Line data Source code
1 : mod compaction;
2 : pub mod delete;
3 : mod eviction_task;
4 : mod init;
5 : pub mod layer_manager;
6 : pub(crate) mod logical_size;
7 : pub mod span;
8 : pub mod uninit;
9 : mod walreceiver;
10 :
11 : use anyhow::{anyhow, bail, ensure, Context, Result};
12 : use arc_swap::ArcSwap;
13 : use bytes::Bytes;
14 : use camino::Utf8Path;
15 : use enumset::EnumSet;
16 : use fail::fail_point;
17 : use once_cell::sync::Lazy;
18 : use pageserver_api::{
19 : key::AUX_FILES_KEY,
20 : keyspace::KeySpaceAccum,
21 : models::{
22 : CompactionAlgorithm, DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest,
23 : EvictionPolicy, InMemoryLayerInfo, LayerMapInfo, TimelineState,
24 : },
25 : reltag::BlockNumber,
26 : shard::{ShardIdentity, TenantShardId},
27 : };
28 : use rand::Rng;
29 : use serde_with::serde_as;
30 : use storage_broker::BrokerClientChannel;
31 : use tokio::{
32 : runtime::Handle,
33 : sync::{oneshot, watch},
34 : };
35 : use tokio_util::sync::CancellationToken;
36 : use tracing::*;
37 : use utils::{
38 : bin_ser::BeSer,
39 : sync::gate::{Gate, GateGuard},
40 : vec_map::VecMap,
41 : };
42 :
43 : use std::ops::{Deref, Range};
44 : use std::pin::pin;
45 : use std::sync::atomic::Ordering as AtomicOrdering;
46 : use std::sync::{Arc, Mutex, RwLock, Weak};
47 : use std::time::{Duration, Instant, SystemTime};
48 : use std::{
49 : array,
50 : collections::{BTreeMap, HashMap, HashSet},
51 : sync::atomic::AtomicU64,
52 : };
53 : use std::{
54 : cmp::{max, min, Ordering},
55 : ops::ControlFlow,
56 : };
57 :
58 : use crate::deletion_queue::DeletionQueueClient;
59 : use crate::tenant::timeline::logical_size::CurrentLogicalSize;
60 : use crate::tenant::{
61 : layer_map::{LayerMap, SearchResult},
62 : metadata::TimelineMetadata,
63 : };
64 : use crate::{
65 : context::{DownloadBehavior, RequestContext},
66 : disk_usage_eviction_task::DiskUsageEvictionInfo,
67 : pgdatadir_mapping::CollectKeySpaceError,
68 : };
69 : use crate::{
70 : disk_usage_eviction_task::finite_f32,
71 : tenant::storage_layer::{
72 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
73 : LayerAccessStatsReset, LayerFileName, ResidentLayer, ValueReconstructResult,
74 : ValueReconstructState, ValuesReconstructState,
75 : },
76 : };
77 : use crate::{
78 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
79 : };
80 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
81 : use crate::{
82 : pgdatadir_mapping::{AuxFilesDirectory, DirectoryKind},
83 : virtual_file::{MaybeFatalIo, VirtualFile},
84 : };
85 :
86 : use crate::config::PageServerConf;
87 : use crate::keyspace::{KeyPartitioning, KeySpace};
88 : use crate::metrics::{
89 : TimelineMetrics, MATERIALIZED_PAGE_CACHE_HIT, MATERIALIZED_PAGE_CACHE_HIT_DIRECT,
90 : };
91 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
92 : use crate::tenant::config::TenantConfOpt;
93 : use pageserver_api::key::{is_inherited_key, is_rel_fsm_block_key, is_rel_vm_block_key};
94 : use pageserver_api::reltag::RelTag;
95 : use pageserver_api::shard::ShardIndex;
96 :
97 : use postgres_connection::PgConnectionConfig;
98 : use postgres_ffi::to_pg_timestamp;
99 : use utils::{
100 : completion,
101 : generation::Generation,
102 : id::TimelineId,
103 : lsn::{AtomicLsn, Lsn, RecordLsn},
104 : seqwait::SeqWait,
105 : simple_rcu::{Rcu, RcuReadGuard},
106 : };
107 :
108 : use crate::page_cache;
109 : use crate::repository::GcResult;
110 : use crate::repository::{Key, Value};
111 : use crate::task_mgr;
112 : use crate::task_mgr::TaskKind;
113 : use crate::ZERO_PAGE;
114 :
115 : use self::delete::DeleteTimelineFlow;
116 : pub(super) use self::eviction_task::EvictionTaskTenantState;
117 : use self::eviction_task::EvictionTaskTimelineState;
118 : use self::layer_manager::LayerManager;
119 : use self::logical_size::LogicalSize;
120 : use self::walreceiver::{WalReceiver, WalReceiverConf};
121 :
122 : use super::config::TenantConf;
123 : use super::secondary::heatmap::{HeatMapLayer, HeatMapTimeline};
124 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
125 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
126 : use super::{remote_timeline_client::RemoteTimelineClient, storage_layer::ReadableLayer};
127 :
128 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
129 : pub(super) enum FlushLoopState {
130 : NotStarted,
131 : Running {
132 : #[cfg(test)]
133 : expect_initdb_optimization: bool,
134 : #[cfg(test)]
135 : initdb_optimization_count: usize,
136 : },
137 : Exited,
138 : }
139 :
140 : /// Wrapper for key range to provide reverse ordering by range length for BinaryHeap
141 : #[derive(Debug, Clone, PartialEq, Eq)]
142 : pub(crate) struct Hole {
143 : key_range: Range<Key>,
144 : coverage_size: usize,
145 : }
146 :
147 : impl Ord for Hole {
148 0 : fn cmp(&self, other: &Self) -> Ordering {
149 0 : other.coverage_size.cmp(&self.coverage_size) // inverse order
150 0 : }
151 : }
152 :
153 : impl PartialOrd for Hole {
154 0 : fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
155 0 : Some(self.cmp(other))
156 0 : }
157 : }
158 :
159 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
160 : /// Can be removed after all refactors are done.
161 42 : fn drop_rlock<T>(rlock: tokio::sync::OwnedRwLockReadGuard<T>) {
162 42 : drop(rlock)
163 42 : }
164 :
165 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
166 : /// Can be removed after all refactors are done.
167 646 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
168 646 : drop(rlock)
169 646 : }
170 :
171 : /// The outward-facing resources required to build a Timeline
172 : pub struct TimelineResources {
173 : pub remote_client: Option<RemoteTimelineClient>,
174 : pub deletion_queue_client: DeletionQueueClient,
175 : pub timeline_get_throttle: Arc<
176 : crate::tenant::throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>,
177 : >,
178 : }
179 :
180 : pub(crate) struct AuxFilesState {
181 : pub(crate) dir: Option<AuxFilesDirectory>,
182 : pub(crate) n_deltas: usize,
183 : }
184 :
185 : pub struct Timeline {
186 : conf: &'static PageServerConf,
187 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
188 :
189 : myself: Weak<Self>,
190 :
191 : pub(crate) tenant_shard_id: TenantShardId,
192 : pub timeline_id: TimelineId,
193 :
194 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
195 : /// Never changes for the lifetime of this [`Timeline`] object.
196 : ///
197 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
198 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
199 : pub(crate) generation: Generation,
200 :
201 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
202 : /// to shards, and is constant through the lifetime of this Timeline.
203 : shard_identity: ShardIdentity,
204 :
205 : pub pg_version: u32,
206 :
207 : /// The tuple has two elements.
208 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
209 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
210 : ///
211 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
212 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
213 : ///
214 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
215 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
216 : /// `PersistentLayerDesc`'s.
217 : ///
218 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
219 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
220 : /// runtime, e.g., during page reconstruction.
221 : ///
222 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
223 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
224 : pub(crate) layers: Arc<tokio::sync::RwLock<LayerManager>>,
225 :
226 : last_freeze_at: AtomicLsn,
227 : // Atomic would be more appropriate here.
228 : last_freeze_ts: RwLock<Instant>,
229 :
230 : // WAL redo manager. `None` only for broken tenants.
231 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
232 :
233 : /// Remote storage client.
234 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
235 : pub remote_client: Option<Arc<RemoteTimelineClient>>,
236 :
237 : // What page versions do we hold in the repository? If we get a
238 : // request > last_record_lsn, we need to wait until we receive all
239 : // the WAL up to the request. The SeqWait provides functions for
240 : // that. TODO: If we get a request for an old LSN, such that the
241 : // versions have already been garbage collected away, we should
242 : // throw an error, but we don't track that currently.
243 : //
244 : // last_record_lsn.load().last points to the end of last processed WAL record.
245 : //
246 : // We also remember the starting point of the previous record in
247 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
248 : // first WAL record when the node is started up. But here, we just
249 : // keep track of it.
250 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
251 :
252 : // All WAL records have been processed and stored durably on files on
253 : // local disk, up to this LSN. On crash and restart, we need to re-process
254 : // the WAL starting from this point.
255 : //
256 : // Some later WAL records might have been processed and also flushed to disk
257 : // already, so don't be surprised to see some, but there's no guarantee on
258 : // them yet.
259 : disk_consistent_lsn: AtomicLsn,
260 :
261 : // Parent timeline that this timeline was branched from, and the LSN
262 : // of the branch point.
263 : ancestor_timeline: Option<Arc<Timeline>>,
264 : ancestor_lsn: Lsn,
265 :
266 : pub(super) metrics: TimelineMetrics,
267 :
268 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
269 : // in `crate::page_service` writes these metrics.
270 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
271 :
272 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
273 :
274 : /// Ensures layers aren't frozen by checkpointer between
275 : /// [`Timeline::get_layer_for_write`] and layer reads.
276 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
277 : /// Must always be acquired before the layer map/individual layer lock
278 : /// to avoid deadlock.
279 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
280 :
281 : /// Used to avoid multiple `flush_loop` tasks running
282 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
283 :
284 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
285 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
286 : /// The flush cycle counter is sent back on the layer_flush_done channel when
287 : /// the flush finishes. You can use that to wait for the flush to finish.
288 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
289 : /// read by whoever sends an update
290 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
291 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
292 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
293 :
294 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
295 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
296 :
297 : // List of child timelines and their branch points. This is needed to avoid
298 : // garbage collecting data that is still needed by the child timelines.
299 : pub gc_info: std::sync::RwLock<GcInfo>,
300 :
301 : // It may change across major versions so for simplicity
302 : // keep it after running initdb for a timeline.
303 : // It is needed in checks when we want to error on some operations
304 : // when they are requested for pre-initdb lsn.
305 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
306 : // though let's keep them both for better error visibility.
307 : pub initdb_lsn: Lsn,
308 :
309 : /// When did we last calculate the partitioning?
310 : partitioning: tokio::sync::Mutex<(KeyPartitioning, Lsn)>,
311 :
312 : /// Configuration: how often should the partitioning be recalculated.
313 : repartition_threshold: u64,
314 :
315 : last_image_layer_creation_check_at: AtomicLsn,
316 :
317 : /// Current logical size of the "datadir", at the last LSN.
318 : current_logical_size: LogicalSize,
319 :
320 : /// Information about the last processed message by the WAL receiver,
321 : /// or None if WAL receiver has not received anything for this timeline
322 : /// yet.
323 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
324 : pub walreceiver: Mutex<Option<WalReceiver>>,
325 :
326 : /// Relation size cache
327 : pub rel_size_cache: RwLock<HashMap<RelTag, (Lsn, BlockNumber)>>,
328 :
329 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
330 :
331 : state: watch::Sender<TimelineState>,
332 :
333 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
334 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
335 : pub delete_progress: Arc<tokio::sync::Mutex<DeleteTimelineFlow>>,
336 :
337 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
338 :
339 : /// Load or creation time information about the disk_consistent_lsn and when the loading
340 : /// happened. Used for consumption metrics.
341 : pub(crate) loaded_at: (Lsn, SystemTime),
342 :
343 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
344 : pub(crate) gate: Gate,
345 :
346 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
347 : /// to the timeline should drop out when this token fires.
348 : pub(crate) cancel: CancellationToken,
349 :
350 : /// Make sure we only have one running compaction at a time in tests.
351 : ///
352 : /// Must only be taken in two places:
353 : /// - [`Timeline::compact`] (this file)
354 : /// - [`delete::delete_local_timeline_directory`]
355 : ///
356 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
357 : compaction_lock: tokio::sync::Mutex<()>,
358 :
359 : /// Make sure we only have one running gc at a time.
360 : ///
361 : /// Must only be taken in two places:
362 : /// - [`Timeline::gc`] (this file)
363 : /// - [`delete::delete_local_timeline_directory`]
364 : ///
365 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
366 : gc_lock: tokio::sync::Mutex<()>,
367 :
368 : /// Cloned from [`super::Tenant::timeline_get_throttle`] on construction.
369 : timeline_get_throttle: Arc<
370 : crate::tenant::throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>,
371 : >,
372 :
373 : /// Keep aux directory cache to avoid it's reconstruction on each update
374 : pub(crate) aux_files: tokio::sync::Mutex<AuxFilesState>,
375 : }
376 :
377 : pub struct WalReceiverInfo {
378 : pub wal_source_connconf: PgConnectionConfig,
379 : pub last_received_msg_lsn: Lsn,
380 : pub last_received_msg_ts: u128,
381 : }
382 :
383 : ///
384 : /// Information about how much history needs to be retained, needed by
385 : /// Garbage Collection.
386 : ///
387 : pub struct GcInfo {
388 : /// Specific LSNs that are needed.
389 : ///
390 : /// Currently, this includes all points where child branches have
391 : /// been forked off from. In the future, could also include
392 : /// explicit user-defined snapshot points.
393 : pub retain_lsns: Vec<Lsn>,
394 :
395 : /// In addition to 'retain_lsns', keep everything newer than this
396 : /// point.
397 : ///
398 : /// This is calculated by subtracting 'gc_horizon' setting from
399 : /// last-record LSN
400 : ///
401 : /// FIXME: is this inclusive or exclusive?
402 : pub horizon_cutoff: Lsn,
403 :
404 : /// In addition to 'retain_lsns' and 'horizon_cutoff', keep everything newer than this
405 : /// point.
406 : ///
407 : /// This is calculated by finding a number such that a record is needed for PITR
408 : /// if only if its LSN is larger than 'pitr_cutoff'.
409 : pub pitr_cutoff: Lsn,
410 : }
411 :
412 : /// An error happened in a get() operation.
413 2 : #[derive(thiserror::Error, Debug)]
414 : pub(crate) enum PageReconstructError {
415 : #[error(transparent)]
416 : Other(#[from] anyhow::Error),
417 :
418 : #[error("Ancestor LSN wait error: {0}")]
419 : AncestorLsnTimeout(#[from] WaitLsnError),
420 :
421 : #[error("timeline shutting down")]
422 : Cancelled,
423 :
424 : /// The ancestor of this is being stopped
425 : #[error("ancestor timeline {0} is being stopped")]
426 : AncestorStopping(TimelineId),
427 :
428 : /// An error happened replaying WAL records
429 : #[error(transparent)]
430 : WalRedo(anyhow::Error),
431 : }
432 :
433 : impl PageReconstructError {
434 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
435 0 : pub(crate) fn is_stopping(&self) -> bool {
436 0 : use PageReconstructError::*;
437 0 : match self {
438 0 : Other(_) => false,
439 0 : AncestorLsnTimeout(_) => false,
440 0 : Cancelled | AncestorStopping(_) => true,
441 0 : WalRedo(_) => false,
442 : }
443 0 : }
444 : }
445 :
446 0 : #[derive(thiserror::Error, Debug)]
447 : enum CreateImageLayersError {
448 : #[error("timeline shutting down")]
449 : Cancelled,
450 :
451 : #[error(transparent)]
452 : GetVectoredError(GetVectoredError),
453 :
454 : #[error(transparent)]
455 : PageReconstructError(PageReconstructError),
456 :
457 : #[error(transparent)]
458 : Other(#[from] anyhow::Error),
459 : }
460 :
461 0 : #[derive(thiserror::Error, Debug)]
462 : enum FlushLayerError {
463 : /// Timeline cancellation token was cancelled
464 : #[error("timeline shutting down")]
465 : Cancelled,
466 :
467 : #[error(transparent)]
468 : CreateImageLayersError(CreateImageLayersError),
469 :
470 : #[error(transparent)]
471 : Other(#[from] anyhow::Error),
472 : }
473 :
474 0 : #[derive(thiserror::Error, Debug)]
475 : pub(crate) enum GetVectoredError {
476 : #[error("timeline shutting down")]
477 : Cancelled,
478 :
479 : #[error("Requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
480 : Oversized(u64),
481 :
482 : #[error("Requested at invalid LSN: {0}")]
483 : InvalidLsn(Lsn),
484 :
485 : #[error("Requested key {0} not found")]
486 : MissingKey(Key),
487 :
488 : #[error(transparent)]
489 : GetReadyAncestorError(GetReadyAncestorError),
490 :
491 : #[error(transparent)]
492 : Other(#[from] anyhow::Error),
493 : }
494 :
495 0 : #[derive(thiserror::Error, Debug)]
496 : pub(crate) enum GetReadyAncestorError {
497 : #[error("ancestor timeline {0} is being stopped")]
498 : AncestorStopping(TimelineId),
499 :
500 : #[error("Ancestor LSN wait error: {0}")]
501 : AncestorLsnTimeout(#[from] WaitLsnError),
502 :
503 : #[error("Cancelled")]
504 : Cancelled,
505 :
506 : #[error(transparent)]
507 : Other(#[from] anyhow::Error),
508 : }
509 :
510 : #[derive(Clone, Copy)]
511 : pub enum LogicalSizeCalculationCause {
512 : Initial,
513 : ConsumptionMetricsSyntheticSize,
514 : EvictionTaskImitation,
515 : TenantSizeHandler,
516 : }
517 :
518 : pub enum GetLogicalSizePriority {
519 : User,
520 : Background,
521 : }
522 :
523 0 : #[derive(enumset::EnumSetType)]
524 : pub(crate) enum CompactFlags {
525 : ForceRepartition,
526 : ForceImageLayerCreation,
527 : }
528 :
529 : impl std::fmt::Debug for Timeline {
530 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
531 0 : write!(f, "Timeline<{}>", self.timeline_id)
532 0 : }
533 : }
534 :
535 0 : #[derive(thiserror::Error, Debug)]
536 : pub(crate) enum WaitLsnError {
537 : // Called on a timeline which is shutting down
538 : #[error("Shutdown")]
539 : Shutdown,
540 :
541 : // Called on an timeline not in active state or shutting down
542 : #[error("Bad state (not active)")]
543 : BadState,
544 :
545 : // Timeout expired while waiting for LSN to catch up with goal.
546 : #[error("{0}")]
547 : Timeout(String),
548 : }
549 :
550 : // The impls below achieve cancellation mapping for errors.
551 : // Perhaps there's a way of achieving this with less cruft.
552 :
553 : impl From<CreateImageLayersError> for CompactionError {
554 0 : fn from(e: CreateImageLayersError) -> Self {
555 0 : match e {
556 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
557 0 : _ => CompactionError::Other(e.into()),
558 : }
559 0 : }
560 : }
561 :
562 : impl From<CreateImageLayersError> for FlushLayerError {
563 0 : fn from(e: CreateImageLayersError) -> Self {
564 0 : match e {
565 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
566 0 : any => FlushLayerError::CreateImageLayersError(any),
567 : }
568 0 : }
569 : }
570 :
571 : impl From<PageReconstructError> for CreateImageLayersError {
572 0 : fn from(e: PageReconstructError) -> Self {
573 0 : match e {
574 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
575 0 : _ => CreateImageLayersError::PageReconstructError(e),
576 : }
577 0 : }
578 : }
579 :
580 : impl From<GetVectoredError> for CreateImageLayersError {
581 0 : fn from(e: GetVectoredError) -> Self {
582 0 : match e {
583 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
584 0 : _ => CreateImageLayersError::GetVectoredError(e),
585 : }
586 0 : }
587 : }
588 :
589 : impl From<GetReadyAncestorError> for PageReconstructError {
590 2 : fn from(e: GetReadyAncestorError) -> Self {
591 2 : use GetReadyAncestorError::*;
592 2 : match e {
593 0 : AncestorStopping(tid) => PageReconstructError::AncestorStopping(tid),
594 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
595 0 : Cancelled => PageReconstructError::Cancelled,
596 2 : Other(other) => PageReconstructError::Other(other),
597 : }
598 2 : }
599 : }
600 :
601 : #[derive(
602 : Eq,
603 : PartialEq,
604 : Debug,
605 : Copy,
606 : Clone,
607 138 : strum_macros::EnumString,
608 0 : strum_macros::Display,
609 0 : serde_with::DeserializeFromStr,
610 : serde_with::SerializeDisplay,
611 : )]
612 : #[strum(serialize_all = "kebab-case")]
613 : pub enum GetVectoredImpl {
614 : Sequential,
615 : Vectored,
616 : }
617 :
618 : pub(crate) enum WaitLsnWaiter<'a> {
619 : Timeline(&'a Timeline),
620 : Tenant,
621 : PageService,
622 : }
623 :
624 : /// Argument to [`Timeline::shutdown`].
625 : #[derive(Debug, Clone, Copy)]
626 : pub(crate) enum ShutdownMode {
627 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
628 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
629 : ///
630 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
631 : /// the call to [`Timeline::shutdown`].
632 : FreezeAndFlush,
633 : /// Shut down immediately, without waiting for any open layers to flush.
634 : Hard,
635 : }
636 :
637 : /// Public interface functions
638 : impl Timeline {
639 : /// Get the LSN where this branch was created
640 8 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
641 8 : self.ancestor_lsn
642 8 : }
643 :
644 : /// Get the ancestor's timeline id
645 14 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
646 14 : self.ancestor_timeline
647 14 : .as_ref()
648 14 : .map(|ancestor| ancestor.timeline_id)
649 14 : }
650 :
651 : /// Lock and get timeline's GC cutoff
652 724 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
653 724 : self.latest_gc_cutoff_lsn.read()
654 724 : }
655 :
656 : /// Look up given page version.
657 : ///
658 : /// If a remote layer file is needed, it is downloaded as part of this
659 : /// call.
660 : ///
661 : /// This method enforces [`Self::timeline_get_throttle`] internally.
662 : ///
663 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
664 : /// abstraction above this needs to store suitable metadata to track what
665 : /// data exists with what keys, in separate metadata entries. If a
666 : /// non-existent key is requested, we may incorrectly return a value from
667 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
668 : /// non-existing key.
669 : ///
670 : /// # Cancel-Safety
671 : ///
672 : /// This method is cancellation-safe.
673 : #[inline(always)]
674 501761 : pub(crate) async fn get(
675 501761 : &self,
676 501761 : key: Key,
677 501761 : lsn: Lsn,
678 501761 : ctx: &RequestContext,
679 501761 : ) -> Result<Bytes, PageReconstructError> {
680 501761 : self.timeline_get_throttle.throttle(ctx, 1).await;
681 501761 : self.get_impl(key, lsn, ctx).await
682 501761 : }
683 : /// Not subject to [`Self::timeline_get_throttle`].
