5538 0 : } else {
5539 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5540 0 : }
5541 : }
5542 : };
5543 0 : Ok(())
5544 0 : }
5545 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))
5546 : );
5547 :
5548 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5549 0 : task_id: format!("{task_id}"),
5550 0 : state: DownloadRemoteLayersTaskState::Running,
5551 0 : total_layer_count: 0,
5552 0 : successful_download_count: 0,
5553 0 : failed_download_count: 0,
5554 0 : };
5555 0 : *status_guard = Some(initial_info.clone());
5556 0 :
5557 0 : Ok(initial_info)
5558 0 : }
5559 :
5560 0 : async fn download_all_remote_layers(
5561 0 : self: &Arc<Self>,
5562 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5563 0 : ) {
5564 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5565 :
5566 0 : let remaining = {
5567 0 : let guard = self.layers.read().await;
5568 0 : guard
5569 0 : .layer_map()
5570 0 : .iter_historic_layers()
5571 0 : .map(|desc| guard.get_from_desc(&desc))
5572 0 : .collect::<Vec<_>>()
5573 0 : };
5574 0 : let total_layer_count = remaining.len();
5575 0 :
5576 0 : macro_rules! lock_status {
5577 0 : ($st:ident) => {
5578 0 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5579 0 : let st = st
5580 0 : .as_mut()
5581 0 : .expect("this function is only called after the task has been spawned");
5582 0 : assert_eq!(
5583 0 : st.task_id,
5584 0 : format!(
5585 0 : "{}",
5586 0 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5587 0 : )
5588 0 : );
5589 0 : let $st = st;
5590 0 : };
5591 0 : }
5592 0 :
5593 0 : {
5594 0 : lock_status!(st);
5595 0 : st.total_layer_count = total_layer_count as u64;
5596 0 : }
5597 0 :
5598 0 : let mut remaining = remaining.into_iter();
5599 0 : let mut have_remaining = true;
5600 0 : let mut js = tokio::task::JoinSet::new();
5601 0 :
5602 0 : let cancel = task_mgr::shutdown_token();
5603 0 :
5604 0 : let limit = request.max_concurrent_downloads;
5605 :
5606 : loop {
5607 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5608 0 : let Some(next) = remaining.next() else {
5609 0 : have_remaining = false;
5610 0 : break;
5611 : };
5612 :
5613 0 : let span = tracing::info_span!("download", layer = %next);
5614 :
5615 0 : js.spawn(
5616 0 : async move {
5617 0 : let res = next.download().await;
5618 0 : (next, res)
5619 0 : }
5620 0 : .instrument(span),
5621 0 : );
5622 0 : }
5623 :
5624 0 : while let Some(res) = js.join_next().await {
5625 0 : match res {
5626 : Ok((_, Ok(_))) => {
5627 0 : lock_status!(st);
5628 0 : st.successful_download_count += 1;
5629 : }
5630 0 : Ok((layer, Err(e))) => {
5631 0 : tracing::error!(%layer, "download failed: {e:#}");
5632 0 : lock_status!(st);
5633 0 : st.failed_download_count += 1;
5634 : }
5635 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5636 0 : Err(je) if je.is_panic() => {
5637 0 : lock_status!(st);
5638 0 : st.failed_download_count += 1;
5639 : }
5640 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5641 : }
5642 : }
5643 :
5644 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5645 0 : break;
5646 0 : }
5647 : }
5648 :
5649 : {
5650 0 : lock_status!(st);
5651 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5652 0 : }
5653 0 : }
5654 :
5655 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5656 0 : &self,
5657 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5658 0 : self.download_all_remote_layers_task_info
5659 0 : .read()
5660 0 : .unwrap()
5661 0 : .clone()
5662 0 : }
5663 : }
5664 :
5665 : impl Timeline {
5666 : /// Returns non-remote layers for eviction.
