Line data Source code
1 : mod compaction;
2 : pub mod delete;
3 : pub(crate) mod detach_ancestor;
4 : mod eviction_task;
5 : mod init;
6 : pub mod layer_manager;
7 : pub(crate) mod logical_size;
8 : pub mod span;
9 : pub mod uninit;
10 : mod walreceiver;
11 :
12 : use anyhow::{anyhow, bail, ensure, Context, Result};
13 : use arc_swap::ArcSwap;
14 : use bytes::Bytes;
15 : use camino::Utf8Path;
16 : use enumset::EnumSet;
17 : use fail::fail_point;
18 : use once_cell::sync::Lazy;
19 : use pageserver_api::{
20 : key::{
21 : AUX_FILES_KEY, KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX,
22 : NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE,
23 : },
24 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
25 : models::{
26 : AtomicAuxFilePolicy, AuxFilePolicy, CompactionAlgorithm, DownloadRemoteLayersTaskInfo,
27 : DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy, InMemoryLayerInfo, LayerMapInfo,
28 : TimelineState,
29 : },
30 : reltag::BlockNumber,
31 : shard::{ShardIdentity, ShardNumber, TenantShardId},
32 : };
33 : use rand::Rng;
34 : use serde_with::serde_as;
35 : use storage_broker::BrokerClientChannel;
36 : use tokio::{
37 : runtime::Handle,
38 : sync::{oneshot, watch},
39 : };
40 : use tokio_util::sync::CancellationToken;
41 : use tracing::*;
42 : use utils::{
43 : bin_ser::BeSer,
44 : sync::gate::{Gate, GateGuard},
45 : vec_map::VecMap,
46 : };
47 :
48 : use std::ops::{Deref, Range};
49 : use std::pin::pin;
50 : use std::sync::atomic::Ordering as AtomicOrdering;
51 : use std::sync::{Arc, Mutex, RwLock, Weak};
52 : use std::time::{Duration, Instant, SystemTime};
53 : use std::{
54 : array,
55 : collections::{BTreeMap, HashMap, HashSet},
56 : sync::atomic::AtomicU64,
57 : };
58 : use std::{
59 : cmp::{max, min, Ordering},
60 : ops::ControlFlow,
61 : };
62 :
63 : use crate::{
64 : aux_file::AuxFileSizeEstimator,
65 : tenant::{
66 : layer_map::{LayerMap, SearchResult},
67 : metadata::TimelineMetadata,
68 : },
69 : };
70 : use crate::{
71 : context::{DownloadBehavior, RequestContext},
72 : disk_usage_eviction_task::DiskUsageEvictionInfo,
73 : pgdatadir_mapping::CollectKeySpaceError,
74 : };
75 : use crate::{deletion_queue::DeletionQueueClient, metrics::GetKind};
76 : use crate::{
77 : disk_usage_eviction_task::finite_f32,
78 : tenant::storage_layer::{
79 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
80 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructResult,
81 : ValueReconstructState, ValuesReconstructState,
82 : },
83 : };
84 : use crate::{
85 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
86 : };
87 : use crate::{
88 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
89 : };
90 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
91 : use crate::{
92 : pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::timeline::init::LocalLayerFileMetadata,
93 : };
94 : use crate::{
95 : pgdatadir_mapping::{AuxFilesDirectory, DirectoryKind},
96 : virtual_file::{MaybeFatalIo, VirtualFile},
97 : };
98 :
99 : use crate::config::PageServerConf;
100 : use crate::keyspace::{KeyPartitioning, KeySpace};
101 : use crate::metrics::{
102 : TimelineMetrics, MATERIALIZED_PAGE_CACHE_HIT, MATERIALIZED_PAGE_CACHE_HIT_DIRECT,
103 : };
104 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
105 : use crate::tenant::config::TenantConfOpt;
106 : use pageserver_api::key::{is_inherited_key, is_rel_fsm_block_key, is_rel_vm_block_key};
107 : use pageserver_api::reltag::RelTag;
108 : use pageserver_api::shard::ShardIndex;
109 :
110 : use postgres_connection::PgConnectionConfig;
111 : use postgres_ffi::to_pg_timestamp;
112 : use utils::{
113 : completion,
114 : generation::Generation,
115 : id::TimelineId,
116 : lsn::{AtomicLsn, Lsn, RecordLsn},
117 : seqwait::SeqWait,
118 : simple_rcu::{Rcu, RcuReadGuard},
119 : };
120 :
121 : use crate::page_cache;
122 : use crate::repository::GcResult;
123 : use crate::repository::{Key, Value};
124 : use crate::task_mgr;
125 : use crate::task_mgr::TaskKind;
126 : use crate::ZERO_PAGE;
127 :
128 : use self::delete::DeleteTimelineFlow;
129 : pub(super) use self::eviction_task::EvictionTaskTenantState;
130 : use self::eviction_task::EvictionTaskTimelineState;
131 : use self::layer_manager::LayerManager;
132 : use self::logical_size::LogicalSize;
133 : use self::walreceiver::{WalReceiver, WalReceiverConf};
134 :
135 : use super::secondary::heatmap::{HeatMapLayer, HeatMapTimeline};
136 : use super::{config::TenantConf, storage_layer::VectoredValueReconstructState};
137 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
138 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
139 : use super::{remote_timeline_client::RemoteTimelineClient, storage_layer::ReadableLayer};
140 :
141 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
142 : pub(super) enum FlushLoopState {
143 : NotStarted,
144 : Running {
145 : #[cfg(test)]
146 : expect_initdb_optimization: bool,
147 : #[cfg(test)]
148 : initdb_optimization_count: usize,
149 : },
150 : Exited,
151 : }
152 :
153 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
154 : pub enum ImageLayerCreationMode {
155 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
156 : Try,
157 : /// Force creating the image layers if possible. For now, no image layers will be created
158 : /// for metadata keys. Used in compaction code path with force flag enabled.
159 : Force,
160 : /// Initial ingestion of the data, and no data should be dropped in this function. This
161 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
162 : /// code path.
163 : Initial,
164 : }
165 :
166 : impl std::fmt::Display for ImageLayerCreationMode {
167 480 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
168 480 : write!(f, "{:?}", self)
169 480 : }
170 : }
171 :
172 : /// Wrapper for key range to provide reverse ordering by range length for BinaryHeap
173 : #[derive(Debug, Clone, PartialEq, Eq)]
174 : pub(crate) struct Hole {
175 : key_range: Range<Key>,
176 : coverage_size: usize,
177 : }
178 :
179 : impl Ord for Hole {
180 0 : fn cmp(&self, other: &Self) -> Ordering {
181 0 : other.coverage_size.cmp(&self.coverage_size) // inverse order
182 0 : }
183 : }
184 :
185 : impl PartialOrd for Hole {
186 0 : fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
187 0 : Some(self.cmp(other))
188 0 : }
189 : }
190 :
191 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
192 : /// Can be removed after all refactors are done.
193 28 : fn drop_rlock<T>(rlock: tokio::sync::OwnedRwLockReadGuard<T>) {
194 28 : drop(rlock)
195 28 : }
196 :
197 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
198 : /// Can be removed after all refactors are done.
199 508 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
200 508 : drop(rlock)
201 508 : }
202 :
203 : /// The outward-facing resources required to build a Timeline
204 : pub struct TimelineResources {
205 : pub remote_client: RemoteTimelineClient,
206 : pub deletion_queue_client: DeletionQueueClient,
207 : pub timeline_get_throttle: Arc<
208 : crate::tenant::throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>,
209 : >,
210 : }
211 :
212 : pub(crate) struct AuxFilesState {
213 : pub(crate) dir: Option<AuxFilesDirectory>,
214 : pub(crate) n_deltas: usize,
215 : }
216 :
217 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
218 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
219 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
220 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
221 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
222 : pub(crate) struct RelSizeCache {
223 : pub(crate) complete_as_of: Lsn,
224 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
225 : }
226 :
227 : pub struct Timeline {
228 : conf: &'static PageServerConf,
229 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
230 :
231 : myself: Weak<Self>,
232 :
233 : pub(crate) tenant_shard_id: TenantShardId,
234 : pub timeline_id: TimelineId,
235 :
236 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
237 : /// Never changes for the lifetime of this [`Timeline`] object.
238 : ///
239 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
240 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
241 : pub(crate) generation: Generation,
242 :
243 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
244 : /// to shards, and is constant through the lifetime of this Timeline.
245 : shard_identity: ShardIdentity,
246 :
247 : pub pg_version: u32,
248 :
249 : /// The tuple has two elements.
250 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
251 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
252 : ///
253 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
254 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
255 : ///
256 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
257 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
258 : /// `PersistentLayerDesc`'s.
259 : ///
260 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
261 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
262 : /// runtime, e.g., during page reconstruction.
263 : ///
264 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
265 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
266 : pub(crate) layers: Arc<tokio::sync::RwLock<LayerManager>>,
267 :
268 : last_freeze_at: AtomicLsn,
269 : // Atomic would be more appropriate here.
270 : last_freeze_ts: RwLock<Instant>,
271 :
272 : pub(crate) standby_horizon: AtomicLsn,
273 :
274 : // WAL redo manager. `None` only for broken tenants.
275 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
276 :
277 : /// Remote storage client.
278 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
279 : pub remote_client: Arc<RemoteTimelineClient>,
280 :
281 : // What page versions do we hold in the repository? If we get a
282 : // request > last_record_lsn, we need to wait until we receive all
283 : // the WAL up to the request. The SeqWait provides functions for
284 : // that. TODO: If we get a request for an old LSN, such that the
285 : // versions have already been garbage collected away, we should
286 : // throw an error, but we don't track that currently.
287 : //
288 : // last_record_lsn.load().last points to the end of last processed WAL record.
289 : //
290 : // We also remember the starting point of the previous record in
291 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
292 : // first WAL record when the node is started up. But here, we just
293 : // keep track of it.
294 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
295 :
296 : // All WAL records have been processed and stored durably on files on
297 : // local disk, up to this LSN. On crash and restart, we need to re-process
298 : // the WAL starting from this point.
299 : //
300 : // Some later WAL records might have been processed and also flushed to disk
301 : // already, so don't be surprised to see some, but there's no guarantee on
302 : // them yet.
303 : disk_consistent_lsn: AtomicLsn,
304 :
305 : // Parent timeline that this timeline was branched from, and the LSN
306 : // of the branch point.
307 : ancestor_timeline: Option<Arc<Timeline>>,
308 : ancestor_lsn: Lsn,
309 :
310 : pub(super) metrics: TimelineMetrics,
311 :
312 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
313 : // in `crate::page_service` writes these metrics.
314 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
315 :
316 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
317 :
318 : /// Ensures layers aren't frozen by checkpointer between
319 : /// [`Timeline::get_layer_for_write`] and layer reads.
320 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
321 : /// Must always be acquired before the layer map/individual layer lock
322 : /// to avoid deadlock.
323 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
324 :
325 : /// Used to avoid multiple `flush_loop` tasks running
326 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
327 :
328 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
329 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
330 : /// The flush cycle counter is sent back on the layer_flush_done channel when
331 : /// the flush finishes. You can use that to wait for the flush to finish.
332 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
333 : /// read by whoever sends an update
334 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
335 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
336 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
337 :
338 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
339 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
340 :
341 : // List of child timelines and their branch points. This is needed to avoid
342 : // garbage collecting data that is still needed by the child timelines.
343 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
344 :
345 : // It may change across major versions so for simplicity
346 : // keep it after running initdb for a timeline.
347 : // It is needed in checks when we want to error on some operations
348 : // when they are requested for pre-initdb lsn.
349 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
350 : // though let's keep them both for better error visibility.
351 : pub initdb_lsn: Lsn,
352 :
353 : /// When did we last calculate the partitioning? Make it pub to test cases.
354 : pub(super) partitioning: tokio::sync::Mutex<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
355 :
356 : /// Configuration: how often should the partitioning be recalculated.
357 : repartition_threshold: u64,
358 :
359 : last_image_layer_creation_check_at: AtomicLsn,
360 :
361 : /// Current logical size of the "datadir", at the last LSN.
362 : current_logical_size: LogicalSize,
363 :
364 : /// Information about the last processed message by the WAL receiver,
365 : /// or None if WAL receiver has not received anything for this timeline
366 : /// yet.
367 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
368 : pub walreceiver: Mutex<Option<WalReceiver>>,
369 :
370 : /// Relation size cache
371 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
372 :
373 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
374 :
375 : state: watch::Sender<TimelineState>,
376 :
377 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
378 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
379 : pub delete_progress: Arc<tokio::sync::Mutex<DeleteTimelineFlow>>,
380 :
381 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
382 :
383 : /// Load or creation time information about the disk_consistent_lsn and when the loading
384 : /// happened. Used for consumption metrics.
385 : pub(crate) loaded_at: (Lsn, SystemTime),
386 :
387 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
388 : pub(crate) gate: Gate,
389 :
390 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
391 : /// to the timeline should drop out when this token fires.
392 : pub(crate) cancel: CancellationToken,
393 :
394 : /// Make sure we only have one running compaction at a time in tests.
395 : ///
396 : /// Must only be taken in two places:
397 : /// - [`Timeline::compact`] (this file)
398 : /// - [`delete::delete_local_timeline_directory`]
399 : ///
400 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
401 : compaction_lock: tokio::sync::Mutex<()>,
402 :
403 : /// Make sure we only have one running gc at a time.
404 : ///
405 : /// Must only be taken in two places:
406 : /// - [`Timeline::gc`] (this file)
407 : /// - [`delete::delete_local_timeline_directory`]
408 : ///
409 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
410 : gc_lock: tokio::sync::Mutex<()>,
411 :
412 : /// Cloned from [`super::Tenant::timeline_get_throttle`] on construction.
413 : timeline_get_throttle: Arc<
414 : crate::tenant::throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>,
415 : >,
416 :
417 : /// Keep aux directory cache to avoid it's reconstruction on each update
418 : pub(crate) aux_files: tokio::sync::Mutex<AuxFilesState>,
419 :
420 : /// Size estimator for aux file v2
421 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
422 :
423 : /// Indicate whether aux file v2 storage is enabled.
424 : pub(crate) last_aux_file_policy: AtomicAuxFilePolicy,
425 : }
426 :
427 : pub struct WalReceiverInfo {
428 : pub wal_source_connconf: PgConnectionConfig,
429 : pub last_received_msg_lsn: Lsn,
430 : pub last_received_msg_ts: u128,
431 : }
432 :
433 : /// Information about how much history needs to be retained, needed by
434 : /// Garbage Collection.
435 : #[derive(Default)]
436 : pub(crate) struct GcInfo {
437 : /// Specific LSNs that are needed.
438 : ///
439 : /// Currently, this includes all points where child branches have
440 : /// been forked off from. In the future, could also include
441 : /// explicit user-defined snapshot points.
442 : pub(crate) retain_lsns: Vec<Lsn>,
443 :
444 : /// The cutoff coordinates, which are combined by selecting the minimum.
445 : pub(crate) cutoffs: GcCutoffs,
446 : }
447 :
448 : impl GcInfo {
449 224 : pub(crate) fn min_cutoff(&self) -> Lsn {
450 224 : self.cutoffs.select_min()
451 224 : }
452 : }
453 :
454 : /// The `GcInfo` component describing which Lsns need to be retained.
455 : #[derive(Debug)]
456 : pub(crate) struct GcCutoffs {
457 : /// Keep everything newer than this point.
458 : ///
459 : /// This is calculated by subtracting 'gc_horizon' setting from
460 : /// last-record LSN
461 : ///
462 : /// FIXME: is this inclusive or exclusive?
463 : pub(crate) horizon: Lsn,
464 :
465 : /// In addition to 'retain_lsns' and 'horizon_cutoff', keep everything newer than this
466 : /// point.
467 : ///
468 : /// This is calculated by finding a number such that a record is needed for PITR
469 : /// if only if its LSN is larger than 'pitr_cutoff'.
470 : pub(crate) pitr: Lsn,
471 : }
472 :
473 : impl Default for GcCutoffs {
474 352 : fn default() -> Self {
475 352 : Self {
476 352 : horizon: Lsn::INVALID,
477 352 : pitr: Lsn::INVALID,
478 352 : }
479 352 : }
480 : }
481 :
482 : impl GcCutoffs {
483 224 : fn select_min(&self) -> Lsn {
484 224 : std::cmp::min(self.horizon, self.pitr)
485 224 : }
486 : }
487 :
488 : pub(crate) struct TimelineVisitOutcome {
489 : completed_keyspace: KeySpace,
490 : image_covered_keyspace: KeySpace,
491 : }
492 :
493 : /// An error happened in a get() operation.
494 2 : #[derive(thiserror::Error, Debug)]
495 : pub(crate) enum PageReconstructError {
496 : #[error(transparent)]
497 : Other(#[from] anyhow::Error),
498 :
499 : #[error("Ancestor LSN wait error: {0}")]
500 : AncestorLsnTimeout(#[from] WaitLsnError),
501 :
502 : #[error("timeline shutting down")]
503 : Cancelled,
504 :
505 : /// The ancestor of this is being stopped
506 : #[error("ancestor timeline {0} is being stopped")]
507 : AncestorStopping(TimelineId),
508 :
509 : /// An error happened replaying WAL records
510 : #[error(transparent)]
511 : WalRedo(anyhow::Error),
512 :
513 : #[error("{0}")]
514 : MissingKey(MissingKeyError),
515 : }
516 :
517 : impl GetVectoredError {
518 : #[cfg(test)]
519 6 : pub(crate) fn is_missing_key_error(&self) -> bool {
520 6 : matches!(self, Self::MissingKey(_))
521 6 : }
522 : }
523 :
524 : #[derive(Debug)]
525 : pub struct MissingKeyError {
526 : key: Key,
527 : shard: ShardNumber,
528 : cont_lsn: Lsn,
529 : request_lsn: Lsn,
530 : ancestor_lsn: Option<Lsn>,
531 : traversal_path: Vec<TraversalPathItem>,
532 : backtrace: Option<std::backtrace::Backtrace>,
533 : }
534 :
535 : impl std::fmt::Display for MissingKeyError {
536 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
537 0 : write!(
538 0 : f,
539 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
540 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
541 0 : )?;
542 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
543 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
544 0 : }
545 :
546 0 : if !self.traversal_path.is_empty() {
547 0 : writeln!(f)?;
548 0 : }
549 :
550 0 : for (r, c, l) in &self.traversal_path {
551 0 : writeln!(
552 0 : f,
553 0 : "layer traversal: result {:?}, cont_lsn {}, layer: {}",
554 0 : r, c, l,
555 0 : )?;
556 : }
557 :
558 0 : if let Some(ref backtrace) = self.backtrace {
559 0 : write!(f, "\n{}", backtrace)?;
560 0 : }
561 :
562 0 : Ok(())
563 0 : }
564 : }
565 :
566 : impl PageReconstructError {
567 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
568 0 : pub(crate) fn is_stopping(&self) -> bool {
569 0 : use PageReconstructError::*;
570 0 : match self {
571 0 : Other(_) => false,
572 0 : AncestorLsnTimeout(_) => false,
573 0 : Cancelled | AncestorStopping(_) => true,
574 0 : WalRedo(_) => false,
575 0 : MissingKey { .. } => false,
576 : }
577 0 : }
578 : }
579 :
580 0 : #[derive(thiserror::Error, Debug)]
581 : enum CreateImageLayersError {
582 : #[error("timeline shutting down")]
583 : Cancelled,
584 :
585 : #[error(transparent)]
586 : GetVectoredError(GetVectoredError),
587 :
588 : #[error(transparent)]
589 : PageReconstructError(PageReconstructError),
590 :
591 : #[error(transparent)]
592 : Other(#[from] anyhow::Error),
593 : }
594 :
595 0 : #[derive(thiserror::Error, Debug)]
596 : enum FlushLayerError {
597 : /// Timeline cancellation token was cancelled
598 : #[error("timeline shutting down")]
599 : Cancelled,
600 :
601 : #[error(transparent)]
602 : CreateImageLayersError(CreateImageLayersError),
603 :
604 : #[error(transparent)]
605 : Other(#[from] anyhow::Error),
606 : }
607 :
608 0 : #[derive(thiserror::Error, Debug)]
609 : pub(crate) enum GetVectoredError {
610 : #[error("timeline shutting down")]
611 : Cancelled,
612 :
613 : #[error("Requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
614 : Oversized(u64),
615 :
616 : #[error("Requested at invalid LSN: {0}")]
617 : InvalidLsn(Lsn),
618 :
619 : #[error("Requested key not found: {0}")]
620 : MissingKey(MissingKeyError),
621 :
622 : #[error(transparent)]
623 : GetReadyAncestorError(GetReadyAncestorError),
624 :
625 : #[error(transparent)]
626 : Other(#[from] anyhow::Error),
627 : }
628 :
629 0 : #[derive(thiserror::Error, Debug)]
630 : pub(crate) enum GetReadyAncestorError {
631 : #[error("ancestor timeline {0} is being stopped")]
632 : AncestorStopping(TimelineId),
633 :
634 : #[error("Ancestor LSN wait error: {0}")]
635 : AncestorLsnTimeout(#[from] WaitLsnError),
636 :
637 : #[error("Cancelled")]
638 : Cancelled,
639 :
640 : #[error(transparent)]
641 : Other(#[from] anyhow::Error),
642 : }
643 :
644 : #[derive(Clone, Copy)]
645 : pub enum LogicalSizeCalculationCause {
646 : Initial,
647 : ConsumptionMetricsSyntheticSize,
648 : EvictionTaskImitation,
649 : TenantSizeHandler,
650 : }
651 :
652 : pub enum GetLogicalSizePriority {
653 : User,
654 : Background,
655 : }
656 :
657 0 : #[derive(enumset::EnumSetType)]
658 : pub(crate) enum CompactFlags {
659 : ForceRepartition,
660 : ForceImageLayerCreation,
661 : }
662 :
663 : impl std::fmt::Debug for Timeline {
664 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
665 0 : write!(f, "Timeline<{}>", self.timeline_id)
666 0 : }
667 : }
668 :
669 0 : #[derive(thiserror::Error, Debug)]
670 : pub(crate) enum WaitLsnError {
671 : // Called on a timeline which is shutting down
672 : #[error("Shutdown")]
673 : Shutdown,
674 :
675 : // Called on an timeline not in active state or shutting down
676 : #[error("Bad state (not active)")]
677 : BadState,
678 :
679 : // Timeout expired while waiting for LSN to catch up with goal.
680 : #[error("{0}")]
681 : Timeout(String),
682 : }
683 :
684 : // The impls below achieve cancellation mapping for errors.
685 : // Perhaps there's a way of achieving this with less cruft.
686 :
687 : impl From<CreateImageLayersError> for CompactionError {
688 0 : fn from(e: CreateImageLayersError) -> Self {
689 0 : match e {
690 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
691 0 : _ => CompactionError::Other(e.into()),
692 : }
693 0 : }
694 : }
695 :
696 : impl From<CreateImageLayersError> for FlushLayerError {
697 0 : fn from(e: CreateImageLayersError) -> Self {
698 0 : match e {
699 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
700 0 : any => FlushLayerError::CreateImageLayersError(any),
701 : }
702 0 : }
703 : }
704 :
705 : impl From<PageReconstructError> for CreateImageLayersError {
706 0 : fn from(e: PageReconstructError) -> Self {
707 0 : match e {
708 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
709 0 : _ => CreateImageLayersError::PageReconstructError(e),
710 : }
711 0 : }
712 : }
713 :
714 : impl From<GetVectoredError> for CreateImageLayersError {
715 0 : fn from(e: GetVectoredError) -> Self {
716 0 : match e {
717 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
718 0 : _ => CreateImageLayersError::GetVectoredError(e),
719 : }
720 0 : }
721 : }
722 :
723 : impl From<GetVectoredError> for PageReconstructError {
724 0 : fn from(e: GetVectoredError) -> Self {
725 0 : match e {
726 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
727 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
728 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
729 0 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
730 0 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
731 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
732 : }
733 0 : }
734 : }
735 :
736 : impl From<GetReadyAncestorError> for PageReconstructError {
737 2 : fn from(e: GetReadyAncestorError) -> Self {
738 2 : use GetReadyAncestorError::*;
739 2 : match e {
740 0 : AncestorStopping(tid) => PageReconstructError::AncestorStopping(tid),
741 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
742 0 : Cancelled => PageReconstructError::Cancelled,
743 2 : Other(other) => PageReconstructError::Other(other),
744 : }
745 2 : }
746 : }
747 :
748 : #[derive(
749 : Eq,
750 : PartialEq,
751 : Debug,
752 : Copy,
753 : Clone,
754 160 : strum_macros::EnumString,
755 0 : strum_macros::Display,
756 0 : serde_with::DeserializeFromStr,
757 : serde_with::SerializeDisplay,
758 : )]
759 : #[strum(serialize_all = "kebab-case")]
760 : pub enum GetVectoredImpl {
761 : Sequential,
762 : Vectored,
763 : }
764 :
765 : #[derive(
766 : Eq,
767 : PartialEq,
768 : Debug,
769 : Copy,
770 : Clone,
771 160 : strum_macros::EnumString,
772 0 : strum_macros::Display,
773 0 : serde_with::DeserializeFromStr,
774 : serde_with::SerializeDisplay,
775 : )]
776 : #[strum(serialize_all = "kebab-case")]
777 : pub enum GetImpl {
778 : Legacy,
779 : Vectored,
780 : }
781 :
782 : pub(crate) enum WaitLsnWaiter<'a> {
783 : Timeline(&'a Timeline),
784 : Tenant,
785 : PageService,
786 : }
787 :
788 : /// Argument to [`Timeline::shutdown`].