684 502833 : async fn get_impl(
685 502833 : &self,
686 502833 : key: Key,
687 502833 : lsn: Lsn,
688 502833 : ctx: &RequestContext,
689 502833 : ) -> Result<Bytes, PageReconstructError> {
690 502833 : if !lsn.is_valid() {
691 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
692 502833 : }
693 502833 :
694 502833 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
695 502833 : // already checked the key against the shard_identity when looking up the Timeline from
696 502833 : // page_service.
697 502833 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
698 :
699 : // XXX: structured stats collection for layer eviction here.
700 502833 : trace!(
701 0 : "get page request for {}@{} from task kind {:?}",
702 0 : key,
703 0 : lsn,
704 0 : ctx.task_kind()
705 0 : );
706 :
707 : // Check the page cache. We will get back the most recent page with lsn <= `lsn`.
708 : // The cached image can be returned directly if there is no WAL between the cached image
709 : // and requested LSN. The cached image can also be used to reduce the amount of WAL needed
710 : // for redo.
711 502833 : let cached_page_img = match self.lookup_cached_page(&key, lsn, ctx).await {
712 0 : Some((cached_lsn, cached_img)) => {
713 0 : match cached_lsn.cmp(&lsn) {
714 0 : Ordering::Less => {} // there might be WAL between cached_lsn and lsn, we need to check
715 : Ordering::Equal => {
716 0 : MATERIALIZED_PAGE_CACHE_HIT_DIRECT.inc();
717 0 : return Ok(cached_img); // exact LSN match, return the image
718 : }
719 : Ordering::Greater => {
720 0 : unreachable!("the returned lsn should never be after the requested lsn")
721 : }
722 : }
723 0 : Some((cached_lsn, cached_img))
724 : }
725 502833 : None => None,
726 : };
727 :
728 502833 : let mut reconstruct_state = ValueReconstructState {
729 502833 : records: Vec::new(),
730 502833 : img: cached_page_img,
731 502833 : };
732 502833 :
733 502833 : let timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME.start_timer();
734 502833 : let path = self
735 502833 : .get_reconstruct_data(key, lsn, &mut reconstruct_state, ctx)
736 34104 : .await?;
737 502723 : timer.stop_and_record();
738 502723 :
739 502723 : let start = Instant::now();
740 502723 : let res = self.reconstruct_value(key, lsn, reconstruct_state).await;
741 502723 : let elapsed = start.elapsed();
742 502723 : crate::metrics::RECONSTRUCT_TIME
743 502723 : .for_result(&res)
744 502723 : .observe(elapsed.as_secs_f64());
745 502723 :
746 502723 : if cfg!(feature = "testing") && res.is_err() {
747 : // it can only be walredo issue
748 : use std::fmt::Write;
749 :
750 0 : let mut msg = String::new();
751 0 :
752 0 : path.into_iter().for_each(|(res, cont_lsn, layer)| {
753 0 : writeln!(
754 0 : msg,
755 0 : "- layer traversal: result {res:?}, cont_lsn {cont_lsn}, layer: {}",
756 0 : layer(),
757 0 : )
758 0 : .expect("string grows")
759 0 : });
760 0 :
761 0 : // this is to rule out or provide evidence that we could in some cases read a duplicate
762 0 : // walrecord
763 0 : tracing::info!("walredo failed, path:\n{msg}");
764 502723 : }
765 :
766 502723 : res
767 502833 : }
768 :
769 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
770 :
771 : /// Look up multiple page versions at a given LSN
772 : ///
773 : /// This naive implementation will be replaced with a more efficient one
774 : /// which actually vectorizes the read path.
775 568 : pub(crate) async fn get_vectored(
776 568 : &self,
777 568 : keyspace: KeySpace,
778 568 : lsn: Lsn,
779 568 : ctx: &RequestContext,
780 568 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
781 568 : if !lsn.is_valid() {
782 0 : return Err(GetVectoredError::InvalidLsn(lsn));
783 568 : }
784 568 :
785 568 : let key_count = keyspace.total_size().try_into().unwrap();
786 568 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
787 0 : return Err(GetVectoredError::Oversized(key_count));
788 568 : }
789 :
790 1136 : for range in &keyspace.ranges {
791 568 : let mut key = range.start;
792 1320 : while key != range.end {
793 752 : assert!(!self.shard_identity.is_key_disposable(&key));
794 752 : key = key.next();
795 : }
796 : }
797 :
798 568 : trace!(
799 0 : "get vectored request for {:?}@{} from task kind {:?} will use {} implementation",
800 0 : keyspace,
801 0 : lsn,
802 0 : ctx.task_kind(),
803 0 : self.conf.get_vectored_impl
804 0 : );
805 :
806 568 : let start = crate::metrics::GET_VECTORED_LATENCY
807 568 : .for_task_kind(ctx.task_kind())
808 568 : .map(|metric| (metric, Instant::now()));
809 :
810 : // start counting after throttle so that throttle time
811 : // is always less than observation time
812 568 : let throttled = self
813 568 : .timeline_get_throttle
814 568 : .throttle(ctx, key_count as usize)
815 0 : .await;
816 :
817 568 : let res = match self.conf.get_vectored_impl {
818 : GetVectoredImpl::Sequential => {
819 568 : self.get_vectored_sequential_impl(keyspace, lsn, ctx).await
820 : }
821 : GetVectoredImpl::Vectored => {
822 0 : let vectored_res = self.get_vectored_impl(keyspace.clone(), lsn, ctx).await;
823 :
824 0 : if self.conf.validate_vectored_get {
825 0 : self.validate_get_vectored_impl(&vectored_res, keyspace, lsn, ctx)
826 0 : .await;
827 0 : }
828 :
829 0 : vectored_res
830 : }
831 : };
832 :
833 568 : if let Some((metric, start)) = start {
834 0 : let elapsed = start.elapsed();
835 0 : let ex_throttled = if let Some(throttled) = throttled {
836 0 : elapsed.checked_sub(throttled)
837 : } else {
838 0 : Some(elapsed)
839 : };
840 :
841 0 : if let Some(ex_throttled) = ex_throttled {
842 0 : metric.observe(ex_throttled.as_secs_f64());
843 0 : } else {
844 0 : use utils::rate_limit::RateLimit;
845 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
846 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
847 0 : let mut rate_limit = LOGGED.lock().unwrap();
848 0 : rate_limit.call(|| {
849 0 : warn!("error deducting time spent throttled; this message is logged at a global rate limit");
850 0 : });
851 0 : }
852 568 : }
853 :
854 568 : res
855 568 : }
856 :
857 : /// Not subject to [`Self::timeline_get_throttle`].
858 578 : pub(super) async fn get_vectored_sequential_impl(
859 578 : &self,
860 578 : keyspace: KeySpace,
861 578 : lsn: Lsn,
862 578 : ctx: &RequestContext,
863 578 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
864 578 : let mut values = BTreeMap::new();
865 1156 : for range in keyspace.ranges {
866 578 : let mut key = range.start;
867 1650 : while key != range.end {
868 1072 : let block = self.get_impl(key, lsn, ctx).await;
869 :
870 : use PageReconstructError::*;
871 0 : match block {
872 : Err(Cancelled | AncestorStopping(_)) => {
873 0 : return Err(GetVectoredError::Cancelled)
874 : }
875 0 : Err(Other(err)) if err.to_string().contains("could not find data for key") => {
876 0 : return Err(GetVectoredError::MissingKey(key))
877 : }
878 1072 : _ => {
879 1072 : values.insert(key, block);
880 1072 : key = key.next();
881 1072 : }
882 : }
883 : }
884 : }
885 :
886 578 : Ok(values)
887 578 : }
888 :
889 12 : pub(super) async fn get_vectored_impl(
890 12 : &self,
891 12 : keyspace: KeySpace,
892 12 : lsn: Lsn,
893 12 : ctx: &RequestContext,
894 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
895 12 : let mut reconstruct_state = ValuesReconstructState::new();
896 12 :
897 12 : self.get_vectored_reconstruct_data(keyspace, lsn, &mut reconstruct_state, ctx)
898 41 : .await?;
899 :
900 12 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
901 374 : for (key, res) in reconstruct_state.keys {
902 362 : match res {
903 0 : Err(err) => {
904 0 : results.insert(key, Err(err));
905 0 : }
906 362 : Ok(state) => {
907 362 : let state = ValueReconstructState::from(state);
908 :
909 362 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
910 362 : results.insert(key, reconstruct_res);
911 : }
912 : }
913 : }
914 :
915 12 : Ok(results)
916 12 : }
917 :
918 : /// Not subject to [`Self::timeline_get_throttle`].
919 10 : pub(super) async fn validate_get_vectored_impl(
920 10 : &self,
921 10 : vectored_res: &Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError>,
922 10 : keyspace: KeySpace,
923 10 : lsn: Lsn,
924 10 : ctx: &RequestContext,
925 10 : ) {
926 10 : let sequential_res = self
927 10 : .get_vectored_sequential_impl(keyspace.clone(), lsn, ctx)
928 20 : .await;
929 :
930 0 : fn errors_match(lhs: &GetVectoredError, rhs: &GetVectoredError) -> bool {
931 0 : use GetVectoredError::*;
932 0 : match (lhs, rhs) {
933 0 : (Oversized(l), Oversized(r)) => l == r,
934 0 : (InvalidLsn(l), InvalidLsn(r)) => l == r,
935 0 : (MissingKey(l), MissingKey(r)) => l == r,
936 0 : (GetReadyAncestorError(_), GetReadyAncestorError(_)) => true,
937 0 : (Other(_), Other(_)) => true,
938 0 : _ => false,
939 : }
940 0 : }
941 :
942 10 : match (&sequential_res, vectored_res) {
943 0 : (Err(GetVectoredError::Cancelled), _) => {},
944 0 : (_, Err(GetVectoredError::Cancelled)) => {},
945 0 : (Err(seq_err), Ok(_)) => {
946 0 : panic!(concat!("Sequential get failed with {}, but vectored get did not",
947 0 : " - keyspace={:?} lsn={}"),
948 0 : seq_err, keyspace, lsn) },
949 0 : (Ok(_), Err(vec_err)) => {
950 0 : panic!(concat!("Vectored get failed with {}, but sequential get did not",
951 0 : " - keyspace={:?} lsn={}"),
952 0 : vec_err, keyspace, lsn) },
953 0 : (Err(seq_err), Err(vec_err)) => {
954 0 : assert!(errors_match(seq_err, vec_err),
955 0 : "Mismatched errors: {seq_err} != {vec_err} - keyspace={keyspace:?} lsn={lsn}")},
956 10 : (Ok(seq_values), Ok(vec_values)) => {
957 320 : seq_values.iter().zip(vec_values.iter()).for_each(|((seq_key, seq_res), (vec_key, vec_res))| {
958 320 : assert_eq!(seq_key, vec_key);
959 320 : match (seq_res, vec_res) {
960 320 : (Ok(seq_blob), Ok(vec_blob)) => {
961 320 : Self::validate_key_equivalence(seq_key, &keyspace, lsn, seq_blob, vec_blob);
962 320 : },
963 0 : (Err(err), Ok(_)) => {
964 0 : panic!(
965 0 : concat!("Sequential get failed with {} for key {}, but vectored get did not",
966 0 : " - keyspace={:?} lsn={}"),
967 0 : err, seq_key, keyspace, lsn) },
968 0 : (Ok(_), Err(err)) => {
969 0 : panic!(
970 0 : concat!("Vectored get failed with {} for key {}, but sequential get did not",
971 0 : " - keyspace={:?} lsn={}"),
972 0 : err, seq_key, keyspace, lsn) },
973 0 : (Err(_), Err(_)) => {}
974 : }
975 320 : })
976 : }
977 : }
978 10 : }
979 :
980 320 : fn validate_key_equivalence(
981 320 : key: &Key,
982 320 : keyspace: &KeySpace,
983 320 : lsn: Lsn,
984 320 : seq: &Bytes,
985 320 : vec: &Bytes,
986 320 : ) {
987 320 : if *key == AUX_FILES_KEY {
988 : // The value reconstruct of AUX_FILES_KEY from records is not deterministic
989 : // since it uses a hash map under the hood. Hence, deserialise both results
990 : // before comparing.
991 0 : let seq_aux_dir_res = AuxFilesDirectory::des(seq);
992 0 : let vec_aux_dir_res = AuxFilesDirectory::des(vec);
993 0 : match (&seq_aux_dir_res, &vec_aux_dir_res) {
994 0 : (Ok(seq_aux_dir), Ok(vec_aux_dir)) => {
995 0 : assert_eq!(
996 : seq_aux_dir, vec_aux_dir,
997 0 : "Mismatch for key {} - keyspace={:?} lsn={}",
998 : key, keyspace, lsn
999 : );
1000 : }
1001 0 : (Err(_), Err(_)) => {}
1002 : _ => {
1003 0 : panic!("Mismatch for {key}: {seq_aux_dir_res:?} != {vec_aux_dir_res:?}");
1004 : }
1005 : }
1006 : } else {
1007 : // All other keys should reconstruct deterministically, so we simply compare the blobs.
1008 320 : assert_eq!(
1009 : seq, vec,
1010 0 : "Image mismatch for key {key} - keyspace={keyspace:?} lsn={lsn}"
1011 : );
1012 : }
1013 320 : }
1014 :
1015 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1016 3921526 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1017 3921526 : self.last_record_lsn.load().last
1018 3921526 : }
1019 :
1020 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1021 0 : self.last_record_lsn.load().prev
1022 0 : }
1023 :
1024 : /// Atomically get both last and prev.
1025 214 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1026 214 : self.last_record_lsn.load()
1027 214 : }
1028 :
1029 1330 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1030 1330 : self.disk_consistent_lsn.load()
1031 1330 : }
1032 :
1033 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1034 : /// not validated with control plane yet.
1035 : /// See [`Self::get_remote_consistent_lsn_visible`].
1036 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1037 0 : if let Some(remote_client) = &self.remote_client {
1038 0 : remote_client.remote_consistent_lsn_projected()
1039 : } else {
1040 0 : None
1041 : }
1042 0 : }
1043 :
1044 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1045 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1046 : /// generation validation in the deletion queue.
1047 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1048 0 : if let Some(remote_client) = &self.remote_client {
1049 0 : remote_client.remote_consistent_lsn_visible()
1050 : } else {
1051 0 : None
1052 : }
1053 0 : }
1054 :
1055 : /// The sum of the file size of all historic layers in the layer map.
1056 : /// This method makes no distinction between local and remote layers.
1057 : /// Hence, the result **does not represent local filesystem usage**.
1058 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1059 0 : let guard = self.layers.read().await;
1060 0 : let layer_map = guard.layer_map();
1061 0 : let mut size = 0;
1062 0 : for l in layer_map.iter_historic_layers() {
1063 0 : size += l.file_size();
1064 0 : }
1065 0 : size
1066 0 : }
1067 :
1068 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1069 0 : self.metrics.resident_physical_size_get()
1070 0 : }
1071 :
1072 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1073 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1074 0 : }
1075 :
1076 : ///
1077 : /// Wait until WAL has been received and processed up to this LSN.
1078 : ///
1079 : /// You should call this before any of the other get_* or list_* functions. Calling
1080 : /// those functions with an LSN that has been processed yet is an error.
1081 : ///
1082 226805 : pub(crate) async fn wait_lsn(
1083 226805 : &self,
1084 226805 : lsn: Lsn,
1085 226805 : who_is_waiting: WaitLsnWaiter<'_>,
1086 226805 : ctx: &RequestContext, /* Prepare for use by cancellation */
1087 226805 : ) -> Result<(), WaitLsnError> {
1088 226805 : if self.cancel.is_cancelled() {
1089 0 : return Err(WaitLsnError::Shutdown);
1090 226805 : } else if !self.is_active() {
1091 0 : return Err(WaitLsnError::BadState);
1092 226805 : }
1093 226805 :
1094 226805 : if cfg!(debug_assertions) {
1095 226805 : match ctx.task_kind() {
1096 : TaskKind::WalReceiverManager
1097 : | TaskKind::WalReceiverConnectionHandler
1098 : | TaskKind::WalReceiverConnectionPoller => {
1099 0 : let is_myself = match who_is_waiting {
1100 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1101 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1102 : };
1103 0 : if is_myself {
1104 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1105 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1106 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1107 0 : }
1108 0 : } else {
1109 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1110 0 : // our walreceiver task can make progress independent of theirs
1111 0 : }
1112 : }
1113 226805 : _ => {}
1114 : }
1115 0 : }
1116 :
1117 226805 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1118 226805 :
1119 226805 : match self
1120 226805 : .last_record_lsn
1121 226805 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1122 0 : .await
1123 : {
1124 226805 : Ok(()) => Ok(()),
1125 0 : Err(e) => {
1126 0 : use utils::seqwait::SeqWaitError::*;
1127 0 : match e {
1128 0 : Shutdown => Err(WaitLsnError::Shutdown),
1129 : Timeout => {
1130 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1131 0 : drop(_timer);
1132 0 : let walreceiver_status = self.walreceiver_status();
1133 0 : Err(WaitLsnError::Timeout(format!(
1134 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1135 0 : lsn,
1136 0 : self.get_last_record_lsn(),
1137 0 : self.get_disk_consistent_lsn(),
1138 0 : walreceiver_status,
1139 0 : )))
1140 : }
1141 : }
1142 : }
1143 : }
1144 226805 : }
1145 :
1146 0 : pub(crate) fn walreceiver_status(&self) -> String {
1147 0 : match &*self.walreceiver.lock().unwrap() {
1148 0 : None => "stopping or stopped".to_string(),
1149 0 : Some(walreceiver) => match walreceiver.status() {
1150 0 : Some(status) => status.to_human_readable_string(),
1151 0 : None => "Not active".to_string(),
1152 : },
1153 : }
1154 0 : }
1155 :
1156 : /// Check that it is valid to request operations with that lsn.
1157 218 : pub(crate) fn check_lsn_is_in_scope(
1158 218 : &self,
1159 218 : lsn: Lsn,
1160 218 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1161 218 : ) -> anyhow::Result<()> {
1162 218 : ensure!(
1163 218 : lsn >= **latest_gc_cutoff_lsn,
1164 4 : "LSN {} is earlier than latest GC horizon {} (we might've already garbage collected needed data)",
1165 4 : lsn,
1166 4 : **latest_gc_cutoff_lsn,
1167 : );
1168 214 : Ok(())
1169 218 : }
1170 :
1171 : /// Flush to disk all data that was written with the put_* functions
1172 1396 : #[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))]
1173 : pub(crate) async fn freeze_and_flush(&self) -> anyhow::Result<()> {
1174 : let to_lsn = self.freeze_inmem_layer(false).await;
1175 : self.flush_frozen_layers_and_wait(to_lsn).await
1176 : }
1177 :
1178 : /// If there is no writer, and conditions for rolling the latest layer are met, then freeze it.
1179 : ///
1180 : /// This is for use in background housekeeping, to provide guarantees of layers closing eventually
1181 : /// even if there are no ongoing writes to drive that.
1182 510 : async fn maybe_freeze_ephemeral_layer(&self) {
1183 510 : let Ok(_write_guard) = self.write_lock.try_lock() else {
1184 : // If the write lock is held, there is an active wal receiver: rolling open layers
1185 : // is their responsibility while they hold this lock.
1186 0 : return;
1187 : };
1188 :
1189 510 : let Ok(layers_guard) = self.layers.try_read() else {
1190 : // Don't block if the layer lock is busy
1191 0 : return;
1192 : };
1193 :
1194 510 : let Some(open_layer) = &layers_guard.layer_map().open_layer else {
1195 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1196 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1197 : // that didn't result in writes to this shard.
1198 :
1199 : // Must not hold the layers lock while waiting for a flush.
1200 510 : drop(layers_guard);
1201 510 :
1202 510 : let last_record_lsn = self.get_last_record_lsn();
1203 510 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1204 510 : if last_record_lsn > disk_consistent_lsn {
1205 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1206 : // we are a sharded tenant and have skipped some WAL
1207 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1208 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1209 : // This should be somewhat rare, so we log it at INFO level.
1210 : //
1211 : // We checked for checkpoint timeout so that a shard without any
1212 : // data ingested (yet) doesn't write a remote index as soon as it
1213 : // sees its LSN advance: we only do this if we've been layer-less
1214 : // for some time.
1215 0 : tracing::info!(
1216 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1217 0 : disk_consistent_lsn,
1218 0 : last_record_lsn
1219 0 : );
1220 :
1221 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1222 0 : self.flush_frozen_layers_and_wait(last_record_lsn)
1223 0 : .await
1224 0 : .ok();
1225 0 : }
1226 510 : }
1227 :
1228 510 : return;
1229 : };
1230 :
1231 0 : let Some(current_size) = open_layer.try_len() else {
1232 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1233 : // read lock to get size should always succeed.
1234 0 : tracing::warn!("Lock conflict while reading size of open layer");
1235 0 : return;
1236 : };
1237 :
1238 0 : let current_lsn = self.get_last_record_lsn();
1239 :
1240 0 : let checkpoint_distance_override = open_layer.tick().await;
1241 :
1242 0 : if let Some(size_override) = checkpoint_distance_override {
1243 0 : if current_size > size_override {
1244 : // This is not harmful, but it only happens in relatively rare cases where
1245 : // time-based checkpoints are not happening fast enough to keep the amount of
1246 : // ephemeral data within configured limits. It's a sign of stress on the system.
1247 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1248 0 : }
1249 0 : }
1250 :
1251 0 : let checkpoint_distance =
1252 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1253 0 :
1254 0 : if self.should_roll(
1255 0 : current_size,
1256 0 : current_size,
1257 0 : checkpoint_distance,
1258 0 : self.get_last_record_lsn(),
1259 0 : self.last_freeze_at.load(),
1260 0 : open_layer.get_opened_at(),
1261 0 : ) {
1262 0 : match open_layer.info() {
1263 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1264 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1265 0 : // happens asynchronously in the background.
1266 0 : tracing::debug!(
1267 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1268 0 : );
1269 : }
1270 : InMemoryLayerInfo::Open { .. } => {
1271 : // Upgrade to a write lock and freeze the layer
1272 0 : drop(layers_guard);
1273 0 : let mut layers_guard = self.layers.write().await;
1274 0 : layers_guard
1275 0 : .try_freeze_in_memory_layer(current_lsn, &self.last_freeze_at)
1276 0 : .await;
1277 : }
1278 : }
1279 0 : self.flush_frozen_layers();
1280 0 : }
1281 510 : }
1282 :
1283 : /// Outermost timeline compaction operation; downloads needed layers.
1284 510 : pub(crate) async fn compact(
1285 510 : self: &Arc<Self>,
1286 510 : cancel: &CancellationToken,
1287 510 : flags: EnumSet<CompactFlags>,
1288 510 : ctx: &RequestContext,
1289 510 : ) -> Result<(), CompactionError> {
1290 510 : // most likely the cancellation token is from background task, but in tests it could be the
1291 510 : // request task as well.
1292 510 :
1293 510 : let prepare = async move {
1294 510 : let guard = self.compaction_lock.lock().await;
1295 :
1296 510 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1297 510 : BackgroundLoopKind::Compaction,
1298 510 : ctx,
1299 510 : )
1300 0 : .await;
1301 :
1302 510 : (guard, permit)
1303 510 : };
1304 :
1305 : // Prior to compaction, check if an open ephemeral layer should be closed: this provides
1306 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1307 : // an ephemeral layer open forever when idle.