5667 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5668 0 : let guard = self.layers.read().await;
5669 0 : let mut max_layer_size: Option<u64> = None;
5670 0 :
5671 0 : let resident_layers = guard
5672 0 : .likely_resident_layers()
5673 0 : .map(|layer| {
5674 0 : let file_size = layer.layer_desc().file_size;
5675 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5676 0 :
5677 0 : let last_activity_ts = layer.access_stats().latest_activity();
5678 0 :
5679 0 : EvictionCandidate {
5680 0 : layer: layer.into(),
5681 0 : last_activity_ts,
5682 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5683 0 : }
5684 0 : })
5685 0 : .collect();
5686 0 :
5687 0 : DiskUsageEvictionInfo {
5688 0 : max_layer_size,
5689 0 : resident_layers,
5690 0 : }
5691 0 : }
5692 :
5693 1638 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5694 1638 : ShardIndex {
5695 1638 : shard_number: self.tenant_shard_id.shard_number,
5696 1638 : shard_count: self.tenant_shard_id.shard_count,
5697 1638 : }
5698 1638 : }
5699 :
5700 : #[cfg(test)]
5701 30 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5702 30 : self.last_record_lsn.advance(new_lsn);
5703 30 : }
5704 :
5705 : #[cfg(test)]
5706 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
5707 2 : self.disk_consistent_lsn.store(new_value);
5708 2 : }
5709 :
5710 : /// Force create an image layer and place it into the layer map.
5711 : ///
5712 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5713 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are placed into the layer map in one run.
5714 : #[cfg(test)]
5715 42 : pub(super) async fn force_create_image_layer(
5716 42 : self: &Arc<Timeline>,
5717 42 : lsn: Lsn,
5718 42 : mut images: Vec<(Key, Bytes)>,
5719 42 : check_start_lsn: Option<Lsn>,
5720 42 : ctx: &RequestContext,
5721 42 : ) -> anyhow::Result<()> {
5722 42 : let last_record_lsn = self.get_last_record_lsn();
5723 42 : assert!(
5724 42 : lsn <= last_record_lsn,
5725 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5726 : );
5727 42 : if let Some(check_start_lsn) = check_start_lsn {
5728 42 : assert!(lsn >= check_start_lsn);
5729 0 : }
5730 74 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5731 42 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5732 42 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5733 42 : let mut image_layer_writer = ImageLayerWriter::new(
5734 42 : self.conf,
5735 42 : self.timeline_id,
5736 42 : self.tenant_shard_id,
5737 42 : &(min_key..end_key),
5738 42 : lsn,
5739 42 : ctx,
5740 42 : )
5741 21 : .await?;
5742 158 : for (key, img) in images {
5743 116 : image_layer_writer.put_image(key, img, ctx).await?;
5744 : }
5745 84 : let image_layer = image_layer_writer.finish(self, ctx).await?;
5746 :
5747 42 : {
5748 42 : let mut guard = self.layers.write().await;
5749 42 : guard.force_insert_layer(image_layer);
5750 42 : }
5751 42 :
5752 42 : Ok(())
5753 42 : }
5754 :
5755 : /// Force create a delta layer and place it into the layer map.
5756 : ///
5757 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5758 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are placed into the layer map in one run.
5759 : #[cfg(test)]
5760 46 : pub(super) async fn force_create_delta_layer(
5761 46 : self: &Arc<Timeline>,
5762 46 : mut deltas: DeltaLayerTestDesc,
5763 46 : check_start_lsn: Option<Lsn>,
5764 46 : ctx: &RequestContext,
5765 46 : ) -> anyhow::Result<()> {
5766 46 : let last_record_lsn = self.get_last_record_lsn();
5767 46 : deltas
5768 46 : .data
5769 62 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5770 46 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5771 46 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5772 154 : for (_, lsn, _) in &deltas.data {
5773 108 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5774 : }
5775 46 : assert!(
5776 46 : deltas.lsn_range.end <= last_record_lsn,
5777 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
5778 : deltas.lsn_range.end,
5779 : last_record_lsn
5780 : );
5781 46 : if let Some(check_start_lsn) = check_start_lsn {
5782 46 : assert!(deltas.lsn_range.start >= check_start_lsn);
5783 0 : }
5784 : // check if the delta layer does not violate the LSN invariant, the legacy compaction should always produce a batch of
5785 : // layers of the same start/end LSN, and so should the force inserted layer
5786 : {
5787 : /// Checks if a overlaps with b, assume a/b = [start, end).