789 : #[derive(Debug, Clone, Copy)]
790 : pub(crate) enum ShutdownMode {
791 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
792 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
793 : ///
794 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
795 : /// the call to [`Timeline::shutdown`].
796 : FreezeAndFlush,
797 : /// Shut down immediately, without waiting for any open layers to flush.
798 : Hard,
799 : }
800 :
801 : struct ImageLayerCreationOutcome {
802 : image: Option<ResidentLayer>,
803 : next_start_key: Key,
804 : }
805 :
806 : /// Public interface functions
807 : impl Timeline {
808 : /// Get the LSN where this branch was created
809 8 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
810 8 : self.ancestor_lsn
811 8 : }
812 :
813 : /// Get the ancestor's timeline id
814 3314 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
815 3314 : self.ancestor_timeline
816 3314 : .as_ref()
817 3314 : .map(|ancestor| ancestor.timeline_id)
818 3314 : }
819 :
820 : /// Lock and get timeline's GC cutoff
821 983 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
822 983 : self.latest_gc_cutoff_lsn.read()
823 983 : }
824 :
825 : /// Look up given page version.
826 : ///
827 : /// If a remote layer file is needed, it is downloaded as part of this
828 : /// call.
829 : ///
830 : /// This method enforces [`Self::timeline_get_throttle`] internally.
831 : ///
832 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
833 : /// abstraction above this needs to store suitable metadata to track what
834 : /// data exists with what keys, in separate metadata entries. If a
835 : /// non-existent key is requested, we may incorrectly return a value from
836 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
837 : /// non-existing key.
838 : ///
839 : /// # Cancel-Safety
840 : ///
841 : /// This method is cancellation-safe.
842 : #[inline(always)]
843 624021 : pub(crate) async fn get(
844 624021 : &self,
845 624021 : key: Key,
846 624021 : lsn: Lsn,
847 624021 : ctx: &RequestContext,
848 624021 : ) -> Result<Bytes, PageReconstructError> {
849 624021 : if !lsn.is_valid() {
850 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
851 624021 : }
852 624021 :
853 624021 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
854 624021 : // already checked the key against the shard_identity when looking up the Timeline from
855 624021 : // page_service.
856 624021 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
857 :
858 624021 : self.timeline_get_throttle.throttle(ctx, 1).await;
859 :
860 : // Check the page cache. We will get back the most recent page with lsn <= `lsn`.
861 : // The cached image can be returned directly if there is no WAL between the cached image
862 : // and requested LSN. The cached image can also be used to reduce the amount of WAL needed
863 : // for redo.
864 624021 : let cached_page_img = match self.lookup_cached_page(&key, lsn, ctx).await {
865 0 : Some((cached_lsn, cached_img)) => {
866 0 : match cached_lsn.cmp(&lsn) {
867 0 : Ordering::Less => {} // there might be WAL between cached_lsn and lsn, we need to check
868 : Ordering::Equal => {
869 0 : MATERIALIZED_PAGE_CACHE_HIT_DIRECT.inc();
870 0 : return Ok(cached_img); // exact LSN match, return the image
871 : }
872 : Ordering::Greater => {
873 0 : unreachable!("the returned lsn should never be after the requested lsn")
874 : }
875 : }
876 0 : Some((cached_lsn, cached_img))
877 : }
878 624021 : None => None,
879 : };
880 :
881 624021 : match self.conf.get_impl {
882 : GetImpl::Legacy => {
883 624021 : let reconstruct_state = ValueReconstructState {
884 624021 : records: Vec::new(),
885 624021 : img: cached_page_img,
886 624021 : };
887 624021 :
888 624021 : self.get_impl(key, lsn, reconstruct_state, ctx).await
889 : }
890 : GetImpl::Vectored => {
891 0 : let keyspace = KeySpace {
892 0 : ranges: vec![key..key.next()],
893 0 : };
894 0 :
895 0 : // Initialise the reconstruct state for the key with the cache
896 0 : // entry returned above.
897 0 : let mut reconstruct_state = ValuesReconstructState::new();
898 0 :
899 0 : // Only add the cached image to the reconstruct state when it exists.
900 0 : if cached_page_img.is_some() {
901 0 : let mut key_state = VectoredValueReconstructState::default();
902 0 : key_state.img = cached_page_img;
903 0 : reconstruct_state.keys.insert(key, Ok(key_state));
904 0 : }
905 :
906 0 : let vectored_res = self
907 0 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
908 0 : .await;
909 :
910 0 : if self.conf.validate_vectored_get {
911 0 : self.validate_get_vectored_impl(&vectored_res, keyspace, lsn, ctx)
912 0 : .await;
913 0 : }
914 :
915 0 : let key_value = vectored_res?.pop_first();
916 0 : match key_value {
917 0 : Some((got_key, value)) => {
918 0 : if got_key != key {
919 0 : error!(
920 0 : "Expected {}, but singular vectored get returned {}",
921 : key, got_key
922 : );
923 0 : Err(PageReconstructError::Other(anyhow!(
924 0 : "Singular vectored get returned wrong key"
925 0 : )))
926 : } else {
927 0 : value
928 : }
929 : }
930 0 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
931 0 : key,
932 0 : shard: self.shard_identity.get_shard_number(&key),
933 0 : cont_lsn: Lsn(0),
934 0 : request_lsn: lsn,
935 0 : ancestor_lsn: None,
936 0 : traversal_path: Vec::new(),
937 0 : backtrace: None,
938 0 : })),
939 : }
940 : }
941 : }
942 624021 : }
943 :
944 : /// Not subject to [`Self::timeline_get_throttle`].
945 625369 : async fn get_impl(
946 625369 : &self,
947 625369 : key: Key,
948 625369 : lsn: Lsn,
949 625369 : mut reconstruct_state: ValueReconstructState,
950 625369 : ctx: &RequestContext,
951 625369 : ) -> Result<Bytes, PageReconstructError> {
952 625369 : // XXX: structured stats collection for layer eviction here.
953 625369 : trace!(
954 0 : "get page request for {}@{} from task kind {:?}",
955 0 : key,
956 0 : lsn,
957 0 : ctx.task_kind()
958 : );
959 :
960 625369 : let timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
961 625369 : .for_get_kind(GetKind::Singular)
962 625369 : .start_timer();
963 625369 : let path = self
964 625369 : .get_reconstruct_data(key, lsn, &mut reconstruct_state, ctx)
965 43615 : .await?;
966 625245 : timer.stop_and_record();
967 625245 :
968 625245 : let start = Instant::now();
969 625245 : let res = self.reconstruct_value(key, lsn, reconstruct_state).await;
970 625245 : let elapsed = start.elapsed();
971 625245 : crate::metrics::RECONSTRUCT_TIME
972 625245 : .for_get_kind(GetKind::Singular)
973 625245 : .observe(elapsed.as_secs_f64());
974 625245 :
975 625245 : if cfg!(feature = "testing") && res.is_err() {
976 : // it can only be walredo issue
977 : use std::fmt::Write;
978 :
979 0 : let mut msg = String::new();
980 0 :
981 0 : path.into_iter().for_each(|(res, cont_lsn, layer)| {
982 0 : writeln!(
983 0 : msg,
984 0 : "- layer traversal: result {res:?}, cont_lsn {cont_lsn}, layer: {}",
985 0 : layer,
986 0 : )
987 0 : .expect("string grows")
988 0 : });
989 0 :
990 0 : // this is to rule out or provide evidence that we could in some cases read a duplicate
991 0 : // walrecord
992 0 : tracing::info!("walredo failed, path:\n{msg}");
993 625245 : }
994 :
995 625245 : res
996 625369 : }
997 :
998 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
999 :
1000 : /// Look up multiple page versions at a given LSN
1001 : ///
1002 : /// This naive implementation will be replaced with a more efficient one
1003 : /// which actually vectorizes the read path.
1004 774 : pub(crate) async fn get_vectored(
1005 774 : &self,
1006 774 : keyspace: KeySpace,
1007 774 : lsn: Lsn,
1008 774 : ctx: &RequestContext,
1009 774 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1010 774 : if !lsn.is_valid() {
1011 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1012 774 : }
1013 774 :
1014 774 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1015 774 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1016 0 : return Err(GetVectoredError::Oversized(key_count));
1017 774 : }
1018 :
1019 1548 : for range in &keyspace.ranges {
1020 774 : let mut key = range.start;
1021 1800 : while key != range.end {
1022 1026 : assert!(!self.shard_identity.is_key_disposable(&key));
1023 1026 : key = key.next();
1024 : }
1025 : }
1026 :
1027 774 : trace!(
1028 0 : "get vectored request for {:?}@{} from task kind {:?} will use {} implementation",
1029 0 : keyspace,
1030 0 : lsn,
1031 0 : ctx.task_kind(),
1032 : self.conf.get_vectored_impl
1033 : );
1034 :
1035 774 : let start = crate::metrics::GET_VECTORED_LATENCY
1036 774 : .for_task_kind(ctx.task_kind())
1037 774 : .map(|metric| (metric, Instant::now()));
1038 :
1039 : // start counting after throttle so that throttle time
1040 : // is always less than observation time
1041 774 : let throttled = self
1042 774 : .timeline_get_throttle
1043 774 : .throttle(ctx, key_count as usize)
1044 0 : .await;
1045 :
1046 774 : let res = match self.conf.get_vectored_impl {
1047 : GetVectoredImpl::Sequential => {
1048 774 : self.get_vectored_sequential_impl(keyspace, lsn, ctx).await
1049 : }
1050 : GetVectoredImpl::Vectored => {
1051 0 : let vectored_res = self
1052 0 : .get_vectored_impl(
1053 0 : keyspace.clone(),
1054 0 : lsn,
1055 0 : &mut ValuesReconstructState::new(),
1056 0 : ctx,
1057 0 : )
1058 0 : .await;
1059 :
1060 0 : if self.conf.validate_vectored_get {
1061 0 : self.validate_get_vectored_impl(&vectored_res, keyspace, lsn, ctx)
1062 0 : .await;
1063 0 : }
1064 :
1065 0 : vectored_res
1066 : }
1067 : };
1068 :
1069 774 : if let Some((metric, start)) = start {
1070 0 : let elapsed = start.elapsed();
1071 0 : let ex_throttled = if let Some(throttled) = throttled {
1072 0 : elapsed.checked_sub(throttled)
1073 : } else {
1074 0 : Some(elapsed)
1075 : };
1076 :
1077 0 : if let Some(ex_throttled) = ex_throttled {
1078 0 : metric.observe(ex_throttled.as_secs_f64());
1079 0 : } else {
1080 0 : use utils::rate_limit::RateLimit;
1081 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1082 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
1083 0 : let mut rate_limit = LOGGED.lock().unwrap();
1084 0 : rate_limit.call(|| {
1085 0 : warn!("error deducting time spent throttled; this message is logged at a global rate limit");
1086 0 : });
1087 0 : }
1088 774 : }
1089 :
1090 774 : res
1091 774 : }
1092 :
1093 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1094 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1095 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1096 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1097 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1098 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1099 : /// the scan operation will not cause OOM in the future.
1100 : #[allow(dead_code)]
1101 8 : pub(crate) async fn scan(
1102 8 : &self,
1103 8 : keyspace: KeySpace,
1104 8 : lsn: Lsn,
1105 8 : ctx: &RequestContext,
1106 8 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1107 8 : if !lsn.is_valid() {
1108 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1109 8 : }
1110 8 :
1111 8 : trace!(
1112 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1113 0 : keyspace,
1114 0 : lsn,
1115 0 : ctx.task_kind()
1116 : );
1117 :
1118 : // We should generalize this into Keyspace::contains in the future.
1119 16 : for range in &keyspace.ranges {
1120 8 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1121 8 : || range.end.field1 > METADATA_KEY_END_PREFIX
1122 : {
1123 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1124 0 : "only metadata keyspace can be scanned"
1125 0 : )));
1126 8 : }
1127 : }
1128 :
1129 8 : let start = crate::metrics::SCAN_LATENCY
1130 8 : .for_task_kind(ctx.task_kind())
1131 8 : .map(ScanLatencyOngoingRecording::start_recording);
1132 :
1133 : // start counting after throttle so that throttle time
1134 : // is always less than observation time
1135 8 : let throttled = self
1136 8 : .timeline_get_throttle
1137 8 : // assume scan = 1 quota for now until we find a better way to process this
1138 8 : .throttle(ctx, 1)
1139 0 : .await;
1140 :
1141 8 : let vectored_res = self
1142 8 : .get_vectored_impl(
1143 8 : keyspace.clone(),
1144 8 : lsn,
1145 8 : &mut ValuesReconstructState::default(),
1146 8 : ctx,
1147 8 : )
1148 0 : .await;
1149 :
1150 8 : if let Some(recording) = start {
1151 0 : recording.observe(throttled);
1152 8 : }
1153 :
1154 8 : vectored_res
1155 8 : }
1156 :
1157 : /// Not subject to [`Self::timeline_get_throttle`].
1158 786 : pub(super) async fn get_vectored_sequential_impl(
1159 786 : &self,
1160 786 : keyspace: KeySpace,
1161 786 : lsn: Lsn,
1162 786 : ctx: &RequestContext,
1163 786 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1164 786 : let mut values = BTreeMap::new();
1165 :
1166 1572 : for range in keyspace.ranges {
1167 786 : let mut key = range.start;
1168 2134 : while key != range.end {
1169 1348 : let block = self
1170 1348 : .get_impl(key, lsn, ValueReconstructState::default(), ctx)
1171 37 : .await;
1172 :
1173 : use PageReconstructError::*;
1174 0 : match block {
1175 : Err(Cancelled | AncestorStopping(_)) => {
1176 0 : return Err(GetVectoredError::Cancelled)
1177 : }
1178 : Err(MissingKey(_))
1179 2 : if NON_INHERITED_RANGE.contains(&key)
1180 2 : || NON_INHERITED_SPARSE_RANGE.contains(&key) =>
1181 2 : {
1182 2 : // Ignore missing key error for aux key range. TODO: currently, we assume non_inherited_range == aux_key_range.
1183 2 : // When we add more types of keys into the page server, we should revisit this part of code and throw errors
1184 2 : // accordingly.
1185 2 : key = key.next();
1186 2 : }
1187 0 : Err(MissingKey(err)) => {
1188 0 : return Err(GetVectoredError::MissingKey(err));
1189 : }
1190 0 : Err(Other(err))
1191 0 : if err
1192 0 : .to_string()
1193 0 : .contains("downloading evicted layer file failed") =>
1194 0 : {
1195 0 : return Err(GetVectoredError::Other(err))
1196 : }
1197 0 : Err(Other(err))
1198 0 : if err
1199 0 : .chain()
1200 0 : .any(|cause| cause.to_string().contains("layer loading failed")) =>
1201 0 : {
1202 0 : // The intent here is to achieve error parity with the vectored read path.
1203 0 : // When vectored read fails to load a layer it fails the whole read, hence
1204 0 : // we mimic this behaviour here to keep the validation happy.
1205 0 : return Err(GetVectoredError::Other(err));
1206 : }
1207 1346 : _ => {
1208 1346 : values.insert(key, block);
1209 1346 : key = key.next();
1210 1346 : }
1211 : }
1212 : }
1213 : }
1214 :
1215 786 : Ok(values)
1216 786 : }
1217 :
1218 470 : pub(super) async fn get_vectored_impl(
1219 470 : &self,
1220 470 : keyspace: KeySpace,
1221 470 : lsn: Lsn,
1222 470 : reconstruct_state: &mut ValuesReconstructState,
1223 470 : ctx: &RequestContext,
1224 470 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1225 470 : let get_kind = if keyspace.total_raw_size() == 1 {
1226 56 : GetKind::Singular
1227 : } else {
1228 414 : GetKind::Vectored
1229 : };
1230 :
1231 470 : let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
1232 470 : .for_get_kind(get_kind)
1233 470 : .start_timer();
1234 470 : self.get_vectored_reconstruct_data(keyspace, lsn, reconstruct_state, ctx)
1235 11519 : .await?;
1236 460 : get_data_timer.stop_and_record();
1237 460 :
1238 460 : let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
1239 460 : .for_get_kind(get_kind)
1240 460 : .start_timer();
1241 460 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
1242 460 : let layers_visited = reconstruct_state.get_layers_visited();
1243 :
1244 78418 : for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
1245 78418 : match res {
1246 0 : Err(err) => {
1247 0 : results.insert(key, Err(err));
1248 0 : }
1249 78418 : Ok(state) => {
1250 78418 : let state = ValueReconstructState::from(state);
1251 :
1252 78418 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
1253 78418 : results.insert(key, reconstruct_res);
1254 : }
1255 : }
1256 : }
1257 460 : reconstruct_timer.stop_and_record();
1258 460 :
1259 460 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1260 460 : // when they're missing. Instead they are omitted from the resulting btree
1261 460 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1262 460 : // to avoid infinite results.
1263 460 : if !results.is_empty() {
1264 138 : // Note that this is an approximation. Tracking the exact number of layers visited
1265 138 : // per key requires virtually unbounded memory usage and is inefficient
1266 138 : // (i.e. segment tree tracking each range queried from a layer)
1267 138 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED
1268 138 : .observe(layers_visited as f64 / results.len() as f64);
1269 322 : }
1270 :
1271 460 : Ok(results)
1272 470 : }
1273 :
1274 : /// Not subject to [`Self::timeline_get_throttle`].
1275 12 : pub(super) async fn validate_get_vectored_impl(
1276 12 : &self,
1277 12 : vectored_res: &Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError>,
1278 12 : keyspace: KeySpace,
1279 12 : lsn: Lsn,
1280 12 : ctx: &RequestContext,
1281 12 : ) {
1282 12 : if keyspace.overlaps(&Key::metadata_key_range()) {
1283 : // skip validation for metadata key range
1284 0 : return;
1285 12 : }
1286 :
1287 12 : let sequential_res = self
1288 12 : .get_vectored_sequential_impl(keyspace.clone(), lsn, ctx)
1289 19 : .await;
1290 :
1291 0 : fn errors_match(lhs: &GetVectoredError, rhs: &GetVectoredError) -> bool {
1292 0 : use GetVectoredError::*;
1293 0 : match (lhs, rhs) {
1294 0 : (Oversized(l), Oversized(r)) => l == r,
1295 0 : (InvalidLsn(l), InvalidLsn(r)) => l == r,
1296 0 : (MissingKey(l), MissingKey(r)) => l.key == r.key,
1297 0 : (GetReadyAncestorError(_), GetReadyAncestorError(_)) => true,
1298 0 : (Other(_), Other(_)) => true,
1299 0 : _ => false,
1300 : }
1301 0 : }
1302 :
1303 12 : match (&sequential_res, vectored_res) {
1304 0 : (Err(GetVectoredError::Cancelled), _) => {},
1305 0 : (_, Err(GetVectoredError::Cancelled)) => {},
1306 0 : (Err(seq_err), Ok(_)) => {
1307 0 : panic!(concat!("Sequential get failed with {}, but vectored get did not",
1308 0 : " - keyspace={:?} lsn={}"),
1309 0 : seq_err, keyspace, lsn) },
1310 0 : (Ok(_), Err(GetVectoredError::GetReadyAncestorError(GetReadyAncestorError::AncestorLsnTimeout(_)))) => {
1311 0 : // Sequential get runs after vectored get, so it is possible for the later
1312 0 : // to time out while waiting for its ancestor's Lsn to become ready and for the
1313 0 : // former to succeed (it essentially has a doubled wait time).
1314 0 : },
1315 0 : (Ok(_), Err(vec_err)) => {
1316 0 : panic!(concat!("Vectored get failed with {}, but sequential get did not",
1317 0 : " - keyspace={:?} lsn={}"),
1318 0 : vec_err, keyspace, lsn) },
1319 0 : (Err(seq_err), Err(vec_err)) => {
1320 0 : assert!(errors_match(seq_err, vec_err),
1321 0 : "Mismatched errors: {seq_err} != {vec_err} - keyspace={keyspace:?} lsn={lsn}")},
1322 12 : (Ok(seq_values), Ok(vec_values)) => {
1323 320 : seq_values.iter().zip(vec_values.iter()).for_each(|((seq_key, seq_res), (vec_key, vec_res))| {
1324 320 : assert_eq!(seq_key, vec_key);
1325 320 : match (seq_res, vec_res) {
1326 320 : (Ok(seq_blob), Ok(vec_blob)) => {
1327 320 : Self::validate_key_equivalence(seq_key, &keyspace, lsn, seq_blob, vec_blob);
1328 320 : },
1329 0 : (Err(err), Ok(_)) => {
1330 0 : panic!(
1331 0 : concat!("Sequential get failed with {} for key {}, but vectored get did not",
1332 0 : " - keyspace={:?} lsn={}"),
1333 0 : err, seq_key, keyspace, lsn) },
1334 0 : (Ok(_), Err(err)) => {
1335 0 : panic!(
1336 0 : concat!("Vectored get failed with {} for key {}, but sequential get did not",
1337 0 : " - keyspace={:?} lsn={}"),
1338 0 : err, seq_key, keyspace, lsn) },
1339 0 : (Err(_), Err(_)) => {}
1340 : }
1341 320 : })
1342 : }
1343 : }
1344 12 : }
1345 :
1346 320 : fn validate_key_equivalence(
1347 320 : key: &Key,
1348 320 : keyspace: &KeySpace,
1349 320 : lsn: Lsn,
1350 320 : seq: &Bytes,
1351 320 : vec: &Bytes,
1352 320 : ) {
1353 320 : if *key == AUX_FILES_KEY {
1354 : // The value reconstruct of AUX_FILES_KEY from records is not deterministic
1355 : // since it uses a hash map under the hood. Hence, deserialise both results
1356 : // before comparing.
1357 0 : let seq_aux_dir_res = AuxFilesDirectory::des(seq);
1358 0 : let vec_aux_dir_res = AuxFilesDirectory::des(vec);
1359 0 : match (&seq_aux_dir_res, &vec_aux_dir_res) {
1360 0 : (Ok(seq_aux_dir), Ok(vec_aux_dir)) => {
1361 0 : assert_eq!(
1362 : seq_aux_dir, vec_aux_dir,
1363 0 : "Mismatch for key {} - keyspace={:?} lsn={}",
1364 : key, keyspace, lsn
1365 : );
1366 : }
1367 0 : (Err(_), Err(_)) => {}
1368 : _ => {
1369 0 : panic!("Mismatch for {key}: {seq_aux_dir_res:?} != {vec_aux_dir_res:?}");
1370 : }
1371 : }
1372 : } else {
1373 : // All other keys should reconstruct deterministically, so we simply compare the blobs.
1374 320 : assert_eq!(
1375 : seq, vec,
1376 0 : "Image mismatch for key {key} - keyspace={keyspace:?} lsn={lsn}"
1377 : );
1378 : }
1379 320 : }
1380 :
1381 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1382 5078176 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1383 5078176 : self.last_record_lsn.load().last
1384 5078176 : }
1385 :
1386 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1387 0 : self.last_record_lsn.load().prev
1388 0 : }
1389 :
1390 : /// Atomically get both last and prev.
1391 224 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1392 224 : self.last_record_lsn.load()
1393 224 : }
1394 :
1395 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1396 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1397 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1398 0 : self.last_record_lsn.status_receiver()
1399 0 : }
1400 :
1401 1099 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1402 1099 : self.disk_consistent_lsn.load()
1403 1099 : }
1404 :
1405 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1406 : /// not validated with control plane yet.
1407 : /// See [`Self::get_remote_consistent_lsn_visible`].
1408 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1409 0 : self.remote_client.remote_consistent_lsn_projected()
1410 0 : }
1411 :
1412 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1413 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1414 : /// generation validation in the deletion queue.
1415 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1416 0 : self.remote_client.remote_consistent_lsn_visible()
1417 0 : }
1418 :
1419 : /// The sum of the file size of all historic layers in the layer map.
1420 : /// This method makes no distinction between local and remote layers.
1421 : /// Hence, the result **does not represent local filesystem usage**.
1422 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1423 0 : let guard = self.layers.read().await;
1424 0 : let layer_map = guard.layer_map();
1425 0 : let mut size = 0;
1426 0 : for l in layer_map.iter_historic_layers() {
1427 0 : size += l.file_size();
1428 0 : }
1429 0 : size
1430 0 : }
1431 :
1432 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1433 0 : self.metrics.resident_physical_size_get()
1434 0 : }
1435 :
1436 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1437 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1438 0 : }
1439 :
1440 : ///
1441 : /// Wait until WAL has been received and processed up to this LSN.
1442 : ///
1443 : /// You should call this before any of the other get_* or list_* functions. Calling
1444 : /// those functions with an LSN that has been processed yet is an error.