1308 510 : self.maybe_freeze_ephemeral_layer().await;
1309 :
1310 : // this wait probably never needs any "long time spent" logging, because we already nag if
1311 : // compaction task goes over it's period (20s) which is quite often in production.
1312 510 : let (_guard, _permit) = tokio::select! {
1313 510 : tuple = prepare => { tuple },
1314 : _ = self.cancel.cancelled() => return Ok(()),
1315 : _ = cancel.cancelled() => return Ok(()),
1316 : };
1317 :
1318 510 : let last_record_lsn = self.get_last_record_lsn();
1319 510 :
1320 510 : // Last record Lsn could be zero in case the timeline was just created
1321 510 : if !last_record_lsn.is_valid() {
1322 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1323 0 : return Ok(());
1324 510 : }
1325 510 :
1326 510 : match self.get_compaction_algorithm() {
1327 0 : CompactionAlgorithm::Tiered => self.compact_tiered(cancel, ctx).await,
1328 89659 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, flags, ctx).await,
1329 : }
1330 510 : }
1331 :
1332 : /// Mutate the timeline with a [`TimelineWriter`].
1333 3977036 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1334 3977036 : TimelineWriter {
1335 3977036 : tl: self,
1336 3977036 : write_guard: self.write_lock.lock().await,
1337 : }
1338 3977036 : }
1339 :
1340 0 : pub(crate) fn activate(
1341 0 : self: &Arc<Self>,
1342 0 : parent: Arc<crate::tenant::Tenant>,
1343 0 : broker_client: BrokerClientChannel,
1344 0 : background_jobs_can_start: Option<&completion::Barrier>,
1345 0 : ctx: &RequestContext,
1346 0 : ) {
1347 0 : if self.tenant_shard_id.is_shard_zero() {
1348 0 : // Logical size is only maintained accurately on shard zero.
1349 0 : self.spawn_initial_logical_size_computation_task(ctx);
1350 0 : }
1351 0 : self.launch_wal_receiver(ctx, broker_client);
1352 0 : self.set_state(TimelineState::Active);
1353 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1354 0 : }
1355 :
1356 : /// After this function returns, there are no timeline-scoped tasks are left running.
1357 : ///
1358 : /// The preferred pattern for is:
1359 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1360 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1361 : /// go the extra mile and keep track of JoinHandles
1362 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1363 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1364 : ///
1365 : /// For legacy reasons, we still have multiple tasks spawned using
1366 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1367 : /// We refer to these as "timeline-scoped task_mgr tasks".
1368 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1369 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1370 : /// or [`task_mgr::shutdown_watcher`].
1371 : /// We want to gradually convert the code base away from these.
1372 : ///
1373 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1374 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1375 : /// ones that aren't mentioned here):
1376 : /// - [`TaskKind::TimelineDeletionWorker`]
1377 : /// - NB: also used for tenant deletion
1378 : /// - [`TaskKind::RemoteUploadTask`]`
1379 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1380 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1381 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1382 : /// - [`TaskKind::Eviction`]
1383 : /// - [`TaskKind::LayerFlushTask`]
1384 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1385 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1386 8 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1387 8 : debug_assert_current_span_has_tenant_and_timeline_id();
1388 :
1389 8 : let try_freeze_and_flush = match mode {
1390 6 : ShutdownMode::FreezeAndFlush => true,
1391 2 : ShutdownMode::Hard => false,
1392 : };
1393 :
1394 : // Regardless of whether we're going to try_freeze_and_flush
1395 : // or not, stop ingesting any more data. Walreceiver only provides
1396 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1397 : // So, only after the self.gate.close() below will we know for sure that
1398 : // no walreceiver tasks are left.
1399 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1400 : // data during the call to `self.freeze_and_flush()` below.
1401 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1402 : // which is what we'd need to properly model early shutdown of the walreceiver
1403 : // task sub-tree before the other Timeline task sub-trees.
1404 8 : let walreceiver = self.walreceiver.lock().unwrap().take();
1405 8 : tracing::debug!(
1406 0 : is_some = walreceiver.is_some(),
1407 0 : "Waiting for WalReceiverManager..."
1408 0 : );
1409 8 : if let Some(walreceiver) = walreceiver {
1410 0 : walreceiver.cancel();
1411 8 : }
1412 : // ... and inform any waiters for newer LSNs that there won't be any.
1413 8 : self.last_record_lsn.shutdown();
1414 8 :
1415 8 : if try_freeze_and_flush {
1416 : // we shut down walreceiver above, so, we won't add anything more
1417 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1418 : // to reach the disk & upload queue, then shut the upload queue and
1419 : // wait for it to drain.
1420 6 : match self.freeze_and_flush().await {
1421 : Ok(_) => {
1422 : // drain the upload queue
1423 6 : if let Some(client) = self.remote_client.as_ref() {
1424 : // if we did not wait for completion here, it might be our shutdown process
1425 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1426 : // be spawned.
1427 : //
1428 : // what is problematic is the shutting down of RemoteTimelineClient, because
1429 : // obviously it does not make sense to stop while we wait for it, but what
1430 : // about corner cases like s3 suddenly hanging up?
1431 6 : client.shutdown().await;
1432 0 : }
1433 : }
1434 0 : Err(e) => {
1435 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1436 0 : // we have some extra WAL replay to do next time the timeline starts.
1437 0 : warn!("failed to freeze and flush: {e:#}");
1438 : }
1439 : }
1440 2 : }
1441 :
1442 : // Signal any subscribers to our cancellation token to drop out
1443 8 : tracing::debug!("Cancelling CancellationToken");
1444 8 : self.cancel.cancel();
1445 :
1446 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1447 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1448 8 : if let Some(remote_client) = self.remote_client.as_ref() {
1449 8 : remote_client.stop();
1450 8 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1451 8 : // to shut down the upload queue tasks.
1452 8 : // TODO: fix that, task management should be encapsulated inside remote_client.
1453 8 : task_mgr::shutdown_tasks(
1454 8 : Some(TaskKind::RemoteUploadTask),
1455 8 : Some(self.tenant_shard_id),
1456 8 : Some(self.timeline_id),
1457 8 : )
1458 0 : .await;
1459 0 : }
1460 :
1461 : // TODO: work toward making this a no-op. See this funciton's doc comment for more context.
1462 8 : tracing::debug!("Waiting for tasks...");
1463 8 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1464 :
1465 : // Finally wait until any gate-holders are complete.
1466 : //
1467 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1468 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1469 8 : self.gate.close().await;
1470 :
1471 8 : self.metrics.shutdown();
1472 8 : }
1473 :
1474 322 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1475 322 : match (self.current_state(), new_state) {
1476 322 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1477 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1478 : }
1479 0 : (st, TimelineState::Loading) => {
1480 0 : error!("ignoring transition from {st:?} into Loading state");
1481 : }
1482 0 : (TimelineState::Broken { .. }, new_state) => {
1483 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1484 : }
1485 : (TimelineState::Stopping, TimelineState::Active) => {
1486 0 : error!("Not activating a Stopping timeline");
1487 : }
1488 320 : (_, new_state) => {
1489 320 : self.state.send_replace(new_state);
1490 320 : }
1491 : }
1492 322 : }
1493 :
1494 2 : pub(crate) fn set_broken(&self, reason: String) {
1495 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1496 2 : let broken_state = TimelineState::Broken {
1497 2 : reason,
1498 2 : backtrace: backtrace_str,
1499 2 : };
1500 2 : self.set_state(broken_state);
1501 2 :
1502 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1503 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1504 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1505 2 : self.cancel.cancel();
1506 2 : }
1507 :
1508 228361 : pub(crate) fn current_state(&self) -> TimelineState {
1509 228361 : self.state.borrow().clone()
1510 228361 : }
1511 :
1512 6 : pub(crate) fn is_broken(&self) -> bool {
1513 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1514 6 : }
1515 :
1516 227023 : pub(crate) fn is_active(&self) -> bool {
1517 227023 : self.current_state() == TimelineState::Active
1518 227023 : }
1519 :
1520 1016 : pub(crate) fn is_stopping(&self) -> bool {
1521 1016 : self.current_state() == TimelineState::Stopping
1522 1016 : }
1523 :
1524 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1525 0 : self.state.subscribe()
1526 0 : }
1527 :
1528 226807 : pub(crate) async fn wait_to_become_active(
1529 226807 : &self,
1530 226807 : _ctx: &RequestContext, // Prepare for use by cancellation
1531 226807 : ) -> Result<(), TimelineState> {
1532 226807 : let mut receiver = self.state.subscribe();
1533 226807 : loop {
1534 226807 : let current_state = receiver.borrow().clone();
1535 226807 : match current_state {
1536 : TimelineState::Loading => {
1537 0 : receiver
1538 0 : .changed()
1539 0 : .await
1540 0 : .expect("holding a reference to self");
1541 : }
1542 : TimelineState::Active { .. } => {
1543 226805 : return Ok(());
1544 : }
1545 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1546 : // There's no chance the timeline can transition back into ::Active
1547 2 : return Err(current_state);
1548 : }
1549 : }
1550 : }
1551 226807 : }
1552 :
1553 0 : pub(crate) async fn layer_map_info(&self, reset: LayerAccessStatsReset) -> LayerMapInfo {
1554 0 : let guard = self.layers.read().await;
1555 0 : let layer_map = guard.layer_map();
1556 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
1557 0 : if let Some(open_layer) = &layer_map.open_layer {
1558 0 : in_memory_layers.push(open_layer.info());
1559 0 : }
1560 0 : for frozen_layer in &layer_map.frozen_layers {
1561 0 : in_memory_layers.push(frozen_layer.info());
1562 0 : }
1563 :
1564 0 : let mut historic_layers = Vec::new();
1565 0 : for historic_layer in layer_map.iter_historic_layers() {
1566 0 : let historic_layer = guard.get_from_desc(&historic_layer);
1567 0 : historic_layers.push(historic_layer.info(reset));
1568 0 : }
1569 :
1570 0 : LayerMapInfo {
1571 0 : in_memory_layers,
1572 0 : historic_layers,
1573 0 : }
1574 0 : }
1575 :
1576 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))]
1577 : pub(crate) async fn download_layer(
1578 : &self,
1579 : layer_file_name: &str,
1580 : ) -> anyhow::Result<Option<bool>> {
1581 : let Some(layer) = self.find_layer(layer_file_name).await else {
1582 : return Ok(None);
1583 : };
1584 :
1585 : if self.remote_client.is_none() {
1586 : return Ok(Some(false));
1587 : }
1588 :
1589 : layer.download().await?;
1590 :
1591 : Ok(Some(true))
1592 : }
1593 :
1594 : /// Evict just one layer.
1595 : ///
1596 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
1597 0 : pub(crate) async fn evict_layer(&self, layer_file_name: &str) -> anyhow::Result<Option<bool>> {
1598 0 : let _gate = self
1599 0 : .gate
1600 0 : .enter()
1601 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
1602 :
1603 0 : let Some(local_layer) = self.find_layer(layer_file_name).await else {
1604 0 : return Ok(None);
1605 : };
1606 :
1607 : // curl has this by default
1608 0 : let timeout = std::time::Duration::from_secs(120);
1609 0 :
1610 0 : match local_layer.evict_and_wait(timeout).await {
1611 0 : Ok(()) => Ok(Some(true)),
1612 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
1613 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
1614 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
1615 : }
1616 0 : }
1617 :
1618 56 : fn should_roll(
1619 56 : &self,
1620 56 : layer_size: u64,
1621 56 : projected_layer_size: u64,
1622 56 : checkpoint_distance: u64,
1623 56 : projected_lsn: Lsn,
1624 56 : last_freeze_at: Lsn,
1625 56 : opened_at: Instant,
1626 56 : ) -> bool {
1627 56 : let distance = projected_lsn.widening_sub(last_freeze_at);
1628 56 :
1629 56 : // Rolling the open layer can be triggered by:
1630 56 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
1631 56 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
1632 56 : // account for how writes are distributed across shards: we expect each node to consume
1633 56 : // 1/count of the LSN on average.
1634 56 : // 2. The size of the currently open layer.
1635 56 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
1636 56 : // up and suspend activity.
1637 56 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
1638 0 : info!(
1639 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
1640 0 : projected_lsn, layer_size, distance
1641 0 : );
1642 :
1643 0 : true
1644 56 : } else if projected_layer_size >= checkpoint_distance {
1645 0 : info!(
1646 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
1647 0 : projected_lsn, layer_size, projected_layer_size
1648 0 : );
1649 :
1650 0 : true
1651 56 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
1652 0 : info!(
1653 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
1654 0 : projected_lsn,
1655 0 : layer_size,
1656 0 : opened_at.elapsed()
1657 0 : );
1658 :
1659 0 : true
1660 : } else {
1661 56 : false
1662 : }
1663 56 : }
1664 : }
1665 :
1666 : /// Number of times we will compute partition within a checkpoint distance.
1667 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
1668 :
1669 : // Private functions
1670 : impl Timeline {
1671 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
1672 0 : let tenant_conf = self.tenant_conf.load();
1673 0 : tenant_conf
1674 0 : .tenant_conf
1675 0 : .lazy_slru_download
1676 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
1677 0 : }
1678 :
1679 1408 : fn get_checkpoint_distance(&self) -> u64 {
1680 1408 : let tenant_conf = self.tenant_conf.load();
1681 1408 : tenant_conf
1682 1408 : .tenant_conf
1683 1408 : .checkpoint_distance
1684 1408 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
1685 1408 : }
1686 :
1687 56 : fn get_checkpoint_timeout(&self) -> Duration {
1688 56 : let tenant_conf = self.tenant_conf.load();
1689 56 : tenant_conf
1690 56 : .tenant_conf
1691 56 : .checkpoint_timeout
1692 56 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
1693 56 : }
1694 :
1695 604 : fn get_compaction_target_size(&self) -> u64 {
1696 604 : let tenant_conf = self.tenant_conf.load();
1697 604 : tenant_conf
1698 604 : .tenant_conf
1699 604 : .compaction_target_size
1700 604 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
1701 604 : }
1702 :
1703 510 : fn get_compaction_threshold(&self) -> usize {
1704 510 : let tenant_conf = self.tenant_conf.load();
1705 510 : tenant_conf
1706 510 : .tenant_conf
1707 510 : .compaction_threshold
1708 510 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
1709 510 : }
1710 :
1711 2 : fn get_image_creation_threshold(&self) -> usize {
1712 2 : let tenant_conf = self.tenant_conf.load();
1713 2 : tenant_conf
1714 2 : .tenant_conf
1715 2 : .image_creation_threshold
1716 2 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
1717 2 : }
1718 :
1719 510 : fn get_compaction_algorithm(&self) -> CompactionAlgorithm {
1720 510 : let tenant_conf = &self.tenant_conf.load();
1721 510 : tenant_conf
1722 510 : .tenant_conf
1723 510 : .compaction_algorithm
1724 510 : .unwrap_or(self.conf.default_tenant_conf.compaction_algorithm)
1725 510 : }
1726 :
1727 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
1728 0 : let tenant_conf = self.tenant_conf.load();
1729 0 : tenant_conf
1730 0 : .tenant_conf
1731 0 : .eviction_policy
1732 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
1733 0 : }
1734 :
1735 320 : fn get_evictions_low_residence_duration_metric_threshold(
1736 320 : tenant_conf: &TenantConfOpt,
1737 320 : default_tenant_conf: &TenantConf,
1738 320 : ) -> Duration {
1739 320 : tenant_conf
1740 320 : .evictions_low_residence_duration_metric_threshold
1741 320 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
1742 320 : }
1743 :
1744 522 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
1745 522 : let tenant_conf = self.tenant_conf.load();
1746 522 : tenant_conf
1747 522 : .tenant_conf
1748 522 : .image_layer_creation_check_threshold
1749 522 : .unwrap_or(
1750 522 : self.conf
1751 522 : .default_tenant_conf
1752 522 : .image_layer_creation_check_threshold,
1753 522 : )
1754 522 : }
1755 :
1756 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
1757 0 : // NB: Most tenant conf options are read by background loops, so,
1758 0 : // changes will automatically be picked up.
1759 0 :
1760 0 : // The threshold is embedded in the metric. So, we need to update it.
1761 0 : {
1762 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
1763 0 : new_conf,
1764 0 : &self.conf.default_tenant_conf,
1765 0 : );
1766 0 :
1767 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
1768 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
1769 0 :
1770 0 : let timeline_id_str = self.timeline_id.to_string();
1771 0 : self.metrics
1772 0 : .evictions_with_low_residence_duration
1773 0 : .write()
1774 0 : .unwrap()
1775 0 : .change_threshold(
1776 0 : &tenant_id_str,
1777 0 : &shard_id_str,
1778 0 : &timeline_id_str,
1779 0 : new_threshold,
1780 0 : );
1781 0 : }
1782 0 : }
1783 :
1784 : /// Open a Timeline handle.
1785 : ///
1786 : /// Loads the metadata for the timeline into memory, but not the layer map.
1787 : #[allow(clippy::too_many_arguments)]
1788 320 : pub(super) fn new(
1789 320 : conf: &'static PageServerConf,
1790 320 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
1791 320 : metadata: &TimelineMetadata,
1792 320 : ancestor: Option<Arc<Timeline>>,
1793 320 : timeline_id: TimelineId,
1794 320 : tenant_shard_id: TenantShardId,
1795 320 : generation: Generation,
1796 320 : shard_identity: ShardIdentity,
1797 320 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
1798 320 : resources: TimelineResources,
1799 320 : pg_version: u32,
1800 320 : state: TimelineState,
1801 320 : cancel: CancellationToken,
1802 320 : ) -> Arc<Self> {
1803 320 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
1804 320 : let (state, _) = watch::channel(state);
1805 320 :
1806 320 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
1807 320 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
1808 320 :
1809 320 : let evictions_low_residence_duration_metric_threshold = {
1810 320 : let loaded_tenant_conf = tenant_conf.load();
1811 320 : Self::get_evictions_low_residence_duration_metric_threshold(
1812 320 : &loaded_tenant_conf.tenant_conf,
1813 320 : &conf.default_tenant_conf,
1814 320 : )
1815 320 : };
1816 320 :
1817 320 : Arc::new_cyclic(|myself| {
1818 320 : let mut result = Timeline {
1819 320 : conf,
1820 320 : tenant_conf,
1821 320 : myself: myself.clone(),
1822 320 : timeline_id,
1823 320 : tenant_shard_id,
1824 320 : generation,
1825 320 : shard_identity,
1826 320 : pg_version,
1827 320 : layers: Default::default(),
1828 320 :
1829 320 : walredo_mgr,
1830 320 : walreceiver: Mutex::new(None),
1831 320 :
1832 320 : remote_client: resources.remote_client.map(Arc::new),
1833 320 :
1834 320 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
1835 320 : last_record_lsn: SeqWait::new(RecordLsn {
1836 320 : last: disk_consistent_lsn,
1837 320 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
1838 320 : }),
1839 320 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
1840 320 :
1841 320 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
1842 320 : last_freeze_ts: RwLock::new(Instant::now()),
1843 320 :
1844 320 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
1845 320 :
1846 320 : ancestor_timeline: ancestor,
1847 320 : ancestor_lsn: metadata.ancestor_lsn(),
1848 320 :
1849 320 : metrics: TimelineMetrics::new(
1850 320 : &tenant_shard_id,
1851 320 : &timeline_id,
1852 320 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
1853 320 : "mtime",
1854 320 : evictions_low_residence_duration_metric_threshold,
1855 320 : ),
1856 320 : ),
1857 320 :
1858 320 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
1859 320 : &tenant_shard_id,
1860 320 : &timeline_id,
1861 320 : ),
1862 320 :
1863 2240 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
1864 320 :
1865 320 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
1866 320 :
1867 320 : layer_flush_start_tx,
1868 320 : layer_flush_done_tx,
1869 320 :
1870 320 : write_lock: tokio::sync::Mutex::new(None),
1871 320 :
1872 320 : gc_info: std::sync::RwLock::new(GcInfo {
1873 320 : retain_lsns: Vec::new(),
1874 320 : horizon_cutoff: Lsn(0),
1875 320 : pitr_cutoff: Lsn(0),
1876 320 : }),
1877 320 :
1878 320 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
1879 320 : initdb_lsn: metadata.initdb_lsn(),
1880 320 :
1881 320 : current_logical_size: if disk_consistent_lsn.is_valid() {
1882 : // we're creating timeline data with some layer files existing locally,
1883 : // need to recalculate timeline's logical size based on data in the layers.
1884 220 : LogicalSize::deferred_initial(disk_consistent_lsn)
1885 : } else {
1886 : // we're creating timeline data without any layers existing locally,
1887 : // initial logical size is 0.
1888 100 : LogicalSize::empty_initial()
1889 : },
1890 320 : partitioning: tokio::sync::Mutex::new((KeyPartitioning::new(), Lsn(0))),
1891 320 : repartition_threshold: 0,
1892 320 : last_image_layer_creation_check_at: AtomicLsn::new(0),
1893 320 :
1894 320 : last_received_wal: Mutex::new(None),
1895 320 : rel_size_cache: RwLock::new(HashMap::new()),
1896 320 :
1897 320 : download_all_remote_layers_task_info: RwLock::new(None),
1898 320 :
1899 320 : state,
1900 320 :
1901 320 : eviction_task_timeline_state: tokio::sync::Mutex::new(
1902 320 : EvictionTaskTimelineState::default(),
1903 320 : ),
1904 320 : delete_progress: Arc::new(tokio::sync::Mutex::new(DeleteTimelineFlow::default())),
1905 320 :
1906 320 : cancel,
1907 320 : gate: Gate::default(),
1908 320 :
1909 320 : compaction_lock: tokio::sync::Mutex::default(),
1910 320 : gc_lock: tokio::sync::Mutex::default(),
1911 320 :
1912 320 : timeline_get_throttle: resources.timeline_get_throttle,
1913 320 :
1914 320 : aux_files: tokio::sync::Mutex::new(AuxFilesState {
1915 320 : dir: None,
1916 320 : n_deltas: 0,
1917 320 : }),
1918 320 : };
1919 320 : result.repartition_threshold =
1920 320 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
1921 320 :
1922 320 : result
1923 320 : .metrics
1924 320 : .last_record_gauge
1925 320 : .set(disk_consistent_lsn.0 as i64);
1926 320 : result
1927 320 : })
1928 320 : }
1929 :
1930 410 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
1931 410 : let Ok(guard) = self.gate.enter() else {
1932 0 : info!("cannot start flush loop when the timeline gate has already been closed");
1933 0 : return;
1934 : };
1935 410 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
1936 410 : match *flush_loop_state {
1937 316 : FlushLoopState::NotStarted => (),
1938 : FlushLoopState::Running { .. } => {
1939 94 : info!(
1940 94 : "skipping attempt to start flush_loop twice {}/{}",
1941 94 : self.tenant_shard_id, self.timeline_id
1942 94 : );
1943 94 : return;
1944 : }
1945 : FlushLoopState::Exited => {
1946 0 : warn!(
1947 0 : "ignoring attempt to restart exited flush_loop {}/{}",
1948 0 : self.tenant_shard_id, self.timeline_id
1949 0 : );
1950 0 : return;
1951 : }
1952 : }
1953 :
1954 316 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
1955 316 : let self_clone = Arc::clone(self);
1956 316 :
1957 316 : debug!("spawning flush loop");
1958 316 : *flush_loop_state = FlushLoopState::Running {
1959 316 : #[cfg(test)]
1960 316 : expect_initdb_optimization: false,
1961 316 : #[cfg(test)]
1962 316 : initdb_optimization_count: 0,
1963 316 : };
1964 316 : task_mgr::spawn(
1965 316 : task_mgr::BACKGROUND_RUNTIME.handle(),
1966 316 : task_mgr::TaskKind::LayerFlushTask,
1967 316 : Some(self.tenant_shard_id),
1968 316 : Some(self.timeline_id),
1969 316 : "layer flush task",
1970 : false,
1971 316 : async move {
1972 316 : let _guard = guard;
1973 316 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
1974 61914 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
1975 8 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
1976 8 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{ ..}));
1977 8 : *flush_loop_state = FlushLoopState::Exited;
1978 8 : Ok(())
1979 8 : }
1980 316 : .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))
1981 : );
1982 410 : }
1983 :
1984 : /// Creates and starts the wal receiver.