5788 48 : pub fn overlaps_with<T: Ord>(a: &Range<T>, b: &Range<T>) -> bool {
5789 48 : !(a.end <= b.start || b.end <= a.start)
5790 48 : }
5791 :
5792 46 : let guard = self.layers.read().await;
5793 88 : for layer in guard.layer_map().iter_historic_layers() {
5794 88 : if layer.is_delta()
5795 48 : && overlaps_with(&layer.lsn_range, &deltas.lsn_range)
5796 14 : && layer.lsn_range != deltas.lsn_range
5797 : {
5798 : // If a delta layer overlaps with another delta layer AND their LSN range is not the same, panic
5799 0 : panic!(
5800 0 : "inserted layer violates delta layer LSN invariant: current_lsn_range={}..{}, conflict_lsn_range={}..{}",
5801 0 : deltas.lsn_range.start, deltas.lsn_range.end, layer.lsn_range.start, layer.lsn_range.end
5802 0 : );
5803 88 : }
5804 : }
5805 : }
5806 46 : let mut delta_layer_writer = DeltaLayerWriter::new(
5807 46 : self.conf,
5808 46 : self.timeline_id,
5809 46 : self.tenant_shard_id,
5810 46 : deltas.key_range.start,
5811 46 : deltas.lsn_range,
5812 46 : ctx,
5813 46 : )
5814 23 : .await?;
5815 154 : for (key, lsn, val) in deltas.data {
5816 108 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5817 : }
5818 46 : let delta_layer = delta_layer_writer
5819 46 : .finish(deltas.key_range.end, self, ctx)
5820 115 : .await?;
5821 :
5822 46 : {
5823 46 : let mut guard = self.layers.write().await;
5824 46 : guard.force_insert_layer(delta_layer);
5825 46 : }
5826 46 :
5827 46 : Ok(())
5828 46 : }
5829 :
5830 : /// Return all keys at the LSN in the image layers
5831 : #[cfg(test)]
5832 6 : pub(crate) async fn inspect_image_layers(
5833 6 : self: &Arc<Timeline>,
5834 6 : lsn: Lsn,
5835 6 : ctx: &RequestContext,
5836 6 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
5837 6 : let mut all_data = Vec::new();
5838 6 : let guard = self.layers.read().await;
5839 34 : for layer in guard.layer_map().iter_historic_layers() {
5840 34 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
5841 8 : let layer = guard.get_from_desc(&layer);
5842 8 : let mut reconstruct_data = ValuesReconstructState::default();
5843 8 : layer
5844 8 : .get_values_reconstruct_data(
5845 8 : KeySpace::single(Key::MIN..Key::MAX),
5846 8 : lsn..Lsn(lsn.0 + 1),
5847 8 : &mut reconstruct_data,
5848 8 : ctx,
5849 8 : )
5850 13 : .await?;
5851 80 : for (k, v) in reconstruct_data.keys {
5852 72 : all_data.push((k, v?.img.unwrap().1));
5853 : }
5854 26 : }
5855 : }
5856 6 : all_data.sort();
5857 6 : Ok(all_data)
5858 6 : }
5859 :
5860 : /// Get all historic layer descriptors in the layer map
5861 : #[cfg(test)]
5862 2 : pub(crate) async fn inspect_historic_layers(
5863 2 : self: &Arc<Timeline>,
5864 2 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
5865 2 : let mut layers = Vec::new();
5866 2 : let guard = self.layers.read().await;
5867 6 : for layer in guard.layer_map().iter_historic_layers() {
5868 6 : layers.push(layer.key());
5869 6 : }
5870 2 : Ok(layers)
5871 2 : }
5872 :
5873 : #[cfg(test)]
5874 10 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
5875 10 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
5876 10 : keyspace.merge(&ks);
5877 10 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
5878 10 : }
5879 : }
5880 :
5881 : type TraversalPathItem = (ValueReconstructResult, Lsn, TraversalId);
5882 :
5883 : /// Tracking writes ingestion does to a particular in-memory layer.