1445 : ///
1446 227444 : pub(crate) async fn wait_lsn(
1447 227444 : &self,
1448 227444 : lsn: Lsn,
1449 227444 : who_is_waiting: WaitLsnWaiter<'_>,
1450 227444 : ctx: &RequestContext, /* Prepare for use by cancellation */
1451 227444 : ) -> Result<(), WaitLsnError> {
1452 227444 : if self.cancel.is_cancelled() {
1453 0 : return Err(WaitLsnError::Shutdown);
1454 227444 : } else if !self.is_active() {
1455 0 : return Err(WaitLsnError::BadState);
1456 227444 : }
1457 227444 :
1458 227444 : if cfg!(debug_assertions) {
1459 227444 : match ctx.task_kind() {
1460 : TaskKind::WalReceiverManager
1461 : | TaskKind::WalReceiverConnectionHandler
1462 : | TaskKind::WalReceiverConnectionPoller => {
1463 0 : let is_myself = match who_is_waiting {
1464 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1465 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1466 : };
1467 0 : if is_myself {
1468 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1469 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1470 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1471 0 : }
1472 0 : } else {
1473 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1474 0 : // our walreceiver task can make progress independent of theirs
1475 0 : }
1476 : }
1477 227444 : _ => {}
1478 : }
1479 0 : }
1480 :
1481 227444 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1482 227444 :
1483 227444 : match self
1484 227444 : .last_record_lsn
1485 227444 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1486 0 : .await
1487 : {
1488 227444 : Ok(()) => Ok(()),
1489 0 : Err(e) => {
1490 0 : use utils::seqwait::SeqWaitError::*;
1491 0 : match e {
1492 0 : Shutdown => Err(WaitLsnError::Shutdown),
1493 : Timeout => {
1494 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1495 0 : drop(_timer);
1496 0 : let walreceiver_status = self.walreceiver_status();
1497 0 : Err(WaitLsnError::Timeout(format!(
1498 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1499 0 : lsn,
1500 0 : self.get_last_record_lsn(),
1501 0 : self.get_disk_consistent_lsn(),
1502 0 : walreceiver_status,
1503 0 : )))
1504 : }
1505 : }
1506 : }
1507 : }
1508 227444 : }
1509 :
1510 0 : pub(crate) fn walreceiver_status(&self) -> String {
1511 0 : match &*self.walreceiver.lock().unwrap() {
1512 0 : None => "stopping or stopped".to_string(),
1513 0 : Some(walreceiver) => match walreceiver.status() {
1514 0 : Some(status) => status.to_human_readable_string(),
1515 0 : None => "Not active".to_string(),
1516 : },
1517 : }
1518 0 : }
1519 :
1520 : /// Check that it is valid to request operations with that lsn.
1521 228 : pub(crate) fn check_lsn_is_in_scope(
1522 228 : &self,
1523 228 : lsn: Lsn,
1524 228 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1525 228 : ) -> anyhow::Result<()> {
1526 228 : ensure!(
1527 228 : lsn >= **latest_gc_cutoff_lsn,
1528 4 : "LSN {} is earlier than latest GC horizon {} (we might've already garbage collected needed data)",
1529 4 : lsn,
1530 4 : **latest_gc_cutoff_lsn,
1531 : );
1532 224 : Ok(())
1533 228 : }
1534 :
1535 : /// Flush to disk all data that was written with the put_* functions
1536 2076 : #[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))]
1537 : pub(crate) async fn freeze_and_flush(&self) -> anyhow::Result<()> {
1538 : self.freeze_and_flush0().await
1539 : }
1540 :
1541 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1542 : // polluting the span hierarchy.
1543 1038 : pub(crate) async fn freeze_and_flush0(&self) -> anyhow::Result<()> {
1544 1038 : let to_lsn = self.freeze_inmem_layer(false).await;
1545 1038 : self.flush_frozen_layers_and_wait(to_lsn).await
1546 1038 : }
1547 :
1548 : // Check if an open ephemeral layer should be closed: this provides
1549 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1550 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1551 : // ephemeral layer bytes has been breached.
1552 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1553 0 : let Ok(_write_guard) = self.write_lock.try_lock() else {
1554 : // If the write lock is held, there is an active wal receiver: rolling open layers
1555 : // is their responsibility while they hold this lock.
1556 0 : return;
1557 : };
1558 :
1559 0 : let Ok(layers_guard) = self.layers.try_read() else {
1560 : // Don't block if the layer lock is busy
1561 0 : return;
1562 : };
1563 :
1564 0 : let Some(open_layer) = &layers_guard.layer_map().open_layer else {
1565 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1566 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1567 : // that didn't result in writes to this shard.
1568 :
1569 : // Must not hold the layers lock while waiting for a flush.
1570 0 : drop(layers_guard);
1571 0 :
1572 0 : let last_record_lsn = self.get_last_record_lsn();
1573 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1574 0 : if last_record_lsn > disk_consistent_lsn {
1575 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1576 : // we are a sharded tenant and have skipped some WAL
1577 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1578 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1579 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1580 : // without any data ingested (yet) doesn't write a remote index as soon as it
1581 : // sees its LSN advance: we only do this if we've been layer-less
1582 : // for some time.
1583 0 : tracing::debug!(
1584 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1585 : disk_consistent_lsn,
1586 : last_record_lsn
1587 : );
1588 :
1589 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1590 0 : self.flush_frozen_layers_and_wait(last_record_lsn)
1591 0 : .await
1592 0 : .ok();
1593 0 : }
1594 0 : }
1595 :
1596 0 : return;
1597 : };
1598 :
1599 0 : let Some(current_size) = open_layer.try_len() else {
1600 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1601 : // read lock to get size should always succeed.
1602 0 : tracing::warn!("Lock conflict while reading size of open layer");
1603 0 : return;
1604 : };
1605 :
1606 0 : let current_lsn = self.get_last_record_lsn();
1607 :
1608 0 : let checkpoint_distance_override = open_layer.tick().await;
1609 :
1610 0 : if let Some(size_override) = checkpoint_distance_override {
1611 0 : if current_size > size_override {
1612 : // This is not harmful, but it only happens in relatively rare cases where
1613 : // time-based checkpoints are not happening fast enough to keep the amount of
1614 : // ephemeral data within configured limits. It's a sign of stress on the system.
1615 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1616 0 : }
1617 0 : }
1618 :
1619 0 : let checkpoint_distance =
1620 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1621 0 :
1622 0 : if self.should_roll(
1623 0 : current_size,
1624 0 : current_size,
1625 0 : checkpoint_distance,
1626 0 : self.get_last_record_lsn(),
1627 0 : self.last_freeze_at.load(),
1628 0 : open_layer.get_opened_at(),
1629 0 : ) {
1630 0 : match open_layer.info() {
1631 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1632 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1633 0 : // happens asynchronously in the background.
1634 0 : tracing::debug!(
1635 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1636 : );
1637 : }
1638 : InMemoryLayerInfo::Open { .. } => {
1639 : // Upgrade to a write lock and freeze the layer
1640 0 : drop(layers_guard);
1641 0 : let mut layers_guard = self.layers.write().await;
1642 0 : layers_guard
1643 0 : .try_freeze_in_memory_layer(current_lsn, &self.last_freeze_at)
1644 0 : .await;
1645 : }
1646 : }
1647 0 : self.flush_frozen_layers();
1648 0 : }
1649 0 : }
1650 :
1651 : /// Outermost timeline compaction operation; downloads needed layers.
1652 366 : pub(crate) async fn compact(
1653 366 : self: &Arc<Self>,
1654 366 : cancel: &CancellationToken,
1655 366 : flags: EnumSet<CompactFlags>,
1656 366 : ctx: &RequestContext,
1657 366 : ) -> Result<(), CompactionError> {
1658 366 : // most likely the cancellation token is from background task, but in tests it could be the
1659 366 : // request task as well.
1660 366 :
1661 366 : let prepare = async move {
1662 366 : let guard = self.compaction_lock.lock().await;
1663 :
1664 366 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1665 366 : BackgroundLoopKind::Compaction,
1666 366 : ctx,
1667 366 : )
1668 0 : .await;
1669 :
1670 366 : (guard, permit)
1671 366 : };
1672 :
1673 : // this wait probably never needs any "long time spent" logging, because we already nag if
1674 : // compaction task goes over it's period (20s) which is quite often in production.
1675 366 : let (_guard, _permit) = tokio::select! {
1676 : tuple = prepare => { tuple },
1677 : _ = self.cancel.cancelled() => return Ok(()),
1678 : _ = cancel.cancelled() => return Ok(()),
1679 : };
1680 :
1681 366 : let last_record_lsn = self.get_last_record_lsn();
1682 366 :
1683 366 : // Last record Lsn could be zero in case the timeline was just created
1684 366 : if !last_record_lsn.is_valid() {
1685 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1686 0 : return Ok(());
1687 366 : }
1688 366 :
1689 366 : match self.get_compaction_algorithm() {
1690 0 : CompactionAlgorithm::Tiered => self.compact_tiered(cancel, ctx).await,
1691 70424 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, flags, ctx).await,
1692 : }
1693 366 : }
1694 :
1695 : /// Mutate the timeline with a [`TimelineWriter`].
1696 5133116 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1697 5133116 : TimelineWriter {
1698 5133116 : tl: self,
1699 5133116 : write_guard: self.write_lock.lock().await,
1700 : }
1701 5133116 : }
1702 :
1703 0 : pub(crate) fn activate(
1704 0 : self: &Arc<Self>,
1705 0 : parent: Arc<crate::tenant::Tenant>,
1706 0 : broker_client: BrokerClientChannel,
1707 0 : background_jobs_can_start: Option<&completion::Barrier>,
1708 0 : ctx: &RequestContext,
1709 0 : ) {
1710 0 : if self.tenant_shard_id.is_shard_zero() {
1711 0 : // Logical size is only maintained accurately on shard zero.
1712 0 : self.spawn_initial_logical_size_computation_task(ctx);
1713 0 : }
1714 0 : self.launch_wal_receiver(ctx, broker_client);
1715 0 : self.set_state(TimelineState::Active);
1716 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1717 0 : }
1718 :
1719 : /// After this function returns, there are no timeline-scoped tasks are left running.
1720 : ///
1721 : /// The preferred pattern for is:
1722 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1723 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1724 : /// go the extra mile and keep track of JoinHandles
1725 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1726 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1727 : ///
1728 : /// For legacy reasons, we still have multiple tasks spawned using
1729 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1730 : /// We refer to these as "timeline-scoped task_mgr tasks".
1731 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1732 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1733 : /// or [`task_mgr::shutdown_watcher`].
1734 : /// We want to gradually convert the code base away from these.
1735 : ///
1736 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1737 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1738 : /// ones that aren't mentioned here):
1739 : /// - [`TaskKind::TimelineDeletionWorker`]
1740 : /// - NB: also used for tenant deletion
1741 : /// - [`TaskKind::RemoteUploadTask`]`
1742 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1743 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1744 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1745 : /// - [`TaskKind::Eviction`]
1746 : /// - [`TaskKind::LayerFlushTask`]
1747 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1748 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1749 8 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1750 8 : debug_assert_current_span_has_tenant_and_timeline_id();
1751 :
1752 8 : let try_freeze_and_flush = match mode {
1753 6 : ShutdownMode::FreezeAndFlush => true,
1754 2 : ShutdownMode::Hard => false,
1755 : };
1756 :
1757 : // Regardless of whether we're going to try_freeze_and_flush
1758 : // or not, stop ingesting any more data. Walreceiver only provides
1759 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1760 : // So, only after the self.gate.close() below will we know for sure that
1761 : // no walreceiver tasks are left.
1762 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1763 : // data during the call to `self.freeze_and_flush()` below.
1764 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1765 : // which is what we'd need to properly model early shutdown of the walreceiver
1766 : // task sub-tree before the other Timeline task sub-trees.
1767 8 : let walreceiver = self.walreceiver.lock().unwrap().take();
1768 8 : tracing::debug!(
1769 0 : is_some = walreceiver.is_some(),
1770 0 : "Waiting for WalReceiverManager..."
1771 : );
1772 8 : if let Some(walreceiver) = walreceiver {
1773 0 : walreceiver.cancel();
1774 8 : }
1775 : // ... and inform any waiters for newer LSNs that there won't be any.
1776 8 : self.last_record_lsn.shutdown();
1777 8 :
1778 8 : if try_freeze_and_flush {
1779 : // we shut down walreceiver above, so, we won't add anything more
1780 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1781 : // to reach the disk & upload queue, then shut the upload queue and
1782 : // wait for it to drain.
1783 6 : match self.freeze_and_flush().await {
1784 : Ok(_) => {
1785 : // drain the upload queue
1786 : // if we did not wait for completion here, it might be our shutdown process
1787 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1788 : // be spawned.
1789 : //
1790 : // what is problematic is the shutting down of RemoteTimelineClient, because
1791 : // obviously it does not make sense to stop while we wait for it, but what
1792 : // about corner cases like s3 suddenly hanging up?
1793 6 : self.remote_client.shutdown().await;
1794 : }
1795 0 : Err(e) => {
1796 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1797 0 : // we have some extra WAL replay to do next time the timeline starts.
1798 0 : warn!("failed to freeze and flush: {e:#}");
1799 : }
1800 : }
1801 2 : }
1802 :
1803 : // Signal any subscribers to our cancellation token to drop out
1804 8 : tracing::debug!("Cancelling CancellationToken");
1805 8 : self.cancel.cancel();
1806 8 :
1807 8 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1808 8 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1809 8 : self.remote_client.stop();
1810 8 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1811 8 : // to shut down the upload queue tasks.
1812 8 : // TODO: fix that, task management should be encapsulated inside remote_client.
1813 8 : task_mgr::shutdown_tasks(
1814 8 : Some(TaskKind::RemoteUploadTask),
1815 8 : Some(self.tenant_shard_id),
1816 8 : Some(self.timeline_id),
1817 8 : )
1818 0 : .await;
1819 :
1820 : // TODO: work toward making this a no-op. See this funciton's doc comment for more context.
1821 8 : tracing::debug!("Waiting for tasks...");
1822 8 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1823 :
1824 : // Finally wait until any gate-holders are complete.
1825 : //
1826 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1827 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1828 8 : self.gate.close().await;
1829 :
1830 8 : self.metrics.shutdown();
1831 8 : }
1832 :
1833 352 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1834 352 : match (self.current_state(), new_state) {
1835 352 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1836 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1837 : }
1838 0 : (st, TimelineState::Loading) => {
1839 0 : error!("ignoring transition from {st:?} into Loading state");
1840 : }
1841 0 : (TimelineState::Broken { .. }, new_state) => {
1842 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1843 : }
1844 : (TimelineState::Stopping, TimelineState::Active) => {
1845 0 : error!("Not activating a Stopping timeline");
1846 : }
1847 350 : (_, new_state) => {
1848 350 : self.state.send_replace(new_state);
1849 350 : }
1850 : }
1851 352 : }
1852 :
1853 2 : pub(crate) fn set_broken(&self, reason: String) {
1854 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1855 2 : let broken_state = TimelineState::Broken {
1856 2 : reason,
1857 2 : backtrace: backtrace_str,
1858 2 : };
1859 2 : self.set_state(broken_state);
1860 2 :
1861 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1862 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1863 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1864 2 : self.cancel.cancel();
1865 2 : }
1866 :
1867 229137 : pub(crate) fn current_state(&self) -> TimelineState {
1868 229137 : self.state.borrow().clone()
1869 229137 : }
1870 :
1871 6 : pub(crate) fn is_broken(&self) -> bool {
1872 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1873 6 : }
1874 :
1875 227664 : pub(crate) fn is_active(&self) -> bool {
1876 227664 : self.current_state() == TimelineState::Active
1877 227664 : }
1878 :
1879 1121 : pub(crate) fn is_stopping(&self) -> bool {
1880 1121 : self.current_state() == TimelineState::Stopping
1881 1121 : }
1882 :
1883 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1884 0 : self.state.subscribe()
1885 0 : }
1886 :
1887 227446 : pub(crate) async fn wait_to_become_active(
1888 227446 : &self,
1889 227446 : _ctx: &RequestContext, // Prepare for use by cancellation
1890 227446 : ) -> Result<(), TimelineState> {
1891 227446 : let mut receiver = self.state.subscribe();
1892 227446 : loop {
1893 227446 : let current_state = receiver.borrow().clone();
1894 227446 : match current_state {
1895 : TimelineState::Loading => {
1896 0 : receiver
1897 0 : .changed()
1898 0 : .await
1899 0 : .expect("holding a reference to self");
1900 : }
1901 : TimelineState::Active { .. } => {
1902 227444 : return Ok(());
1903 : }
1904 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1905 : // There's no chance the timeline can transition back into ::Active
1906 2 : return Err(current_state);
1907 : }
1908 : }
1909 : }
1910 227446 : }
1911 :
1912 0 : pub(crate) async fn layer_map_info(&self, reset: LayerAccessStatsReset) -> LayerMapInfo {
1913 0 : let guard = self.layers.read().await;
1914 0 : let layer_map = guard.layer_map();
1915 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
1916 0 : if let Some(open_layer) = &layer_map.open_layer {
1917 0 : in_memory_layers.push(open_layer.info());
1918 0 : }
1919 0 : for frozen_layer in &layer_map.frozen_layers {
1920 0 : in_memory_layers.push(frozen_layer.info());
1921 0 : }
1922 :
1923 0 : let mut historic_layers = Vec::new();
1924 0 : for historic_layer in layer_map.iter_historic_layers() {
1925 0 : let historic_layer = guard.get_from_desc(&historic_layer);
1926 0 : historic_layers.push(historic_layer.info(reset));
1927 0 : }
1928 :
1929 0 : LayerMapInfo {
1930 0 : in_memory_layers,
1931 0 : historic_layers,
1932 0 : }
1933 0 : }
1934 :
1935 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))]
1936 : pub(crate) async fn download_layer(
1937 : &self,
1938 : layer_file_name: &LayerName,
1939 : ) -> anyhow::Result<Option<bool>> {
1940 : let Some(layer) = self.find_layer(layer_file_name).await else {
1941 : return Ok(None);
1942 : };
1943 :
1944 : layer.download().await?;
1945 :
1946 : Ok(Some(true))
1947 : }
1948 :
1949 : /// Evict just one layer.
1950 : ///
1951 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
1952 0 : pub(crate) async fn evict_layer(
1953 0 : &self,
1954 0 : layer_file_name: &LayerName,
1955 0 : ) -> anyhow::Result<Option<bool>> {
1956 0 : let _gate = self
1957 0 : .gate
1958 0 : .enter()
1959 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
1960 :
1961 0 : let Some(local_layer) = self.find_layer(layer_file_name).await else {
1962 0 : return Ok(None);
1963 : };
1964 :
1965 : // curl has this by default
1966 0 : let timeout = std::time::Duration::from_secs(120);
1967 0 :
1968 0 : match local_layer.evict_and_wait(timeout).await {
1969 0 : Ok(()) => Ok(Some(true)),
1970 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
1971 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
1972 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
1973 : }
1974 0 : }
1975 :
1976 58 : fn should_roll(
1977 58 : &self,
1978 58 : layer_size: u64,
1979 58 : projected_layer_size: u64,
1980 58 : checkpoint_distance: u64,
1981 58 : projected_lsn: Lsn,
1982 58 : last_freeze_at: Lsn,
1983 58 : opened_at: Instant,
1984 58 : ) -> bool {
1985 58 : let distance = projected_lsn.widening_sub(last_freeze_at);
1986 58 :
1987 58 : // Rolling the open layer can be triggered by:
1988 58 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
1989 58 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
1990 58 : // account for how writes are distributed across shards: we expect each node to consume
1991 58 : // 1/count of the LSN on average.
1992 58 : // 2. The size of the currently open layer.
1993 58 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
1994 58 : // up and suspend activity.
1995 58 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
1996 0 : info!(
1997 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
1998 : projected_lsn, layer_size, distance
1999 : );
2000 :
2001 0 : true
2002 58 : } else if projected_layer_size >= checkpoint_distance {
2003 0 : info!(
2004 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2005 : projected_lsn, layer_size, projected_layer_size
2006 : );
2007 :
2008 0 : true
2009 58 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2010 0 : info!(
2011 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2012 0 : projected_lsn,
2013 0 : layer_size,
2014 0 : opened_at.elapsed()
2015 : );
2016 :
2017 0 : true
2018 : } else {
2019 58 : false
2020 : }
2021 58 : }
2022 : }
2023 :
2024 : /// Number of times we will compute partition within a checkpoint distance.
2025 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2026 :
2027 : // Private functions
2028 : impl Timeline {
2029 166 : pub(crate) fn get_switch_aux_file_policy(&self) -> AuxFilePolicy {
2030 166 : let tenant_conf = self.tenant_conf.load();
2031 166 : tenant_conf
2032 166 : .tenant_conf
2033 166 : .switch_aux_file_policy
2034 166 : .unwrap_or(self.conf.default_tenant_conf.switch_aux_file_policy)
2035 166 : }
2036 :
2037 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2038 0 : let tenant_conf = self.tenant_conf.load();
2039 0 : tenant_conf
2040 0 : .tenant_conf
2041 0 : .lazy_slru_download
2042 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2043 0 : }
2044 :
2045 1256 : fn get_checkpoint_distance(&self) -> u64 {
2046 1256 : let tenant_conf = self.tenant_conf.load();
2047 1256 : tenant_conf
2048 1256 : .tenant_conf
2049 1256 : .checkpoint_distance
2050 1256 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2051 1256 : }
2052 :
2053 58 : fn get_checkpoint_timeout(&self) -> Duration {
2054 58 : let tenant_conf = self.tenant_conf.load();
2055 58 : tenant_conf
2056 58 : .tenant_conf
2057 58 : .checkpoint_timeout
2058 58 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2059 58 : }
2060 :
2061 480 : fn get_compaction_target_size(&self) -> u64 {
2062 480 : let tenant_conf = self.tenant_conf.load();
2063 480 : tenant_conf
2064 480 : .tenant_conf
2065 480 : .compaction_target_size
2066 480 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2067 480 : }
2068 :
2069 366 : fn get_compaction_threshold(&self) -> usize {
2070 366 : let tenant_conf = self.tenant_conf.load();
2071 366 : tenant_conf
2072 366 : .tenant_conf
2073 366 : .compaction_threshold
2074 366 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2075 366 : }
2076 :
2077 14 : fn get_image_creation_threshold(&self) -> usize {
2078 14 : let tenant_conf = self.tenant_conf.load();
2079 14 : tenant_conf
2080 14 : .tenant_conf
2081 14 : .image_creation_threshold
2082 14 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2083 14 : }
2084 :
2085 366 : fn get_compaction_algorithm(&self) -> CompactionAlgorithm {
2086 366 : let tenant_conf = &self.tenant_conf.load();
2087 366 : tenant_conf
2088 366 : .tenant_conf
2089 366 : .compaction_algorithm
2090 366 : .unwrap_or(self.conf.default_tenant_conf.compaction_algorithm)
2091 366 : }
2092 :
2093 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2094 0 : let tenant_conf = self.tenant_conf.load();
2095 0 : tenant_conf
2096 0 : .tenant_conf
2097 0 : .eviction_policy
2098 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2099 0 : }
2100 :
2101 360 : fn get_evictions_low_residence_duration_metric_threshold(
2102 360 : tenant_conf: &TenantConfOpt,
2103 360 : default_tenant_conf: &TenantConf,
2104 360 : ) -> Duration {
2105 360 : tenant_conf
2106 360 : .evictions_low_residence_duration_metric_threshold
2107 360 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2108 360 : }
2109 :
2110 480 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2111 480 : let tenant_conf = self.tenant_conf.load();
2112 480 : tenant_conf
2113 480 : .tenant_conf
2114 480 : .image_layer_creation_check_threshold
2115 480 : .unwrap_or(
2116 480 : self.conf
2117 480 : .default_tenant_conf
2118 480 : .image_layer_creation_check_threshold,
2119 480 : )
2120 480 : }
2121 :
2122 8 : pub(super) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2123 8 : // NB: Most tenant conf options are read by background loops, so,
2124 8 : // changes will automatically be picked up.
2125 8 :
2126 8 : // The threshold is embedded in the metric. So, we need to update it.
2127 8 : {
2128 8 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2129 8 : new_conf,
2130 8 : &self.conf.default_tenant_conf,
2131 8 : );
2132 8 :
2133 8 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2134 8 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2135 8 :
2136 8 : let timeline_id_str = self.timeline_id.to_string();
2137 8 : self.metrics
2138 8 : .evictions_with_low_residence_duration
2139 8 : .write()
2140 8 : .unwrap()
2141 8 : .change_threshold(
2142 8 : &tenant_id_str,
2143 8 : &shard_id_str,
2144 8 : &timeline_id_str,
2145 8 : new_threshold,
2146 8 : );
2147 8 : }
2148 8 : }
2149 :
2150 : /// Open a Timeline handle.
2151 : ///
2152 : /// Loads the metadata for the timeline into memory, but not the layer map.