1985 : ///
1986 : /// This function is expected to be called at most once per Timeline's lifecycle
1987 : /// when the timeline is activated.
1988 0 : fn launch_wal_receiver(
1989 0 : self: &Arc<Self>,
1990 0 : ctx: &RequestContext,
1991 0 : broker_client: BrokerClientChannel,
1992 0 : ) {
1993 0 : info!(
1994 0 : "launching WAL receiver for timeline {} of tenant {}",
1995 0 : self.timeline_id, self.tenant_shard_id
1996 0 : );
1997 :
1998 0 : let tenant_conf = self.tenant_conf.load();
1999 0 : let wal_connect_timeout = tenant_conf
2000 0 : .tenant_conf
2001 0 : .walreceiver_connect_timeout
2002 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2003 0 : let lagging_wal_timeout = tenant_conf
2004 0 : .tenant_conf
2005 0 : .lagging_wal_timeout
2006 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2007 0 : let max_lsn_wal_lag = tenant_conf
2008 0 : .tenant_conf
2009 0 : .max_lsn_wal_lag
2010 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2011 0 :
2012 0 : let mut guard = self.walreceiver.lock().unwrap();
2013 0 : assert!(
2014 0 : guard.is_none(),
2015 0 : "multiple launches / re-launches of WAL receiver are not supported"
2016 : );
2017 0 : *guard = Some(WalReceiver::start(
2018 0 : Arc::clone(self),
2019 0 : WalReceiverConf {
2020 0 : wal_connect_timeout,
2021 0 : lagging_wal_timeout,
2022 0 : max_lsn_wal_lag,
2023 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2024 0 : availability_zone: self.conf.availability_zone.clone(),
2025 0 : ingest_batch_size: self.conf.ingest_batch_size,
2026 0 : },
2027 0 : broker_client,
2028 0 : ctx,
2029 0 : ));
2030 0 : }
2031 :
2032 : /// Initialize with an empty layer map. Used when creating a new timeline.
2033 314 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2034 314 : let mut layers = self.layers.try_write().expect(
2035 314 : "in the context where we call this function, no other task has access to the object",
2036 314 : );
2037 314 : layers.initialize_empty(Lsn(start_lsn.0));
2038 314 : }
2039 :
2040 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2041 : /// files.
2042 6 : pub(super) async fn load_layer_map(
2043 6 : &self,
2044 6 : disk_consistent_lsn: Lsn,
2045 6 : index_part: Option<IndexPart>,
2046 6 : ) -> anyhow::Result<()> {
2047 : use init::{Decision::*, Discovered, DismissedLayer};
2048 : use LayerFileName::*;
2049 :
2050 6 : let mut guard = self.layers.write().await;
2051 :
2052 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2053 6 :
2054 6 : // Scan timeline directory and create ImageFileName and DeltaFilename
2055 6 : // structs representing all files on disk
2056 6 : let timeline_path = self
2057 6 : .conf
2058 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2059 6 : let conf = self.conf;
2060 6 : let span = tracing::Span::current();
2061 6 :
2062 6 : // Copy to move into the task we're about to spawn
2063 6 : let generation = self.generation;
2064 6 : let shard = self.get_shard_index();
2065 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2066 :
2067 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2068 6 : move || {
2069 6 : let _g = span.entered();
2070 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2071 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2072 6 : let mut unrecognized_files = Vec::new();
2073 6 :
2074 6 : let mut path = timeline_path;
2075 :
2076 22 : for discovered in discovered {
2077 16 : let (name, kind) = match discovered {
2078 16 : Discovered::Layer(file_name, file_size) => {
2079 16 : discovered_layers.push((file_name, file_size));
2080 16 : continue;
2081 : }
2082 : Discovered::Metadata => {
2083 0 : warn!("found legacy metadata file, these should have been removed in load_tenant_config");
2084 0 : continue;
2085 : }
2086 : Discovered::IgnoredBackup => {
2087 0 : continue;
2088 : }
2089 0 : Discovered::Unknown(file_name) => {
2090 0 : // we will later error if there are any
2091 0 : unrecognized_files.push(file_name);
2092 0 : continue;
2093 : }
2094 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2095 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2096 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2097 : };
2098 0 : path.push(Utf8Path::new(&name));
2099 0 : init::cleanup(&path, kind)?;
2100 0 : path.pop();
2101 : }
2102 :
2103 6 : if !unrecognized_files.is_empty() {
2104 : // assume that if there are any there are many many.
2105 0 : let n = unrecognized_files.len();
2106 0 : let first = &unrecognized_files[..n.min(10)];
2107 0 : anyhow::bail!(
2108 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2109 0 : );
2110 6 : }
2111 6 :
2112 6 : let decided = init::reconcile(
2113 6 : discovered_layers,
2114 6 : index_part.as_ref(),
2115 6 : disk_consistent_lsn,
2116 6 : generation,
2117 6 : shard,
2118 6 : );
2119 6 :
2120 6 : let mut loaded_layers = Vec::new();
2121 6 : let mut needs_cleanup = Vec::new();
2122 6 : let mut total_physical_size = 0;
2123 :
2124 22 : for (name, decision) in decided {
2125 16 : let decision = match decision {
2126 0 : Ok(UseRemote { local, remote }) => {
2127 0 : // Remote is authoritative, but we may still choose to retain
2128 0 : // the local file if the contents appear to match
2129 0 : if local.file_size() == remote.file_size() {
2130 : // Use the local file, but take the remote metadata so that we pick up
2131 : // the correct generation.
2132 0 : UseLocal(remote)
2133 : } else {
2134 0 : path.push(name.file_name());
2135 0 : init::cleanup_local_file_for_remote(&path, &local, &remote)?;
2136 0 : path.pop();
2137 0 : UseRemote { local, remote }
2138 : }
2139 : }
2140 16 : Ok(decision) => decision,
2141 0 : Err(DismissedLayer::Future { local }) => {
2142 0 : if local.is_some() {
2143 0 : path.push(name.file_name());
2144 0 : init::cleanup_future_layer(&path, &name, disk_consistent_lsn)?;
2145 0 : path.pop();
2146 0 : }
2147 0 : needs_cleanup.push(name);
2148 0 : continue;
2149 : }
2150 0 : Err(DismissedLayer::LocalOnly(local)) => {
2151 0 : path.push(name.file_name());
2152 0 : init::cleanup_local_only_file(&path, &name, &local)?;
2153 0 : path.pop();
2154 0 : // this file never existed remotely, we will have to do rework
2155 0 : continue;
2156 : }
2157 : };
2158 :
2159 16 : match &name {
2160 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2161 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2162 : }
2163 :
2164 16 : tracing::debug!(layer=%name, ?decision, "applied");
2165 :
2166 16 : let layer = match decision {
2167 16 : UseLocal(m) => {
2168 16 : total_physical_size += m.file_size();
2169 16 : Layer::for_resident(conf, &this, name, m).drop_eviction_guard()
2170 : }
2171 0 : Evicted(remote) | UseRemote { remote, .. } => {
2172 0 : Layer::for_evicted(conf, &this, name, remote)
2173 : }
2174 : };
2175 :
2176 16 : loaded_layers.push(layer);
2177 : }
2178 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2179 6 : }
2180 6 : })
2181 6 : .await
2182 6 : .map_err(anyhow::Error::new)
2183 6 : .and_then(|x| x)?;
2184 :
2185 6 : let num_layers = loaded_layers.len();
2186 6 :
2187 6 : guard.initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2188 :
2189 6 : if let Some(rtc) = self.remote_client.as_ref() {
2190 6 : rtc.schedule_layer_file_deletion(&needs_cleanup)?;
2191 6 : rtc.schedule_index_upload_for_file_changes()?;
2192 : // This barrier orders above DELETEs before any later operations.
2193 : // This is critical because code executing after the barrier might
2194 : // create again objects with the same key that we just scheduled for deletion.
2195 : // For example, if we just scheduled deletion of an image layer "from the future",
2196 : // later compaction might run again and re-create the same image layer.
2197 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2198 : // "same" here means same key range and LSN.
2199 : //
2200 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2201 : // the upload queue may execute the PUT first, then the DELETE.
2202 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2203 : //
2204 : // 1. a future image layer is created and uploaded
2205 : // 2. ps restart
2206 : // 3. the future layer from (1) is deleted during load layer map
2207 : // 4. image layer is re-created and uploaded
2208 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2209 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2210 : //
2211 : // See https://github.com/neondatabase/neon/issues/5878
2212 : //
2213 : // NB: generation numbers naturally protect against this because they disambiguate
2214 : // (1) and (4)
2215 6 : rtc.schedule_barrier()?;
2216 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2217 : // on retry.
2218 0 : }
2219 :
2220 6 : info!(
2221 6 : "loaded layer map with {} layers at {}, total physical size: {}",
2222 6 : num_layers, disk_consistent_lsn, total_physical_size
2223 6 : );
2224 :
2225 6 : timer.stop_and_record();
2226 6 : Ok(())
2227 6 : }
2228 :
2229 : /// Retrieve current logical size of the timeline.
2230 : ///
2231 : /// The size could be lagging behind the actual number, in case
2232 : /// the initial size calculation has not been run (gets triggered on the first size access).
2233 : ///
2234 : /// return size and boolean flag that shows if the size is exact
2235 0 : pub(crate) fn get_current_logical_size(
2236 0 : self: &Arc<Self>,
2237 0 : priority: GetLogicalSizePriority,
2238 0 : ctx: &RequestContext,
2239 0 : ) -> logical_size::CurrentLogicalSize {
2240 0 : if !self.tenant_shard_id.is_shard_zero() {
2241 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2242 : // when HTTP API is serving a GET for timeline zero, return zero
2243 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2244 0 : }
2245 0 :
2246 0 : let current_size = self.current_logical_size.current_size();
2247 0 : debug!("Current size: {current_size:?}");
2248 :
2249 0 : match (current_size.accuracy(), priority) {
2250 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2251 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2252 0 : // background task will eventually deliver an exact value, we're in no rush
2253 0 : }
2254 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2255 : // background task is not ready, but user is asking for it now;
2256 : // => make the background task skip the line
2257 : // (The alternative would be to calculate the size here, but,
2258 : // it can actually take a long time if the user has a lot of rels.
2259 : // And we'll inevitable need it again; So, let the background task do the work.)
2260 0 : match self
2261 0 : .current_logical_size
2262 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2263 0 : .get()
2264 : {
2265 0 : Some(cancel) => cancel.cancel(),
2266 : None => {
2267 0 : let state = self.current_state();
2268 0 : if matches!(
2269 0 : state,
2270 : TimelineState::Broken { .. } | TimelineState::Stopping
2271 0 : ) {
2272 0 :
2273 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2274 0 : // Don't make noise.
2275 0 : } else {
2276 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2277 : }
2278 : }
2279 : };
2280 : }
2281 : }
2282 :
2283 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2284 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2285 0 : let first = self
2286 0 : .current_logical_size
2287 0 : .did_return_approximate_to_walreceiver
2288 0 : .compare_exchange(
2289 0 : false,
2290 0 : true,
2291 0 : AtomicOrdering::Relaxed,
2292 0 : AtomicOrdering::Relaxed,
2293 0 : )
2294 0 : .is_ok();
2295 0 : if first {
2296 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2297 0 : }
2298 0 : }
2299 0 : }
2300 :
2301 0 : current_size
2302 0 : }
2303 :
2304 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2305 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2306 : // nothing to do for freshly created timelines;
2307 0 : assert_eq!(
2308 0 : self.current_logical_size.current_size().accuracy(),
2309 0 : logical_size::Accuracy::Exact,
2310 0 : );
2311 0 : self.current_logical_size.initialized.add_permits(1);
2312 0 : return;
2313 : };
2314 :
2315 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2316 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2317 0 : self.current_logical_size
2318 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2319 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2320 0 :
2321 0 : let self_clone = Arc::clone(self);
2322 0 : let background_ctx = ctx.detached_child(
2323 0 : TaskKind::InitialLogicalSizeCalculation,
2324 0 : DownloadBehavior::Download,
2325 0 : );
2326 0 : task_mgr::spawn(
2327 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2328 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2329 0 : Some(self.tenant_shard_id),
2330 0 : Some(self.timeline_id),
2331 0 : "initial size calculation",
2332 : false,
2333 : // NB: don't log errors here, task_mgr will do that.
2334 0 : async move {
2335 0 : let cancel = task_mgr::shutdown_token();
2336 0 : self_clone
2337 0 : .initial_logical_size_calculation_task(
2338 0 : initial_part_end,
2339 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2340 0 : cancel,
2341 0 : background_ctx,
2342 0 : )
2343 0 : .await;
2344 0 : Ok(())
2345 0 : }
2346 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)),
2347 : );
2348 0 : }
2349 :
2350 0 : async fn initial_logical_size_calculation_task(
2351 0 : self: Arc<Self>,
2352 0 : initial_part_end: Lsn,
2353 0 : skip_concurrency_limiter: CancellationToken,
2354 0 : cancel: CancellationToken,
2355 0 : background_ctx: RequestContext,
2356 0 : ) {
2357 0 : scopeguard::defer! {
2358 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2359 0 : self.current_logical_size.initialized.add_permits(1);
2360 0 : }
2361 :
2362 : enum BackgroundCalculationError {
2363 : Cancelled,
2364 : Other(anyhow::Error),
2365 : }
2366 :
2367 0 : let try_once = |attempt: usize| {
2368 0 : let background_ctx = &background_ctx;
2369 0 : let self_ref = &self;
2370 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2371 0 : async move {
2372 0 : let cancel = task_mgr::shutdown_token();
2373 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2374 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2375 0 : background_ctx,
2376 0 : );
2377 :
2378 : use crate::metrics::initial_logical_size::StartCircumstances;
2379 0 : let (_maybe_permit, circumstances) = tokio::select! {
2380 0 : permit = wait_for_permit => {
2381 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2382 : }
2383 : _ = self_ref.cancel.cancelled() => {
2384 : return Err(BackgroundCalculationError::Cancelled);
2385 : }
2386 : _ = cancel.cancelled() => {
2387 : return Err(BackgroundCalculationError::Cancelled);
2388 : },
2389 : () = skip_concurrency_limiter.cancelled() => {
2390 : // Some action that is part of a end user interaction requested logical size
2391 : // => break out of the rate limit
2392 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2393 : // but then again what happens if they cancel; also, we should just be using
2394 : // one runtime across the entire process, so, let's leave this for now.
2395 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2396 : }
2397 : };
2398 :
2399 0 : let metrics_guard = if attempt == 1 {
2400 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2401 : } else {
2402 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2403 : };
2404 :
2405 0 : match self_ref
2406 0 : .logical_size_calculation_task(
2407 0 : initial_part_end,
2408 0 : LogicalSizeCalculationCause::Initial,
2409 0 : background_ctx,
2410 0 : )
2411 0 : .await
2412 : {
2413 0 : Ok(calculated_size) => Ok((calculated_size, metrics_guard)),
2414 : Err(CalculateLogicalSizeError::Cancelled) => {
2415 0 : Err(BackgroundCalculationError::Cancelled)
2416 : }
2417 0 : Err(CalculateLogicalSizeError::Other(err)) => {
2418 0 : if let Some(PageReconstructError::AncestorStopping(_)) =
2419 0 : err.root_cause().downcast_ref()
2420 : {
2421 0 : Err(BackgroundCalculationError::Cancelled)
2422 : } else {
2423 0 : Err(BackgroundCalculationError::Other(err))
2424 : }
2425 : }
2426 : }
2427 0 : }
2428 0 : };
2429 :
2430 0 : let retrying = async {
2431 0 : let mut attempt = 0;
2432 0 : loop {
2433 0 : attempt += 1;
2434 0 :
2435 0 : match try_once(attempt).await {
2436 0 : Ok(res) => return ControlFlow::Continue(res),
2437 0 : Err(BackgroundCalculationError::Cancelled) => return ControlFlow::Break(()),
2438 0 : Err(BackgroundCalculationError::Other(e)) => {
2439 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2440 : // exponential back-off doesn't make sense at these long intervals;
2441 : // use fixed retry interval with generous jitter instead
2442 0 : let sleep_duration = Duration::from_secs(
2443 0 : u64::try_from(
2444 0 : // 1hour base
2445 0 : (60_i64 * 60_i64)
2446 0 : // 10min jitter
2447 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2448 0 : )
2449 0 : .expect("10min < 1hour"),
2450 0 : );
2451 0 : tokio::time::sleep(sleep_duration).await;
2452 : }
2453 : }
2454 : }
2455 0 : };
2456 :
2457 0 : let (calculated_size, metrics_guard) = tokio::select! {
2458 0 : res = retrying => {
2459 : match res {
2460 : ControlFlow::Continue(calculated_size) => calculated_size,
2461 : ControlFlow::Break(()) => return,
2462 : }
2463 : }
2464 : _ = cancel.cancelled() => {
2465 : return;
2466 : }
2467 : };
2468 :
2469 : // we cannot query current_logical_size.current_size() to know the current
2470 : // *negative* value, only truncated to u64.
2471 0 : let added = self
2472 0 : .current_logical_size
2473 0 : .size_added_after_initial
2474 0 : .load(AtomicOrdering::Relaxed);
2475 0 :
2476 0 : let sum = calculated_size.saturating_add_signed(added);
2477 0 :
2478 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2479 0 : self.metrics.current_logical_size_gauge.set(sum);
2480 0 :
2481 0 : self.current_logical_size
2482 0 : .initial_logical_size
2483 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
2484 0 : .ok()
2485 0 : .expect("only this task sets it");
2486 0 : }
2487 :
2488 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
2489 0 : self: &Arc<Self>,
2490 0 : lsn: Lsn,
2491 0 : cause: LogicalSizeCalculationCause,
2492 0 : ctx: RequestContext,
2493 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
2494 0 : let (sender, receiver) = oneshot::channel();
2495 0 : let self_clone = Arc::clone(self);
2496 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
2497 0 : // we should stop the size calculation work and return an error.
2498 0 : // That would require restructuring this function's API to
2499 0 : // return the result directly, instead of a Receiver for the result.
2500 0 : let ctx = ctx.detached_child(
2501 0 : TaskKind::OndemandLogicalSizeCalculation,
2502 0 : DownloadBehavior::Download,
2503 0 : );
2504 0 : task_mgr::spawn(
2505 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2506 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
2507 0 : Some(self.tenant_shard_id),
2508 0 : Some(self.timeline_id),
2509 0 : "ondemand logical size calculation",
2510 0 : false,
2511 0 : async move {
2512 0 : let res = self_clone
2513 0 : .logical_size_calculation_task(lsn, cause, &ctx)
2514 0 : .await;
2515 0 : let _ = sender.send(res).ok();
2516 0 : Ok(()) // Receiver is responsible for handling errors
2517 0 : }
2518 0 : .in_current_span(),
2519 0 : );
2520 0 : receiver
2521 0 : }
2522 :
2523 : /// # Cancel-Safety
2524 : ///
2525 : /// This method is cancellation-safe.
2526 0 : #[instrument(skip_all)]
2527 : async fn logical_size_calculation_task(
2528 : self: &Arc<Self>,
2529 : lsn: Lsn,
2530 : cause: LogicalSizeCalculationCause,
2531 : ctx: &RequestContext,
2532 : ) -> Result<u64, CalculateLogicalSizeError> {
2533 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
2534 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
2535 : // accurate relation sizes, and they do not emit consumption metrics.
2536 : debug_assert!(self.tenant_shard_id.is_shard_zero());
2537 :
2538 : let guard = self
2539 : .gate
2540 : .enter()
2541 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
2542 :
2543 : let self_calculation = Arc::clone(self);
2544 :
2545 0 : let mut calculation = pin!(async {
2546 0 : let ctx = ctx.attached_child();
2547 0 : self_calculation
2548 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
2549 0 : .await
2550 0 : });
2551 :
2552 0 : tokio::select! {
2553 0 : res = &mut calculation => { res }
2554 : _ = self.cancel.cancelled() => {
2555 0 : debug!("cancelling logical size calculation for timeline shutdown");
2556 : calculation.await
2557 : }
2558 : }
2559 : }
2560 :
2561 : /// Calculate the logical size of the database at the latest LSN.
2562 : ///
2563 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
2564 : /// especially if we need to download remote layers.
2565 : ///
2566 : /// # Cancel-Safety
2567 : ///
2568 : /// This method is cancellation-safe.
2569 0 : async fn calculate_logical_size(
2570 0 : &self,
2571 0 : up_to_lsn: Lsn,
2572 0 : cause: LogicalSizeCalculationCause,
2573 0 : _guard: &GateGuard,
2574 0 : ctx: &RequestContext,
2575 0 : ) -> Result<u64, CalculateLogicalSizeError> {
2576 0 : info!(
2577 0 : "Calculating logical size for timeline {} at {}",
2578 0 : self.timeline_id, up_to_lsn
2579 0 : );
2580 :
2581 0 : pausable_failpoint!("timeline-calculate-logical-size-pause");
2582 :
2583 : // See if we've already done the work for initial size calculation.
2584 : // This is a short-cut for timelines that are mostly unused.
2585 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
2586 0 : return Ok(size);
2587 0 : }
2588 0 : let storage_time_metrics = match cause {
2589 : LogicalSizeCalculationCause::Initial
2590 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
2591 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
2592 : LogicalSizeCalculationCause::EvictionTaskImitation => {
2593 0 : &self.metrics.imitate_logical_size_histo
2594 : }
2595 : };
2596 0 : let timer = storage_time_metrics.start_timer();
2597 0 : let logical_size = self
2598 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
2599 0 : .await?;
2600 0 : debug!("calculated logical size: {logical_size}");
2601 0 : timer.stop_and_record();
2602 0 : Ok(logical_size)
2603 0 : }
2604 :
2605 : /// Update current logical size, adding `delta' to the old value.
2606 270570 : fn update_current_logical_size(&self, delta: i64) {
2607 270570 : let logical_size = &self.current_logical_size;
2608 270570 : logical_size.increment_size(delta);
2609 270570 :
2610 270570 : // Also set the value in the prometheus gauge. Note that
2611 270570 : // there is a race condition here: if this is is called by two
2612 270570 : // threads concurrently, the prometheus gauge might be set to
2613 270570 : // one value while current_logical_size is set to the
2614 270570 : // other.