5884 : ///
5885 : /// Cleared upon freezing a layer.
5886 : struct TimelineWriterState {
5887 : open_layer: Arc<InMemoryLayer>,
5888 : current_size: u64,
5889 : // Previous Lsn which passed through
5890 : prev_lsn: Option<Lsn>,
5891 : // Largest Lsn which passed through the current writer
5892 : max_lsn: Option<Lsn>,
5893 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
5894 : cached_last_freeze_at: Lsn,
5895 : }
5896 :
5897 : impl TimelineWriterState {
5898 1256 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
5899 1256 : Self {
5900 1256 : open_layer,
5901 1256 : current_size,
5902 1256 : prev_lsn: None,
5903 1256 : max_lsn: None,
5904 1256 : cached_last_freeze_at: last_freeze_at,
5905 1256 : }
5906 1256 : }
5907 : }
5908 :
5909 : /// Various functions to mutate the timeline.
5910 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
5911 : // This is probably considered a bad practice in Rust and should be fixed eventually,
5912 : // but will cause large code changes.
5913 : pub(crate) struct TimelineWriter<'a> {
5914 : tl: &'a Timeline,
5915 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
5916 : }
5917 :
5918 : impl Deref for TimelineWriter<'_> {
5919 : type Target = Timeline;
5920 :
5921 4807302 : fn deref(&self) -> &Self::Target {
5922 4807302 : self.tl
5923 4807302 : }
5924 : }
5925 :
5926 : #[derive(PartialEq)]
5927 : enum OpenLayerAction {
5928 : Roll,
5929 : Open,
5930 : None,
5931 : }
5932 :
5933 : impl<'a> TimelineWriter<'a> {
5934 : /// Put a new page version that can be constructed from a WAL record
5935 : ///
5936 : /// This will implicitly extend the relation, if the page is beyond the
5937 : /// current end-of-file.
5938 5090546 : pub(crate) async fn put(
5939 5090546 : &mut self,
5940 5090546 : key: Key,
5941 5090546 : lsn: Lsn,
5942 5090546 : value: &Value,
5943 5090546 : ctx: &RequestContext,
5944 5090546 : ) -> anyhow::Result<()> {
5945 5090546 : // Avoid doing allocations for "small" values.
5946 5090546 : // In the regression test suite, the limit of 256 avoided allocations in 95% of cases:
5947 5090546 : // https://github.com/neondatabase/neon/pull/5056#discussion_r1301975061
5948 5090546 : let mut buf = smallvec::SmallVec::<[u8; 256]>::new();
5949 5090546 : value.ser_into(&mut buf)?;
5950 5090546 : let buf_size: u64 = buf.len().try_into().expect("oversized value buf");
5951 5090546 :
5952 5090546 : let action = self.get_open_layer_action(lsn, buf_size);
5953 5090546 : let layer = self.handle_open_layer_action(lsn, action, ctx).await?;
5954 5090546 : let res = layer.put_value(key, lsn, &buf, ctx).await;
5955 :
5956 5090546 : if res.is_ok() {
5957 5090546 : // Update the current size only when the entire write was ok.
5958 5090546 : // In case of failures, we may have had partial writes which
5959 5090546 : // render the size tracking out of sync. That's ok because
5960 5090546 : // the checkpoint distance should be significantly smaller
5961 5090546 : // than the S3 single shot upload limit of 5GiB.