2153 : #[allow(clippy::too_many_arguments)]
2154 352 : pub(super) fn new(
2155 352 : conf: &'static PageServerConf,
2156 352 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2157 352 : metadata: &TimelineMetadata,
2158 352 : ancestor: Option<Arc<Timeline>>,
2159 352 : timeline_id: TimelineId,
2160 352 : tenant_shard_id: TenantShardId,
2161 352 : generation: Generation,
2162 352 : shard_identity: ShardIdentity,
2163 352 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2164 352 : resources: TimelineResources,
2165 352 : pg_version: u32,
2166 352 : state: TimelineState,
2167 352 : aux_file_policy: Option<AuxFilePolicy>,
2168 352 : cancel: CancellationToken,
2169 352 : ) -> Arc<Self> {
2170 352 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2171 352 : let (state, _) = watch::channel(state);
2172 352 :
2173 352 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2174 352 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2175 352 :
2176 352 : let evictions_low_residence_duration_metric_threshold = {
2177 352 : let loaded_tenant_conf = tenant_conf.load();
2178 352 : Self::get_evictions_low_residence_duration_metric_threshold(
2179 352 : &loaded_tenant_conf.tenant_conf,
2180 352 : &conf.default_tenant_conf,
2181 352 : )
2182 352 : };
2183 352 :
2184 352 : Arc::new_cyclic(|myself| {
2185 352 : let metrics = TimelineMetrics::new(
2186 352 : &tenant_shard_id,
2187 352 : &timeline_id,
2188 352 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2189 352 : "mtime",
2190 352 : evictions_low_residence_duration_metric_threshold,
2191 352 : ),
2192 352 : );
2193 352 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2194 :
2195 352 : let mut result = Timeline {
2196 352 : conf,
2197 352 : tenant_conf,
2198 352 : myself: myself.clone(),
2199 352 : timeline_id,
2200 352 : tenant_shard_id,
2201 352 : generation,
2202 352 : shard_identity,
2203 352 : pg_version,
2204 352 : layers: Default::default(),
2205 352 :
2206 352 : walredo_mgr,
2207 352 : walreceiver: Mutex::new(None),
2208 352 :
2209 352 : remote_client: Arc::new(resources.remote_client),
2210 352 :
2211 352 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2212 352 : last_record_lsn: SeqWait::new(RecordLsn {
2213 352 : last: disk_consistent_lsn,
2214 352 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2215 352 : }),
2216 352 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2217 352 :
2218 352 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2219 352 : last_freeze_ts: RwLock::new(Instant::now()),
2220 352 :
2221 352 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2222 352 :
2223 352 : ancestor_timeline: ancestor,
2224 352 : ancestor_lsn: metadata.ancestor_lsn(),
2225 352 :
2226 352 : metrics,
2227 352 :
2228 352 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2229 352 : &tenant_shard_id,
2230 352 : &timeline_id,
2231 352 : ),
2232 352 :
2233 2464 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2234 352 :
2235 352 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2236 352 :
2237 352 : layer_flush_start_tx,
2238 352 : layer_flush_done_tx,
2239 352 :
2240 352 : write_lock: tokio::sync::Mutex::new(None),
2241 352 :
2242 352 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2243 352 :
2244 352 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2245 352 : initdb_lsn: metadata.initdb_lsn(),
2246 352 :
2247 352 : current_logical_size: if disk_consistent_lsn.is_valid() {
2248 : // we're creating timeline data with some layer files existing locally,
2249 : // need to recalculate timeline's logical size based on data in the layers.
2250 230 : LogicalSize::deferred_initial(disk_consistent_lsn)
2251 : } else {
2252 : // we're creating timeline data without any layers existing locally,
2253 : // initial logical size is 0.
2254 122 : LogicalSize::empty_initial()
2255 : },
2256 352 : partitioning: tokio::sync::Mutex::new((
2257 352 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2258 352 : Lsn(0),
2259 352 : )),
2260 352 : repartition_threshold: 0,
2261 352 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2262 352 :
2263 352 : last_received_wal: Mutex::new(None),
2264 352 : rel_size_cache: RwLock::new(RelSizeCache {
2265 352 : complete_as_of: disk_consistent_lsn,
2266 352 : map: HashMap::new(),
2267 352 : }),
2268 352 :
2269 352 : download_all_remote_layers_task_info: RwLock::new(None),
2270 352 :
2271 352 : state,
2272 352 :
2273 352 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2274 352 : EvictionTaskTimelineState::default(),
2275 352 : ),
2276 352 : delete_progress: Arc::new(tokio::sync::Mutex::new(DeleteTimelineFlow::default())),
2277 352 :
2278 352 : cancel,
2279 352 : gate: Gate::default(),
2280 352 :
2281 352 : compaction_lock: tokio::sync::Mutex::default(),
2282 352 : gc_lock: tokio::sync::Mutex::default(),
2283 352 :
2284 352 : standby_horizon: AtomicLsn::new(0),
2285 352 :
2286 352 : timeline_get_throttle: resources.timeline_get_throttle,
2287 352 :
2288 352 : aux_files: tokio::sync::Mutex::new(AuxFilesState {
2289 352 : dir: None,
2290 352 : n_deltas: 0,
2291 352 : }),
2292 352 :
2293 352 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2294 352 :
2295 352 : last_aux_file_policy: AtomicAuxFilePolicy::new(aux_file_policy),
2296 352 : };
2297 352 : result.repartition_threshold =
2298 352 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2299 352 :
2300 352 : result
2301 352 : .metrics
2302 352 : .last_record_gauge
2303 352 : .set(disk_consistent_lsn.0 as i64);
2304 352 : result
2305 352 : })
2306 352 : }
2307 :
2308 460 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2309 460 : let Ok(guard) = self.gate.enter() else {
2310 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2311 0 : return;
2312 : };
2313 460 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2314 460 : match *flush_loop_state {
2315 346 : FlushLoopState::NotStarted => (),
2316 : FlushLoopState::Running { .. } => {
2317 114 : info!(
2318 0 : "skipping attempt to start flush_loop twice {}/{}",
2319 0 : self.tenant_shard_id, self.timeline_id
2320 : );
2321 114 : return;
2322 : }
2323 : FlushLoopState::Exited => {
2324 0 : warn!(
2325 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2326 0 : self.tenant_shard_id, self.timeline_id
2327 : );
2328 0 : return;
2329 : }
2330 : }
2331 :
2332 346 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2333 346 : let self_clone = Arc::clone(self);
2334 346 :
2335 346 : debug!("spawning flush loop");
2336 346 : *flush_loop_state = FlushLoopState::Running {
2337 346 : #[cfg(test)]
2338 346 : expect_initdb_optimization: false,
2339 346 : #[cfg(test)]
2340 346 : initdb_optimization_count: 0,
2341 346 : };
2342 346 : task_mgr::spawn(
2343 346 : task_mgr::BACKGROUND_RUNTIME.handle(),
2344 346 : task_mgr::TaskKind::LayerFlushTask,
2345 346 : Some(self.tenant_shard_id),
2346 346 : Some(self.timeline_id),
2347 346 : "layer flush task",
2348 : false,
2349 346 : async move {
2350 346 : let _guard = guard;
2351 346 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2352 59255 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2353 8 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2354 8 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{ ..}));
2355 8 : *flush_loop_state = FlushLoopState::Exited;
2356 8 : Ok(())
2357 8 : }
2358 346 : .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))
2359 : );
2360 460 : }
2361 :
2362 : /// Creates and starts the wal receiver.
2363 : ///
2364 : /// This function is expected to be called at most once per Timeline's lifecycle
2365 : /// when the timeline is activated.
2366 0 : fn launch_wal_receiver(
2367 0 : self: &Arc<Self>,
2368 0 : ctx: &RequestContext,
2369 0 : broker_client: BrokerClientChannel,
2370 0 : ) {
2371 0 : info!(
2372 0 : "launching WAL receiver for timeline {} of tenant {}",
2373 0 : self.timeline_id, self.tenant_shard_id
2374 : );
2375 :
2376 0 : let tenant_conf = self.tenant_conf.load();
2377 0 : let wal_connect_timeout = tenant_conf
2378 0 : .tenant_conf
2379 0 : .walreceiver_connect_timeout
2380 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2381 0 : let lagging_wal_timeout = tenant_conf
2382 0 : .tenant_conf
2383 0 : .lagging_wal_timeout
2384 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2385 0 : let max_lsn_wal_lag = tenant_conf
2386 0 : .tenant_conf
2387 0 : .max_lsn_wal_lag
2388 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2389 0 :
2390 0 : let mut guard = self.walreceiver.lock().unwrap();
2391 0 : assert!(
2392 0 : guard.is_none(),
2393 0 : "multiple launches / re-launches of WAL receiver are not supported"
2394 : );
2395 0 : *guard = Some(WalReceiver::start(
2396 0 : Arc::clone(self),
2397 0 : WalReceiverConf {
2398 0 : wal_connect_timeout,
2399 0 : lagging_wal_timeout,
2400 0 : max_lsn_wal_lag,
2401 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2402 0 : availability_zone: self.conf.availability_zone.clone(),
2403 0 : ingest_batch_size: self.conf.ingest_batch_size,
2404 0 : },
2405 0 : broker_client,
2406 0 : ctx,
2407 0 : ));
2408 0 : }
2409 :
2410 : /// Initialize with an empty layer map. Used when creating a new timeline.
2411 346 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2412 346 : let mut layers = self.layers.try_write().expect(
2413 346 : "in the context where we call this function, no other task has access to the object",
2414 346 : );
2415 346 : layers.initialize_empty(Lsn(start_lsn.0));
2416 346 : }
2417 :
2418 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2419 : /// files.
2420 6 : pub(super) async fn load_layer_map(
2421 6 : &self,
2422 6 : disk_consistent_lsn: Lsn,
2423 6 : index_part: Option<IndexPart>,
2424 6 : ) -> anyhow::Result<()> {
2425 : use init::{Decision::*, Discovered, DismissedLayer};
2426 : use LayerName::*;
2427 :
2428 6 : let mut guard = self.layers.write().await;
2429 :
2430 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2431 6 :
2432 6 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2433 6 : // structs representing all files on disk
2434 6 : let timeline_path = self
2435 6 : .conf
2436 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2437 6 : let conf = self.conf;
2438 6 : let span = tracing::Span::current();
2439 6 :
2440 6 : // Copy to move into the task we're about to spawn
2441 6 : let generation = self.generation;
2442 6 : let shard = self.get_shard_index();
2443 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2444 :
2445 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2446 6 : move || {
2447 6 : let _g = span.entered();
2448 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2449 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2450 6 : let mut unrecognized_files = Vec::new();
2451 6 :
2452 6 : let mut path = timeline_path;
2453 :
2454 22 : for discovered in discovered {
2455 16 : let (name, kind) = match discovered {
2456 16 : Discovered::Layer(layer_file_name, local_path, file_size) => {
2457 16 : discovered_layers.push((layer_file_name, local_path, file_size));
2458 16 : continue;
2459 : }
2460 : Discovered::IgnoredBackup => {
2461 0 : continue;
2462 : }
2463 0 : Discovered::Unknown(file_name) => {
2464 0 : // we will later error if there are any
2465 0 : unrecognized_files.push(file_name);
2466 0 : continue;
2467 : }
2468 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2469 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2470 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2471 : };
2472 0 : path.push(Utf8Path::new(&name));
2473 0 : init::cleanup(&path, kind)?;
2474 0 : path.pop();
2475 : }
2476 :
2477 6 : if !unrecognized_files.is_empty() {
2478 : // assume that if there are any there are many many.
2479 0 : let n = unrecognized_files.len();
2480 0 : let first = &unrecognized_files[..n.min(10)];
2481 0 : anyhow::bail!(
2482 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2483 0 : );
2484 6 : }
2485 6 :
2486 6 : let decided = init::reconcile(
2487 6 : discovered_layers,
2488 6 : index_part.as_ref(),
2489 6 : disk_consistent_lsn,
2490 6 : generation,
2491 6 : shard,
2492 6 : );
2493 6 :
2494 6 : let mut loaded_layers = Vec::new();
2495 6 : let mut needs_cleanup = Vec::new();
2496 6 : let mut total_physical_size = 0;
2497 :
2498 22 : for (name, decision) in decided {
2499 16 : let decision = match decision {
2500 0 : Ok(UseRemote { local, remote }) => {
2501 0 : // Remote is authoritative, but we may still choose to retain
2502 0 : // the local file if the contents appear to match
2503 0 : if local.metadata.file_size() == remote.file_size() {
2504 : // Use the local file, but take the remote metadata so that we pick up
2505 : // the correct generation.
2506 0 : UseLocal(LocalLayerFileMetadata {
2507 0 : metadata: remote,
2508 0 : local_path: local.local_path,
2509 0 : })
2510 : } else {
2511 0 : init::cleanup_local_file_for_remote(&local, &remote)?;
2512 0 : UseRemote { local, remote }
2513 : }
2514 : }
2515 16 : Ok(decision) => decision,
2516 0 : Err(DismissedLayer::Future { local }) => {
2517 0 : if let Some(local) = local {
2518 0 : init::cleanup_future_layer(
2519 0 : &local.local_path,
2520 0 : &name,
2521 0 : disk_consistent_lsn,
2522 0 : )?;
2523 0 : }
2524 0 : needs_cleanup.push(name);
2525 0 : continue;
2526 : }
2527 0 : Err(DismissedLayer::LocalOnly(local)) => {
2528 0 : init::cleanup_local_only_file(&name, &local)?;
2529 : // this file never existed remotely, we will have to do rework
2530 0 : continue;
2531 : }
2532 : };
2533 :
2534 16 : match &name {
2535 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2536 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2537 : }
2538 :
2539 16 : tracing::debug!(layer=%name, ?decision, "applied");
2540 :
2541 16 : let layer = match decision {
2542 16 : UseLocal(local) => {
2543 16 : total_physical_size += local.metadata.file_size();
2544 16 : Layer::for_resident(conf, &this, local.local_path, name, local.metadata)
2545 16 : .drop_eviction_guard()
2546 : }
2547 0 : Evicted(remote) | UseRemote { remote, .. } => {
2548 0 : Layer::for_evicted(conf, &this, name, remote)
2549 : }
2550 : };
2551 :
2552 16 : loaded_layers.push(layer);
2553 : }
2554 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2555 6 : }
2556 6 : })
2557 5 : .await
2558 6 : .map_err(anyhow::Error::new)
2559 6 : .and_then(|x| x)?;
2560 :
2561 6 : let num_layers = loaded_layers.len();
2562 6 :
2563 6 : guard.initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2564 6 :
2565 6 : self.remote_client
2566 6 : .schedule_layer_file_deletion(&needs_cleanup)?;
2567 6 : self.remote_client
2568 6 : .schedule_index_upload_for_file_changes()?;
2569 : // This barrier orders above DELETEs before any later operations.
2570 : // This is critical because code executing after the barrier might
2571 : // create again objects with the same key that we just scheduled for deletion.
2572 : // For example, if we just scheduled deletion of an image layer "from the future",
2573 : // later compaction might run again and re-create the same image layer.
2574 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2575 : // "same" here means same key range and LSN.
2576 : //
2577 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2578 : // the upload queue may execute the PUT first, then the DELETE.
2579 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2580 : //
2581 : // 1. a future image layer is created and uploaded
2582 : // 2. ps restart
2583 : // 3. the future layer from (1) is deleted during load layer map
2584 : // 4. image layer is re-created and uploaded
2585 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2586 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2587 : //
2588 : // See https://github.com/neondatabase/neon/issues/5878
2589 : //
2590 : // NB: generation numbers naturally protect against this because they disambiguate
2591 : // (1) and (4)
2592 6 : self.remote_client.schedule_barrier()?;
2593 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2594 : // on retry.
2595 :
2596 6 : info!(
2597 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2598 : num_layers, disk_consistent_lsn, total_physical_size
2599 : );
2600 :
2601 6 : timer.stop_and_record();
2602 6 : Ok(())
2603 6 : }
2604 :
2605 : /// Retrieve current logical size of the timeline.
2606 : ///
2607 : /// The size could be lagging behind the actual number, in case
2608 : /// the initial size calculation has not been run (gets triggered on the first size access).
2609 : ///
2610 : /// return size and boolean flag that shows if the size is exact
2611 0 : pub(crate) fn get_current_logical_size(
2612 0 : self: &Arc<Self>,
2613 0 : priority: GetLogicalSizePriority,
2614 0 : ctx: &RequestContext,
2615 0 : ) -> logical_size::CurrentLogicalSize {
2616 0 : if !self.tenant_shard_id.is_shard_zero() {
2617 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2618 : // when HTTP API is serving a GET for timeline zero, return zero
2619 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2620 0 : }
2621 0 :
2622 0 : let current_size = self.current_logical_size.current_size();
2623 0 : debug!("Current size: {current_size:?}");
2624 :
2625 0 : match (current_size.accuracy(), priority) {
2626 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2627 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2628 0 : // background task will eventually deliver an exact value, we're in no rush
2629 0 : }
2630 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2631 : // background task is not ready, but user is asking for it now;
2632 : // => make the background task skip the line
2633 : // (The alternative would be to calculate the size here, but,
2634 : // it can actually take a long time if the user has a lot of rels.
2635 : // And we'll inevitable need it again; So, let the background task do the work.)
2636 0 : match self
2637 0 : .current_logical_size
2638 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2639 0 : .get()
2640 : {
2641 0 : Some(cancel) => cancel.cancel(),
2642 : None => {
2643 0 : let state = self.current_state();
2644 0 : if matches!(
2645 0 : state,
2646 : TimelineState::Broken { .. } | TimelineState::Stopping
2647 0 : ) {
2648 0 :
2649 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2650 0 : // Don't make noise.
2651 0 : } else {
2652 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2653 0 : debug_assert!(false);
2654 : }
2655 : }
2656 : };
2657 : }
2658 : }
2659 :
2660 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2661 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2662 0 : let first = self
2663 0 : .current_logical_size
2664 0 : .did_return_approximate_to_walreceiver
2665 0 : .compare_exchange(
2666 0 : false,
2667 0 : true,
2668 0 : AtomicOrdering::Relaxed,
2669 0 : AtomicOrdering::Relaxed,
2670 0 : )
2671 0 : .is_ok();
2672 0 : if first {
2673 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2674 0 : }
2675 0 : }
2676 0 : }
2677 :
2678 0 : current_size
2679 0 : }
2680 :
2681 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2682 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2683 : // nothing to do for freshly created timelines;
2684 0 : assert_eq!(
2685 0 : self.current_logical_size.current_size().accuracy(),
2686 0 : logical_size::Accuracy::Exact,
2687 0 : );
2688 0 : self.current_logical_size.initialized.add_permits(1);
2689 0 : return;
2690 : };
2691 :
2692 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2693 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2694 0 : self.current_logical_size
2695 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2696 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2697 0 :
2698 0 : let self_clone = Arc::clone(self);
2699 0 : let background_ctx = ctx.detached_child(
2700 0 : TaskKind::InitialLogicalSizeCalculation,
2701 0 : DownloadBehavior::Download,
2702 0 : );
2703 0 : task_mgr::spawn(
2704 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2705 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2706 0 : Some(self.tenant_shard_id),
2707 0 : Some(self.timeline_id),
2708 0 : "initial size calculation",
2709 : false,
2710 : // NB: don't log errors here, task_mgr will do that.
2711 0 : async move {
2712 0 : let cancel = task_mgr::shutdown_token();
2713 0 : self_clone
2714 0 : .initial_logical_size_calculation_task(
2715 0 : initial_part_end,
2716 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2717 0 : cancel,
2718 0 : background_ctx,
2719 0 : )
2720 0 : .await;
2721 0 : Ok(())
2722 0 : }
2723 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)),
2724 : );
2725 0 : }
2726 :
2727 0 : async fn initial_logical_size_calculation_task(
2728 0 : self: Arc<Self>,
2729 0 : initial_part_end: Lsn,
2730 0 : skip_concurrency_limiter: CancellationToken,
2731 0 : cancel: CancellationToken,
2732 0 : background_ctx: RequestContext,
2733 0 : ) {
2734 : scopeguard::defer! {
2735 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2736 : self.current_logical_size.initialized.add_permits(1);
2737 : }
2738 :
2739 : enum BackgroundCalculationError {
2740 : Cancelled,
2741 : Other(anyhow::Error),
2742 : }
2743 :
2744 0 : let try_once = |attempt: usize| {
2745 0 : let background_ctx = &background_ctx;
2746 0 : let self_ref = &self;
2747 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2748 0 : async move {
2749 0 : let cancel = task_mgr::shutdown_token();
2750 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2751 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2752 0 : background_ctx,
2753 0 : );
2754 :
2755 : use crate::metrics::initial_logical_size::StartCircumstances;
2756 0 : let (_maybe_permit, circumstances) = tokio::select! {
2757 : permit = wait_for_permit => {
2758 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2759 : }
2760 : _ = self_ref.cancel.cancelled() => {
2761 : return Err(BackgroundCalculationError::Cancelled);
2762 : }
2763 : _ = cancel.cancelled() => {
2764 : return Err(BackgroundCalculationError::Cancelled);
2765 : },
2766 : () = skip_concurrency_limiter.cancelled() => {
2767 : // Some action that is part of a end user interaction requested logical size
2768 : // => break out of the rate limit
2769 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2770 : // but then again what happens if they cancel; also, we should just be using
2771 : // one runtime across the entire process, so, let's leave this for now.
2772 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2773 : }
2774 : };
2775 :
2776 0 : let metrics_guard = if attempt == 1 {
2777 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2778 : } else {
2779 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2780 : };
2781 :
2782 0 : match self_ref
2783 0 : .logical_size_calculation_task(
2784 0 : initial_part_end,
2785 0 : LogicalSizeCalculationCause::Initial,
2786 0 : background_ctx,
2787 0 : )
2788 0 : .await
2789 : {
2790 0 : Ok(calculated_size) => Ok((calculated_size, metrics_guard)),
2791 : Err(CalculateLogicalSizeError::Cancelled) => {
2792 0 : Err(BackgroundCalculationError::Cancelled)
2793 : }
2794 0 : Err(CalculateLogicalSizeError::Other(err)) => {
2795 0 : if let Some(PageReconstructError::AncestorStopping(_)) =
2796 0 : err.root_cause().downcast_ref()
2797 : {
2798 0 : Err(BackgroundCalculationError::Cancelled)
2799 : } else {
2800 0 : Err(BackgroundCalculationError::Other(err))
2801 : }
2802 : }
2803 : }
2804 0 : }
2805 0 : };
2806 :
2807 0 : let retrying = async {
2808 0 : let mut attempt = 0;
2809 0 : loop {
2810 0 : attempt += 1;
2811 0 :
2812 0 : match try_once(attempt).await {
2813 0 : Ok(res) => return ControlFlow::Continue(res),
2814 0 : Err(BackgroundCalculationError::Cancelled) => return ControlFlow::Break(()),
2815 0 : Err(BackgroundCalculationError::Other(e)) => {
2816 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2817 : // exponential back-off doesn't make sense at these long intervals;
2818 : // use fixed retry interval with generous jitter instead
2819 0 : let sleep_duration = Duration::from_secs(
2820 0 : u64::try_from(
2821 0 : // 1hour base
2822 0 : (60_i64 * 60_i64)
2823 0 : // 10min jitter
2824 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2825 0 : )
2826 0 : .expect("10min < 1hour"),
2827 0 : );
2828 0 : tokio::time::sleep(sleep_duration).await;
2829 : }
2830 : }
2831 : }
2832 0 : };
2833 :
2834 0 : let (calculated_size, metrics_guard) = tokio::select! {
2835 : res = retrying => {
2836 : match res {
2837 : ControlFlow::Continue(calculated_size) => calculated_size,
2838 : ControlFlow::Break(()) => return,
2839 : }
2840 : }
2841 : _ = cancel.cancelled() => {
2842 : return;
2843 : }
2844 : };
2845 :
2846 : // we cannot query current_logical_size.current_size() to know the current
2847 : // *negative* value, only truncated to u64.
2848 0 : let added = self
2849 0 : .current_logical_size
2850 0 : .size_added_after_initial
2851 0 : .load(AtomicOrdering::Relaxed);
2852 0 :
2853 0 : let sum = calculated_size.saturating_add_signed(added);
2854 0 :
2855 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2856 0 : self.metrics.current_logical_size_gauge.set(sum);
2857 0 :
2858 0 : self.current_logical_size
2859 0 : .initial_logical_size
2860 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
2861 0 : .ok()
2862 0 : .expect("only this task sets it");
2863 0 : }
2864 :
2865 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
2866 0 : self: &Arc<Self>,
2867 0 : lsn: Lsn,
2868 0 : cause: LogicalSizeCalculationCause,
2869 0 : ctx: RequestContext,
2870 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
2871 0 : let (sender, receiver) = oneshot::channel();
2872 0 : let self_clone = Arc::clone(self);
2873 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
2874 0 : // we should stop the size calculation work and return an error.
2875 0 : // That would require restructuring this function's API to
2876 0 : // return the result directly, instead of a Receiver for the result.
2877 0 : let ctx = ctx.detached_child(
2878 0 : TaskKind::OndemandLogicalSizeCalculation,
2879 0 : DownloadBehavior::Download,
2880 0 : );
2881 0 : task_mgr::spawn(
2882 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2883 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
2884 0 : Some(self.tenant_shard_id),
2885 0 : Some(self.timeline_id),
2886 0 : "ondemand logical size calculation",
2887 0 : false,
2888 0 : async move {
2889 0 : let res = self_clone
2890 0 : .logical_size_calculation_task(lsn, cause, &ctx)
2891 0 : .await;
2892 0 : let _ = sender.send(res).ok();
2893 0 : Ok(()) // Receiver is responsible for handling errors
2894 0 : }
2895 0 : .in_current_span(),
2896 0 : );
2897 0 : receiver
2898 0 : }
2899 :
2900 : /// # Cancel-Safety
2901 : ///
2902 : /// This method is cancellation-safe.
2903 0 : #[instrument(skip_all)]
2904 : async fn logical_size_calculation_task(
2905 : self: &Arc<Self>,
2906 : lsn: Lsn,
2907 : cause: LogicalSizeCalculationCause,
2908 : ctx: &RequestContext,
2909 : ) -> Result<u64, CalculateLogicalSizeError> {
2910 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
2911 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
2912 : // accurate relation sizes, and they do not emit consumption metrics.
2913 : debug_assert!(self.tenant_shard_id.is_shard_zero());
2914 :
2915 : let guard = self
2916 : .gate
2917 : .enter()
2918 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
2919 :
2920 : let self_calculation = Arc::clone(self);
2921 :
2922 0 : let mut calculation = pin!(async {
2923 0 : let ctx = ctx.attached_child();
2924 0 : self_calculation
2925 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
2926 0 : .await
2927 0 : });
2928 :
2929 : tokio::select! {
2930 : res = &mut calculation => { res }
2931 : _ = self.cancel.cancelled() => {
2932 : debug!("cancelling logical size calculation for timeline shutdown");
2933 : calculation.await
2934 : }
2935 : }
2936 : }
2937 :
2938 : /// Calculate the logical size of the database at the latest LSN.