2615 270570 : match logical_size.current_size() {
2616 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
2617 270570 : .metrics
2618 270570 : .current_logical_size_gauge
2619 270570 : .set(new_current_size.into()),
2620 0 : CurrentLogicalSize::Approximate(_) => {
2621 0 : // don't update the gauge yet, this allows us not to update the gauge back and
2622 0 : // forth between the initial size calculation task.
2623 0 : }
2624 : }
2625 270570 : }
2626 :
2627 2544 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
2628 2544 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
2629 2544 : let aux_metric =
2630 2544 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
2631 2544 :
2632 2544 : let sum_of_entries = self
2633 2544 : .directory_metrics
2634 2544 : .iter()
2635 17808 : .map(|v| v.load(AtomicOrdering::Relaxed))
2636 2544 : .sum();
2637 2544 : // Set a high general threshold and a lower threshold for the auxiliary files,
2638 2544 : // as we can have large numbers of relations in the db directory.
2639 2544 : const SUM_THRESHOLD: u64 = 5000;
2640 2544 : const AUX_THRESHOLD: u64 = 1000;
2641 2544 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
2642 0 : self.metrics
2643 0 : .directory_entries_count_gauge
2644 0 : .set(sum_of_entries);
2645 2544 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
2646 0 : metric.set(sum_of_entries);
2647 2544 : }
2648 2544 : }
2649 :
2650 0 : async fn find_layer(&self, layer_file_name: &str) -> Option<Layer> {
2651 0 : let guard = self.layers.read().await;
2652 0 : for historic_layer in guard.layer_map().iter_historic_layers() {
2653 0 : let historic_layer_name = historic_layer.filename().file_name();
2654 0 : if layer_file_name == historic_layer_name {
2655 0 : return Some(guard.get_from_desc(&historic_layer));
2656 0 : }
2657 : }
2658 :
2659 0 : None
2660 0 : }
2661 :
2662 : /// The timeline heatmap is a hint to secondary locations from the primary location,
2663 : /// indicating which layers are currently on-disk on the primary.
2664 : ///
2665 : /// None is returned if the Timeline is in a state where uploading a heatmap
2666 : /// doesn't make sense, such as shutting down or initializing. The caller
2667 : /// should treat this as a cue to simply skip doing any heatmap uploading
2668 : /// for this timeline.
2669 0 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
2670 : // no point in heatmaps without remote client
2671 0 : let _remote_client = self.remote_client.as_ref()?;
2672 :
2673 0 : if !self.is_active() {
2674 0 : return None;
2675 0 : }
2676 :
2677 0 : let guard = self.layers.read().await;
2678 :
2679 0 : let resident = guard.likely_resident_layers().map(|layer| {
2680 0 : let last_activity_ts = layer.access_stats().latest_activity_or_now();
2681 0 :
2682 0 : HeatMapLayer::new(
2683 0 : layer.layer_desc().filename(),
2684 0 : layer.metadata().into(),
2685 0 : last_activity_ts,
2686 0 : )
2687 0 : });
2688 0 :
2689 0 : let layers = resident.collect();
2690 0 :
2691 0 : Some(HeatMapTimeline::new(self.timeline_id, layers))
2692 0 : }
2693 : }
2694 :
2695 : type TraversalId = String;
2696 :
2697 : trait TraversalLayerExt {
2698 : fn traversal_id(&self) -> TraversalId;
2699 : }
2700 :
2701 : impl TraversalLayerExt for Layer {
2702 146 : fn traversal_id(&self) -> TraversalId {
2703 146 : self.local_path().to_string()
2704 146 : }
2705 : }
2706 :
2707 : impl TraversalLayerExt for Arc<InMemoryLayer> {
2708 4 : fn traversal_id(&self) -> TraversalId {
2709 4 : format!("timeline {} in-memory {self}", self.get_timeline_id())
2710 4 : }
2711 : }
2712 :
2713 : impl Timeline {
2714 : ///
2715 : /// Get a handle to a Layer for reading.
2716 : ///
2717 : /// The returned Layer might be from an ancestor timeline, if the
2718 : /// segment hasn't been updated on this timeline yet.
2719 : ///
2720 : /// This function takes the current timeline's locked LayerMap as an argument,
2721 : /// so callers can avoid potential race conditions.
2722 : ///
2723 : /// # Cancel-Safety
2724 : ///
2725 : /// This method is cancellation-safe.
2726 502833 : async fn get_reconstruct_data(
2727 502833 : &self,
2728 502833 : key: Key,
2729 502833 : request_lsn: Lsn,
2730 502833 : reconstruct_state: &mut ValueReconstructState,
2731 502833 : ctx: &RequestContext,
2732 502833 : ) -> Result<Vec<TraversalPathItem>, PageReconstructError> {
2733 502833 : // Start from the current timeline.
2734 502833 : let mut timeline_owned;
2735 502833 : let mut timeline = self;
2736 502833 :
2737 502833 : let mut read_count = scopeguard::guard(0, |cnt| {
2738 502833 : crate::metrics::READ_NUM_FS_LAYERS.observe(cnt as f64)
2739 502833 : });
2740 502833 :
2741 502833 : // For debugging purposes, collect the path of layers that we traversed
2742 502833 : // through. It's included in the error message if we fail to find the key.
2743 502833 : let mut traversal_path = Vec::<TraversalPathItem>::new();
2744 :
2745 502833 : let cached_lsn = if let Some((cached_lsn, _)) = &reconstruct_state.img {
2746 0 : *cached_lsn
2747 : } else {
2748 502833 : Lsn(0)
2749 : };
2750 :
2751 : // 'prev_lsn' tracks the last LSN that we were at in our search. It's used
2752 : // to check that each iteration make some progress, to break infinite
2753 : // looping if something goes wrong.
2754 502833 : let mut prev_lsn = None;
2755 502833 :
2756 502833 : let mut result = ValueReconstructResult::Continue;
2757 502833 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
2758 :
2759 1356959 : 'outer: loop {
2760 1356959 : if self.cancel.is_cancelled() {
2761 0 : return Err(PageReconstructError::Cancelled);
2762 1356959 : }
2763 1356959 :
2764 1356959 : // The function should have updated 'state'
2765 1356959 : //info!("CALLED for {} at {}: {:?} with {} records, cached {}", key, cont_lsn, result, reconstruct_state.records.len(), cached_lsn);
2766 1356959 : match result {
2767 502723 : ValueReconstructResult::Complete => return Ok(traversal_path),
2768 : ValueReconstructResult::Continue => {
2769 : // If we reached an earlier cached page image, we're done.
2770 854230 : if cont_lsn == cached_lsn + 1 {
2771 0 : MATERIALIZED_PAGE_CACHE_HIT.inc_by(1);
2772 0 : return Ok(traversal_path);
2773 854230 : }
2774 854230 : if let Some(prev) = prev_lsn {
2775 124594 : if prev <= cont_lsn {
2776 : // Didn't make any progress in last iteration. Error out to avoid
2777 : // getting stuck in the loop.
2778 102 : return Err(layer_traversal_error(format!(
2779 102 : "could not find layer with more data for key {} at LSN {}, request LSN {}, ancestor {}",
2780 102 : key,
2781 102 : Lsn(cont_lsn.0 - 1),
2782 102 : request_lsn,
2783 102 : timeline.ancestor_lsn
2784 102 : ), traversal_path));
2785 124492 : }
2786 729636 : }
2787 854128 : prev_lsn = Some(cont_lsn);
2788 : }
2789 : ValueReconstructResult::Missing => {
2790 : return Err(layer_traversal_error(
2791 6 : if cfg!(test) {
2792 6 : format!(
2793 6 : "could not find data for key {} (shard {:?}) at LSN {}, for request at LSN {}\n{}",
2794 6 : key, self.shard_identity.get_shard_number(&key), cont_lsn, request_lsn, std::backtrace::Backtrace::force_capture(),
2795 6 : )
2796 : } else {
2797 0 : format!(
2798 0 : "could not find data for key {} (shard {:?}) at LSN {}, for request at LSN {}",
2799 0 : key, self.shard_identity.get_shard_number(&key), cont_lsn, request_lsn
2800 0 : )
2801 : },
2802 6 : traversal_path,
2803 : ));
2804 : }
2805 : }
2806 :
2807 : // Recurse into ancestor if needed
2808 854128 : if is_inherited_key(key) && Lsn(cont_lsn.0 - 1) <= timeline.ancestor_lsn {
2809 226805 : trace!(
2810 0 : "going into ancestor {}, cont_lsn is {}",
2811 0 : timeline.ancestor_lsn,
2812 0 : cont_lsn
2813 0 : );
2814 :
2815 226805 : timeline_owned = timeline.get_ready_ancestor_timeline(ctx).await?;
2816 226803 : timeline = &*timeline_owned;
2817 226803 : prev_lsn = None;
2818 226803 : continue 'outer;
2819 627323 : }
2820 :
2821 627323 : let guard = timeline.layers.read().await;
2822 627323 : let layers = guard.layer_map();
2823 :
2824 : // Check the open and frozen in-memory layers first, in order from newest
2825 : // to oldest.
2826 627323 : if let Some(open_layer) = &layers.open_layer {
2827 556003 : let start_lsn = open_layer.get_lsn_range().start;
2828 556003 : if cont_lsn > start_lsn {
2829 : //info!("CHECKING for {} at {} on open layer {}", key, cont_lsn, open_layer.filename().display());
2830 : // Get all the data needed to reconstruct the page version from this layer.
2831 : // But if we have an older cached page image, no need to go past that.
2832 502139 : let lsn_floor = max(cached_lsn + 1, start_lsn);
2833 502139 :
2834 502139 : let open_layer = open_layer.clone();
2835 502139 : drop(guard);
2836 502139 :
2837 502139 : result = match open_layer
2838 502139 : .get_value_reconstruct_data(
2839 502139 : key,
2840 502139 : lsn_floor..cont_lsn,
2841 502139 : reconstruct_state,
2842 502139 : ctx,
2843 502139 : )
2844 7232 : .await
2845 : {
2846 502139 : Ok(result) => result,
2847 0 : Err(e) => return Err(PageReconstructError::from(e)),
2848 : };
2849 502139 : cont_lsn = lsn_floor;
2850 502139 : // metrics: open_layer does not count as fs access, so we are not updating `read_count`
2851 502139 : traversal_path.push((
2852 502139 : result,
2853 502139 : cont_lsn,
2854 502139 : Box::new(move || open_layer.traversal_id()),
2855 502139 : ));
2856 502139 : continue 'outer;
2857 53864 : }
2858 71320 : }
2859 125184 : for frozen_layer in layers.frozen_layers.iter().rev() {
2860 1316 : let start_lsn = frozen_layer.get_lsn_range().start;
2861 1316 : if cont_lsn > start_lsn {
2862 : //info!("CHECKING for {} at {} on frozen layer {}", key, cont_lsn, frozen_layer.filename().display());
2863 1316 : let lsn_floor = max(cached_lsn + 1, start_lsn);
2864 1316 :
2865 1316 : let frozen_layer = frozen_layer.clone();
2866 1316 : drop(guard);
2867 1316 :
2868 1316 : result = match frozen_layer
2869 1316 : .get_value_reconstruct_data(
2870 1316 : key,
2871 1316 : lsn_floor..cont_lsn,
2872 1316 : reconstruct_state,
2873 1316 : ctx,
2874 1316 : )
2875 0 : .await
2876 : {
2877 1316 : Ok(result) => result,
2878 0 : Err(e) => return Err(PageReconstructError::from(e)),
2879 : };
2880 1316 : cont_lsn = lsn_floor;
2881 1316 : // metrics: open_layer does not count as fs access, so we are not updating `read_count`
2882 1316 : traversal_path.push((
2883 1316 : result,
2884 1316 : cont_lsn,
2885 1316 : Box::new(move || frozen_layer.traversal_id()),
2886 1316 : ));
2887 1316 : continue 'outer;
2888 0 : }
2889 : }
2890 :
2891 123868 : if let Some(SearchResult { lsn_floor, layer }) = layers.search(key, cont_lsn) {
2892 123764 : let layer = guard.get_from_desc(&layer);
2893 123764 : drop(guard);
2894 123764 :
2895 123764 : // Get all the data needed to reconstruct the page version from this layer.
2896 123764 : // But if we have an older cached page image, no need to go past that.
2897 123764 : let lsn_floor = max(cached_lsn + 1, lsn_floor);
2898 123764 : result = match layer
2899 123764 : .get_value_reconstruct_data(key, lsn_floor..cont_lsn, reconstruct_state, ctx)
2900 23436 : .await
2901 : {
2902 123764 : Ok(result) => result,
2903 0 : Err(e) => return Err(PageReconstructError::from(e)),
2904 : };
2905 123764 : cont_lsn = lsn_floor;
2906 123764 : *read_count += 1;
2907 123764 : traversal_path.push((
2908 123764 : result,
2909 123764 : cont_lsn,
2910 123764 : Box::new({
2911 123764 : let layer = layer.to_owned();
2912 123764 : move || layer.traversal_id()
2913 123764 : }),
2914 123764 : ));
2915 123764 : continue 'outer;
2916 104 : } else if timeline.ancestor_timeline.is_some() {
2917 : // Nothing on this timeline. Traverse to parent
2918 102 : result = ValueReconstructResult::Continue;
2919 102 : cont_lsn = Lsn(timeline.ancestor_lsn.0 + 1);
2920 102 : continue 'outer;
2921 : } else {
2922 : // Nothing found
2923 2 : result = ValueReconstructResult::Missing;
2924 2 : continue 'outer;
2925 : }
2926 : }
2927 502833 : }
2928 :
2929 : /// Get the data needed to reconstruct all keys in the provided keyspace
2930 : ///
2931 : /// The algorithm is as follows:
2932 : /// 1. While some keys are still not done and there's a timeline to visit:
2933 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
2934 : /// 2.1: Build the fringe for the current keyspace
2935 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
2936 : /// intersects
2937 : /// 2.3. Pop the timeline from the fringe
2938 : /// 2.4. If the fringe is empty, go back to 1
2939 12 : async fn get_vectored_reconstruct_data(
2940 12 : &self,
2941 12 : mut keyspace: KeySpace,
2942 12 : request_lsn: Lsn,
2943 12 : reconstruct_state: &mut ValuesReconstructState,
2944 12 : ctx: &RequestContext,
2945 12 : ) -> Result<(), GetVectoredError> {
2946 12 : let mut timeline_owned: Arc<Timeline>;
2947 12 : let mut timeline = self;
2948 12 :
2949 12 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
2950 :
2951 : loop {
2952 14 : if self.cancel.is_cancelled() {
2953 0 : return Err(GetVectoredError::Cancelled);
2954 14 : }
2955 :
2956 14 : let completed = Self::get_vectored_reconstruct_data_timeline(
2957 14 : timeline,
2958 14 : keyspace.clone(),
2959 14 : cont_lsn,
2960 14 : reconstruct_state,
2961 14 : &self.cancel,
2962 14 : ctx,
2963 14 : )
2964 41 : .await?;
2965 :
2966 14 : keyspace.remove_overlapping_with(&completed);
2967 14 : if keyspace.total_size() == 0 || timeline.ancestor_timeline.is_none() {
2968 12 : break;
2969 2 : }
2970 2 :
2971 2 : cont_lsn = Lsn(timeline.ancestor_lsn.0 + 1);
2972 2 : timeline_owned = timeline
2973 2 : .get_ready_ancestor_timeline(ctx)
2974 0 : .await
2975 2 : .map_err(GetVectoredError::GetReadyAncestorError)?;
2976 2 : timeline = &*timeline_owned;
2977 : }
2978 :
2979 12 : if keyspace.total_size() != 0 {
2980 0 : return Err(GetVectoredError::MissingKey(keyspace.start().unwrap()));
2981 12 : }
2982 12 :
2983 12 : Ok(())
2984 12 : }
2985 :
2986 : /// Collect the reconstruct data for a ketspace from the specified timeline.
2987 : ///
2988 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
2989 : /// the current keyspace. The current keyspace of the search at any given timeline
2990 : /// is the original keyspace minus all the keys that have been completed minus
2991 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
2992 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
2993 : ///
2994 : /// This is basically a depth-first search visitor implementation where a vertex
2995 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
2996 : ///
2997 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
2998 : /// and get all the required reconstruct data from the layer in one go.
2999 14 : async fn get_vectored_reconstruct_data_timeline(
3000 14 : timeline: &Timeline,
3001 14 : keyspace: KeySpace,
3002 14 : mut cont_lsn: Lsn,
3003 14 : reconstruct_state: &mut ValuesReconstructState,
3004 14 : cancel: &CancellationToken,
3005 14 : ctx: &RequestContext,
3006 14 : ) -> Result<KeySpace, GetVectoredError> {
3007 14 : let mut unmapped_keyspace = keyspace.clone();
3008 14 : let mut fringe = LayerFringe::new();
3009 14 :
3010 14 : let mut completed_keyspace = KeySpace::default();
3011 :
3012 32 : loop {
3013 32 : if cancel.is_cancelled() {
3014 0 : return Err(GetVectoredError::Cancelled);
3015 32 : }
3016 32 :
3017 32 : let keys_done_last_step = reconstruct_state.consume_done_keys();
3018 32 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3019 32 : completed_keyspace.merge(&keys_done_last_step);
3020 :
3021 32 : let guard = timeline.layers.read().await;
3022 32 : let layers = guard.layer_map();
3023 32 :
3024 32 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3025 0 : let start_lsn = l.get_lsn_range().start;
3026 0 : cont_lsn > start_lsn
3027 32 : });
3028 32 :
3029 32 : match in_memory_layer {
3030 0 : Some(l) => {
3031 0 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3032 0 : fringe.update(
3033 0 : ReadableLayer::InMemoryLayer(l),
3034 0 : unmapped_keyspace.clone(),
3035 0 : lsn_range,
3036 0 : );
3037 0 : }
3038 : None => {
3039 32 : for range in unmapped_keyspace.ranges.iter() {
3040 18 : let results = layers.range_search(range.clone(), cont_lsn);
3041 18 :
3042 18 : results
3043 18 : .found
3044 18 : .into_iter()
3045 18 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3046 18 : (
3047 18 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3048 18 : keyspace_accum.to_keyspace(),
3049 18 : lsn_floor..cont_lsn,
3050 18 : )
3051 18 : })
3052 18 : .for_each(|(layer, keyspace, lsn_range)| {
3053 18 : fringe.update(layer, keyspace, lsn_range)
3054 18 : });
3055 18 : }
3056 : }
3057 : }
3058 :
3059 : // It's safe to drop the layer map lock after planning the next round of reads.
3060 : // The fringe keeps readable handles for the layers which are safe to read even
3061 : // if layers were compacted or flushed.
3062 : //
3063 : // The more interesting consideration is: "Why is the read algorithm still correct
3064 : // if the layer map changes while it is operating?". Doing a vectored read on a
3065 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3066 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3067 : // range at *a particular point in time*. It is fine for the answer to be different
3068 : // at two different time points.
3069 32 : drop(guard);
3070 :
3071 32 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3072 18 : let next_cont_lsn = lsn_range.start;
3073 18 : layer_to_read
3074 18 : .get_values_reconstruct_data(
3075 18 : keyspace_to_read.clone(),
3076 18 : lsn_range,
3077 18 : reconstruct_state,
3078 18 : ctx,
3079 18 : )
3080 41 : .await?;
3081 :
3082 18 : unmapped_keyspace = keyspace_to_read;
3083 18 : cont_lsn = next_cont_lsn;
3084 : } else {
3085 14 : break;
3086 14 : }
3087 14 : }
3088 14 :
3089 14 : Ok(completed_keyspace)
3090 14 : }
3091 :
3092 : /// # Cancel-safety
3093 : ///
3094 : /// This method is cancellation-safe.
3095 502833 : async fn lookup_cached_page(
3096 502833 : &self,
3097 502833 : key: &Key,
3098 502833 : lsn: Lsn,
3099 502833 : ctx: &RequestContext,
3100 502833 : ) -> Option<(Lsn, Bytes)> {
3101 502833 : let cache = page_cache::get();
3102 :
3103 : // FIXME: It's pointless to check the cache for things that are not 8kB pages.
3104 : // We should look at the key to determine if it's a cacheable object
3105 502833 : let (lsn, read_guard) = cache
3106 502833 : .lookup_materialized_page(self.tenant_shard_id, self.timeline_id, key, lsn, ctx)
3107 502833 : .await?;
3108 0 : let img = Bytes::from(read_guard.to_vec());
3109 0 : Some((lsn, img))
3110 502833 : }
3111 :
3112 226807 : async fn get_ready_ancestor_timeline(
3113 226807 : &self,
3114 226807 : ctx: &RequestContext,
3115 226807 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3116 226807 : let ancestor = match self.get_ancestor_timeline() {
3117 226807 : Ok(timeline) => timeline,
3118 0 : Err(e) => return Err(GetReadyAncestorError::from(e)),
3119 : };
3120 :
3121 : // It's possible that the ancestor timeline isn't active yet, or
3122 : // is active but hasn't yet caught up to the branch point. Wait
3123 : // for it.
3124 : //
3125 : // This cannot happen while the pageserver is running normally,
3126 : // because you cannot create a branch from a point that isn't
3127 : // present in the pageserver yet. However, we don't wait for the
3128 : // branch point to be uploaded to cloud storage before creating
3129 : // a branch. I.e., the branch LSN need not be remote consistent
3130 : // for the branching operation to succeed.
3131 : //
3132 : // Hence, if we try to load a tenant in such a state where
3133 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3134 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3135 : // then we will need to wait for the ancestor timeline to
3136 : // re-stream WAL up to branch_lsn before we access it.
3137 : //
3138 : // How can a tenant get in such a state?
3139 : // - ungraceful pageserver process exit
3140 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3141 : //
3142 : // NB: this could be avoided by requiring
3143 : // branch_lsn >= remote_consistent_lsn
3144 : // during branch creation.
3145 226807 : match ancestor.wait_to_become_active(ctx).await {
3146 226805 : Ok(()) => {}
3147 : Err(TimelineState::Stopping) => {
3148 0 : return Err(GetReadyAncestorError::AncestorStopping(
3149 0 : ancestor.timeline_id,
3150 0 : ));
3151 : }
3152 2 : Err(state) => {
3153 2 : return Err(GetReadyAncestorError::Other(anyhow::anyhow!(
3154 2 : "Timeline {} will not become active. Current state: {:?}",
3155 2 : ancestor.timeline_id,
3156 2 : &state,
3157 2 : )));
3158 : }
3159 : }
3160 226805 : ancestor
3161 226805 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3162 0 : .await
3163 226805 : .map_err(|e| match e {
3164 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3165 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3166 0 : e @ WaitLsnError::BadState => GetReadyAncestorError::Other(anyhow::anyhow!(e)),
3167 226805 : })?;
3168 :
3169 226805 : Ok(ancestor)
3170 226807 : }
3171 :
3172 226807 : fn get_ancestor_timeline(&self) -> anyhow::Result<Arc<Timeline>> {
3173 226807 : let ancestor = self.ancestor_timeline.as_ref().with_context(|| {
3174 0 : format!(
3175 0 : "Ancestor is missing. Timeline id: {} Ancestor id {:?}",
3176 0 : self.timeline_id,
3177 0 : self.get_ancestor_timeline_id(),
3178 0 : )
3179 226807 : })?;
3180 226807 : Ok(Arc::clone(ancestor))
3181 226807 : }
3182 :
3183 5452 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3184 5452 : &self.shard_identity
3185 5452 : }
3186 :
3187 : ///
3188 : /// Get a handle to the latest layer for appending.