5962 5090546 : let state = self.write_guard.as_mut().unwrap();
5963 5090546 :
5964 5090546 : state.current_size += buf_size;
5965 5090546 : state.prev_lsn = Some(lsn);
5966 5090546 : state.max_lsn = std::cmp::max(state.max_lsn, Some(lsn));
5967 5090546 : }
5968 :
5969 5090546 : res
5970 5090546 : }
5971 :
5972 5090548 : async fn handle_open_layer_action(
5973 5090548 : &mut self,
5974 5090548 : at: Lsn,
5975 5090548 : action: OpenLayerAction,
5976 5090548 : ctx: &RequestContext,
5977 5090548 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
5978 5090548 : match action {
5979 : OpenLayerAction::Roll => {
5980 80 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
5981 80 : self.roll_layer(freeze_at).await?;
5982 80 : self.open_layer(at, ctx).await?;
5983 : }
5984 1176 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
5985 : OpenLayerAction::None => {
5986 5089292 : assert!(self.write_guard.is_some());
5987 : }
5988 : }
5989 :
5990 5090548 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
5991 5090548 : }
5992 :
5993 1256 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
5994 1256 : let layer = self.tl.get_layer_for_write(at, ctx).await?;
5995 1256 : let initial_size = layer.size().await?;
5996 :
5997 1256 : let last_freeze_at = self.last_freeze_at.load();
5998 1256 : self.write_guard.replace(TimelineWriterState::new(
5999 1256 : layer,
6000 1256 : initial_size,
6001 1256 : last_freeze_at,
6002 1256 : ));
6003 1256 :
6004 1256 : Ok(())
6005 1256 : }
6006 :
6007 80 : async fn roll_layer(&mut self, freeze_at: Lsn) -> anyhow::Result<()> {
6008 80 : let current_size = self.write_guard.as_ref().unwrap().current_size;
6009 80 :
6010 80 : // self.write_guard will be taken by the freezing
6011 80 : self.tl
6012 80 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
6013 10 : .await;
6014 :
6015 80 : self.tl.flush_frozen_layers(freeze_at)?;
6016 :
6017 80 : if current_size >= self.get_checkpoint_distance() * 2 {
6018 0 : warn!("Flushed oversized open layer with size {}", current_size)
6019 80 : }
6020 :
6021 80 : Ok(())
6022 80 : }
6023 :
6024 5090548 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
6025 5090548 : let state = &*self.write_guard;
6026 5090548 : let Some(state) = &state else {
6027 1176 : return OpenLayerAction::Open;
6028 : };
6029 :
6030 : #[cfg(feature = "testing")]
6031 5089372 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
6032 : // this check and assertion are not really needed because
6033 : // LayerManager::try_freeze_in_memory_layer will always clear out the
6034 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
6035 : // is no TimelineWriterState.
6036 0 : assert!(
6037 0 : state.open_layer.end_lsn.get().is_some(),
6038 0 : "our open_layer must be outdated"
6039 : );
6040 :
6041 : // this would be a memory leak waiting to happen because the in-memory layer always has
6042 : // an index
6043 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
6044 5089372 : }
6045 5089372 :
6046 5089372 : if state.prev_lsn == Some(lsn) {
6047 : // Rolling mid LSN is not supported by [downstream code].
6048 : // Hence, only roll at LSN boundaries.
6049 : //
6050 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
6051 286346 : return OpenLayerAction::None;
6052 4803026 : }
6053 4803026 :
6054 4803026 : if state.current_size == 0 {
6055 : // Don't roll empty layers
6056 0 : return OpenLayerAction::None;
6057 4803026 : }
6058 4803026 :
6059 4803026 : if self.tl.should_roll(
6060 4803026 : state.current_size,
6061 4803026 : state.current_size + new_value_size,
6062 4803026 : self.get_checkpoint_distance(),
6063 4803026 : lsn,
6064 4803026 : state.cached_last_freeze_at,
6065 4803026 : state.open_layer.get_opened_at(),
6066 4803026 : ) {
6067 80 : OpenLayerAction::Roll
6068 : } else {
6069 4802946 : OpenLayerAction::None
6070 : }
6071 5090548 : }
6072 :
6073 : /// Put a batch of keys at the specified Lsns.
6074 : ///
6075 : /// The batch is sorted by Lsn (enforced by usage of [`utils::vec_map::VecMap`].