2939 : ///
2940 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
2941 : /// especially if we need to download remote layers.
2942 : ///
2943 : /// # Cancel-Safety
2944 : ///
2945 : /// This method is cancellation-safe.
2946 0 : async fn calculate_logical_size(
2947 0 : &self,
2948 0 : up_to_lsn: Lsn,
2949 0 : cause: LogicalSizeCalculationCause,
2950 0 : _guard: &GateGuard,
2951 0 : ctx: &RequestContext,
2952 0 : ) -> Result<u64, CalculateLogicalSizeError> {
2953 0 : info!(
2954 0 : "Calculating logical size for timeline {} at {}",
2955 : self.timeline_id, up_to_lsn
2956 : );
2957 :
2958 : pausable_failpoint!("timeline-calculate-logical-size-pause");
2959 :
2960 : // See if we've already done the work for initial size calculation.
2961 : // This is a short-cut for timelines that are mostly unused.
2962 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
2963 0 : return Ok(size);
2964 0 : }
2965 0 : let storage_time_metrics = match cause {
2966 : LogicalSizeCalculationCause::Initial
2967 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
2968 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
2969 : LogicalSizeCalculationCause::EvictionTaskImitation => {
2970 0 : &self.metrics.imitate_logical_size_histo
2971 : }
2972 : };
2973 0 : let timer = storage_time_metrics.start_timer();
2974 0 : let logical_size = self
2975 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
2976 0 : .await?;
2977 0 : debug!("calculated logical size: {logical_size}");
2978 0 : timer.stop_and_record();
2979 0 : Ok(logical_size)
2980 0 : }
2981 :
2982 : /// Update current logical size, adding `delta' to the old value.
2983 270570 : fn update_current_logical_size(&self, delta: i64) {
2984 270570 : let logical_size = &self.current_logical_size;
2985 270570 : logical_size.increment_size(delta);
2986 270570 :
2987 270570 : // Also set the value in the prometheus gauge. Note that
2988 270570 : // there is a race condition here: if this is is called by two
2989 270570 : // threads concurrently, the prometheus gauge might be set to
2990 270570 : // one value while current_logical_size is set to the
2991 270570 : // other.
2992 270570 : match logical_size.current_size() {
2993 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
2994 270570 : .metrics
2995 270570 : .current_logical_size_gauge
2996 270570 : .set(new_current_size.into()),
2997 0 : CurrentLogicalSize::Approximate(_) => {
2998 0 : // don't update the gauge yet, this allows us not to update the gauge back and
2999 0 : // forth between the initial size calculation task.
3000 0 : }
3001 : }
3002 270570 : }
3003 :
3004 2674 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3005 2674 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3006 2674 : let aux_metric =
3007 2674 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3008 2674 :
3009 2674 : let sum_of_entries = self
3010 2674 : .directory_metrics
3011 2674 : .iter()
3012 18718 : .map(|v| v.load(AtomicOrdering::Relaxed))
3013 2674 : .sum();
3014 2674 : // Set a high general threshold and a lower threshold for the auxiliary files,
3015 2674 : // as we can have large numbers of relations in the db directory.
3016 2674 : const SUM_THRESHOLD: u64 = 5000;
3017 2674 : const AUX_THRESHOLD: u64 = 1000;
3018 2674 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3019 0 : self.metrics
3020 0 : .directory_entries_count_gauge
3021 0 : .set(sum_of_entries);
3022 2674 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3023 0 : metric.set(sum_of_entries);
3024 2674 : }
3025 2674 : }
3026 :
3027 0 : async fn find_layer(&self, layer_name: &LayerName) -> Option<Layer> {
3028 0 : let guard = self.layers.read().await;
3029 0 : for historic_layer in guard.layer_map().iter_historic_layers() {
3030 0 : let historic_layer_name = historic_layer.layer_name();
3031 0 : if layer_name == &historic_layer_name {
3032 0 : return Some(guard.get_from_desc(&historic_layer));
3033 0 : }
3034 : }
3035 :
3036 0 : None
3037 0 : }
3038 :
3039 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3040 : /// indicating which layers are currently on-disk on the primary.
3041 : ///
3042 : /// None is returned if the Timeline is in a state where uploading a heatmap
3043 : /// doesn't make sense, such as shutting down or initializing. The caller
3044 : /// should treat this as a cue to simply skip doing any heatmap uploading
3045 : /// for this timeline.
3046 0 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3047 0 : if !self.is_active() {
3048 0 : return None;
3049 0 : }
3050 :
3051 0 : let guard = self.layers.read().await;
3052 :
3053 0 : let resident = guard.likely_resident_layers().map(|layer| {
3054 0 : let last_activity_ts = layer.access_stats().latest_activity_or_now();
3055 0 :
3056 0 : HeatMapLayer::new(
3057 0 : layer.layer_desc().layer_name(),
3058 0 : (&layer.metadata()).into(),
3059 0 : last_activity_ts,
3060 0 : )
3061 0 : });
3062 0 :
3063 0 : let layers = resident.collect();
3064 0 :
3065 0 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3066 0 : }
3067 :
3068 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3069 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3070 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3071 0 : // branchpoint in the value in IndexPart::lineage
3072 0 : self.ancestor_lsn == lsn
3073 0 : || (self.ancestor_lsn == Lsn::INVALID
3074 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3075 0 : }
3076 : }
3077 :
3078 : type TraversalId = Arc<str>;
3079 :
3080 : trait TraversalLayerExt {
3081 : fn traversal_id(&self) -> TraversalId;
3082 : }
3083 :
3084 : impl TraversalLayerExt for Layer {
3085 210377 : fn traversal_id(&self) -> TraversalId {
3086 210377 : Arc::clone(self.debug_str())
3087 210377 : }
3088 : }
3089 :
3090 : impl TraversalLayerExt for Arc<InMemoryLayer> {
3091 605871 : fn traversal_id(&self) -> TraversalId {
3092 605871 : Arc::clone(self.local_path_str())
3093 605871 : }
3094 : }
3095 :
3096 : impl Timeline {
3097 : ///
3098 : /// Get a handle to a Layer for reading.
3099 : ///
3100 : /// The returned Layer might be from an ancestor timeline, if the
3101 : /// segment hasn't been updated on this timeline yet.
3102 : ///
3103 : /// This function takes the current timeline's locked LayerMap as an argument,
3104 : /// so callers can avoid potential race conditions.
3105 : ///
3106 : /// # Cancel-Safety
3107 : ///
3108 : /// This method is cancellation-safe.
3109 625369 : async fn get_reconstruct_data(
3110 625369 : &self,
3111 625369 : key: Key,
3112 625369 : request_lsn: Lsn,
3113 625369 : reconstruct_state: &mut ValueReconstructState,
3114 625369 : ctx: &RequestContext,
3115 625369 : ) -> Result<Vec<TraversalPathItem>, PageReconstructError> {
3116 625369 : // Start from the current timeline.
3117 625369 : let mut timeline_owned;
3118 625369 : let mut timeline = self;
3119 625369 :
3120 625369 : let mut read_count = scopeguard::guard(0, |cnt| {
3121 625369 : crate::metrics::READ_NUM_LAYERS_VISITED.observe(cnt as f64)
3122 625369 : });
3123 625369 :
3124 625369 : // For debugging purposes, collect the path of layers that we traversed
3125 625369 : // through. It's included in the error message if we fail to find the key.
3126 625369 : let mut traversal_path = Vec::<TraversalPathItem>::new();
3127 :
3128 625369 : let cached_lsn = if let Some((cached_lsn, _)) = &reconstruct_state.img {
3129 0 : *cached_lsn
3130 : } else {
3131 625369 : Lsn(0)
3132 : };
3133 :
3134 : // 'prev_lsn' tracks the last LSN that we were at in our search. It's used
3135 : // to check that each iteration make some progress, to break infinite
3136 : // looping if something goes wrong.
3137 625369 : let mut prev_lsn = None;
3138 625369 :
3139 625369 : let mut result = ValueReconstructResult::Continue;
3140 625369 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3141 :
3142 1669145 : 'outer: loop {
3143 1669145 : if self.cancel.is_cancelled() {
3144 0 : return Err(PageReconstructError::Cancelled);
3145 1669145 : }
3146 1669145 :
3147 1669145 : // The function should have updated 'state'
3148 1669145 : //info!("CALLED for {} at {}: {:?} with {} records, cached {}", key, cont_lsn, result, reconstruct_state.records.len(), cached_lsn);
3149 1669145 : match result {
3150 625245 : ValueReconstructResult::Complete => return Ok(traversal_path),
3151 : ValueReconstructResult::Continue => {
3152 : // If we reached an earlier cached page image, we're done.
3153 1043886 : if cont_lsn == cached_lsn + 1 {
3154 0 : MATERIALIZED_PAGE_CACHE_HIT.inc_by(1);
3155 0 : return Ok(traversal_path);
3156 1043886 : }
3157 1043886 : if let Some(prev) = prev_lsn {
3158 191107 : if prev <= cont_lsn {
3159 : // Didn't make any progress in last iteration. Error out to avoid
3160 : // getting stuck in the loop.
3161 108 : return Err(PageReconstructError::MissingKey(MissingKeyError {
3162 108 : key,
3163 108 : shard: self.shard_identity.get_shard_number(&key),
3164 108 : cont_lsn: Lsn(cont_lsn.0 - 1),
3165 108 : request_lsn,
3166 108 : ancestor_lsn: Some(timeline.ancestor_lsn),
3167 108 : traversal_path,
3168 108 : backtrace: None,
3169 108 : }));
3170 190999 : }
3171 852779 : }
3172 1043778 : prev_lsn = Some(cont_lsn);
3173 : }
3174 : ValueReconstructResult::Missing => {
3175 : return Err(PageReconstructError::MissingKey(MissingKeyError {
3176 14 : key,
3177 14 : shard: self.shard_identity.get_shard_number(&key),
3178 14 : cont_lsn,
3179 14 : request_lsn,
3180 14 : ancestor_lsn: None,
3181 14 : traversal_path,
3182 14 : backtrace: if cfg!(test) {
3183 14 : Some(std::backtrace::Backtrace::force_capture())
3184 : } else {
3185 0 : None
3186 : },
3187 : }));
3188 : }
3189 : }
3190 :
3191 : // Recurse into ancestor if needed
3192 1043778 : if is_inherited_key(key) && Lsn(cont_lsn.0 - 1) <= timeline.ancestor_lsn {
3193 227412 : trace!(
3194 0 : "going into ancestor {}, cont_lsn is {}",
3195 : timeline.ancestor_lsn,
3196 : cont_lsn
3197 : );
3198 :
3199 227412 : timeline_owned = timeline.get_ready_ancestor_timeline(ctx).await?;
3200 227410 : timeline = &*timeline_owned;
3201 227410 : prev_lsn = None;
3202 227410 : continue 'outer;
3203 816366 : }
3204 :
3205 816366 : let guard = timeline.layers.read().await;
3206 816366 : let layers = guard.layer_map();
3207 :
3208 : // Check the open and frozen in-memory layers first, in order from newest
3209 : // to oldest.
3210 816366 : if let Some(open_layer) = &layers.open_layer {
3211 717384 : let start_lsn = open_layer.get_lsn_range().start;
3212 717384 : if cont_lsn > start_lsn {
3213 : //info!("CHECKING for {} at {} on open layer {}", key, cont_lsn, open_layer.layer_name().display());
3214 : // Get all the data needed to reconstruct the page version from this layer.
3215 : // But if we have an older cached page image, no need to go past that.
3216 604275 : let lsn_floor = max(cached_lsn + 1, start_lsn);
3217 604275 :
3218 604275 : let open_layer = open_layer.clone();
3219 604275 : drop(guard);
3220 604275 :
3221 604275 : result = match open_layer
3222 604275 : .get_value_reconstruct_data(
3223 604275 : key,
3224 604275 : lsn_floor..cont_lsn,
3225 604275 : reconstruct_state,
3226 604275 : ctx,
3227 604275 : )
3228 10491 : .await
3229 : {
3230 604275 : Ok(result) => result,
3231 0 : Err(e) => return Err(PageReconstructError::from(e)),
3232 : };
3233 604275 : cont_lsn = lsn_floor;
3234 604275 : *read_count += 1;
3235 604275 : traversal_path.push((result, cont_lsn, open_layer.traversal_id()));
3236 604275 : continue 'outer;
3237 113109 : }
3238 98982 : }
3239 212091 : for frozen_layer in layers.frozen_layers.iter().rev() {
3240 1596 : let start_lsn = frozen_layer.get_lsn_range().start;
3241 1596 : if cont_lsn > start_lsn {
3242 : //info!("CHECKING for {} at {} on frozen layer {}", key, cont_lsn, frozen_layer.layer_name().display());
3243 1596 : let lsn_floor = max(cached_lsn + 1, start_lsn);
3244 1596 :
3245 1596 : let frozen_layer = frozen_layer.clone();
3246 1596 : drop(guard);
3247 1596 :
3248 1596 : result = match frozen_layer
3249 1596 : .get_value_reconstruct_data(
3250 1596 : key,
3251 1596 : lsn_floor..cont_lsn,
3252 1596 : reconstruct_state,
3253 1596 : ctx,
3254 1596 : )
3255 0 : .await
3256 : {
3257 1596 : Ok(result) => result,
3258 0 : Err(e) => return Err(PageReconstructError::from(e)),
3259 : };
3260 1596 : cont_lsn = lsn_floor;
3261 1596 : *read_count += 1;
3262 1596 : traversal_path.push((result, cont_lsn, frozen_layer.traversal_id()));
3263 1596 : continue 'outer;
3264 0 : }
3265 : }
3266 :
3267 210495 : if let Some(SearchResult { lsn_floor, layer }) = layers.search(key, cont_lsn) {
3268 210377 : let layer = guard.get_from_desc(&layer);
3269 210377 : drop(guard);
3270 210377 : // Get all the data needed to reconstruct the page version from this layer.
3271 210377 : // But if we have an older cached page image, no need to go past that.
3272 210377 : let lsn_floor = max(cached_lsn + 1, lsn_floor);
3273 210377 : result = match layer
3274 210377 : .get_value_reconstruct_data(key, lsn_floor..cont_lsn, reconstruct_state, ctx)
3275 29695 : .await
3276 : {
3277 210377 : Ok(result) => result,
3278 0 : Err(e) => return Err(PageReconstructError::from(e)),
3279 : };
3280 210377 : cont_lsn = lsn_floor;
3281 210377 : *read_count += 1;
3282 210377 : traversal_path.push((result, cont_lsn, layer.traversal_id()));
3283 210377 : continue 'outer;
3284 118 : } else if timeline.ancestor_timeline.is_some() {
3285 : // Nothing on this timeline. Traverse to parent
3286 108 : result = ValueReconstructResult::Continue;
3287 108 : cont_lsn = Lsn(timeline.ancestor_lsn.0 + 1);
3288 108 : continue 'outer;
3289 : } else {
3290 : // Nothing found
3291 10 : result = ValueReconstructResult::Missing;
3292 10 : continue 'outer;
3293 : }
3294 : }
3295 625369 : }
3296 :
3297 : /// Get the data needed to reconstruct all keys in the provided keyspace
3298 : ///
3299 : /// The algorithm is as follows:
3300 : /// 1. While some keys are still not done and there's a timeline to visit:
3301 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3302 : /// 2.1: Build the fringe for the current keyspace
3303 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3304 : /// intersects
3305 : /// 2.3. Pop the timeline from the fringe
3306 : /// 2.4. If the fringe is empty, go back to 1
3307 470 : async fn get_vectored_reconstruct_data(
3308 470 : &self,
3309 470 : mut keyspace: KeySpace,
3310 470 : request_lsn: Lsn,
3311 470 : reconstruct_state: &mut ValuesReconstructState,
3312 470 : ctx: &RequestContext,
3313 470 : ) -> Result<(), GetVectoredError> {
3314 470 : let mut timeline_owned: Arc<Timeline>;
3315 470 : let mut timeline = self;
3316 470 :
3317 470 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3318 :
3319 470 : let missing_keyspace = loop {
3320 504 : if self.cancel.is_cancelled() {
3321 0 : return Err(GetVectoredError::Cancelled);
3322 504 : }
3323 :
3324 : let TimelineVisitOutcome {
3325 504 : completed_keyspace: completed,
3326 504 : image_covered_keyspace,
3327 504 : } = Self::get_vectored_reconstruct_data_timeline(
3328 504 : timeline,
3329 504 : keyspace.clone(),
3330 504 : cont_lsn,
3331 504 : reconstruct_state,
3332 504 : &self.cancel,
3333 504 : ctx,
3334 504 : )
3335 11519 : .await?;
3336 :
3337 504 : keyspace.remove_overlapping_with(&completed);
3338 504 :
3339 504 : // Do not descend into the ancestor timeline for aux files.
3340 504 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3341 504 : // stalling compaction.
3342 504 : keyspace.remove_overlapping_with(&KeySpace {
3343 504 : ranges: vec![NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE],
3344 504 : });
3345 504 :
3346 504 : // Keyspace is fully retrieved
3347 504 : if keyspace.is_empty() {
3348 460 : break None;
3349 44 : }
3350 44 :
3351 44 : // Not fully retrieved but no ancestor timeline.
3352 44 : if timeline.ancestor_timeline.is_none() {
3353 10 : break Some(keyspace);
3354 34 : }
3355 34 :
3356 34 : // Now we see if there are keys covered by the image layer but does not exist in the
3357 34 : // image layer, which means that the key does not exist.
3358 34 :
3359 34 : // The block below will stop the vectored search if any of the keys encountered an image layer
3360 34 : // which did not contain a snapshot for said key. Since we have already removed all completed
3361 34 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3362 34 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3363 34 : // and stop the search as a result of that.
3364 34 : let removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3365 34 : if !removed.is_empty() {
3366 0 : break Some(removed);
3367 34 : }
3368 34 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3369 34 : // keyspace.
3370 34 :
3371 34 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3372 34 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3373 34 : timeline_owned = timeline
3374 34 : .get_ready_ancestor_timeline(ctx)
3375 0 : .await
3376 34 : .map_err(GetVectoredError::GetReadyAncestorError)?;
3377 34 : timeline = &*timeline_owned;
3378 : };
3379 :
3380 470 : if let Some(missing_keyspace) = missing_keyspace {
3381 10 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3382 10 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3383 10 : shard: self
3384 10 : .shard_identity
3385 10 : .get_shard_number(&missing_keyspace.start().unwrap()),
3386 10 : cont_lsn,
3387 10 : request_lsn,
3388 10 : ancestor_lsn: Some(timeline.ancestor_lsn),
3389 10 : traversal_path: vec![],
3390 10 : backtrace: None,
3391 10 : }));
3392 460 : }
3393 460 :
3394 460 : Ok(())
3395 470 : }
3396 :
3397 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3398 : ///
3399 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3400 : /// the current keyspace. The current keyspace of the search at any given timeline
3401 : /// is the original keyspace minus all the keys that have been completed minus
3402 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3403 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3404 : ///
3405 : /// This is basically a depth-first search visitor implementation where a vertex
3406 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3407 : ///
3408 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3409 : /// and get all the required reconstruct data from the layer in one go.
3410 : ///
3411 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3412 : /// decides how to deal with these two keyspaces.
3413 504 : async fn get_vectored_reconstruct_data_timeline(
3414 504 : timeline: &Timeline,
3415 504 : keyspace: KeySpace,
3416 504 : mut cont_lsn: Lsn,
3417 504 : reconstruct_state: &mut ValuesReconstructState,
3418 504 : cancel: &CancellationToken,
3419 504 : ctx: &RequestContext,
3420 504 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3421 504 : let mut unmapped_keyspace = keyspace.clone();
3422 504 : let mut fringe = LayerFringe::new();
3423 504 :
3424 504 : let mut completed_keyspace = KeySpace::default();
3425 504 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3426 :
3427 868 : loop {
3428 868 : if cancel.is_cancelled() {
3429 0 : return Err(GetVectoredError::Cancelled);
3430 868 : }
3431 868 :
3432 868 : let (keys_done_last_step, keys_with_image_coverage) =
3433 868 : reconstruct_state.consume_done_keys();
3434 868 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3435 868 : completed_keyspace.merge(&keys_done_last_step);
3436 868 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3437 102 : unmapped_keyspace
3438 102 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3439 102 : image_covered_keyspace.add_range(keys_with_image_coverage);
3440 766 : }
3441 :
3442 : // Do not descent any further if the last layer we visited
3443 : // completed all keys in the keyspace it inspected. This is not
3444 : // required for correctness, but avoids visiting extra layers
3445 : // which turns out to be a perf bottleneck in some cases.
3446 868 : if !unmapped_keyspace.is_empty() {
3447 722 : let guard = timeline.layers.read().await;
3448 722 : let layers = guard.layer_map();
3449 722 :
3450 722 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3451 20 : let start_lsn = l.get_lsn_range().start;
3452 20 : cont_lsn > start_lsn
3453 722 : });
3454 722 :
3455 722 : match in_memory_layer {
3456 10 : Some(l) => {
3457 10 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3458 10 : fringe.update(
3459 10 : ReadableLayer::InMemoryLayer(l),
3460 10 : unmapped_keyspace.clone(),
3461 10 : lsn_range,
3462 10 : );
3463 10 : }
3464 : None => {
3465 72508 : for range in unmapped_keyspace.ranges.iter() {
3466 72508 : let results = layers.range_search(range.clone(), cont_lsn);
3467 72508 :
3468 72508 : results
3469 72508 : .found
3470 72508 : .into_iter()
3471 72508 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3472 64130 : (
3473 64130 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3474 64130 : keyspace_accum.to_keyspace(),
3475 64130 : lsn_floor..cont_lsn,
3476 64130 : )
3477 72508 : })
3478 72508 : .for_each(|(layer, keyspace, lsn_range)| {
3479 64130 : fringe.update(layer, keyspace, lsn_range)
3480 72508 : });
3481 72508 : }
3482 : }
3483 : }
3484 :
3485 : // It's safe to drop the layer map lock after planning the next round of reads.
3486 : // The fringe keeps readable handles for the layers which are safe to read even
3487 : // if layers were compacted or flushed.
3488 : //
3489 : // The more interesting consideration is: "Why is the read algorithm still correct
3490 : // if the layer map changes while it is operating?". Doing a vectored read on a
3491 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3492 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3493 : // range at *a particular point in time*. It is fine for the answer to be different
3494 : // at two different time points.
3495 722 : drop(guard);
3496 146 : }
3497 :
3498 868 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3499 364 : let next_cont_lsn = lsn_range.start;
3500 364 : layer_to_read
3501 364 : .get_values_reconstruct_data(
3502 364 : keyspace_to_read.clone(),
3503 364 : lsn_range,
3504 364 : reconstruct_state,
3505 364 : ctx,
3506 364 : )
3507 11516 : .await?;
3508 :
3509 364 : unmapped_keyspace = keyspace_to_read;
3510 364 : cont_lsn = next_cont_lsn;
3511 364 :
3512 364 : reconstruct_state.on_layer_visited(&layer_to_read);
3513 : } else {
3514 504 : break;
3515 504 : }
3516 504 : }
3517 504 :
3518 504 : Ok(TimelineVisitOutcome {
3519 504 : completed_keyspace,
3520 504 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3521 504 : })
3522 504 : }
3523 :
3524 : /// # Cancel-safety
3525 : ///
3526 : /// This method is cancellation-safe.
3527 624021 : async fn lookup_cached_page(
3528 624021 : &self,
3529 624021 : key: &Key,
3530 624021 : lsn: Lsn,
3531 624021 : ctx: &RequestContext,
3532 624021 : ) -> Option<(Lsn, Bytes)> {
3533 624021 : let cache = page_cache::get();
3534 :
3535 : // FIXME: It's pointless to check the cache for things that are not 8kB pages.
3536 : // We should look at the key to determine if it's a cacheable object
3537 624021 : let (lsn, read_guard) = cache
3538 624021 : .lookup_materialized_page(self.tenant_shard_id, self.timeline_id, key, lsn, ctx)
3539 624021 : .await?;
3540 0 : let img = Bytes::from(read_guard.to_vec());
3541 0 : Some((lsn, img))
3542 624021 : }
3543 :
3544 227446 : async fn get_ready_ancestor_timeline(
3545 227446 : &self,
3546 227446 : ctx: &RequestContext,
3547 227446 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3548 227446 : let ancestor = match self.get_ancestor_timeline() {
3549 227446 : Ok(timeline) => timeline,
3550 0 : Err(e) => return Err(GetReadyAncestorError::from(e)),
3551 : };
3552 :
3553 : // It's possible that the ancestor timeline isn't active yet, or
3554 : // is active but hasn't yet caught up to the branch point. Wait
3555 : // for it.
3556 : //
3557 : // This cannot happen while the pageserver is running normally,
3558 : // because you cannot create a branch from a point that isn't
3559 : // present in the pageserver yet. However, we don't wait for the
3560 : // branch point to be uploaded to cloud storage before creating
3561 : // a branch. I.e., the branch LSN need not be remote consistent
3562 : // for the branching operation to succeed.
3563 : //
3564 : // Hence, if we try to load a tenant in such a state where
3565 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3566 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3567 : // then we will need to wait for the ancestor timeline to
3568 : // re-stream WAL up to branch_lsn before we access it.
3569 : //
3570 : // How can a tenant get in such a state?
3571 : // - ungraceful pageserver process exit
3572 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3573 : //
3574 : // NB: this could be avoided by requiring
3575 : // branch_lsn >= remote_consistent_lsn
3576 : // during branch creation.