3189 : ///
3190 3648020 : async fn get_layer_for_write(&self, lsn: Lsn) -> anyhow::Result<Arc<InMemoryLayer>> {
3191 3648020 : let mut guard = self.layers.write().await;
3192 3648020 : let layer = guard
3193 3648020 : .get_layer_for_write(
3194 3648020 : lsn,
3195 3648020 : self.get_last_record_lsn(),
3196 3648020 : self.conf,
3197 3648020 : self.timeline_id,
3198 3648020 : self.tenant_shard_id,
3199 3648020 : )
3200 455 : .await?;
3201 3648020 : Ok(layer)
3202 3648020 : }
3203 :
3204 4122936 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3205 4122936 : assert!(new_lsn.is_aligned());
3206 :
3207 4122936 : self.metrics.last_record_gauge.set(new_lsn.0 as i64);
3208 4122936 : self.last_record_lsn.advance(new_lsn);
3209 4122936 : }
3210 :
3211 : /// Whether there was a layer to freeze or not, return the value of get_last_record_lsn
3212 : /// before we attempted the freeze: this guarantees that ingested data is frozen up to this lsn (inclusive).
3213 698 : async fn freeze_inmem_layer(&self, write_lock_held: bool) -> Lsn {
3214 : // Freeze the current open in-memory layer. It will be written to disk on next
3215 : // iteration.
3216 :
3217 698 : let _write_guard = if write_lock_held {
3218 0 : None
3219 : } else {
3220 698 : Some(self.write_lock.lock().await)
3221 : };
3222 :
3223 698 : let to_lsn = self.get_last_record_lsn();
3224 698 : self.freeze_inmem_layer_at(to_lsn).await;
3225 698 : to_lsn
3226 698 : }
3227 :
3228 698 : async fn freeze_inmem_layer_at(&self, at: Lsn) {
3229 698 : let mut guard = self.layers.write().await;
3230 698 : guard
3231 698 : .try_freeze_in_memory_layer(at, &self.last_freeze_at)
3232 2 : .await;
3233 698 : }
3234 :
3235 : /// Layer flusher task's main loop.
3236 316 : async fn flush_loop(
3237 316 : self: &Arc<Self>,
3238 316 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3239 316 : ctx: &RequestContext,
3240 316 : ) {
3241 316 : info!("started flush loop");
3242 698 : loop {
3243 1630 : tokio::select! {
3244 1630 : _ = self.cancel.cancelled() => {
3245 1630 : info!("shutting down layer flush task due to Timeline::cancel");
3246 1630 : break;
3247 1630 : },
3248 1630 : _ = layer_flush_start_rx.changed() => {}
3249 1630 : }
3250 698 : trace!("waking up");
3251 698 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3252 698 :
3253 698 : // The highest LSN to which we flushed in the loop over frozen layers
3254 698 : let mut flushed_to_lsn = Lsn(0);
3255 :
3256 698 : let result = loop {
3257 1390 : if self.cancel.is_cancelled() {
3258 0 : info!("dropping out of flush loop for timeline shutdown");
3259 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3260 : // anyone waiting on that will respect self.cancel as well: they will stop
3261 : // waiting at the same time we as drop out of this loop.
3262 0 : return;
3263 1390 : }
3264 1390 :
3265 1390 : let timer = self.metrics.flush_time_histo.start_timer();
3266 :
3267 1390 : let layer_to_flush = {
3268 1390 : let guard = self.layers.read().await;
3269 1390 : guard.layer_map().frozen_layers.front().cloned()
3270 : // drop 'layers' lock to allow concurrent reads and writes
3271 : };
3272 1390 : let Some(layer_to_flush) = layer_to_flush else {
3273 698 : break Ok(());
3274 : };
3275 61298 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3276 692 : Ok(this_layer_to_lsn) => {
3277 692 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3278 692 : }
3279 : Err(FlushLayerError::Cancelled) => {
3280 0 : info!("dropping out of flush loop for timeline shutdown");
3281 0 : return;
3282 : }
3283 0 : err @ Err(
3284 : FlushLayerError::Other(_) | FlushLayerError::CreateImageLayersError(_),
3285 : ) => {
3286 0 : error!("could not flush frozen layer: {err:?}");
3287 0 : break err.map(|_| ());
3288 : }
3289 : }
3290 692 : timer.stop_and_record();
3291 : };
3292 :
3293 : // Unsharded tenants should never advance their LSN beyond the end of the
3294 : // highest layer they write: such gaps between layer data and the frozen LSN
3295 : // are only legal on sharded tenants.
3296 698 : debug_assert!(
3297 698 : self.shard_identity.count.count() > 1
3298 698 : || flushed_to_lsn >= frozen_to_lsn
3299 6 : || !flushed_to_lsn.is_valid()
3300 : );
3301 :
3302 698 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3303 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3304 : // to us via layer_flush_start_rx, then advance it here.
3305 : //
3306 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3307 : // never encounter a gap in the wal.
3308 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3309 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3310 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3311 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3312 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3313 0 : }
3314 0 : }
3315 698 : }
3316 :
3317 : // Notify any listeners that we're done
3318 698 : let _ = self
3319 698 : .layer_flush_done_tx
3320 698 : .send_replace((flush_counter, result));
3321 : }
3322 8 : }
3323 :
3324 : /// Request the flush loop to write out all frozen layers up to `to_lsn` as Delta L0 files to disk.
3325 : /// The caller is responsible for the freezing, e.g., [`Self::freeze_inmem_layer`].
3326 : ///
3327 : /// `last_record_lsn` may be higher than the highest LSN of a frozen layer: if this is the case,
3328 : /// it means no data will be written between the top of the highest frozen layer and to_lsn,
3329 : /// e.g. because this tenant shard has ingested up to to_lsn and not written any data locally for that part of the WAL.
3330 698 : async fn flush_frozen_layers_and_wait(&self, last_record_lsn: Lsn) -> anyhow::Result<()> {
3331 698 : let mut rx = self.layer_flush_done_tx.subscribe();
3332 698 :
3333 698 : // Increment the flush cycle counter and wake up the flush task.
3334 698 : // Remember the new value, so that when we listen for the flush
3335 698 : // to finish, we know when the flush that we initiated has
3336 698 : // finished, instead of some other flush that was started earlier.
3337 698 : let mut my_flush_request = 0;
3338 698 :
3339 698 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3340 698 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3341 0 : anyhow::bail!("cannot flush frozen layers when flush_loop is not running, state is {flush_loop_state:?}")
3342 698 : }
3343 698 :
3344 698 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3345 698 : my_flush_request = *counter + 1;
3346 698 : *counter = my_flush_request;
3347 698 : *lsn = std::cmp::max(last_record_lsn, *lsn);
3348 698 : });
3349 :
3350 1394 : loop {
3351 1394 : {
3352 1394 : let (last_result_counter, last_result) = &*rx.borrow();
3353 1394 : if *last_result_counter >= my_flush_request {
3354 698 : if let Err(_err) = last_result {
3355 : // We already logged the original error in
3356 : // flush_loop. We cannot propagate it to the caller
3357 : // here, because it might not be Cloneable
3358 0 : anyhow::bail!(
3359 0 : "Could not flush frozen layer. Request id: {}",
3360 0 : my_flush_request
3361 0 : );
3362 : } else {
3363 698 : return Ok(());
3364 : }
3365 696 : }
3366 696 : }
3367 696 : trace!("waiting for flush to complete");
3368 1392 : tokio::select! {
3369 696 : rx_e = rx.changed() => {
3370 : rx_e?;
3371 : },
3372 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3373 : // the notification from [`flush_loop`] that it completed.
3374 : _ = self.cancel.cancelled() => {
3375 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3376 : return Ok(())
3377 : }
3378 : };
3379 696 : trace!("done")
3380 : }
3381 698 : }
3382 :
3383 0 : fn flush_frozen_layers(&self) {
3384 0 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3385 0 : *counter += 1;
3386 0 :
3387 0 : *lsn = std::cmp::max(*lsn, Lsn(self.last_freeze_at.load().0 - 1));
3388 0 : });
3389 0 : }
3390 :
3391 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3392 : ///
3393 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3394 1384 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3395 : async fn flush_frozen_layer(
3396 : self: &Arc<Self>,
3397 : frozen_layer: Arc<InMemoryLayer>,
3398 : ctx: &RequestContext,
3399 : ) -> Result<Lsn, FlushLayerError> {
3400 : debug_assert_current_span_has_tenant_and_timeline_id();
3401 :
3402 : // As a special case, when we have just imported an image into the repository,
3403 : // instead of writing out a L0 delta layer, we directly write out image layer
3404 : // files instead. This is possible as long as *all* the data imported into the
3405 : // repository have the same LSN.
3406 : let lsn_range = frozen_layer.get_lsn_range();
3407 : let (layers_to_upload, delta_layer_to_add) =
3408 : if lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1) {
3409 : #[cfg(test)]
3410 : match &mut *self.flush_loop_state.lock().unwrap() {
3411 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3412 : panic!("flush loop not running")
3413 : }
3414 : FlushLoopState::Running {
3415 : initdb_optimization_count,
3416 : ..
3417 : } => {
3418 : *initdb_optimization_count += 1;
3419 : }
3420 : }
3421 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3422 : // require downloading anything during initial import.
3423 : let (partitioning, _lsn) = self
3424 : .repartition(
3425 : self.initdb_lsn,
3426 : self.get_compaction_target_size(),
3427 : EnumSet::empty(),
3428 : ctx,
3429 : )
3430 : .await?;
3431 :
3432 : if self.cancel.is_cancelled() {
3433 : return Err(FlushLayerError::Cancelled);
3434 : }
3435 :
3436 : // For image layers, we add them immediately into the layer map.
3437 : (
3438 : self.create_image_layers(&partitioning, self.initdb_lsn, true, ctx)
3439 : .await?,
3440 : None,
3441 : )
3442 : } else {
3443 : #[cfg(test)]
3444 : match &mut *self.flush_loop_state.lock().unwrap() {
3445 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3446 : panic!("flush loop not running")
3447 : }
3448 : FlushLoopState::Running {
3449 : expect_initdb_optimization,
3450 : ..
3451 : } => {
3452 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3453 : }
3454 : }
3455 : // Normal case, write out a L0 delta layer file.
3456 : // `create_delta_layer` will not modify the layer map.
3457 : // We will remove frozen layer and add delta layer in one atomic operation later.
3458 : let layer = self.create_delta_layer(&frozen_layer, ctx).await?;
3459 : (
3460 : // FIXME: even though we have a single image and single delta layer assumption
3461 : // we push them to vec
3462 : vec![layer.clone()],
3463 : Some(layer),
3464 : )
3465 : };
3466 :
3467 692 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
3468 :
3469 : if self.cancel.is_cancelled() {
3470 : return Err(FlushLayerError::Cancelled);
3471 : }
3472 :
3473 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
3474 :
3475 : // The new on-disk layers are now in the layer map. We can remove the
3476 : // in-memory layer from the map now. The flushed layer is stored in
3477 : // the mapping in `create_delta_layer`.
3478 : {
3479 : let mut guard = self.layers.write().await;
3480 :
3481 : if self.cancel.is_cancelled() {
3482 : return Err(FlushLayerError::Cancelled);
3483 : }
3484 :
3485 : guard.finish_flush_l0_layer(delta_layer_to_add.as_ref(), &frozen_layer, &self.metrics);
3486 :
3487 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
3488 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
3489 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)?;
3490 : }
3491 : // release lock on 'layers'
3492 : };
3493 :
3494 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
3495 : // a compaction can delete the file and then it won't be available for uploads any more.
3496 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
3497 : // race situation.
3498 : // See https://github.com/neondatabase/neon/issues/4526
3499 692 : pausable_failpoint!("flush-frozen-pausable");
3500 :
3501 : // This failpoint is used by another test case `test_pageserver_recovery`.
3502 0 : fail_point!("flush-frozen-exit");
3503 :
3504 : Ok(Lsn(lsn_range.end.0 - 1))
3505 : }
3506 :
3507 : /// Return true if the value changed
3508 : ///
3509 : /// This function must only be used from the layer flush task, and may not be called concurrently.
3510 692 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
3511 692 : // We do a simple load/store cycle: that's why this function isn't safe for concurrent use.
3512 692 : let old_value = self.disk_consistent_lsn.load();
3513 692 : if new_value != old_value {
3514 692 : assert!(new_value >= old_value);
3515 692 : self.disk_consistent_lsn.store(new_value);
3516 692 : true
3517 : } else {
3518 0 : false
3519 : }
3520 692 : }
3521 :
3522 : /// Update metadata file
3523 692 : fn schedule_uploads(
3524 692 : &self,
3525 692 : disk_consistent_lsn: Lsn,
3526 692 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
3527 692 : ) -> anyhow::Result<TimelineMetadata> {
3528 692 : // We can only save a valid 'prev_record_lsn' value on disk if we
3529 692 : // flushed *all* in-memory changes to disk. We only track
3530 692 : // 'prev_record_lsn' in memory for the latest processed record, so we
3531 692 : // don't remember what the correct value that corresponds to some old
3532 692 : // LSN is. But if we flush everything, then the value corresponding
3533 692 : // current 'last_record_lsn' is correct and we can store it on disk.
3534 692 : let RecordLsn {
3535 692 : last: last_record_lsn,
3536 692 : prev: prev_record_lsn,
3537 692 : } = self.last_record_lsn.load();
3538 692 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
3539 692 : Some(prev_record_lsn)
3540 : } else {
3541 0 : None
3542 : };
3543 :
3544 692 : let ancestor_timeline_id = self
3545 692 : .ancestor_timeline
3546 692 : .as_ref()
3547 692 : .map(|ancestor| ancestor.timeline_id);
3548 692 :
3549 692 : let metadata = TimelineMetadata::new(
3550 692 : disk_consistent_lsn,
3551 692 : ondisk_prev_record_lsn,
3552 692 : ancestor_timeline_id,
3553 692 : self.ancestor_lsn,
3554 692 : *self.latest_gc_cutoff_lsn.read(),
3555 692 : self.initdb_lsn,
3556 692 : self.pg_version,
3557 692 : );
3558 692 :
3559 692 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
3560 0 : "{}",
3561 0 : x.unwrap()
3562 692 : ));
3563 :
3564 692 : if let Some(remote_client) = &self.remote_client {
3565 1398 : for layer in layers_to_upload {
3566 706 : remote_client.schedule_layer_file_upload(layer)?;
3567 : }
3568 692 : remote_client.schedule_index_upload_for_metadata_update(&metadata)?;
3569 0 : }
3570 :
3571 692 : Ok(metadata)
3572 692 : }
3573 :
3574 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
3575 0 : if let Some(remote_client) = &self.remote_client {
3576 0 : remote_client
3577 0 : .preserve_initdb_archive(
3578 0 : &self.tenant_shard_id.tenant_id,
3579 0 : &self.timeline_id,
3580 0 : &self.cancel,
3581 0 : )
3582 0 : .await?;
3583 : } else {
3584 0 : bail!("No remote storage configured, but was asked to backup the initdb archive for {} / {}", self.tenant_shard_id.tenant_id, self.timeline_id);
3585 : }
3586 0 : Ok(())
3587 0 : }
3588 :
3589 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
3590 : // in layer map immediately. The caller is responsible to put it into the layer map.
3591 598 : async fn create_delta_layer(
3592 598 : self: &Arc<Self>,
3593 598 : frozen_layer: &Arc<InMemoryLayer>,
3594 598 : ctx: &RequestContext,
3595 598 : ) -> anyhow::Result<ResidentLayer> {
3596 598 : let self_clone = Arc::clone(self);
3597 598 : let frozen_layer = Arc::clone(frozen_layer);
3598 598 : let ctx = ctx.attached_child();
3599 598 : let work = async move {
3600 83340 : let new_delta = frozen_layer.write_to_disk(&self_clone, &ctx).await?;
3601 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
3602 : // We just need to fsync the directory in which these inodes are linked,
3603 : // which we know to be the timeline directory.
3604 : //
3605 : // We use fatal_err() below because the after write_to_disk returns with success,
3606 : // the in-memory state of the filesystem already has the layer file in its final place,
3607 : // and subsequent pageserver code could think it's durable while it really isn't.
3608 598 : let timeline_dir = VirtualFile::open(
3609 598 : &self_clone
3610 598 : .conf
3611 598 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
3612 598 : )
3613 299 : .await
3614 598 : .fatal_err("VirtualFile::open for timeline dir fsync");
3615 598 : timeline_dir
3616 598 : .sync_all()
3617 299 : .await
3618 598 : .fatal_err("VirtualFile::sync_all timeline dir");
3619 598 : anyhow::Ok(new_delta)
3620 598 : };
3621 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
3622 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
3623 : use crate::virtual_file::io_engine::IoEngine;
3624 598 : match crate::virtual_file::io_engine::get() {
3625 0 : IoEngine::NotSet => panic!("io engine not set"),
3626 : IoEngine::StdFs => {
3627 299 : let span = tracing::info_span!("blocking");
3628 299 : tokio::task::spawn_blocking({
3629 299 : move || Handle::current().block_on(work.instrument(span))
3630 299 : })
3631 299 : .await
3632 299 : .context("spawn_blocking")
3633 299 : .and_then(|x| x)
3634 : }
3635 : #[cfg(target_os = "linux")]
3636 58411 : IoEngine::TokioEpollUring => work.await,
3637 : }
3638 598 : }
3639 :
3640 604 : async fn repartition(
3641 604 : &self,
3642 604 : lsn: Lsn,
3643 604 : partition_size: u64,
3644 604 : flags: EnumSet<CompactFlags>,
3645 604 : ctx: &RequestContext,
3646 604 : ) -> anyhow::Result<(KeyPartitioning, Lsn)> {
3647 604 : let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
3648 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
3649 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
3650 : // and hence before the compaction task starts.
3651 0 : anyhow::bail!("repartition() called concurrently, this should not happen");
3652 : };
3653 604 : if lsn < partitioning_guard.1 {
3654 0 : anyhow::bail!("repartition() called with LSN going backwards, this should not happen");
3655 604 : }
3656 604 :
3657 604 : let distance = lsn.0 - partitioning_guard.1 .0;
3658 604 : if partitioning_guard.1 != Lsn(0)
3659 408 : && distance <= self.repartition_threshold
3660 408 : && !flags.contains(CompactFlags::ForceRepartition)
3661 : {
3662 408 : debug!(
3663 0 : distance,
3664 0 : threshold = self.repartition_threshold,
3665 0 : "no repartitioning needed"
3666 0 : );
3667 408 : return Ok((partitioning_guard.0.clone(), partitioning_guard.1));
3668 196 : }
3669 :
3670 13300 : let keyspace = self.collect_keyspace(lsn, ctx).await?;
3671 196 : let partitioning = keyspace.partition(partition_size);
3672 196 :
3673 196 : *partitioning_guard = (partitioning, lsn);
3674 196 :
3675 196 : Ok((partitioning_guard.0.clone(), partitioning_guard.1))
3676 604 : }
3677 :
3678 : // Is it time to create a new image layer for the given partition?
3679 522 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
3680 522 : let last = self.last_image_layer_creation_check_at.load();
3681 522 : if lsn != Lsn(0) {
3682 522 : let distance = lsn
3683 522 : .checked_sub(last)
3684 522 : .expect("Attempt to compact with LSN going backwards");
3685 522 :
3686 522 : let min_distance = self.get_image_layer_creation_check_threshold() as u64
3687 522 : * self.get_checkpoint_distance();
3688 522 :
3689 522 : // Skip the expensive delta layer counting below if we've not ingested
3690 522 : // sufficient WAL since the last check.
3691 522 : if distance.0 < min_distance {
3692 520 : return false;
3693 2 : }
3694 0 : }
3695 :
3696 2 : self.last_image_layer_creation_check_at.store(lsn);
3697 2 :
3698 2 : let threshold = self.get_image_creation_threshold();
3699 :
3700 2 : let guard = self.layers.read().await;
3701 2 : let layers = guard.layer_map();
3702 2 :
3703 2 : let mut max_deltas = 0;
3704 4 : for part_range in &partition.ranges {
3705 2 : let image_coverage = layers.image_coverage(part_range, lsn);
3706 4 : for (img_range, last_img) in image_coverage {
3707 2 : let img_lsn = if let Some(last_img) = last_img {
3708 0 : last_img.get_lsn_range().end
3709 : } else {
3710 2 : Lsn(0)
3711 : };
3712 : // Let's consider an example:
3713 : //
3714 : // delta layer with LSN range 71-81
3715 : // delta layer with LSN range 81-91
3716 : // delta layer with LSN range 91-101
3717 : // image layer at LSN 100
3718 : //
3719 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
3720 : // there's no need to create a new one. We check this case explicitly, to avoid passing
3721 : // a bogus range to count_deltas below, with start > end. It's even possible that there
3722 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
3723 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
3724 2 : if img_lsn < lsn {
3725 2 : let num_deltas =
3726 2 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
3727 2 :
3728 2 : max_deltas = max_deltas.max(num_deltas);
3729 2 : if num_deltas >= threshold {
3730 0 : debug!(
3731 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
3732 0 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
3733 0 : );
3734 0 : return true;
3735 2 : }
3736 0 : }
3737 : }
3738 : }
3739 :
3740 2 : debug!(
3741 0 : max_deltas,
3742 0 : "none of the partitioned ranges had >= {threshold} deltas"
3743 0 : );
3744 2 : false
3745 522 : }
3746 :
3747 1208 : #[tracing::instrument(skip_all, fields(%lsn, %force))]
3748 : async fn create_image_layers(
3749 : self: &Arc<Timeline>,
3750 : partitioning: &KeyPartitioning,
3751 : lsn: Lsn,
3752 : force: bool,
3753 : ctx: &RequestContext,
3754 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
3755 : let timer = self.metrics.create_images_time_histo.start_timer();
3756 : let mut image_layers = Vec::new();
3757 :
3758 : // We need to avoid holes between generated image layers.
3759 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
3760 : // image layer with hole between them. In this case such layer can not be utilized by GC.
3761 : //
3762 : // How such hole between partitions can appear?