6076 414046 : pub(crate) async fn put_batch(
6077 414046 : &mut self,
6078 414046 : batch: VecMap<Lsn, (Key, Value)>,
6079 414046 : ctx: &RequestContext,
6080 414046 : ) -> anyhow::Result<()> {
6081 1114438 : for (lsn, (key, val)) in batch {
6082 700392 : self.put(key, lsn, &val, ctx).await?
6083 : }
6084 :
6085 414046 : Ok(())
6086 414046 : }
6087 :
6088 2 : pub(crate) async fn delete_batch(
6089 2 : &mut self,
6090 2 : batch: &[(Range<Key>, Lsn)],
6091 2 : ctx: &RequestContext,
6092 2 : ) -> anyhow::Result<()> {
6093 2 : if let Some((_, lsn)) = batch.first() {
6094 2 : let action = self.get_open_layer_action(*lsn, 0);
6095 2 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
6096 2 : layer.put_tombstones(batch).await?;
6097 0 : }
6098 :
6099 2 : Ok(())
6100 2 : }
6101 :
6102 : /// Track the end of the latest digested WAL record.
6103 : /// Remember the (end of) last valid WAL record remembered in the timeline.
6104 : ///
6105 : /// Call this after you have finished writing all the WAL up to 'lsn'.
6106 : ///
6107 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
6108 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
6109 : /// the latest and can be read.
6110 5279062 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
6111 5279062 : self.tl.finish_write(new_lsn);
6112 5279062 : }
6113 :
6114 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
6115 270570 : self.tl.update_current_logical_size(delta)
6116 270570 : }
6117 : }
6118 :
6119 : // We need TimelineWriter to be send in upcoming conversion of
6120 : // Timeline::layers to tokio::sync::RwLock.
6121 : #[test]
6122 2 : fn is_send() {
6123 2 : fn _assert_send<T: Send>() {}
6124 2 : _assert_send::<TimelineWriter<'_>>();
6125 2 : }
6126 :
6127 : #[cfg(test)]
6128 : mod tests {
6129 : use utils::{id::TimelineId, lsn::Lsn};
6130 :
6131 : use crate::tenant::{
6132 : harness::TenantHarness, storage_layer::Layer, timeline::EvictionError, Timeline,
6133 : };
6134 :
6135 : #[tokio::test]
6136 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
6137 2 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6138 2 : .await
6139 2 : .unwrap();
6140 2 :
6141 8 : let (tenant, ctx) = harness.load().await;
6142 2 : let timeline = tenant
6143 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6144 4 : .await
6145 2 : .unwrap();
6146 2 :
6147 2 : let layer = find_some_layer(&timeline).await;
6148 2 : let layer = layer
6149 2 : .keep_resident()
6150 2 : .await
6151 2 : .expect("no download => no downloading errors")
6152 2 : .drop_eviction_guard();
6153 2 :
6154 2 : let forever = std::time::Duration::from_secs(120);
6155 2 :
6156 2 : let first = layer.evict_and_wait(forever);
6157 2 : let second = layer.evict_and_wait(forever);
6158 2 :
6159 2 : let (first, second) = tokio::join!(first, second);
6160 2 :
6161 2 : let res = layer.keep_resident().await;
6162 2 : assert!(res.is_none(), "{res:?}");
6163 2 :
6164 2 : match (first, second) {
6165 2 : (Ok(()), Ok(())) => {
6166 2 : // because there are no more timeline locks being taken on eviction path, we can
6167 2 : // witness all three outcomes here.
6168 2 : }
6169 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
6170 0 : // if one completes before the other, this is fine just as well.
6171 0 : }
6172 2 : other => unreachable!("unexpected {:?}", other),
6173 2 : }
6174 2 : }
6175 :
6176 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
6177 2 : let layers = timeline.layers.read().await;
6178 2 : let desc = layers
6179 2 : .layer_map()
6180 2 : .iter_historic_layers()
6181 2 : .next()
6182 2 : .expect("must find one layer to evict");
6183 2 :
6184 2 : layers.get_from_desc(&desc)
6185 2 : }
6186 : }
|