3577 227446 : match ancestor.wait_to_become_active(ctx).await {
3578 227444 : Ok(()) => {}
3579 : Err(TimelineState::Stopping) => {
3580 0 : return Err(GetReadyAncestorError::AncestorStopping(
3581 0 : ancestor.timeline_id,
3582 0 : ));
3583 : }
3584 2 : Err(state) => {
3585 2 : return Err(GetReadyAncestorError::Other(anyhow::anyhow!(
3586 2 : "Timeline {} will not become active. Current state: {:?}",
3587 2 : ancestor.timeline_id,
3588 2 : &state,
3589 2 : )));
3590 : }
3591 : }
3592 227444 : ancestor
3593 227444 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3594 0 : .await
3595 227444 : .map_err(|e| match e {
3596 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3597 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3598 0 : e @ WaitLsnError::BadState => GetReadyAncestorError::Other(anyhow::anyhow!(e)),
3599 227444 : })?;
3600 :
3601 227444 : Ok(ancestor)
3602 227446 : }
3603 :
3604 227446 : pub(crate) fn get_ancestor_timeline(&self) -> anyhow::Result<Arc<Timeline>> {
3605 227446 : let ancestor = self.ancestor_timeline.as_ref().with_context(|| {
3606 0 : format!(
3607 0 : "Ancestor is missing. Timeline id: {} Ancestor id {:?}",
3608 0 : self.timeline_id,
3609 0 : self.get_ancestor_timeline_id(),
3610 0 : )
3611 227446 : })?;
3612 227446 : Ok(Arc::clone(ancestor))
3613 227446 : }
3614 :
3615 5452 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3616 5452 : &self.shard_identity
3617 5452 : }
3618 :
3619 : ///
3620 : /// Get a handle to the latest layer for appending.
3621 : ///
3622 4804100 : async fn get_layer_for_write(
3623 4804100 : &self,
3624 4804100 : lsn: Lsn,
3625 4804100 : ctx: &RequestContext,
3626 4804100 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3627 4804100 : let mut guard = self.layers.write().await;
3628 4804100 : let layer = guard
3629 4804100 : .get_layer_for_write(
3630 4804100 : lsn,
3631 4804100 : self.get_last_record_lsn(),
3632 4804100 : self.conf,
3633 4804100 : self.timeline_id,
3634 4804100 : self.tenant_shard_id,
3635 4804100 : ctx,
3636 4804100 : )
3637 626 : .await?;
3638 4804100 : Ok(layer)
3639 4804100 : }
3640 :
3641 5279018 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3642 5279018 : assert!(new_lsn.is_aligned());
3643 :
3644 5279018 : self.metrics.last_record_gauge.set(new_lsn.0 as i64);
3645 5279018 : self.last_record_lsn.advance(new_lsn);
3646 5279018 : }
3647 :
3648 : /// Whether there was a layer to freeze or not, return the value of get_last_record_lsn
3649 : /// before we attempted the freeze: this guarantees that ingested data is frozen up to this lsn (inclusive).
3650 1038 : async fn freeze_inmem_layer(&self, write_lock_held: bool) -> Lsn {
3651 : // Freeze the current open in-memory layer. It will be written to disk on next
3652 : // iteration.
3653 :
3654 1038 : let _write_guard = if write_lock_held {
3655 0 : None
3656 : } else {
3657 1038 : Some(self.write_lock.lock().await)
3658 : };
3659 :
3660 1038 : let to_lsn = self.get_last_record_lsn();
3661 1038 : self.freeze_inmem_layer_at(to_lsn).await;
3662 1038 : to_lsn
3663 1038 : }
3664 :
3665 1038 : async fn freeze_inmem_layer_at(&self, at: Lsn) {
3666 1038 : let mut guard = self.layers.write().await;
3667 1038 : guard
3668 1038 : .try_freeze_in_memory_layer(at, &self.last_freeze_at)
3669 4 : .await;
3670 1038 : }
3671 :
3672 : /// Layer flusher task's main loop.
3673 346 : async fn flush_loop(
3674 346 : self: &Arc<Self>,
3675 346 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3676 346 : ctx: &RequestContext,
3677 346 : ) {
3678 346 : info!("started flush loop");
3679 1038 : loop {
3680 1038 : tokio::select! {
3681 : _ = self.cancel.cancelled() => {
3682 : info!("shutting down layer flush task due to Timeline::cancel");
3683 : break;
3684 : },
3685 : _ = layer_flush_start_rx.changed() => {}
3686 : }
3687 1038 : trace!("waking up");
3688 1038 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3689 1038 :
3690 1038 : // The highest LSN to which we flushed in the loop over frozen layers
3691 1038 : let mut flushed_to_lsn = Lsn(0);
3692 :
3693 1038 : let result = loop {
3694 2048 : if self.cancel.is_cancelled() {
3695 0 : info!("dropping out of flush loop for timeline shutdown");
3696 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3697 : // anyone waiting on that will respect self.cancel as well: they will stop
3698 : // waiting at the same time we as drop out of this loop.
3699 0 : return;
3700 2048 : }
3701 2048 :
3702 2048 : let timer = self.metrics.flush_time_histo.start_timer();
3703 :
3704 2048 : let layer_to_flush = {
3705 2048 : let guard = self.layers.read().await;
3706 2048 : guard.layer_map().frozen_layers.front().cloned()
3707 : // drop 'layers' lock to allow concurrent reads and writes
3708 : };
3709 2048 : let Some(layer_to_flush) = layer_to_flush else {
3710 1038 : break Ok(());
3711 : };
3712 58312 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3713 1010 : Ok(this_layer_to_lsn) => {
3714 1010 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3715 1010 : }
3716 : Err(FlushLayerError::Cancelled) => {
3717 0 : info!("dropping out of flush loop for timeline shutdown");
3718 0 : return;
3719 : }
3720 0 : err @ Err(
3721 : FlushLayerError::Other(_) | FlushLayerError::CreateImageLayersError(_),
3722 : ) => {
3723 0 : error!("could not flush frozen layer: {err:?}");
3724 0 : break err.map(|_| ());
3725 : }
3726 : }
3727 1010 : timer.stop_and_record();
3728 : };
3729 :
3730 : // Unsharded tenants should never advance their LSN beyond the end of the
3731 : // highest layer they write: such gaps between layer data and the frozen LSN
3732 : // are only legal on sharded tenants.
3733 1038 : debug_assert!(
3734 1038 : self.shard_identity.count.count() > 1
3735 1038 : || flushed_to_lsn >= frozen_to_lsn
3736 28 : || !flushed_to_lsn.is_valid()
3737 : );
3738 :
3739 1038 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3740 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3741 : // to us via layer_flush_start_rx, then advance it here.
3742 : //
3743 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3744 : // never encounter a gap in the wal.
3745 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3746 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3747 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3748 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3749 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3750 0 : }
3751 0 : }
3752 1038 : }
3753 :
3754 : // Notify any listeners that we're done
3755 1038 : let _ = self
3756 1038 : .layer_flush_done_tx
3757 1038 : .send_replace((flush_counter, result));
3758 : }
3759 8 : }
3760 :
3761 : /// Request the flush loop to write out all frozen layers up to `to_lsn` as Delta L0 files to disk.
3762 : /// The caller is responsible for the freezing, e.g., [`Self::freeze_inmem_layer`].
3763 : ///
3764 : /// `last_record_lsn` may be higher than the highest LSN of a frozen layer: if this is the case,
3765 : /// it means no data will be written between the top of the highest frozen layer and to_lsn,
3766 : /// e.g. because this tenant shard has ingested up to to_lsn and not written any data locally for that part of the WAL.
3767 1038 : async fn flush_frozen_layers_and_wait(&self, last_record_lsn: Lsn) -> anyhow::Result<()> {
3768 1038 : let mut rx = self.layer_flush_done_tx.subscribe();
3769 1038 :
3770 1038 : // Increment the flush cycle counter and wake up the flush task.
3771 1038 : // Remember the new value, so that when we listen for the flush
3772 1038 : // to finish, we know when the flush that we initiated has
3773 1038 : // finished, instead of some other flush that was started earlier.
3774 1038 : let mut my_flush_request = 0;
3775 1038 :
3776 1038 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3777 1038 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3778 0 : anyhow::bail!("cannot flush frozen layers when flush_loop is not running, state is {flush_loop_state:?}")
3779 1038 : }
3780 1038 :
3781 1038 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3782 1038 : my_flush_request = *counter + 1;
3783 1038 : *counter = my_flush_request;
3784 1038 : *lsn = std::cmp::max(last_record_lsn, *lsn);
3785 1038 : });
3786 :
3787 2069 : loop {
3788 2069 : {
3789 2069 : let (last_result_counter, last_result) = &*rx.borrow();
3790 2069 : if *last_result_counter >= my_flush_request {
3791 1038 : if let Err(_err) = last_result {
3792 : // We already logged the original error in
3793 : // flush_loop. We cannot propagate it to the caller
3794 : // here, because it might not be Cloneable
3795 0 : anyhow::bail!(
3796 0 : "Could not flush frozen layer. Request id: {}",
3797 0 : my_flush_request
3798 0 : );
3799 : } else {
3800 1038 : return Ok(());
3801 : }
3802 1031 : }
3803 1031 : }
3804 1031 : trace!("waiting for flush to complete");
3805 : tokio::select! {
3806 : rx_e = rx.changed() => {
3807 : rx_e?;
3808 : },
3809 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3810 : // the notification from [`flush_loop`] that it completed.
3811 : _ = self.cancel.cancelled() => {
3812 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3813 : return Ok(())
3814 : }
3815 : };
3816 1031 : trace!("done")
3817 : }
3818 1038 : }
3819 :
3820 0 : fn flush_frozen_layers(&self) {
3821 0 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3822 0 : *counter += 1;
3823 0 :
3824 0 : *lsn = std::cmp::max(*lsn, Lsn(self.last_freeze_at.load().0 - 1));
3825 0 : });
3826 0 : }
3827 :
3828 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3829 : ///
3830 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3831 2020 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3832 : async fn flush_frozen_layer(
3833 : self: &Arc<Self>,
3834 : frozen_layer: Arc<InMemoryLayer>,
3835 : ctx: &RequestContext,
3836 : ) -> Result<Lsn, FlushLayerError> {
3837 : debug_assert_current_span_has_tenant_and_timeline_id();
3838 :
3839 : // As a special case, when we have just imported an image into the repository,
3840 : // instead of writing out a L0 delta layer, we directly write out image layer
3841 : // files instead. This is possible as long as *all* the data imported into the
3842 : // repository have the same LSN.
3843 : let lsn_range = frozen_layer.get_lsn_range();
3844 :
3845 : // Whether to directly create image layers for this flush, or flush them as delta layers
3846 : let create_image_layer =
3847 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3848 :
3849 : #[cfg(test)]
3850 : {
3851 : match &mut *self.flush_loop_state.lock().unwrap() {
3852 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3853 : panic!("flush loop not running")
3854 : }
3855 : FlushLoopState::Running {
3856 : expect_initdb_optimization,
3857 : initdb_optimization_count,
3858 : ..
3859 : } => {
3860 : if create_image_layer {
3861 : *initdb_optimization_count += 1;
3862 : } else {
3863 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3864 : }
3865 : }
3866 : }
3867 : }
3868 :
3869 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3870 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3871 : // require downloading anything during initial import.
3872 : let ((rel_partition, metadata_partition), _lsn) = self
3873 : .repartition(
3874 : self.initdb_lsn,
3875 : self.get_compaction_target_size(),
3876 : EnumSet::empty(),
3877 : ctx,
3878 : )
3879 : .await?;
3880 :
3881 : if self.cancel.is_cancelled() {
3882 : return Err(FlushLayerError::Cancelled);
3883 : }
3884 :
3885 : // For metadata, always create delta layers.
3886 : let delta_layer = if !metadata_partition.parts.is_empty() {
3887 : assert_eq!(
3888 : metadata_partition.parts.len(),
3889 : 1,
3890 : "currently sparse keyspace should only contain a single aux file keyspace"
3891 : );
3892 : let metadata_keyspace = &metadata_partition.parts[0];
3893 : assert_eq!(
3894 : metadata_keyspace.0.ranges.len(),
3895 : 1,
3896 : "aux file keyspace should be a single range"
3897 : );
3898 : self.create_delta_layer(
3899 : &frozen_layer,
3900 : Some(metadata_keyspace.0.ranges[0].clone()),
3901 : ctx,
3902 : )
3903 : .await?
3904 : } else {
3905 : None
3906 : };
3907 :
3908 : // For image layers, we add them immediately into the layer map.
3909 : let mut layers_to_upload = Vec::new();
3910 : layers_to_upload.extend(
3911 : self.create_image_layers(
3912 : &rel_partition,
3913 : self.initdb_lsn,
3914 : ImageLayerCreationMode::Initial,
3915 : ctx,
3916 : )
3917 : .await?,
3918 : );
3919 :
3920 : if let Some(delta_layer) = delta_layer {
3921 : layers_to_upload.push(delta_layer.clone());
3922 : (layers_to_upload, Some(delta_layer))
3923 : } else {
3924 : (layers_to_upload, None)
3925 : }
3926 : } else {
3927 : // Normal case, write out a L0 delta layer file.
3928 : // `create_delta_layer` will not modify the layer map.
3929 : // We will remove frozen layer and add delta layer in one atomic operation later.
3930 : let Some(layer) = self.create_delta_layer(&frozen_layer, None, ctx).await? else {
3931 : panic!("delta layer cannot be empty if no filter is applied");
3932 : };
3933 : (
3934 : // FIXME: even though we have a single image and single delta layer assumption
3935 : // we push them to vec
3936 : vec![layer.clone()],
3937 : Some(layer),
3938 : )
3939 : };
3940 :
3941 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
3942 :
3943 : if self.cancel.is_cancelled() {
3944 : return Err(FlushLayerError::Cancelled);
3945 : }
3946 :
3947 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
3948 :
3949 : // The new on-disk layers are now in the layer map. We can remove the
3950 : // in-memory layer from the map now. The flushed layer is stored in
3951 : // the mapping in `create_delta_layer`.
3952 : {
3953 : let mut guard = self.layers.write().await;
3954 :
3955 : if self.cancel.is_cancelled() {
3956 : return Err(FlushLayerError::Cancelled);
3957 : }
3958 :
3959 : guard.finish_flush_l0_layer(delta_layer_to_add.as_ref(), &frozen_layer, &self.metrics);
3960 :
3961 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
3962 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
3963 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)?;
3964 : }
3965 : // release lock on 'layers'
3966 : };
3967 :
3968 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
3969 : // a compaction can delete the file and then it won't be available for uploads any more.
3970 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
3971 : // race situation.
3972 : // See https://github.com/neondatabase/neon/issues/4526
3973 : pausable_failpoint!("flush-frozen-pausable");
3974 :
3975 : // This failpoint is used by another test case `test_pageserver_recovery`.
3976 : fail_point!("flush-frozen-exit");
3977 :
3978 : Ok(Lsn(lsn_range.end.0 - 1))
3979 : }
3980 :
3981 : /// Return true if the value changed
3982 : ///
3983 : /// This function must only be used from the layer flush task, and may not be called concurrently.
3984 1010 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
3985 1010 : // We do a simple load/store cycle: that's why this function isn't safe for concurrent use.
3986 1010 : let old_value = self.disk_consistent_lsn.load();
3987 1010 : if new_value != old_value {
3988 1010 : assert!(new_value >= old_value);
3989 1010 : self.disk_consistent_lsn.store(new_value);
3990 1010 : true
3991 : } else {
3992 0 : false
3993 : }
3994 1010 : }
3995 :
3996 : /// Update metadata file
3997 1014 : fn schedule_uploads(
3998 1014 : &self,
3999 1014 : disk_consistent_lsn: Lsn,
4000 1014 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
4001 1014 : ) -> anyhow::Result<()> {
4002 1014 : // We can only save a valid 'prev_record_lsn' value on disk if we
4003 1014 : // flushed *all* in-memory changes to disk. We only track
4004 1014 : // 'prev_record_lsn' in memory for the latest processed record, so we
4005 1014 : // don't remember what the correct value that corresponds to some old
4006 1014 : // LSN is. But if we flush everything, then the value corresponding
4007 1014 : // current 'last_record_lsn' is correct and we can store it on disk.
4008 1014 : let RecordLsn {
4009 1014 : last: last_record_lsn,
4010 1014 : prev: prev_record_lsn,
4011 1014 : } = self.last_record_lsn.load();
4012 1014 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
4013 1014 : Some(prev_record_lsn)
4014 : } else {
4015 0 : None
4016 : };
4017 :
4018 1014 : let update = crate::tenant::metadata::MetadataUpdate::new(
4019 1014 : disk_consistent_lsn,
4020 1014 : ondisk_prev_record_lsn,
4021 1014 : *self.latest_gc_cutoff_lsn.read(),
4022 1014 : );
4023 1014 :
4024 1014 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
4025 0 : "{}",
4026 0 : x.unwrap()
4027 1014 : ));
4028 :
4029 2038 : for layer in layers_to_upload {
4030 1024 : self.remote_client.schedule_layer_file_upload(layer)?;
4031 : }
4032 1014 : self.remote_client
4033 1014 : .schedule_index_upload_for_metadata_update(&update)?;
4034 :
4035 1014 : Ok(())
4036 1014 : }
4037 :
4038 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
4039 0 : self.remote_client
4040 0 : .preserve_initdb_archive(
4041 0 : &self.tenant_shard_id.tenant_id,
4042 0 : &self.timeline_id,
4043 0 : &self.cancel,
4044 0 : )
4045 0 : .await
4046 0 : }
4047 :
4048 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
4049 : // in layer map immediately. The caller is responsible to put it into the layer map.
4050 1010 : async fn create_delta_layer(
4051 1010 : self: &Arc<Self>,
4052 1010 : frozen_layer: &Arc<InMemoryLayer>,
4053 1010 : key_range: Option<Range<Key>>,
4054 1010 : ctx: &RequestContext,
4055 1010 : ) -> anyhow::Result<Option<ResidentLayer>> {
4056 1010 : let self_clone = Arc::clone(self);
4057 1010 : let frozen_layer = Arc::clone(frozen_layer);
4058 1010 : let ctx = ctx.attached_child();
4059 1010 : let work = async move {
4060 1010 : let Some(new_delta) = frozen_layer
4061 1010 : .write_to_disk(&self_clone, &ctx, key_range)
4062 85556 : .await?
4063 : else {
4064 114 : return Ok(None);
4065 : };
4066 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
4067 : // We just need to fsync the directory in which these inodes are linked,
4068 : // which we know to be the timeline directory.
4069 : //
4070 : // We use fatal_err() below because the after write_to_disk returns with success,
4071 : // the in-memory state of the filesystem already has the layer file in its final place,
4072 : // and subsequent pageserver code could think it's durable while it really isn't.
4073 896 : let timeline_dir = VirtualFile::open(
4074 896 : &self_clone
4075 896 : .conf
4076 896 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
4077 896 : &ctx,
4078 896 : )
4079 448 : .await
4080 896 : .fatal_err("VirtualFile::open for timeline dir fsync");
4081 896 : timeline_dir
4082 896 : .sync_all()
4083 448 : .await
4084 896 : .fatal_err("VirtualFile::sync_all timeline dir");
4085 896 : anyhow::Ok(Some(new_delta))
4086 1010 : };
4087 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
4088 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
4089 : use crate::virtual_file::io_engine::IoEngine;
4090 1010 : match crate::virtual_file::io_engine::get() {
4091 0 : IoEngine::NotSet => panic!("io engine not set"),
4092 : IoEngine::StdFs => {
4093 505 : let span = tracing::info_span!("blocking");
4094 505 : tokio::task::spawn_blocking({
4095 505 : move || Handle::current().block_on(work.instrument(span))
4096 505 : })
4097 505 : .await
4098 505 : .context("spawn_blocking")
4099 505 : .and_then(|x| x)
4100 : }
4101 : #[cfg(target_os = "linux")]
4102 54366 : IoEngine::TokioEpollUring => work.await,
4103 : }
4104 1010 : }
4105 :
4106 480 : async fn repartition(
4107 480 : &self,
4108 480 : lsn: Lsn,
4109 480 : partition_size: u64,
4110 480 : flags: EnumSet<CompactFlags>,
4111 480 : ctx: &RequestContext,
4112 480 : ) -> anyhow::Result<((KeyPartitioning, SparseKeyPartitioning), Lsn)> {
4113 480 : let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
4114 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
4115 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
4116 : // and hence before the compaction task starts.
4117 0 : anyhow::bail!("repartition() called concurrently, this should not happen");
4118 : };
4119 480 : let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
4120 480 : if lsn < *partition_lsn {
4121 0 : anyhow::bail!("repartition() called with LSN going backwards, this should not happen");
4122 480 : }
4123 480 :
4124 480 : let distance = lsn.0 - partition_lsn.0;
4125 480 : if *partition_lsn != Lsn(0)
4126 262 : && distance <= self.repartition_threshold
4127 262 : && !flags.contains(CompactFlags::ForceRepartition)
4128 : {
4129 250 : debug!(
4130 : distance,
4131 : threshold = self.repartition_threshold,
4132 0 : "no repartitioning needed"
4133 : );
4134 250 : return Ok((
4135 250 : (dense_partition.clone(), sparse_partition.clone()),
4136 250 : *partition_lsn,
4137 250 : ));
4138 230 : }
4139 :
4140 13419 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4141 230 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4142 230 : let sparse_partitioning = SparseKeyPartitioning {
4143 230 : parts: vec![sparse_ks],
4144 230 : }; // no partitioning for metadata keys for now
4145 230 : *partitioning_guard = ((dense_partitioning, sparse_partitioning), lsn);
4146 230 :
4147 230 : Ok((partitioning_guard.0.clone(), partitioning_guard.1))
4148 480 : }
4149 :
4150 : // Is it time to create a new image layer for the given partition?
4151 14 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4152 14 : let threshold = self.get_image_creation_threshold();
4153 :
4154 14 : let guard = self.layers.read().await;
4155 14 : let layers = guard.layer_map();
4156 14 :
4157 14 : let mut max_deltas = 0;
4158 28 : for part_range in &partition.ranges {
4159 14 : let image_coverage = layers.image_coverage(part_range, lsn);
4160 28 : for (img_range, last_img) in image_coverage {
4161 14 : let img_lsn = if let Some(last_img) = last_img {
4162 0 : last_img.get_lsn_range().end
4163 : } else {
4164 14 : Lsn(0)
4165 : };
4166 : // Let's consider an example:
4167 : //
4168 : // delta layer with LSN range 71-81
4169 : // delta layer with LSN range 81-91
4170 : // delta layer with LSN range 91-101
4171 : // image layer at LSN 100
4172 : //
4173 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4174 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4175 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4176 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4177 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4178 14 : if img_lsn < lsn {
4179 14 : let num_deltas =
4180 14 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4181 14 :
4182 14 : max_deltas = max_deltas.max(num_deltas);
4183 14 : if num_deltas >= threshold {
4184 0 : debug!(
4185 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4186 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4187 : );
4188 0 : return true;
4189 14 : }
4190 0 : }
4191 : }
4192 : }
4193 :
4194 14 : debug!(
4195 : max_deltas,
4196 0 : "none of the partitioned ranges had >= {threshold} deltas"
4197 : );
4198 14 : false
4199 14 : }
4200 :
4201 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4202 : /// so that at most one image layer will be produced from this function.
4203 144 : async fn create_image_layer_for_rel_blocks(
4204 144 : self: &Arc<Self>,
4205 144 : partition: &KeySpace,
4206 144 : mut image_layer_writer: ImageLayerWriter,
4207 144 : lsn: Lsn,
4208 144 : ctx: &RequestContext,
4209 144 : img_range: Range<Key>,
4210 144 : start: Key,
4211 144 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4212 144 : let mut wrote_keys = false;
4213 144 :
4214 144 : let mut key_request_accum = KeySpaceAccum::new();
4215 918 : for range in &partition.ranges {
4216 774 : let mut key = range.start;
4217 1800 : while key < range.end {
4218 : // Decide whether to retain this key: usually we do, but sharded tenants may
4219 : // need to drop keys that don't belong to them. If we retain the key, add it
4220 : // to `key_request_accum` for later issuing a vectored get
4221 1026 : if self.shard_identity.is_key_disposable(&key) {
4222 0 : debug!(
4223 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4224 0 : key,
4225 0 : self.shard_identity.get_shard_number(&key)
4226 : );
4227 1026 : } else {
4228 1026 : key_request_accum.add_key(key);
4229 1026 : }
4230 :
4231 1026 : let last_key_in_range = key.next() == range.end;
4232 1026 : key = key.next();
4233 1026 :
4234 1026 : // Maybe flush `key_rest_accum`
4235 1026 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4236 1026 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4237 : {
4238 774 : let results = self
4239 774 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
4240 18 : .await?;
4241 :
4242 1800 : for (img_key, img) in results {
4243 1026 : let img = match img {
4244 1026 : Ok(img) => img,
4245 0 : Err(err) => {
4246 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4247 0 : // page without losing any actual user data. That seems better
4248 0 : // than failing repeatedly and getting stuck.
4249 0 : //
4250 0 : // We had a bug at one point, where we truncated the FSM and VM
4251 0 : // in the pageserver, but the Postgres didn't know about that
4252 0 : // and continued to generate incremental WAL records for pages
4253 0 : // that didn't exist in the pageserver. Trying to replay those
4254 0 : // WAL records failed to find the previous image of the page.
4255 0 : // This special case allows us to recover from that situation.
4256 0 : // See https://github.com/neondatabase/neon/issues/2601.
4257 0 : //
4258 0 : // Unfortunately we cannot do this for the main fork, or for
4259 0 : // any metadata keys, keys, as that would lead to actual data
4260 0 : // loss.