3763 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
3764 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
3765 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
3766 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
3767 : let mut start = Key::MIN;
3768 :
3769 : for partition in partitioning.parts.iter() {
3770 : let img_range = start..partition.ranges.last().unwrap().end;
3771 : if !force && !self.time_for_new_image_layer(partition, lsn).await {
3772 : start = img_range.end;
3773 : continue;
3774 : }
3775 :
3776 : let mut image_layer_writer = ImageLayerWriter::new(
3777 : self.conf,
3778 : self.timeline_id,
3779 : self.tenant_shard_id,
3780 : &img_range,
3781 : lsn,
3782 : )
3783 : .await?;
3784 :
3785 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
3786 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
3787 0 : "failpoint image-layer-writer-fail-before-finish"
3788 0 : )))
3789 0 : });
3790 :
3791 : let mut wrote_keys = false;
3792 :
3793 : let mut key_request_accum = KeySpaceAccum::new();
3794 : for range in &partition.ranges {
3795 : let mut key = range.start;
3796 : while key < range.end {
3797 : // Decide whether to retain this key: usually we do, but sharded tenants may
3798 : // need to drop keys that don't belong to them. If we retain the key, add it
3799 : // to `key_request_accum` for later issuing a vectored get
3800 : if self.shard_identity.is_key_disposable(&key) {
3801 0 : debug!(
3802 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
3803 0 : key,
3804 0 : self.shard_identity.get_shard_number(&key)
3805 0 : );
3806 : } else {
3807 : key_request_accum.add_key(key);
3808 : }
3809 :
3810 : let last_key_in_range = key.next() == range.end;
3811 : key = key.next();
3812 :
3813 : // Maybe flush `key_rest_accum`
3814 : if key_request_accum.size() >= Timeline::MAX_GET_VECTORED_KEYS
3815 : || last_key_in_range
3816 : {
3817 : let results = self
3818 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
3819 : .await?;
3820 :
3821 : for (img_key, img) in results {
3822 : let img = match img {
3823 : Ok(img) => img,
3824 : Err(err) => {
3825 : // If we fail to reconstruct a VM or FSM page, we can zero the
3826 : // page without losing any actual user data. That seems better
3827 : // than failing repeatedly and getting stuck.
3828 : //
3829 : // We had a bug at one point, where we truncated the FSM and VM
3830 : // in the pageserver, but the Postgres didn't know about that
3831 : // and continued to generate incremental WAL records for pages
3832 : // that didn't exist in the pageserver. Trying to replay those
3833 : // WAL records failed to find the previous image of the page.
3834 : // This special case allows us to recover from that situation.
3835 : // See https://github.com/neondatabase/neon/issues/2601.
3836 : //
3837 : // Unfortunately we cannot do this for the main fork, or for
3838 : // any metadata keys, keys, as that would lead to actual data
3839 : // loss.
3840 : if is_rel_fsm_block_key(img_key) || is_rel_vm_block_key(img_key)
3841 : {
3842 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
3843 : ZERO_PAGE.clone()
3844 : } else {
3845 : return Err(CreateImageLayersError::PageReconstructError(
3846 : err,
3847 : ));
3848 : }
3849 : }
3850 : };
3851 :
3852 : // Write all the keys we just read into our new image layer.
3853 : image_layer_writer.put_image(img_key, img).await?;
3854 : wrote_keys = true;
3855 : }
3856 : }
3857 : }
3858 : }
3859 :
3860 : if wrote_keys {
3861 : // Normal path: we have written some data into the new image layer for this
3862 : // partition, so flush it to disk.
3863 : start = img_range.end;
3864 : let image_layer = image_layer_writer.finish(self).await?;
3865 : image_layers.push(image_layer);
3866 : } else {
3867 : // Special case: the image layer may be empty if this is a sharded tenant and the
3868 : // partition does not cover any keys owned by this shard. In this case, to ensure
3869 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
3870 : // layer we write will cover the key range that we just scanned.
3871 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
3872 : }
3873 : }
3874 :
3875 : // The writer.finish() above already did the fsync of the inodes.
3876 : // We just need to fsync the directory in which these inodes are linked,
3877 : // which we know to be the timeline directory.
3878 : if !image_layers.is_empty() {
3879 : // We use fatal_err() below because the after writer.finish() returns with success,
3880 : // the in-memory state of the filesystem already has the layer file in its final place,
3881 : // and subsequent pageserver code could think it's durable while it really isn't.
3882 : let timeline_dir = VirtualFile::open(
3883 : &self
3884 : .conf
3885 : .timeline_path(&self.tenant_shard_id, &self.timeline_id),
3886 : )
3887 : .await
3888 : .fatal_err("VirtualFile::open for timeline dir fsync");
3889 : timeline_dir
3890 : .sync_all()
3891 : .await
3892 : .fatal_err("VirtualFile::sync_all timeline dir");
3893 : }
3894 :
3895 : let mut guard = self.layers.write().await;
3896 :
3897 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
3898 : // now they are being scheduled outside of write lock
3899 : guard.track_new_image_layers(&image_layers, &self.metrics);
3900 : drop_wlock(guard);
3901 : timer.stop_and_record();
3902 :
3903 : Ok(image_layers)
3904 : }
3905 :
3906 : /// Wait until the background initial logical size calculation is complete, or
3907 : /// this Timeline is shut down. Calling this function will cause the initial
3908 : /// logical size calculation to skip waiting for the background jobs barrier.
3909 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
3910 0 : if let Some(await_bg_cancel) = self
3911 0 : .current_logical_size
3912 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3913 0 : .get()
3914 0 : {
3915 0 : await_bg_cancel.cancel();
3916 0 : } else {
3917 : // We should not wait if we were not able to explicitly instruct
3918 : // the logical size cancellation to skip the concurrency limit semaphore.
3919 : // TODO: this is an unexpected case. We should restructure so that it
3920 : // can't happen.
3921 0 : tracing::info!(
3922 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
3923 0 : );
3924 : }
3925 :
3926 0 : tokio::select!(
3927 0 : _ = self.current_logical_size.initialized.acquire() => {},
3928 0 : _ = self.cancel.cancelled() => {}
3929 0 : )
3930 0 : }
3931 : }
3932 :
3933 : /// Top-level failure to compact.
3934 0 : #[derive(Debug, thiserror::Error)]
3935 : pub(crate) enum CompactionError {
3936 : #[error("The timeline or pageserver is shutting down")]
3937 : ShuttingDown,
3938 : /// Compaction cannot be done right now; page reconstruction and so on.
3939 : #[error(transparent)]
3940 : Other(#[from] anyhow::Error),
3941 : }
3942 :
3943 : impl From<CollectKeySpaceError> for CompactionError {
3944 0 : fn from(err: CollectKeySpaceError) -> Self {
3945 0 : match err {
3946 : CollectKeySpaceError::Cancelled
3947 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
3948 0 : CompactionError::ShuttingDown
3949 : }
3950 0 : e => CompactionError::Other(e.into()),
3951 : }
3952 0 : }
3953 : }
3954 :
3955 : #[serde_as]
3956 294 : #[derive(serde::Serialize)]
3957 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
3958 :
3959 : #[derive(Default)]
3960 : enum DurationRecorder {
3961 : #[default]
3962 : NotStarted,
3963 : Recorded(RecordedDuration, tokio::time::Instant),
3964 : }
3965 :
3966 : impl DurationRecorder {
3967 720 : fn till_now(&self) -> DurationRecorder {
3968 720 : match self {
3969 : DurationRecorder::NotStarted => {
3970 0 : panic!("must only call on recorded measurements")
3971 : }
3972 720 : DurationRecorder::Recorded(_, ended) => {
3973 720 : let now = tokio::time::Instant::now();
3974 720 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
3975 720 : }
3976 720 : }
3977 720 : }
3978 294 : fn into_recorded(self) -> Option<RecordedDuration> {
3979 294 : match self {
3980 0 : DurationRecorder::NotStarted => None,
3981 294 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
3982 : }
3983 294 : }
3984 : }
3985 :
3986 : impl Timeline {
3987 42 : async fn finish_compact_batch(
3988 42 : self: &Arc<Self>,
3989 42 : new_deltas: &[ResidentLayer],
3990 42 : new_images: &[ResidentLayer],
3991 42 : layers_to_remove: &[Layer],
3992 42 : ) -> anyhow::Result<()> {
3993 42 : let mut guard = self.layers.write().await;
3994 :
3995 42 : let mut duplicated_layers = HashSet::new();
3996 42 :
3997 42 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
3998 :
3999 284 : for l in new_deltas {
4000 242 : if guard.contains(l.as_ref()) {
4001 : // expected in tests
4002 0 : tracing::error!(layer=%l, "duplicated L1 layer");
4003 :
4004 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
4005 : // `guard` on self.layers. as of writing this, there are no error returns except
4006 : // for compact_level0_phase1 creating an L0, which does not happen in practice
4007 : // because we have not implemented L0 => L0 compaction.
4008 0 : duplicated_layers.insert(l.layer_desc().key());
4009 242 : } else if LayerMap::is_l0(l.layer_desc()) {
4010 0 : bail!("compaction generates a L0 layer file as output, which will cause infinite compaction.");
4011 242 : } else {
4012 242 : insert_layers.push(l.clone());
4013 242 : }
4014 : }
4015 :
4016 : // only remove those inputs which were not outputs
4017 42 : let remove_layers: Vec<Layer> = layers_to_remove
4018 42 : .iter()
4019 442 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
4020 42 : .cloned()
4021 42 : .collect();
4022 42 :
4023 42 : if !new_images.is_empty() {
4024 0 : guard.track_new_image_layers(new_images, &self.metrics);
4025 42 : }
4026 :
4027 : // deletion will happen later, the layer file manager calls garbage_collect_on_drop
4028 42 : guard.finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
4029 :
4030 42 : if let Some(remote_client) = self.remote_client.as_ref() {
4031 42 : remote_client.schedule_compaction_update(&remove_layers, new_deltas)?;
4032 0 : }
4033 :
4034 42 : drop_wlock(guard);
4035 42 :
4036 42 : Ok(())
4037 42 : }
4038 :
4039 : /// Schedules the uploads of the given image layers
4040 510 : fn upload_new_image_layers(
4041 510 : self: &Arc<Self>,
4042 510 : new_images: impl IntoIterator<Item = ResidentLayer>,
4043 510 : ) -> anyhow::Result<()> {
4044 510 : let Some(remote_client) = &self.remote_client else {
4045 0 : return Ok(());
4046 : };
4047 510 : for layer in new_images {
4048 0 : remote_client.schedule_layer_file_upload(layer)?;
4049 : }
4050 : // should any new image layer been created, not uploading index_part will
4051 : // result in a mismatch between remote_physical_size and layermap calculated
4052 : // size, which will fail some tests, but should not be an issue otherwise.
4053 510 : remote_client.schedule_index_upload_for_file_changes()?;
4054 510 : Ok(())
4055 510 : }
4056 :
4057 : /// Update information about which layer files need to be retained on
4058 : /// garbage collection. This is separate from actually performing the GC,
4059 : /// and is updated more frequently, so that compaction can remove obsolete
4060 : /// page versions more aggressively.
4061 : ///
4062 : /// TODO: that's wishful thinking, compaction doesn't actually do that
4063 : /// currently.
4064 : ///
4065 : /// The caller specifies how much history is needed with the 3 arguments:
4066 : ///
4067 : /// retain_lsns: keep a version of each page at these LSNs
4068 : /// cutoff_horizon: also keep everything newer than this LSN
4069 : /// pitr: the time duration required to keep data for PITR
4070 : ///
4071 : /// The 'retain_lsns' list is currently used to prevent removing files that
4072 : /// are needed by child timelines. In the future, the user might be able to
4073 : /// name additional points in time to retain. The caller is responsible for
4074 : /// collecting that information.
4075 : ///
4076 : /// The 'cutoff_horizon' point is used to retain recent versions that might still be
4077 : /// needed by read-only nodes. (As of this writing, the caller just passes
4078 : /// the latest LSN subtracted by a constant, and doesn't do anything smart
4079 : /// to figure out what read-only nodes might actually need.)
4080 : ///
4081 : /// The 'pitr' duration is used to calculate a 'pitr_cutoff', which can be used to determine
4082 : /// whether a record is needed for PITR.
4083 : ///
4084 : /// NOTE: This function holds a short-lived lock to protect the 'gc_info'
4085 : /// field, so that the three values passed as argument are stored
4086 : /// atomically. But the caller is responsible for ensuring that no new
4087 : /// branches are created that would need to be included in 'retain_lsns',
4088 : /// for example. The caller should hold `Tenant::gc_cs` lock to ensure
4089 : /// that.
4090 : ///
4091 1016 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
4092 : pub(super) async fn update_gc_info(
4093 : &self,
4094 : retain_lsns: Vec<Lsn>,
4095 : cutoff_horizon: Lsn,
4096 : pitr: Duration,
4097 : cancel: &CancellationToken,
4098 : ctx: &RequestContext,
4099 : ) -> anyhow::Result<()> {
4100 : // First, calculate pitr_cutoff_timestamp and then convert it to LSN.
4101 : //
4102 : // Some unit tests depend on garbage-collection working even when
4103 : // CLOG data is missing, so that find_lsn_for_timestamp() doesn't
4104 : // work, so avoid calling it altogether if time-based retention is not
4105 : // configured. It would be pointless anyway.
4106 : let pitr_cutoff = if pitr != Duration::ZERO {
4107 : let now = SystemTime::now();
4108 : if let Some(pitr_cutoff_timestamp) = now.checked_sub(pitr) {
4109 : let pitr_timestamp = to_pg_timestamp(pitr_cutoff_timestamp);
4110 :
4111 : match self
4112 : .find_lsn_for_timestamp(pitr_timestamp, cancel, ctx)
4113 : .await?
4114 : {
4115 : LsnForTimestamp::Present(lsn) => lsn,
4116 : LsnForTimestamp::Future(lsn) => {
4117 : // The timestamp is in the future. That sounds impossible,
4118 : // but what it really means is that there hasn't been
4119 : // any commits since the cutoff timestamp.
4120 : //
4121 : // In this case we should use the LSN of the most recent commit,
4122 : // which is implicitly the last LSN in the log.
4123 0 : debug!("future({})", lsn);
4124 : self.get_last_record_lsn()
4125 : }
4126 : LsnForTimestamp::Past(lsn) => {
4127 0 : debug!("past({})", lsn);
4128 : // conservative, safe default is to remove nothing, when we
4129 : // have no commit timestamp data available
4130 : *self.get_latest_gc_cutoff_lsn()
4131 : }
4132 : LsnForTimestamp::NoData(lsn) => {
4133 0 : debug!("nodata({})", lsn);
4134 : // conservative, safe default is to remove nothing, when we
4135 : // have no commit timestamp data available
4136 : *self.get_latest_gc_cutoff_lsn()
4137 : }
4138 : }
4139 : } else {
4140 : // If we don't have enough data to convert to LSN,
4141 : // play safe and don't remove any layers.
4142 : *self.get_latest_gc_cutoff_lsn()
4143 : }
4144 : } else {
4145 : // No time-based retention was configured. Set time-based cutoff to
4146 : // same as LSN based.
4147 : cutoff_horizon
4148 : };
4149 :
4150 : // Grab the lock and update the values
4151 : *self.gc_info.write().unwrap() = GcInfo {
4152 : retain_lsns,
4153 : horizon_cutoff: cutoff_horizon,
4154 : pitr_cutoff,
4155 : };
4156 :
4157 : Ok(())
4158 : }
4159 :
4160 : /// Garbage collect layer files on a timeline that are no longer needed.
4161 : ///
4162 : /// Currently, we don't make any attempt at removing unneeded page versions
4163 : /// within a layer file. We can only remove the whole file if it's fully
4164 : /// obsolete.
4165 508 : pub(super) async fn gc(&self) -> anyhow::Result<GcResult> {
4166 508 : // this is most likely the background tasks, but it might be the spawned task from
4167 508 : // immediate_gc
4168 508 : let cancel = crate::task_mgr::shutdown_token();
4169 508 : let _g = tokio::select! {
4170 506 : guard = self.gc_lock.lock() => guard,
4171 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
4172 : _ = cancel.cancelled() => return Ok(GcResult::default()),
4173 : };
4174 506 : let timer = self.metrics.garbage_collect_histo.start_timer();
4175 :
4176 0 : fail_point!("before-timeline-gc");
4177 :
4178 : // Is the timeline being deleted?
4179 506 : if self.is_stopping() {
4180 0 : anyhow::bail!("timeline is Stopping");
4181 506 : }
4182 506 :
4183 506 : let (horizon_cutoff, pitr_cutoff, retain_lsns) = {
4184 506 : let gc_info = self.gc_info.read().unwrap();
4185 506 :
4186 506 : let horizon_cutoff = min(gc_info.horizon_cutoff, self.get_disk_consistent_lsn());
4187 506 : let pitr_cutoff = gc_info.pitr_cutoff;
4188 506 : let retain_lsns = gc_info.retain_lsns.clone();
4189 506 : (horizon_cutoff, pitr_cutoff, retain_lsns)
4190 506 : };
4191 506 :
4192 506 : let new_gc_cutoff = Lsn::min(horizon_cutoff, pitr_cutoff);
4193 :
4194 506 : let res = self
4195 506 : .gc_timeline(horizon_cutoff, pitr_cutoff, retain_lsns, new_gc_cutoff)
4196 506 : .instrument(
4197 506 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
4198 : )
4199 0 : .await?;
4200 :
4201 : // only record successes
4202 506 : timer.stop_and_record();
4203 506 :
4204 506 : Ok(res)
4205 508 : }
4206 :
4207 506 : async fn gc_timeline(
4208 506 : &self,
4209 506 : horizon_cutoff: Lsn,
4210 506 : pitr_cutoff: Lsn,
4211 506 : retain_lsns: Vec<Lsn>,
4212 506 : new_gc_cutoff: Lsn,
4213 506 : ) -> anyhow::Result<GcResult> {
4214 506 : let now = SystemTime::now();
4215 506 : let mut result: GcResult = GcResult::default();
4216 506 :
4217 506 : // Nothing to GC. Return early.
4218 506 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
4219 506 : if latest_gc_cutoff >= new_gc_cutoff {
4220 0 : info!(
4221 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
4222 0 : );
4223 0 : return Ok(result);
4224 506 : }
4225 :
4226 : // We need to ensure that no one tries to read page versions or create
4227 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
4228 : // for details. This will block until the old value is no longer in use.
4229 : //
4230 : // The GC cutoff should only ever move forwards.
4231 506 : let waitlist = {
4232 506 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
4233 506 : ensure!(
4234 506 : *write_guard <= new_gc_cutoff,
4235 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
4236 0 : *write_guard,
4237 : new_gc_cutoff
4238 : );
4239 506 : write_guard.store_and_unlock(new_gc_cutoff)
4240 506 : };
4241 506 : waitlist.wait().await;
4242 :
4243 506 : info!("GC starting");
4244 :
4245 506 : debug!("retain_lsns: {:?}", retain_lsns);
4246 :
4247 506 : let mut layers_to_remove = Vec::new();
4248 :
4249 : // Scan all layers in the timeline (remote or on-disk).
4250 : //
4251 : // Garbage collect the layer if all conditions are satisfied:
4252 : // 1. it is older than cutoff LSN;
4253 : // 2. it is older than PITR interval;
4254 : // 3. it doesn't need to be retained for 'retain_lsns';
4255 : // 4. newer on-disk image layers cover the layer's whole key range
4256 : //
4257 : // TODO holding a write lock is too agressive and avoidable
4258 506 : let mut guard = self.layers.write().await;
4259 506 : let layers = guard.layer_map();
4260 3158 : 'outer: for l in layers.iter_historic_layers() {
4261 3158 : result.layers_total += 1;
4262 3158 :
4263 3158 : // 1. Is it newer than GC horizon cutoff point?
4264 3158 : if l.get_lsn_range().end > horizon_cutoff {
4265 506 : debug!(
4266 0 : "keeping {} because it's newer than horizon_cutoff {}",
4267 0 : l.filename(),
4268 0 : horizon_cutoff,
4269 0 : );
4270 506 : result.layers_needed_by_cutoff += 1;
4271 506 : continue 'outer;
4272 2652 : }
4273 2652 :
4274 2652 : // 2. It is newer than PiTR cutoff point?
4275 2652 : if l.get_lsn_range().end > pitr_cutoff {
4276 0 : debug!(
4277 0 : "keeping {} because it's newer than pitr_cutoff {}",
4278 0 : l.filename(),
4279 0 : pitr_cutoff,
4280 0 : );
4281 0 : result.layers_needed_by_pitr += 1;
4282 0 : continue 'outer;
4283 2652 : }
4284 :
4285 : // 3. Is it needed by a child branch?
4286 : // NOTE With that we would keep data that
4287 : // might be referenced by child branches forever.
4288 : // We can track this in child timeline GC and delete parent layers when
4289 : // they are no longer needed. This might be complicated with long inheritance chains.
4290 : //
4291 : // TODO Vec is not a great choice for `retain_lsns`
4292 2652 : for retain_lsn in &retain_lsns {
4293 : // start_lsn is inclusive
4294 8 : if &l.get_lsn_range().start <= retain_lsn {
4295 8 : debug!(
4296 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
4297 0 : l.filename(),
4298 0 : retain_lsn,
4299 0 : l.is_incremental(),
4300 0 : );
4301 8 : result.layers_needed_by_branches += 1;
4302 8 : continue 'outer;
4303 0 : }
4304 : }
4305 :
4306 : // 4. Is there a later on-disk layer for this relation?
4307 : //
4308 : // The end-LSN is exclusive, while disk_consistent_lsn is
4309 : // inclusive. For example, if disk_consistent_lsn is 100, it is
4310 : // OK for a delta layer to have end LSN 101, but if the end LSN
4311 : // is 102, then it might not have been fully flushed to disk
4312 : // before crash.
4313 : //
4314 : // For example, imagine that the following layers exist:
4315 : //
4316 : // 1000 - image (A)
4317 : // 1000-2000 - delta (B)
4318 : // 2000 - image (C)
4319 : // 2000-3000 - delta (D)
4320 : // 3000 - image (E)
4321 : //
4322 : // If GC horizon is at 2500, we can remove layers A and B, but
4323 : // we cannot remove C, even though it's older than 2500, because
4324 : // the delta layer 2000-3000 depends on it.
4325 2644 : if !layers
4326 2644 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
4327 : {
4328 2644 : debug!("keeping {} because it is the latest layer", l.filename());
4329 2644 : result.layers_not_updated += 1;
4330 2644 : continue 'outer;
4331 0 : }
4332 0 :
4333 0 : // We didn't find any reason to keep this file, so remove it.
4334 0 : debug!(
4335 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
4336 0 : l.filename(),
4337 0 : l.is_incremental(),
4338 0 : );
4339 0 : layers_to_remove.push(l);
4340 : }
4341 :
4342 506 : if !layers_to_remove.is_empty() {
4343 : // Persist the new GC cutoff value before we actually remove anything.
4344 : // This unconditionally schedules also an index_part.json update, even though, we will
4345 : // be doing one a bit later with the unlinked gc'd layers.