4261 0 : if is_rel_fsm_block_key(img_key) || is_rel_vm_block_key(img_key) {
4262 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4263 0 : ZERO_PAGE.clone()
4264 : } else {
4265 0 : return Err(CreateImageLayersError::PageReconstructError(err));
4266 : }
4267 : }
4268 : };
4269 :
4270 : // Write all the keys we just read into our new image layer.
4271 1091 : image_layer_writer.put_image(img_key, img, ctx).await?;
4272 1026 : wrote_keys = true;
4273 : }
4274 252 : }
4275 : }
4276 : }
4277 :
4278 144 : if wrote_keys {
4279 : // Normal path: we have written some data into the new image layer for this
4280 : // partition, so flush it to disk.
4281 289 : let image_layer = image_layer_writer.finish(self, ctx).await?;
4282 144 : Ok(ImageLayerCreationOutcome {
4283 144 : image: Some(image_layer),
4284 144 : next_start_key: img_range.end,
4285 144 : })
4286 : } else {
4287 : // Special case: the image layer may be empty if this is a sharded tenant and the
4288 : // partition does not cover any keys owned by this shard. In this case, to ensure
4289 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4290 : // layer we write will cover the key range that we just scanned.
4291 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4292 0 : Ok(ImageLayerCreationOutcome {
4293 0 : image: None,
4294 0 : next_start_key: start,
4295 0 : })
4296 : }
4297 144 : }
4298 :
4299 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4300 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4301 : /// would not be too large to fit in a single image layer.
4302 : #[allow(clippy::too_many_arguments)]
4303 364 : async fn create_image_layer_for_metadata_keys(
4304 364 : self: &Arc<Self>,
4305 364 : partition: &KeySpace,
4306 364 : mut image_layer_writer: ImageLayerWriter,
4307 364 : lsn: Lsn,
4308 364 : ctx: &RequestContext,
4309 364 : img_range: Range<Key>,
4310 364 : mode: ImageLayerCreationMode,
4311 364 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4312 364 : assert!(!matches!(mode, ImageLayerCreationMode::Initial));
4313 :
4314 : // Metadata keys image layer creation.
4315 364 : let mut reconstruct_state = ValuesReconstructState::default();
4316 364 : let data = self
4317 364 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4318 4129 : .await?;
4319 364 : let (data, total_kb_retrieved, total_key_retrieved) = {
4320 364 : let mut new_data = BTreeMap::new();
4321 364 : let mut total_kb_retrieved = 0;
4322 364 : let mut total_key_retrieved = 0;
4323 48368 : for (k, v) in data {
4324 48004 : let v = v.map_err(CreateImageLayersError::PageReconstructError)?;
4325 48004 : total_kb_retrieved += KEY_SIZE + v.len();
4326 48004 : total_key_retrieved += 1;
4327 48004 : new_data.insert(k, v);
4328 : }
4329 364 : (new_data, total_kb_retrieved / 1024, total_key_retrieved)
4330 364 : };
4331 364 : let delta_file_accessed = reconstruct_state.get_delta_layers_visited();
4332 364 :
4333 364 : let trigger_generation = delta_file_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4334 364 : info!(
4335 0 : "generate image layers for metadata keys: trigger_generation={trigger_generation}, \
4336 0 : delta_file_accessed={delta_file_accessed}, total_kb_retrieved={total_kb_retrieved}, \
4337 0 : total_key_retrieved={total_key_retrieved}"
4338 : );
4339 364 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4340 350 : return Ok(ImageLayerCreationOutcome {
4341 350 : image: None,
4342 350 : next_start_key: img_range.end,
4343 350 : });
4344 14 : }
4345 14 : let has_keys = !data.is_empty();
4346 10018 : for (k, v) in data {
4347 : // Even if the value is empty (deleted), we do not delete it for now until we can ensure vectored get
4348 : // considers this situation properly.
4349 : // if v.is_empty() {
4350 : // continue;
4351 : // }
4352 :
4353 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4354 :
4355 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4356 : // on the normal data path either.
4357 10160 : image_layer_writer.put_image(k, v, ctx).await?;
4358 : }
4359 : Ok(ImageLayerCreationOutcome {
4360 14 : image: if has_keys {
4361 29 : let image_layer = image_layer_writer.finish(self, ctx).await?;
4362 14 : Some(image_layer)
4363 : } else {
4364 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4365 0 : None
4366 : },
4367 14 : next_start_key: img_range.end,
4368 : })
4369 364 : }
4370 :
4371 960 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4372 : async fn create_image_layers(
4373 : self: &Arc<Timeline>,
4374 : partitioning: &KeyPartitioning,
4375 : lsn: Lsn,
4376 : mode: ImageLayerCreationMode,
4377 : ctx: &RequestContext,
4378 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4379 : let timer = self.metrics.create_images_time_histo.start_timer();
4380 : let mut image_layers = Vec::new();
4381 :
4382 : // We need to avoid holes between generated image layers.
4383 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4384 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4385 : //
4386 : // How such hole between partitions can appear?
4387 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4388 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4389 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4390 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4391 : let mut start = Key::MIN;
4392 :
4393 : let check_for_image_layers = {
4394 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4395 : let distance = lsn
4396 : .checked_sub(last_checks_at)
4397 : .expect("Attempt to compact with LSN going backwards");
4398 : let min_distance = self.get_image_layer_creation_check_threshold() as u64
4399 : * self.get_checkpoint_distance();
4400 :
4401 : // Skip the expensive delta layer counting if this timeline has not ingested sufficient
4402 : // WAL since the last check.
4403 : distance.0 >= min_distance
4404 : };
4405 :
4406 : if check_for_image_layers {
4407 : self.last_image_layer_creation_check_at.store(lsn);
4408 : }
4409 :
4410 : for partition in partitioning.parts.iter() {
4411 : let img_range = start..partition.ranges.last().unwrap().end;
4412 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4413 : if compact_metadata {
4414 : for range in &partition.ranges {
4415 : assert!(
4416 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4417 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4418 : "metadata keys must be partitioned separately"
4419 : );
4420 : }
4421 : if mode == ImageLayerCreationMode::Initial {
4422 : return Err(CreateImageLayersError::Other(anyhow::anyhow!("no image layer should be created for metadata keys when flushing frozen layers")));
4423 : }
4424 : } else if let ImageLayerCreationMode::Try = mode {
4425 : // check_for_image_layers = false -> skip
4426 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4427 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4428 : start = img_range.end;
4429 : continue;
4430 : }
4431 : }
4432 :
4433 : let image_layer_writer = ImageLayerWriter::new(
4434 : self.conf,
4435 : self.timeline_id,
4436 : self.tenant_shard_id,
4437 : &img_range,
4438 : lsn,
4439 : ctx,
4440 : )
4441 : .await?;
4442 :
4443 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4444 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4445 0 : "failpoint image-layer-writer-fail-before-finish"
4446 0 : )))
4447 0 : });
4448 :
4449 : if !compact_metadata {
4450 : let ImageLayerCreationOutcome {
4451 : image,
4452 : next_start_key,
4453 : } = self
4454 : .create_image_layer_for_rel_blocks(
4455 : partition,
4456 : image_layer_writer,
4457 : lsn,
4458 : ctx,
4459 : img_range,
4460 : start,
4461 : )
4462 : .await?;
4463 :
4464 : start = next_start_key;
4465 : image_layers.extend(image);
4466 : } else {
4467 : let ImageLayerCreationOutcome {
4468 : image,
4469 : next_start_key,
4470 : } = self
4471 : .create_image_layer_for_metadata_keys(
4472 : partition,
4473 : image_layer_writer,
4474 : lsn,
4475 : ctx,
4476 : img_range,
4477 : mode,
4478 : )
4479 : .await?;
4480 : start = next_start_key;
4481 : image_layers.extend(image);
4482 : }
4483 : }
4484 :
4485 : // The writer.finish() above already did the fsync of the inodes.
4486 : // We just need to fsync the directory in which these inodes are linked,
4487 : // which we know to be the timeline directory.
4488 : if !image_layers.is_empty() {
4489 : // We use fatal_err() below because the after writer.finish() returns with success,
4490 : // the in-memory state of the filesystem already has the layer file in its final place,
4491 : // and subsequent pageserver code could think it's durable while it really isn't.
4492 : let timeline_dir = VirtualFile::open(
4493 : &self
4494 : .conf
4495 : .timeline_path(&self.tenant_shard_id, &self.timeline_id),
4496 : ctx,
4497 : )
4498 : .await
4499 : .fatal_err("VirtualFile::open for timeline dir fsync");
4500 : timeline_dir
4501 : .sync_all()
4502 : .await
4503 : .fatal_err("VirtualFile::sync_all timeline dir");
4504 : }
4505 :
4506 : let mut guard = self.layers.write().await;
4507 :
4508 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4509 : // now they are being scheduled outside of write lock
4510 : guard.track_new_image_layers(&image_layers, &self.metrics);
4511 : drop_wlock(guard);
4512 : timer.stop_and_record();
4513 :
4514 : Ok(image_layers)
4515 : }
4516 :
4517 : /// Wait until the background initial logical size calculation is complete, or
4518 : /// this Timeline is shut down. Calling this function will cause the initial
4519 : /// logical size calculation to skip waiting for the background jobs barrier.
4520 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4521 0 : if !self.shard_identity.is_shard_zero() {
4522 : // We don't populate logical size on shard >0: skip waiting for it.
4523 0 : return;
4524 0 : }
4525 0 :
4526 0 : if self.remote_client.is_deleting() {
4527 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4528 0 : return;
4529 0 : }
4530 :
4531 0 : if let Some(await_bg_cancel) = self
4532 0 : .current_logical_size
4533 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4534 0 : .get()
4535 0 : {
4536 0 : await_bg_cancel.cancel();
4537 0 : } else {
4538 : // We should not wait if we were not able to explicitly instruct
4539 : // the logical size cancellation to skip the concurrency limit semaphore.
4540 : // TODO: this is an unexpected case. We should restructure so that it
4541 : // can't happen.
4542 0 : tracing::warn!(
4543 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4544 : );
4545 0 : debug_assert!(false);
4546 : }
4547 :
4548 : tokio::select!(
4549 : _ = self.current_logical_size.initialized.acquire() => {},
4550 : _ = self.cancel.cancelled() => {}
4551 : )
4552 0 : }
4553 :
4554 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4555 : /// Timelines layers up to the ancestor_lsn.
4556 : ///
4557 : /// Requires a timeline that:
4558 : /// - has an ancestor to detach from
4559 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4560 : /// a technical requirement
4561 : ///
4562 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4563 : /// polled again until completion.
4564 : ///
4565 : /// During the operation all timelines sharing the data with this timeline will be reparented
4566 : /// from our ancestor to be branches of this timeline.
4567 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4568 0 : self: &Arc<Timeline>,
4569 0 : tenant: &crate::tenant::Tenant,
4570 0 : options: detach_ancestor::Options,
4571 0 : ctx: &RequestContext,
4572 0 : ) -> Result<
4573 0 : (
4574 0 : completion::Completion,
4575 0 : detach_ancestor::PreparedTimelineDetach,
4576 0 : ),
4577 0 : detach_ancestor::Error,
4578 0 : > {
4579 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4580 0 : }
4581 :
4582 : /// Completes the ancestor detach. This method is to be called while holding the
4583 : /// TenantManager's tenant slot, so during this method we cannot be deleted nor can any
4584 : /// timeline be deleted. After this method returns successfully, tenant must be reloaded.
4585 : ///
4586 : /// Pageserver receiving a SIGKILL during this operation is not supported (yet).
4587 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4588 0 : self: &Arc<Timeline>,
4589 0 : tenant: &crate::tenant::Tenant,
4590 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4591 0 : ctx: &RequestContext,
4592 0 : ) -> Result<Vec<TimelineId>, anyhow::Error> {
4593 0 : detach_ancestor::complete(self, tenant, prepared, ctx).await
4594 0 : }
4595 : }
4596 :
4597 : /// Top-level failure to compact.
4598 0 : #[derive(Debug, thiserror::Error)]
4599 : pub(crate) enum CompactionError {
4600 : #[error("The timeline or pageserver is shutting down")]
4601 : ShuttingDown,
4602 : /// Compaction cannot be done right now; page reconstruction and so on.
4603 : #[error(transparent)]
4604 : Other(#[from] anyhow::Error),
4605 : }
4606 :
4607 : impl From<CollectKeySpaceError> for CompactionError {
4608 0 : fn from(err: CollectKeySpaceError) -> Self {
4609 0 : match err {
4610 : CollectKeySpaceError::Cancelled
4611 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4612 0 : CompactionError::ShuttingDown
4613 : }
4614 0 : e => CompactionError::Other(e.into()),
4615 : }
4616 0 : }
4617 : }
4618 :
4619 : #[serde_as]
4620 196 : #[derive(serde::Serialize)]
4621 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4622 :
4623 : #[derive(Default)]
4624 : enum DurationRecorder {
4625 : #[default]
4626 : NotStarted,
4627 : Recorded(RecordedDuration, tokio::time::Instant),
4628 : }
4629 :
4630 : impl DurationRecorder {
4631 506 : fn till_now(&self) -> DurationRecorder {
4632 506 : match self {
4633 : DurationRecorder::NotStarted => {
4634 0 : panic!("must only call on recorded measurements")
4635 : }
4636 506 : DurationRecorder::Recorded(_, ended) => {
4637 506 : let now = tokio::time::Instant::now();
4638 506 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4639 506 : }
4640 506 : }
4641 506 : }
4642 196 : fn into_recorded(self) -> Option<RecordedDuration> {
4643 196 : match self {
4644 0 : DurationRecorder::NotStarted => None,
4645 196 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4646 : }
4647 196 : }
4648 : }
4649 :
4650 : impl Timeline {
4651 28 : async fn finish_compact_batch(
4652 28 : self: &Arc<Self>,
4653 28 : new_deltas: &[ResidentLayer],
4654 28 : new_images: &[ResidentLayer],
4655 28 : layers_to_remove: &[Layer],
4656 28 : ) -> anyhow::Result<()> {
4657 28 : let mut guard = self.layers.write().await;
4658 :
4659 28 : let mut duplicated_layers = HashSet::new();
4660 28 :
4661 28 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
4662 :
4663 256 : for l in new_deltas {
4664 228 : if guard.contains(l.as_ref()) {
4665 : // expected in tests
4666 0 : tracing::error!(layer=%l, "duplicated L1 layer");
4667 :
4668 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
4669 : // `guard` on self.layers. as of writing this, there are no error returns except
4670 : // for compact_level0_phase1 creating an L0, which does not happen in practice
4671 : // because we have not implemented L0 => L0 compaction.
4672 0 : duplicated_layers.insert(l.layer_desc().key());
4673 228 : } else if LayerMap::is_l0(l.layer_desc()) {
4674 0 : bail!("compaction generates a L0 layer file as output, which will cause infinite compaction.");
4675 228 : } else {
4676 228 : insert_layers.push(l.clone());
4677 228 : }
4678 : }
4679 :
4680 : // only remove those inputs which were not outputs
4681 28 : let remove_layers: Vec<Layer> = layers_to_remove
4682 28 : .iter()
4683 322 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
4684 28 : .cloned()
4685 28 : .collect();
4686 28 :
4687 28 : if !new_images.is_empty() {
4688 0 : guard.track_new_image_layers(new_images, &self.metrics);
4689 28 : }
4690 :
4691 : // deletion will happen later, the layer file manager calls garbage_collect_on_drop
4692 28 : guard.finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
4693 28 :
4694 28 : self.remote_client
4695 28 : .schedule_compaction_update(&remove_layers, new_deltas)?;
4696 :
4697 28 : drop_wlock(guard);
4698 28 :
4699 28 : Ok(())
4700 28 : }
4701 :
4702 0 : async fn rewrite_layers(
4703 0 : self: &Arc<Self>,
4704 0 : replace_layers: Vec<(Layer, ResidentLayer)>,
4705 0 : drop_layers: Vec<Layer>,
4706 0 : ) -> anyhow::Result<()> {
4707 0 : let mut guard = self.layers.write().await;
4708 :
4709 0 : guard.rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
4710 0 :
4711 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
4712 0 :
4713 0 : self.remote_client
4714 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
4715 :
4716 0 : Ok(())
4717 0 : }
4718 :
4719 : /// Schedules the uploads of the given image layers
4720 366 : fn upload_new_image_layers(
4721 366 : self: &Arc<Self>,
4722 366 : new_images: impl IntoIterator<Item = ResidentLayer>,
4723 366 : ) -> anyhow::Result<()> {
4724 396 : for layer in new_images {
4725 30 : self.remote_client.schedule_layer_file_upload(layer)?;
4726 : }
4727 : // should any new image layer been created, not uploading index_part will
4728 : // result in a mismatch between remote_physical_size and layermap calculated
4729 : // size, which will fail some tests, but should not be an issue otherwise.
4730 366 : self.remote_client
4731 366 : .schedule_index_upload_for_file_changes()?;
4732 366 : Ok(())
4733 366 : }
4734 :
4735 : /// Find the Lsns above which layer files need to be retained on
4736 : /// garbage collection. This is separate from actually performing the GC,
4737 : /// and is updated more frequently, so that compaction can remove obsolete
4738 : /// page versions more aggressively.
4739 : ///
4740 : /// TODO: that's wishful thinking, compaction doesn't actually do that
4741 : /// currently.
4742 : ///
4743 : /// The 'cutoff_horizon' point is used to retain recent versions that might still be
4744 : /// needed by read-only nodes. (As of this writing, the caller just passes
4745 : /// the latest LSN subtracted by a constant, and doesn't do anything smart
4746 : /// to figure out what read-only nodes might actually need.)
4747 : ///
4748 : /// The 'pitr' duration is used to calculate a 'pitr_cutoff', which can be used to determine
4749 : /// whether a record is needed for PITR.
4750 1512 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
4751 : pub(super) async fn find_gc_cutoffs(
4752 : &self,
4753 : cutoff_horizon: Lsn,
4754 : pitr: Duration,
4755 : cancel: &CancellationToken,
4756 : ctx: &RequestContext,
4757 : ) -> anyhow::Result<GcCutoffs> {
4758 : let _timer = self
4759 : .metrics
4760 : .find_gc_cutoffs_histo
4761 : .start_timer()
4762 : .record_on_drop();
4763 :
4764 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
4765 :
4766 : // First, calculate pitr_cutoff_timestamp and then convert it to LSN.
4767 : //
4768 : // Some unit tests depend on garbage-collection working even when
4769 : // CLOG data is missing, so that find_lsn_for_timestamp() doesn't
4770 : // work, so avoid calling it altogether if time-based retention is not
4771 : // configured. It would be pointless anyway.
4772 : let pitr_cutoff = if pitr != Duration::ZERO {
4773 : let now = SystemTime::now();
4774 : if let Some(pitr_cutoff_timestamp) = now.checked_sub(pitr) {
4775 : let pitr_timestamp = to_pg_timestamp(pitr_cutoff_timestamp);
4776 :
4777 : match self
4778 : .find_lsn_for_timestamp(pitr_timestamp, cancel, ctx)
4779 : .await?
4780 : {
4781 : LsnForTimestamp::Present(lsn) => lsn,
4782 : LsnForTimestamp::Future(lsn) => {
4783 : // The timestamp is in the future. That sounds impossible,
4784 : // but what it really means is that there hasn't been
4785 : // any commits since the cutoff timestamp.
4786 : //
4787 : // In this case we should use the LSN of the most recent commit,
4788 : // which is implicitly the last LSN in the log.
4789 : debug!("future({})", lsn);
4790 : self.get_last_record_lsn()
4791 : }
4792 : LsnForTimestamp::Past(lsn) => {
4793 : debug!("past({})", lsn);
4794 : // conservative, safe default is to remove nothing, when we
4795 : // have no commit timestamp data available
4796 : *self.get_latest_gc_cutoff_lsn()
4797 : }
4798 : LsnForTimestamp::NoData(lsn) => {
4799 : debug!("nodata({})", lsn);
4800 : // conservative, safe default is to remove nothing, when we
4801 : // have no commit timestamp data available
4802 : *self.get_latest_gc_cutoff_lsn()
4803 : }
4804 : }
4805 : } else {
4806 : // If we don't have enough data to convert to LSN,
4807 : // play safe and don't remove any layers.
4808 : *self.get_latest_gc_cutoff_lsn()
4809 : }
4810 : } else {
4811 : // No time-based retention was configured. Interpret this as "keep no history".
4812 : self.get_last_record_lsn()
4813 : };
4814 :
4815 : Ok(GcCutoffs {
4816 : horizon: cutoff_horizon,
4817 : pitr: pitr_cutoff,
4818 : })
4819 : }
4820 :
4821 : /// Garbage collect layer files on a timeline that are no longer needed.
4822 : ///
4823 : /// Currently, we don't make any attempt at removing unneeded page versions
4824 : /// within a layer file. We can only remove the whole file if it's fully
4825 : /// obsolete.
4826 756 : pub(super) async fn gc(&self) -> anyhow::Result<GcResult> {
4827 : // this is most likely the background tasks, but it might be the spawned task from
4828 : // immediate_gc
4829 755 : let _g = tokio::select! {
4830 : guard = self.gc_lock.lock() => guard,
4831 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
4832 : };
4833 755 : let timer = self.metrics.garbage_collect_histo.start_timer();
4834 :
4835 : fail_point!("before-timeline-gc");
4836 :
4837 : // Is the timeline being deleted?
4838 755 : if self.is_stopping() {
4839 0 : anyhow::bail!("timeline is Stopping");
4840 755 : }
4841 755 :
4842 755 : let (horizon_cutoff, pitr_cutoff, retain_lsns) = {
4843 755 : let gc_info = self.gc_info.read().unwrap();
4844 755 :
4845 755 : let horizon_cutoff = min(gc_info.cutoffs.horizon, self.get_disk_consistent_lsn());
4846 755 : let pitr_cutoff = gc_info.cutoffs.pitr;
4847 755 : let retain_lsns = gc_info.retain_lsns.clone();
4848 755 : (horizon_cutoff, pitr_cutoff, retain_lsns)
4849 755 : };
4850 755 :
4851 755 : let mut new_gc_cutoff = Lsn::min(horizon_cutoff, pitr_cutoff);
4852 755 : let standby_horizon = self.standby_horizon.load();
4853 755 : // Hold GC for the standby, but as a safety guard do it only within some
4854 755 : // reasonable lag.
4855 755 : if standby_horizon != Lsn::INVALID {
4856 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
4857 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
4858 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
4859 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
4860 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
4861 : } else {
4862 0 : warn!(
4863 0 : "standby is lagging for more than {}MB, not holding gc for it",
4864 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
4865 : )
4866 : }
4867 0 : }
4868 755 : }
4869 :
4870 : // Reset standby horizon to ignore it if it is not updated till next GC.
4871 : // It is an easy way to unset it when standby disappears without adding
4872 : // more conf options.
4873 755 : self.standby_horizon.store(Lsn::INVALID);
4874 755 : self.metrics
4875 755 : .standby_horizon_gauge
4876 755 : .set(Lsn::INVALID.0 as i64);
4877 :
4878 755 : let res = self
4879 755 : .gc_timeline(horizon_cutoff, pitr_cutoff, retain_lsns, new_gc_cutoff)
4880 755 : .instrument(
4881 755 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
4882 : )
4883 0 : .await?;
4884 :
4885 : // only record successes
4886 755 : timer.stop_and_record();
4887 755 :
4888 755 : Ok(res)
4889 756 : }
4890 :
4891 755 : async fn gc_timeline(
4892 755 : &self,
4893 755 : horizon_cutoff: Lsn,
4894 755 : pitr_cutoff: Lsn,
4895 755 : retain_lsns: Vec<Lsn>,
4896 755 : new_gc_cutoff: Lsn,
4897 755 : ) -> anyhow::Result<GcResult> {
4898 755 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
4899 755 :
4900 755 : let now = SystemTime::now();
4901 755 : let mut result: GcResult = GcResult::default();
4902 755 :
4903 755 : // Nothing to GC. Return early.
4904 755 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
4905 755 : if latest_gc_cutoff >= new_gc_cutoff {
4906 22 : info!(
4907 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
4908 : );
4909 22 : return Ok(result);
4910 733 : }
4911 :
4912 : // We need to ensure that no one tries to read page versions or create
4913 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
4914 : // for details. This will block until the old value is no longer in use.
4915 : //
4916 : // The GC cutoff should only ever move forwards.
4917 733 : let waitlist = {
4918 733 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
4919 733 : ensure!(
4920 733 : *write_guard <= new_gc_cutoff,
4921 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
4922 0 : *write_guard,
4923 : new_gc_cutoff
4924 : );
4925 733 : write_guard.store_and_unlock(new_gc_cutoff)
4926 733 : };
4927 733 : waitlist.wait().await;
4928 :
4929 733 : info!("GC starting");
4930 :
4931 733 : debug!("retain_lsns: {:?}", retain_lsns);
4932 :
4933 733 : let mut layers_to_remove = Vec::new();
4934 :
4935 : // Scan all layers in the timeline (remote or on-disk).
4936 : //
4937 : // Garbage collect the layer if all conditions are satisfied:
4938 : // 1. it is older than cutoff LSN;
4939 : // 2. it is older than PITR interval;
4940 : // 3. it doesn't need to be retained for 'retain_lsns';
4941 : // 4. newer on-disk image layers cover the layer's whole key range
4942 : //
4943 : // TODO holding a write lock is too agressive and avoidable
4944 733 : let mut guard = self.layers.write().await;
4945 733 : let layers = guard.layer_map();
4946 12411 : 'outer: for l in layers.iter_historic_layers() {
4947 12411 : result.layers_total += 1;
4948 12411 :
4949 12411 : // 1. Is it newer than GC horizon cutoff point?