4346 0 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
4347 0 : self.schedule_uploads(disk_consistent_lsn, None)?;
4348 :
4349 0 : let gc_layers = layers_to_remove
4350 0 : .iter()
4351 0 : .map(|x| guard.get_from_desc(x))
4352 0 : .collect::<Vec<Layer>>();
4353 0 :
4354 0 : result.layers_removed = gc_layers.len() as u64;
4355 :
4356 0 : if let Some(remote_client) = self.remote_client.as_ref() {
4357 0 : remote_client.schedule_gc_update(&gc_layers)?;
4358 0 : }
4359 :
4360 0 : guard.finish_gc_timeline(&gc_layers);
4361 0 :
4362 0 : #[cfg(feature = "testing")]
4363 0 : {
4364 0 : result.doomed_layers = gc_layers;
4365 0 : }
4366 506 : }
4367 :
4368 506 : info!(
4369 506 : "GC completed removing {} layers, cutoff {}",
4370 506 : result.layers_removed, new_gc_cutoff
4371 506 : );
4372 :
4373 506 : result.elapsed = now.elapsed()?;
4374 506 : Ok(result)
4375 506 : }
4376 :
4377 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
4378 503085 : async fn reconstruct_value(
4379 503085 : &self,
4380 503085 : key: Key,
4381 503085 : request_lsn: Lsn,
4382 503085 : mut data: ValueReconstructState,
4383 503085 : ) -> Result<Bytes, PageReconstructError> {
4384 503085 : // Perform WAL redo if needed
4385 503085 : data.records.reverse();
4386 503085 :
4387 503085 : // If we have a page image, and no WAL, we're all set
4388 503085 : if data.records.is_empty() {
4389 503079 : if let Some((img_lsn, img)) = &data.img {
4390 503079 : trace!(
4391 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
4392 0 : key,
4393 0 : img_lsn,
4394 0 : request_lsn,
4395 0 : );
4396 503079 : Ok(img.clone())
4397 : } else {
4398 0 : Err(PageReconstructError::from(anyhow!(
4399 0 : "base image for {key} at {request_lsn} not found"
4400 0 : )))
4401 : }
4402 : } else {
4403 : // We need to do WAL redo.
4404 : //
4405 : // If we don't have a base image, then the oldest WAL record better initialize
4406 : // the page
4407 6 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
4408 0 : Err(PageReconstructError::from(anyhow!(
4409 0 : "Base image for {} at {} not found, but got {} WAL records",
4410 0 : key,
4411 0 : request_lsn,
4412 0 : data.records.len()
4413 0 : )))
4414 : } else {
4415 6 : if data.img.is_some() {
4416 6 : trace!(
4417 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
4418 0 : data.records.len(),
4419 0 : key,
4420 0 : request_lsn
4421 0 : );
4422 : } else {
4423 0 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
4424 : };
4425 :
4426 6 : let last_rec_lsn = data.records.last().unwrap().0;
4427 :
4428 6 : let img = match self
4429 6 : .walredo_mgr
4430 6 : .as_ref()
4431 6 : .context("timeline has no walredo manager")
4432 6 : .map_err(PageReconstructError::WalRedo)?
4433 6 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
4434 0 : .await
4435 6 : .context("reconstruct a page image")
4436 : {
4437 6 : Ok(img) => img,
4438 0 : Err(e) => return Err(PageReconstructError::WalRedo(e)),
4439 : };
4440 :
4441 6 : if img.len() == page_cache::PAGE_SZ {
4442 0 : let cache = page_cache::get();
4443 0 : if let Err(e) = cache
4444 0 : .memorize_materialized_page(
4445 0 : self.tenant_shard_id,
4446 0 : self.timeline_id,
4447 0 : key,
4448 0 : last_rec_lsn,
4449 0 : &img,
4450 0 : )
4451 0 : .await
4452 0 : .context("Materialized page memoization failed")
4453 : {
4454 0 : return Err(PageReconstructError::from(e));
4455 0 : }
4456 6 : }
4457 :
4458 6 : Ok(img)
4459 : }
4460 : }
4461 503085 : }
4462 :
4463 0 : pub(crate) async fn spawn_download_all_remote_layers(
4464 0 : self: Arc<Self>,
4465 0 : request: DownloadRemoteLayersTaskSpawnRequest,
4466 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
4467 0 : use pageserver_api::models::DownloadRemoteLayersTaskState;
4468 0 :
4469 0 : // this is not really needed anymore; it has tests which really check the return value from
4470 0 : // http api. it would be better not to maintain this anymore.
4471 0 :
4472 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
4473 0 : if let Some(st) = &*status_guard {
4474 0 : match &st.state {
4475 : DownloadRemoteLayersTaskState::Running => {
4476 0 : return Err(st.clone());
4477 : }
4478 : DownloadRemoteLayersTaskState::ShutDown
4479 0 : | DownloadRemoteLayersTaskState::Completed => {
4480 0 : *status_guard = None;
4481 0 : }
4482 : }
4483 0 : }
4484 :
4485 0 : let self_clone = Arc::clone(&self);
4486 0 : let task_id = task_mgr::spawn(
4487 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
4488 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
4489 0 : Some(self.tenant_shard_id),
4490 0 : Some(self.timeline_id),
4491 0 : "download all remote layers task",
4492 : false,
4493 0 : async move {
4494 0 : self_clone.download_all_remote_layers(request).await;
4495 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
4496 0 : match &mut *status_guard {
4497 : None => {
4498 0 : warn!("tasks status is supposed to be Some(), since we are running");
4499 : }
4500 0 : Some(st) => {
4501 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
4502 0 : if st.task_id != exp_task_id {
4503 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
4504 0 : } else {
4505 0 : st.state = DownloadRemoteLayersTaskState::Completed;
4506 0 : }
4507 : }
4508 : };
4509 0 : Ok(())
4510 0 : }
4511 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))
4512 : );
4513 :
4514 0 : let initial_info = DownloadRemoteLayersTaskInfo {
4515 0 : task_id: format!("{task_id}"),
4516 0 : state: DownloadRemoteLayersTaskState::Running,
4517 0 : total_layer_count: 0,
4518 0 : successful_download_count: 0,
4519 0 : failed_download_count: 0,
4520 0 : };
4521 0 : *status_guard = Some(initial_info.clone());
4522 0 :
4523 0 : Ok(initial_info)
4524 0 : }
4525 :
4526 0 : async fn download_all_remote_layers(
4527 0 : self: &Arc<Self>,
4528 0 : request: DownloadRemoteLayersTaskSpawnRequest,
4529 0 : ) {
4530 : use pageserver_api::models::DownloadRemoteLayersTaskState;
4531 :
4532 0 : let remaining = {
4533 0 : let guard = self.layers.read().await;
4534 0 : guard
4535 0 : .layer_map()
4536 0 : .iter_historic_layers()
4537 0 : .map(|desc| guard.get_from_desc(&desc))
4538 0 : .collect::<Vec<_>>()
4539 0 : };
4540 0 : let total_layer_count = remaining.len();
4541 0 :
4542 0 : macro_rules! lock_status {
4543 0 : ($st:ident) => {
4544 0 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
4545 0 : let st = st
4546 0 : .as_mut()
4547 0 : .expect("this function is only called after the task has been spawned");
4548 0 : assert_eq!(
4549 0 : st.task_id,
4550 0 : format!(
4551 0 : "{}",
4552 0 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
4553 0 : )
4554 0 : );
4555 0 : let $st = st;
4556 0 : };
4557 0 : }
4558 0 :
4559 0 : {
4560 0 : lock_status!(st);
4561 0 : st.total_layer_count = total_layer_count as u64;
4562 0 : }
4563 0 :
4564 0 : let mut remaining = remaining.into_iter();
4565 0 : let mut have_remaining = true;
4566 0 : let mut js = tokio::task::JoinSet::new();
4567 0 :
4568 0 : let cancel = task_mgr::shutdown_token();
4569 0 :
4570 0 : let limit = request.max_concurrent_downloads;
4571 :
4572 : loop {
4573 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
4574 0 : let Some(next) = remaining.next() else {
4575 0 : have_remaining = false;
4576 0 : break;
4577 : };
4578 :
4579 0 : let span = tracing::info_span!("download", layer = %next);
4580 :
4581 0 : js.spawn(
4582 0 : async move {
4583 0 : let res = next.download().await;
4584 0 : (next, res)
4585 0 : }
4586 0 : .instrument(span),
4587 0 : );
4588 : }
4589 :
4590 0 : while let Some(res) = js.join_next().await {
4591 0 : match res {
4592 : Ok((_, Ok(_))) => {
4593 0 : lock_status!(st);
4594 0 : st.successful_download_count += 1;
4595 : }
4596 0 : Ok((layer, Err(e))) => {
4597 0 : tracing::error!(%layer, "download failed: {e:#}");
4598 0 : lock_status!(st);
4599 0 : st.failed_download_count += 1;
4600 : }
4601 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
4602 0 : Err(je) if je.is_panic() => {
4603 0 : lock_status!(st);
4604 0 : st.failed_download_count += 1;
4605 : }
4606 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
4607 : }
4608 : }
4609 :
4610 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
4611 0 : break;
4612 0 : }
4613 : }
4614 :
4615 : {
4616 0 : lock_status!(st);
4617 0 : st.state = DownloadRemoteLayersTaskState::Completed;
4618 0 : }
4619 0 : }
4620 :
4621 0 : pub(crate) fn get_download_all_remote_layers_task_info(
4622 0 : &self,
4623 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
4624 0 : self.download_all_remote_layers_task_info
4625 0 : .read()
4626 0 : .unwrap()
4627 0 : .clone()
4628 0 : }
4629 : }
4630 :
4631 : impl Timeline {
4632 : /// Returns non-remote layers for eviction.
4633 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
4634 0 : let guard = self.layers.read().await;
4635 0 : let mut max_layer_size: Option<u64> = None;
4636 0 :
4637 0 : let resident_layers = guard
4638 0 : .likely_resident_layers()
4639 0 : .map(|layer| {
4640 0 : let file_size = layer.layer_desc().file_size;
4641 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
4642 0 :
4643 0 : let last_activity_ts = layer.access_stats().latest_activity_or_now();
4644 0 :
4645 0 : EvictionCandidate {
4646 0 : layer: layer.into(),
4647 0 : last_activity_ts,
4648 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
4649 0 : }
4650 0 : })
4651 0 : .collect();
4652 0 :
4653 0 : DiskUsageEvictionInfo {
4654 0 : max_layer_size,
4655 0 : resident_layers,
4656 0 : }
4657 0 : }
4658 :
4659 966 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
4660 966 : ShardIndex {
4661 966 : shard_number: self.tenant_shard_id.shard_number,
4662 966 : shard_count: self.tenant_shard_id.shard_count,
4663 966 : }
4664 966 : }
4665 : }
4666 :
4667 : type TraversalPathItem = (
4668 : ValueReconstructResult,
4669 : Lsn,
4670 : Box<dyn Send + FnOnce() -> TraversalId>,
4671 : );
4672 :
4673 : /// Helper function for get_reconstruct_data() to add the path of layers traversed
4674 : /// to an error, as anyhow context information.
4675 108 : fn layer_traversal_error(msg: String, path: Vec<TraversalPathItem>) -> PageReconstructError {
4676 108 : // We want the original 'msg' to be the outermost context. The outermost context
4677 108 : // is the most high-level information, which also gets propagated to the client.
4678 108 : let mut msg_iter = path
4679 108 : .into_iter()
4680 150 : .map(|(r, c, l)| {
4681 150 : format!(
4682 150 : "layer traversal: result {:?}, cont_lsn {}, layer: {}",
4683 150 : r,
4684 150 : c,
4685 150 : l(),
4686 150 : )
4687 150 : })
4688 108 : .chain(std::iter::once(msg));
4689 108 : // Construct initial message from the first traversed layer
4690 108 : let err = anyhow!(msg_iter.next().unwrap());
4691 108 :
4692 108 : // Append all subsequent traversals, and the error message 'msg', as contexts.
4693 150 : let msg = msg_iter.fold(err, |err, msg| err.context(msg));
4694 108 : PageReconstructError::from(msg)
4695 108 : }
4696 :
4697 : struct TimelineWriterState {
4698 : open_layer: Arc<InMemoryLayer>,
4699 : current_size: u64,
4700 : // Previous Lsn which passed through
4701 : prev_lsn: Option<Lsn>,
4702 : // Largest Lsn which passed through the current writer
4703 : max_lsn: Option<Lsn>,
4704 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
4705 : cached_last_freeze_at: Lsn,
4706 : }
4707 :
4708 : impl TimelineWriterState {
4709 3648020 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
4710 3648020 : Self {
4711 3648020 : open_layer,
4712 3648020 : current_size,
4713 3648020 : prev_lsn: None,
4714 3648020 : max_lsn: None,
4715 3648020 : cached_last_freeze_at: last_freeze_at,
4716 3648020 : }
4717 3648020 : }
4718 : }
4719 :
4720 : /// Various functions to mutate the timeline.
4721 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
4722 : // This is probably considered a bad practice in Rust and should be fixed eventually,
4723 : // but will cause large code changes.
4724 : pub(crate) struct TimelineWriter<'a> {
4725 : tl: &'a Timeline,
4726 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
4727 : }
4728 :
4729 : impl Deref for TimelineWriter<'_> {
4730 : type Target = Timeline;
4731 :
4732 3650620 : fn deref(&self) -> &Self::Target {
4733 3650620 : self.tl
4734 3650620 : }
4735 : }
4736 :
4737 : impl Drop for TimelineWriter<'_> {
4738 3977036 : fn drop(&mut self) {
4739 3977036 : self.write_guard.take();
4740 3977036 : }
4741 : }
4742 :
4743 : #[derive(PartialEq)]
4744 : enum OpenLayerAction {
4745 : Roll,
4746 : Open,
4747 : None,
4748 : }
4749 :
4750 : impl<'a> TimelineWriter<'a> {
4751 : /// Put a new page version that can be constructed from a WAL record
4752 : ///
4753 : /// This will implicitly extend the relation, if the page is beyond the
4754 : /// current end-of-file.
4755 3933972 : pub(crate) async fn put(
4756 3933972 : &mut self,
4757 3933972 : key: Key,
4758 3933972 : lsn: Lsn,
4759 3933972 : value: &Value,
4760 3933972 : ctx: &RequestContext,
4761 3933972 : ) -> anyhow::Result<()> {
4762 3933972 : // Avoid doing allocations for "small" values.
4763 3933972 : // In the regression test suite, the limit of 256 avoided allocations in 95% of cases:
4764 3933972 : // https://github.com/neondatabase/neon/pull/5056#discussion_r1301975061
4765 3933972 : let mut buf = smallvec::SmallVec::<[u8; 256]>::new();
4766 3933972 : value.ser_into(&mut buf)?;
4767 3933972 : let buf_size: u64 = buf.len().try_into().expect("oversized value buf");
4768 3933972 :
4769 3933972 : let action = self.get_open_layer_action(lsn, buf_size);
4770 3933972 : let layer = self.handle_open_layer_action(lsn, action).await?;
4771 3933972 : let res = layer.put_value(key, lsn, &buf, ctx).await;
4772 :
4773 3933972 : if res.is_ok() {
4774 3933972 : // Update the current size only when the entire write was ok.
4775 3933972 : // In case of failures, we may have had partial writes which
4776 3933972 : // render the size tracking out of sync. That's ok because
4777 3933972 : // the checkpoint distance should be significantly smaller
4778 3933972 : // than the S3 single shot upload limit of 5GiB.
4779 3933972 : let state = self.write_guard.as_mut().unwrap();
4780 3933972 :
4781 3933972 : state.current_size += buf_size;
4782 3933972 : state.prev_lsn = Some(lsn);
4783 3933972 : state.max_lsn = std::cmp::max(state.max_lsn, Some(lsn));
4784 3933972 : }
4785 :
4786 3933972 : res
4787 3933972 : }
4788 :
4789 3933974 : async fn handle_open_layer_action(
4790 3933974 : &mut self,
4791 3933974 : at: Lsn,
4792 3933974 : action: OpenLayerAction,
4793 3933974 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
4794 3933974 : match action {
4795 : OpenLayerAction::Roll => {
4796 0 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
4797 0 : self.roll_layer(freeze_at).await?;
4798 0 : self.open_layer(at).await?;
4799 : }
4800 3648020 : OpenLayerAction::Open => self.open_layer(at).await?,
4801 : OpenLayerAction::None => {
4802 285954 : assert!(self.write_guard.is_some());
4803 : }
4804 : }
4805 :
4806 3933974 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
4807 3933974 : }
4808 :
4809 3648020 : async fn open_layer(&mut self, at: Lsn) -> anyhow::Result<()> {
4810 3648020 : let layer = self.tl.get_layer_for_write(at).await?;
4811 3648020 : let initial_size = layer.size().await?;
4812 :
4813 3648020 : let last_freeze_at = self.last_freeze_at.load();
4814 3648020 : self.write_guard.replace(TimelineWriterState::new(
4815 3648020 : layer,
4816 3648020 : initial_size,
4817 3648020 : last_freeze_at,
4818 3648020 : ));
4819 3648020 :
4820 3648020 : Ok(())
4821 3648020 : }
4822 :
4823 0 : async fn roll_layer(&mut self, freeze_at: Lsn) -> anyhow::Result<()> {
4824 0 : assert!(self.write_guard.is_some());
4825 :
4826 0 : self.tl.freeze_inmem_layer_at(freeze_at).await;
4827 :
4828 0 : let now = Instant::now();
4829 0 : *(self.last_freeze_ts.write().unwrap()) = now;
4830 0 :
4831 0 : self.tl.flush_frozen_layers();
4832 0 :
4833 0 : let current_size = self.write_guard.as_ref().unwrap().current_size;
4834 0 : if current_size > self.get_checkpoint_distance() {
4835 0 : warn!("Flushed oversized open layer with size {}", current_size)
4836 0 : }
4837 :
4838 0 : Ok(())
4839 0 : }
4840 :
4841 3933974 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
4842 3933974 : let state = &*self.write_guard;
4843 3933974 : let Some(state) = &state else {
4844 3648020 : return OpenLayerAction::Open;
4845 : };
4846 :
4847 285954 : if state.prev_lsn == Some(lsn) {
4848 : // Rolling mid LSN is not supported by downstream code.
4849 : // Hence, only roll at LSN boundaries.
4850 285898 : return OpenLayerAction::None;
4851 56 : }
4852 56 :
4853 56 : if state.current_size == 0 {
4854 : // Don't roll empty layers
4855 0 : return OpenLayerAction::None;
4856 56 : }
4857 56 :
4858 56 : if self.tl.should_roll(
4859 56 : state.current_size,
4860 56 : state.current_size + new_value_size,
4861 56 : self.get_checkpoint_distance(),
4862 56 : lsn,
4863 56 : state.cached_last_freeze_at,
4864 56 : state.open_layer.get_opened_at(),
4865 56 : ) {
4866 0 : OpenLayerAction::Roll
4867 : } else {
4868 56 : OpenLayerAction::None
4869 : }
4870 3933974 : }
4871 :
4872 : /// Put a batch of keys at the specified Lsns.
4873 : ///
4874 : /// The batch is sorted by Lsn (enforced by usage of [`utils::vec_map::VecMap`].
4875 413960 : pub(crate) async fn put_batch(
4876 413960 : &mut self,
4877 413960 : batch: VecMap<Lsn, (Key, Value)>,
4878 413960 : ctx: &RequestContext,
4879 413960 : ) -> anyhow::Result<()> {
4880 1113816 : for (lsn, (key, val)) in batch {
4881 699856 : self.put(key, lsn, &val, ctx).await?
4882 : }
4883 :
4884 413960 : Ok(())
4885 413960 : }
4886 :
4887 2 : pub(crate) async fn delete_batch(&mut self, batch: &[(Range<Key>, Lsn)]) -> anyhow::Result<()> {
4888 2 : if let Some((_, lsn)) = batch.first() {
4889 2 : let action = self.get_open_layer_action(*lsn, 0);
4890 2 : let layer = self.handle_open_layer_action(*lsn, action).await?;
4891 2 : layer.put_tombstones(batch).await?;
4892 0 : }
4893 :
4894 2 : Ok(())
4895 2 : }
4896 :
4897 : /// Track the end of the latest digested WAL record.
4898 : /// Remember the (end of) last valid WAL record remembered in the timeline.
4899 : ///
4900 : /// Call this after you have finished writing all the WAL up to 'lsn'.
4901 : ///
4902 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
4903 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
4904 : /// the latest and can be read.
4905 4122936 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4906 4122936 : self.tl.finish_write(new_lsn);
4907 4122936 : }
4908 :
4909 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
4910 270570 : self.tl.update_current_logical_size(delta)
4911 270570 : }
4912 : }
4913 :
4914 : // We need TimelineWriter to be send in upcoming conversion of
4915 : // Timeline::layers to tokio::sync::RwLock.
4916 : #[test]
4917 2 : fn is_send() {
4918 2 : fn _assert_send<T: Send>() {}
4919 2 : _assert_send::<TimelineWriter<'_>>();
4920 2 : }
4921 :
4922 : /// Add a suffix to a layer file's name: .{num}.old
4923 : /// Uses the first available num (starts at 0)
4924 0 : fn rename_to_backup(path: &Utf8Path) -> anyhow::Result<()> {
4925 0 : let filename = path
4926 0 : .file_name()
4927 0 : .ok_or_else(|| anyhow!("Path {path} don't have a file name"))?;
4928 0 : let mut new_path = path.to_owned();
4929 :
4930 0 : for i in 0u32.. {
4931 0 : new_path.set_file_name(format!("{filename}.{i}.old"));
4932 0 : if !new_path.exists() {
4933 0 : std::fs::rename(path, &new_path)
4934 0 : .with_context(|| format!("rename {path:?} to {new_path:?}"))?;
4935 0 : return Ok(());
4936 0 : }
4937 : }
4938 :
4939 0 : bail!("couldn't find an unused backup number for {:?}", path)
4940 0 : }
4941 :
4942 : #[cfg(test)]
4943 : mod tests {
4944 : use utils::{id::TimelineId, lsn::Lsn};
4945 :
4946 : use crate::tenant::{
4947 : harness::TenantHarness, storage_layer::Layer, timeline::EvictionError, Timeline,
4948 : };
4949 :
4950 : #[tokio::test]
4951 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
4952 2 : let harness =
4953 2 : TenantHarness::create("two_layer_eviction_attempts_at_the_same_time").unwrap();
4954 2 :
4955 2 : let (tenant, ctx) = harness.load().await;
4956 2 : let timeline = tenant
4957 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
4958 5 : .await
4959 2 : .unwrap();
4960 2 :
4961 2 : let layer = find_some_layer(&timeline).await;
4962 2 : let layer = layer
4963 2 : .keep_resident()
4964 2 : .await
4965 2 : .expect("no download => no downloading errors")
4966 2 : .drop_eviction_guard();
4967 2 :
4968 2 : let forever = std::time::Duration::from_secs(120);
4969 2 :
4970 2 : let first = layer.evict_and_wait(forever);
4971 2 : let second = layer.evict_and_wait(forever);
4972 2 :
4973 4 : let (first, second) = tokio::join!(first, second);
4974 2 :
4975 2 : let res = layer.keep_resident().await;
4976 2 : assert!(res.is_none(), "{res:?}");
4977 2 :
4978 2 : match (first, second) {
4979 2 : (Ok(()), Ok(())) => {
4980 2 : // because there are no more timeline locks being taken on eviction path, we can
4981 2 : // witness all three outcomes here.
4982 2 : }
4983 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
4984 0 : // if one completes before the other, this is fine just as well.
4985 0 : }
4986 2 : other => unreachable!("unexpected {:?}", other),
4987 2 : }
4988 2 : }
4989 :
4990 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
4991 2 : let layers = timeline.layers.read().await;
4992 2 : let desc = layers
4993 2 : .layer_map()
4994 2 : .iter_historic_layers()
4995 2 : .next()
4996 2 : .expect("must find one layer to evict");
4997 2 :
4998 2 : layers.get_from_desc(&desc)
4999 2 : }
5000 : }
|