4950 12411 : if l.get_lsn_range().end > horizon_cutoff {
4951 753 : debug!(
4952 0 : "keeping {} because it's newer than horizon_cutoff {}",
4953 0 : l.layer_name(),
4954 : horizon_cutoff,
4955 : );
4956 753 : result.layers_needed_by_cutoff += 1;
4957 753 : continue 'outer;
4958 11658 : }
4959 11658 :
4960 11658 : // 2. It is newer than PiTR cutoff point?
4961 11658 : if l.get_lsn_range().end > pitr_cutoff {
4962 0 : debug!(
4963 0 : "keeping {} because it's newer than pitr_cutoff {}",
4964 0 : l.layer_name(),
4965 : pitr_cutoff,
4966 : );
4967 0 : result.layers_needed_by_pitr += 1;
4968 0 : continue 'outer;
4969 11658 : }
4970 :
4971 : // 3. Is it needed by a child branch?
4972 : // NOTE With that we would keep data that
4973 : // might be referenced by child branches forever.
4974 : // We can track this in child timeline GC and delete parent layers when
4975 : // they are no longer needed. This might be complicated with long inheritance chains.
4976 : //
4977 : // TODO Vec is not a great choice for `retain_lsns`
4978 11658 : for retain_lsn in &retain_lsns {
4979 : // start_lsn is inclusive
4980 10 : if &l.get_lsn_range().start <= retain_lsn {
4981 10 : debug!(
4982 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
4983 0 : l.layer_name(),
4984 0 : retain_lsn,
4985 0 : l.is_incremental(),
4986 : );
4987 10 : result.layers_needed_by_branches += 1;
4988 10 : continue 'outer;
4989 0 : }
4990 : }
4991 :
4992 : // 4. Is there a later on-disk layer for this relation?
4993 : //
4994 : // The end-LSN is exclusive, while disk_consistent_lsn is
4995 : // inclusive. For example, if disk_consistent_lsn is 100, it is
4996 : // OK for a delta layer to have end LSN 101, but if the end LSN
4997 : // is 102, then it might not have been fully flushed to disk
4998 : // before crash.
4999 : //
5000 : // For example, imagine that the following layers exist:
5001 : //
5002 : // 1000 - image (A)
5003 : // 1000-2000 - delta (B)
5004 : // 2000 - image (C)
5005 : // 2000-3000 - delta (D)
5006 : // 3000 - image (E)
5007 : //
5008 : // If GC horizon is at 2500, we can remove layers A and B, but
5009 : // we cannot remove C, even though it's older than 2500, because
5010 : // the delta layer 2000-3000 depends on it.
5011 11648 : if !layers
5012 11648 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5013 : {
5014 11642 : debug!("keeping {} because it is the latest layer", l.layer_name());
5015 11642 : result.layers_not_updated += 1;
5016 11642 : continue 'outer;
5017 6 : }
5018 6 :
5019 6 : // We didn't find any reason to keep this file, so remove it.
5020 6 : debug!(
5021 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5022 0 : l.layer_name(),
5023 0 : l.is_incremental(),
5024 : );
5025 6 : layers_to_remove.push(l);
5026 : }
5027 :
5028 733 : if !layers_to_remove.is_empty() {
5029 : // Persist the new GC cutoff value before we actually remove anything.
5030 : // This unconditionally schedules also an index_part.json update, even though, we will
5031 : // be doing one a bit later with the unlinked gc'd layers.
5032 4 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5033 4 : self.schedule_uploads(disk_consistent_lsn, None)?;
5034 :
5035 4 : let gc_layers = layers_to_remove
5036 4 : .iter()
5037 6 : .map(|x| guard.get_from_desc(x))
5038 4 : .collect::<Vec<Layer>>();
5039 4 :
5040 4 : result.layers_removed = gc_layers.len() as u64;
5041 4 :
5042 4 : self.remote_client.schedule_gc_update(&gc_layers)?;
5043 :
5044 4 : guard.finish_gc_timeline(&gc_layers);
5045 4 :
5046 4 : #[cfg(feature = "testing")]
5047 4 : {
5048 4 : result.doomed_layers = gc_layers;
5049 4 : }
5050 729 : }
5051 :
5052 733 : info!(
5053 0 : "GC completed removing {} layers, cutoff {}",
5054 : result.layers_removed, new_gc_cutoff
5055 : );
5056 :
5057 733 : result.elapsed = now.elapsed()?;
5058 733 : Ok(result)
5059 755 : }
5060 :
5061 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5062 703663 : async fn reconstruct_value(
5063 703663 : &self,
5064 703663 : key: Key,
5065 703663 : request_lsn: Lsn,
5066 703663 : mut data: ValueReconstructState,
5067 703663 : ) -> Result<Bytes, PageReconstructError> {
5068 703663 : // Perform WAL redo if needed
5069 703663 : data.records.reverse();
5070 703663 :
5071 703663 : // If we have a page image, and no WAL, we're all set
5072 703663 : if data.records.is_empty() {
5073 703651 : if let Some((img_lsn, img)) = &data.img {
5074 703651 : trace!(
5075 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5076 : key,
5077 : img_lsn,
5078 : request_lsn,
5079 : );
5080 703651 : Ok(img.clone())
5081 : } else {
5082 0 : Err(PageReconstructError::from(anyhow!(
5083 0 : "base image for {key} at {request_lsn} not found"
5084 0 : )))
5085 : }
5086 : } else {
5087 : // We need to do WAL redo.
5088 : //
5089 : // If we don't have a base image, then the oldest WAL record better initialize
5090 : // the page
5091 12 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5092 0 : Err(PageReconstructError::from(anyhow!(
5093 0 : "Base image for {} at {} not found, but got {} WAL records",
5094 0 : key,
5095 0 : request_lsn,
5096 0 : data.records.len()
5097 0 : )))
5098 : } else {
5099 12 : if data.img.is_some() {
5100 12 : trace!(
5101 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5102 0 : data.records.len(),
5103 : key,
5104 : request_lsn
5105 : );
5106 : } else {
5107 0 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5108 : };
5109 :
5110 12 : let last_rec_lsn = data.records.last().unwrap().0;
5111 :
5112 12 : let img = match self
5113 12 : .walredo_mgr
5114 12 : .as_ref()
5115 12 : .context("timeline has no walredo manager")
5116 12 : .map_err(PageReconstructError::WalRedo)?
5117 12 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5118 0 : .await
5119 12 : .context("reconstruct a page image")
5120 : {
5121 12 : Ok(img) => img,
5122 0 : Err(e) => return Err(PageReconstructError::WalRedo(e)),
5123 : };
5124 :
5125 12 : if img.len() == page_cache::PAGE_SZ {
5126 0 : let cache = page_cache::get();
5127 0 : if let Err(e) = cache
5128 0 : .memorize_materialized_page(
5129 0 : self.tenant_shard_id,
5130 0 : self.timeline_id,
5131 0 : key,
5132 0 : last_rec_lsn,
5133 0 : &img,
5134 0 : )
5135 0 : .await
5136 0 : .context("Materialized page memoization failed")
5137 : {
5138 0 : return Err(PageReconstructError::from(e));
5139 0 : }
5140 12 : }
5141 :
5142 12 : Ok(img)
5143 : }
5144 : }
5145 703663 : }
5146 :
5147 0 : pub(crate) async fn spawn_download_all_remote_layers(
5148 0 : self: Arc<Self>,
5149 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5150 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5151 0 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5152 0 :
5153 0 : // this is not really needed anymore; it has tests which really check the return value from
5154 0 : // http api. it would be better not to maintain this anymore.
5155 0 :
5156 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5157 0 : if let Some(st) = &*status_guard {
5158 0 : match &st.state {
5159 : DownloadRemoteLayersTaskState::Running => {
5160 0 : return Err(st.clone());
5161 : }
5162 : DownloadRemoteLayersTaskState::ShutDown
5163 0 : | DownloadRemoteLayersTaskState::Completed => {
5164 0 : *status_guard = None;
5165 0 : }
5166 : }
5167 0 : }
5168 :
5169 0 : let self_clone = Arc::clone(&self);
5170 0 : let task_id = task_mgr::spawn(
5171 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5172 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5173 0 : Some(self.tenant_shard_id),
5174 0 : Some(self.timeline_id),
5175 0 : "download all remote layers task",
5176 : false,
5177 0 : async move {
5178 0 : self_clone.download_all_remote_layers(request).await;
5179 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5180 0 : match &mut *status_guard {
5181 : None => {
5182 0 : warn!("tasks status is supposed to be Some(), since we are running");
5183 : }
5184 0 : Some(st) => {
5185 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5186 0 : if st.task_id != exp_task_id {
5187 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5188 0 : } else {
5189 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5190 0 : }
5191 : }
5192 : };
5193 0 : Ok(())
5194 0 : }
5195 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))
5196 : );
5197 :
5198 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5199 0 : task_id: format!("{task_id}"),
5200 0 : state: DownloadRemoteLayersTaskState::Running,
5201 0 : total_layer_count: 0,
5202 0 : successful_download_count: 0,
5203 0 : failed_download_count: 0,
5204 0 : };
5205 0 : *status_guard = Some(initial_info.clone());
5206 0 :
5207 0 : Ok(initial_info)
5208 0 : }
5209 :
5210 0 : async fn download_all_remote_layers(
5211 0 : self: &Arc<Self>,
5212 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5213 0 : ) {
5214 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5215 :
5216 0 : let remaining = {
5217 0 : let guard = self.layers.read().await;
5218 0 : guard
5219 0 : .layer_map()
5220 0 : .iter_historic_layers()
5221 0 : .map(|desc| guard.get_from_desc(&desc))
5222 0 : .collect::<Vec<_>>()
5223 0 : };
5224 0 : let total_layer_count = remaining.len();
5225 0 :
5226 0 : macro_rules! lock_status {
5227 0 : ($st:ident) => {
5228 0 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5229 0 : let st = st
5230 0 : .as_mut()
5231 0 : .expect("this function is only called after the task has been spawned");
5232 0 : assert_eq!(
5233 0 : st.task_id,
5234 0 : format!(
5235 0 : "{}",
5236 0 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5237 0 : )
5238 0 : );
5239 0 : let $st = st;
5240 0 : };
5241 0 : }
5242 0 :
5243 0 : {
5244 0 : lock_status!(st);
5245 0 : st.total_layer_count = total_layer_count as u64;
5246 0 : }
5247 0 :
5248 0 : let mut remaining = remaining.into_iter();
5249 0 : let mut have_remaining = true;
5250 0 : let mut js = tokio::task::JoinSet::new();
5251 0 :
5252 0 : let cancel = task_mgr::shutdown_token();
5253 0 :
5254 0 : let limit = request.max_concurrent_downloads;
5255 :
5256 : loop {
5257 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5258 0 : let Some(next) = remaining.next() else {
5259 0 : have_remaining = false;
5260 0 : break;
5261 : };
5262 :
5263 0 : let span = tracing::info_span!("download", layer = %next);
5264 :
5265 0 : js.spawn(
5266 0 : async move {
5267 0 : let res = next.download().await;
5268 0 : (next, res)
5269 0 : }
5270 0 : .instrument(span),
5271 0 : );
5272 : }
5273 :
5274 0 : while let Some(res) = js.join_next().await {
5275 0 : match res {
5276 : Ok((_, Ok(_))) => {
5277 0 : lock_status!(st);
5278 0 : st.successful_download_count += 1;
5279 : }
5280 0 : Ok((layer, Err(e))) => {
5281 0 : tracing::error!(%layer, "download failed: {e:#}");
5282 0 : lock_status!(st);
5283 0 : st.failed_download_count += 1;
5284 : }
5285 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5286 0 : Err(je) if je.is_panic() => {
5287 0 : lock_status!(st);
5288 0 : st.failed_download_count += 1;
5289 : }
5290 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5291 : }
5292 : }
5293 :
5294 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5295 0 : break;
5296 0 : }
5297 : }
5298 :
5299 : {
5300 0 : lock_status!(st);
5301 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5302 0 : }
5303 0 : }
5304 :
5305 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5306 0 : &self,
5307 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5308 0 : self.download_all_remote_layers_task_info
5309 0 : .read()
5310 0 : .unwrap()
5311 0 : .clone()
5312 0 : }
5313 : }
5314 :
5315 : impl Timeline {
5316 : /// Returns non-remote layers for eviction.
5317 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5318 0 : let guard = self.layers.read().await;
5319 0 : let mut max_layer_size: Option<u64> = None;
5320 0 :
5321 0 : let resident_layers = guard
5322 0 : .likely_resident_layers()
5323 0 : .map(|layer| {
5324 0 : let file_size = layer.layer_desc().file_size;
5325 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5326 0 :
5327 0 : let last_activity_ts = layer.access_stats().latest_activity_or_now();
5328 0 :
5329 0 : EvictionCandidate {
5330 0 : layer: layer.into(),
5331 0 : last_activity_ts,
5332 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5333 0 : }
5334 0 : })
5335 0 : .collect();
5336 0 :
5337 0 : DiskUsageEvictionInfo {
5338 0 : max_layer_size,
5339 0 : resident_layers,
5340 0 : }
5341 0 : }
5342 :
5343 1300 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5344 1300 : ShardIndex {
5345 1300 : shard_number: self.tenant_shard_id.shard_number,
5346 1300 : shard_count: self.tenant_shard_id.shard_count,
5347 1300 : }
5348 1300 : }
5349 : }
5350 :
5351 : type TraversalPathItem = (ValueReconstructResult, Lsn, TraversalId);
5352 :
5353 : struct TimelineWriterState {
5354 : open_layer: Arc<InMemoryLayer>,
5355 : current_size: u64,
5356 : // Previous Lsn which passed through
5357 : prev_lsn: Option<Lsn>,
5358 : // Largest Lsn which passed through the current writer
5359 : max_lsn: Option<Lsn>,
5360 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
5361 : cached_last_freeze_at: Lsn,
5362 : }
5363 :
5364 : impl TimelineWriterState {
5365 4804100 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
5366 4804100 : Self {
5367 4804100 : open_layer,
5368 4804100 : current_size,
5369 4804100 : prev_lsn: None,
5370 4804100 : max_lsn: None,
5371 4804100 : cached_last_freeze_at: last_freeze_at,
5372 4804100 : }
5373 4804100 : }
5374 : }
5375 :
5376 : /// Various functions to mutate the timeline.
5377 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
5378 : // This is probably considered a bad practice in Rust and should be fixed eventually,
5379 : // but will cause large code changes.
5380 : pub(crate) struct TimelineWriter<'a> {
5381 : tl: &'a Timeline,
5382 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
5383 : }
5384 :
5385 : impl Deref for TimelineWriter<'_> {
5386 : type Target = Timeline;
5387 :
5388 4806832 : fn deref(&self) -> &Self::Target {
5389 4806832 : self.tl
5390 4806832 : }
5391 : }
5392 :
5393 : impl Drop for TimelineWriter<'_> {
5394 5133116 : fn drop(&mut self) {
5395 5133116 : self.write_guard.take();
5396 5133116 : }
5397 : }
5398 :
5399 : #[derive(PartialEq)]
5400 : enum OpenLayerAction {
5401 : Roll,
5402 : Open,
5403 : None,
5404 : }
5405 :
5406 : impl<'a> TimelineWriter<'a> {
5407 : /// Put a new page version that can be constructed from a WAL record
5408 : ///
5409 : /// This will implicitly extend the relation, if the page is beyond the
5410 : /// current end-of-file.
5411 5090196 : pub(crate) async fn put(
5412 5090196 : &mut self,
5413 5090196 : key: Key,
5414 5090196 : lsn: Lsn,
5415 5090196 : value: &Value,
5416 5090196 : ctx: &RequestContext,
5417 5090196 : ) -> anyhow::Result<()> {
5418 5090196 : // Avoid doing allocations for "small" values.
5419 5090196 : // In the regression test suite, the limit of 256 avoided allocations in 95% of cases:
5420 5090196 : // https://github.com/neondatabase/neon/pull/5056#discussion_r1301975061
5421 5090196 : let mut buf = smallvec::SmallVec::<[u8; 256]>::new();
5422 5090196 : value.ser_into(&mut buf)?;
5423 5090196 : let buf_size: u64 = buf.len().try_into().expect("oversized value buf");
5424 5090196 :
5425 5090196 : let action = self.get_open_layer_action(lsn, buf_size);
5426 5090196 : let layer = self.handle_open_layer_action(lsn, action, ctx).await?;
5427 5090196 : let res = layer.put_value(key, lsn, &buf, ctx).await;
5428 :
5429 5090196 : if res.is_ok() {
5430 5090196 : // Update the current size only when the entire write was ok.
5431 5090196 : // In case of failures, we may have had partial writes which
5432 5090196 : // render the size tracking out of sync. That's ok because
5433 5090196 : // the checkpoint distance should be significantly smaller
5434 5090196 : // than the S3 single shot upload limit of 5GiB.
5435 5090196 : let state = self.write_guard.as_mut().unwrap();
5436 5090196 :
5437 5090196 : state.current_size += buf_size;
5438 5090196 : state.prev_lsn = Some(lsn);
5439 5090196 : state.max_lsn = std::cmp::max(state.max_lsn, Some(lsn));
5440 5090196 : }
5441 :
5442 5090196 : res
5443 5090196 : }
5444 :
5445 5090198 : async fn handle_open_layer_action(
5446 5090198 : &mut self,
5447 5090198 : at: Lsn,
5448 5090198 : action: OpenLayerAction,
5449 5090198 : ctx: &RequestContext,
5450 5090198 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
5451 5090198 : match action {
5452 : OpenLayerAction::Roll => {
5453 0 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
5454 0 : self.roll_layer(freeze_at).await?;
5455 0 : self.open_layer(at, ctx).await?;
5456 : }
5457 4804100 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
5458 : OpenLayerAction::None => {
5459 286098 : assert!(self.write_guard.is_some());
5460 : }
5461 : }
5462 :
5463 5090198 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
5464 5090198 : }
5465 :
5466 4804100 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
5467 4804100 : let layer = self.tl.get_layer_for_write(at, ctx).await?;
5468 4804100 : let initial_size = layer.size().await?;
5469 :
5470 4804100 : let last_freeze_at = self.last_freeze_at.load();
5471 4804100 : self.write_guard.replace(TimelineWriterState::new(
5472 4804100 : layer,
5473 4804100 : initial_size,
5474 4804100 : last_freeze_at,
5475 4804100 : ));
5476 4804100 :
5477 4804100 : Ok(())
5478 4804100 : }
5479 :
5480 0 : async fn roll_layer(&mut self, freeze_at: Lsn) -> anyhow::Result<()> {
5481 0 : assert!(self.write_guard.is_some());
5482 :
5483 0 : self.tl.freeze_inmem_layer_at(freeze_at).await;
5484 :
5485 0 : let now = Instant::now();
5486 0 : *(self.last_freeze_ts.write().unwrap()) = now;
5487 0 :
5488 0 : self.tl.flush_frozen_layers();
5489 0 :
5490 0 : let current_size = self.write_guard.as_ref().unwrap().current_size;
5491 0 : if current_size > self.get_checkpoint_distance() {
5492 0 : warn!("Flushed oversized open layer with size {}", current_size)
5493 0 : }
5494 :
5495 0 : Ok(())
5496 0 : }
5497 :
5498 5090198 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
5499 5090198 : let state = &*self.write_guard;
5500 5090198 : let Some(state) = &state else {
5501 4804100 : return OpenLayerAction::Open;
5502 : };
5503 :
5504 286098 : if state.prev_lsn == Some(lsn) {
5505 : // Rolling mid LSN is not supported by downstream code.
5506 : // Hence, only roll at LSN boundaries.
5507 286040 : return OpenLayerAction::None;
5508 58 : }
5509 58 :
5510 58 : if state.current_size == 0 {
5511 : // Don't roll empty layers
5512 0 : return OpenLayerAction::None;
5513 58 : }
5514 58 :
5515 58 : if self.tl.should_roll(
5516 58 : state.current_size,
5517 58 : state.current_size + new_value_size,
5518 58 : self.get_checkpoint_distance(),
5519 58 : lsn,
5520 58 : state.cached_last_freeze_at,
5521 58 : state.open_layer.get_opened_at(),
5522 58 : ) {
5523 0 : OpenLayerAction::Roll
5524 : } else {
5525 58 : OpenLayerAction::None
5526 : }
5527 5090198 : }
5528 :
5529 : /// Put a batch of keys at the specified Lsns.
5530 : ///
5531 : /// The batch is sorted by Lsn (enforced by usage of [`utils::vec_map::VecMap`].
5532 413994 : pub(crate) async fn put_batch(
5533 413994 : &mut self,
5534 413994 : batch: VecMap<Lsn, (Key, Value)>,
5535 413994 : ctx: &RequestContext,
5536 413994 : ) -> anyhow::Result<()> {
5537 1114028 : for (lsn, (key, val)) in batch {
5538 700034 : self.put(key, lsn, &val, ctx).await?
5539 : }
5540 :
5541 413994 : Ok(())
5542 413994 : }
5543 :
5544 2 : pub(crate) async fn delete_batch(
5545 2 : &mut self,
5546 2 : batch: &[(Range<Key>, Lsn)],
5547 2 : ctx: &RequestContext,
5548 2 : ) -> anyhow::Result<()> {
5549 2 : if let Some((_, lsn)) = batch.first() {
5550 2 : let action = self.get_open_layer_action(*lsn, 0);
5551 2 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
5552 2 : layer.put_tombstones(batch).await?;
5553 0 : }
5554 :
5555 2 : Ok(())
5556 2 : }
5557 :
5558 : /// Track the end of the latest digested WAL record.
5559 : /// Remember the (end of) last valid WAL record remembered in the timeline.
5560 : ///
5561 : /// Call this after you have finished writing all the WAL up to 'lsn'.
5562 : ///
5563 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
5564 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
5565 : /// the latest and can be read.
5566 5279018 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
5567 5279018 : self.tl.finish_write(new_lsn);
5568 5279018 : }
5569 :
5570 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
5571 270570 : self.tl.update_current_logical_size(delta)
5572 270570 : }
5573 : }
5574 :
5575 : // We need TimelineWriter to be send in upcoming conversion of
5576 : // Timeline::layers to tokio::sync::RwLock.
5577 : #[test]
5578 2 : fn is_send() {
5579 2 : fn _assert_send<T: Send>() {}
5580 2 : _assert_send::<TimelineWriter<'_>>();
5581 2 : }
5582 :
5583 : /// Add a suffix to a layer file's name: .{num}.old
5584 : /// Uses the first available num (starts at 0)
5585 0 : fn rename_to_backup(path: &Utf8Path) -> anyhow::Result<()> {
5586 0 : let filename = path
5587 0 : .file_name()
5588 0 : .ok_or_else(|| anyhow!("Path {path} don't have a file name"))?;
5589 0 : let mut new_path = path.to_owned();
5590 :
5591 0 : for i in 0u32.. {
5592 0 : new_path.set_file_name(format!("{filename}.{i}.old"));
5593 0 : if !new_path.exists() {
5594 0 : std::fs::rename(path, &new_path)
5595 0 : .with_context(|| format!("rename {path:?} to {new_path:?}"))?;
5596 0 : return Ok(());
5597 0 : }
5598 : }
5599 :
5600 0 : bail!("couldn't find an unused backup number for {:?}", path)
5601 0 : }
5602 :
5603 : #[cfg(test)]
5604 : mod tests {
5605 : use utils::{id::TimelineId, lsn::Lsn};
5606 :
5607 : use crate::tenant::{
5608 : harness::TenantHarness, storage_layer::Layer, timeline::EvictionError, Timeline,
5609 : };
5610 :
5611 : #[tokio::test]
5612 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
5613 2 : let harness =
5614 2 : TenantHarness::create("two_layer_eviction_attempts_at_the_same_time").unwrap();
5615 2 :
5616 8 : let (tenant, ctx) = harness.load().await;
5617 2 : let timeline = tenant
5618 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
5619 6 : .await
5620 2 : .unwrap();
5621 2 :
5622 2 : let layer = find_some_layer(&timeline).await;
5623 2 : let layer = layer
5624 2 : .keep_resident()
5625 2 : .await
5626 2 : .expect("no download => no downloading errors")
5627 2 : .drop_eviction_guard();
5628 2 :
5629 2 : let forever = std::time::Duration::from_secs(120);
5630 2 :
5631 2 : let first = layer.evict_and_wait(forever);
5632 2 : let second = layer.evict_and_wait(forever);
5633 2 :
5634 2 : let (first, second) = tokio::join!(first, second);
5635 2 :
5636 2 : let res = layer.keep_resident().await;
5637 2 : assert!(res.is_none(), "{res:?}");
5638 2 :
5639 2 : match (first, second) {
5640 2 : (Ok(()), Ok(())) => {
5641 2 : // because there are no more timeline locks being taken on eviction path, we can
5642 2 : // witness all three outcomes here.
5643 2 : }
5644 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
5645 0 : // if one completes before the other, this is fine just as well.
5646 0 : }
5647 2 : other => unreachable!("unexpected {:?}", other),
5648 2 : }
5649 2 : }
5650 :
5651 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
5652 2 : let layers = timeline.layers.read().await;
5653 2 : let desc = layers
5654 2 : .layer_map()
5655 2 : .iter_historic_layers()
5656 2 : .next()
5657 2 : .expect("must find one layer to evict");
5658 2 :
5659 2 : layers.get_from_desc(&desc)
5660 2 : }
5661 : }
|
