Line data Source code
1 : pub(crate) mod analysis;
2 : pub(crate) mod compaction;
3 : pub mod delete;
4 : pub(crate) mod detach_ancestor;
5 : mod eviction_task;
6 : pub(crate) mod handle;
7 : pub(crate) mod import_pgdata;
8 : mod init;
9 : pub mod layer_manager;
10 : pub(crate) mod logical_size;
11 : pub mod offload;
12 : pub mod span;
13 : pub mod uninit;
14 : mod walreceiver;
15 :
16 : use anyhow::{anyhow, bail, ensure, Context, Result};
17 : use arc_swap::{ArcSwap, ArcSwapOption};
18 : use bytes::Bytes;
19 : use camino::Utf8Path;
20 : use chrono::{DateTime, Utc};
21 : use compaction::CompactionOutcome;
22 : use enumset::EnumSet;
23 : use fail::fail_point;
24 : use futures::{stream::FuturesUnordered, StreamExt};
25 : use handle::ShardTimelineId;
26 : use layer_manager::Shutdown;
27 : use offload::OffloadError;
28 : use once_cell::sync::Lazy;
29 : use pageserver_api::models::PageTraceEvent;
30 : use pageserver_api::{
31 : key::{
32 : KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
33 : SPARSE_RANGE,
34 : },
35 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
36 : models::{
37 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
38 : DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy,
39 : InMemoryLayerInfo, LayerMapInfo, LsnLease, TimelineState,
40 : },
41 : reltag::BlockNumber,
42 : shard::{ShardIdentity, ShardNumber, TenantShardId},
43 : };
44 : use rand::Rng;
45 : use remote_storage::DownloadError;
46 : use serde_with::serde_as;
47 : use storage_broker::BrokerClientChannel;
48 : use tokio::runtime::Handle;
49 : use tokio::sync::mpsc::Sender;
50 : use tokio::sync::{oneshot, watch, Notify};
51 : use tokio_util::sync::CancellationToken;
52 : use tracing::*;
53 : use utils::critical;
54 : use utils::rate_limit::RateLimit;
55 : use utils::{
56 : fs_ext,
57 : guard_arc_swap::GuardArcSwap,
58 : pausable_failpoint,
59 : postgres_client::PostgresClientProtocol,
60 : sync::gate::{Gate, GateGuard},
61 : };
62 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
63 :
64 : use std::array;
65 : use std::cmp::{max, min};
66 : use std::collections::btree_map::Entry;
67 : use std::collections::{BTreeMap, HashMap, HashSet};
68 : use std::ops::{ControlFlow, Deref, Range};
69 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering as AtomicOrdering};
70 : use std::sync::{Arc, Mutex, OnceLock, RwLock, Weak};
71 : use std::time::{Duration, Instant, SystemTime};
72 :
73 : use crate::l0_flush::{self, L0FlushGlobalState};
74 : use crate::tenant::storage_layer::ImageLayerName;
75 : use crate::{
76 : aux_file::AuxFileSizeEstimator,
77 : page_service::TenantManagerTypes,
78 : tenant::{
79 : config::AttachmentMode,
80 : layer_map::{LayerMap, SearchResult},
81 : metadata::TimelineMetadata,
82 : storage_layer::{
83 : inmemory_layer::IndexEntry, BatchLayerWriter, IoConcurrency, PersistentLayerDesc,
84 : ValueReconstructSituation,
85 : },
86 : },
87 : walingest::WalLagCooldown,
88 : walredo,
89 : };
90 : use crate::{
91 : context::{DownloadBehavior, RequestContext},
92 : disk_usage_eviction_task::DiskUsageEvictionInfo,
93 : pgdatadir_mapping::CollectKeySpaceError,
94 : };
95 : use crate::{
96 : disk_usage_eviction_task::finite_f32,
97 : tenant::storage_layer::{
98 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
99 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
100 : ValuesReconstructState,
101 : },
102 : };
103 : use crate::{
104 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
105 : };
106 : use crate::{
107 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
108 : };
109 : use crate::{
110 : pgdatadir_mapping::DirectoryKind,
111 : virtual_file::{MaybeFatalIo, VirtualFile},
112 : };
113 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
114 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
115 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
116 :
117 : use crate::config::PageServerConf;
118 : use crate::keyspace::{KeyPartitioning, KeySpace};
119 : use crate::metrics::{TimelineMetrics, DELTAS_PER_READ_GLOBAL, LAYERS_PER_READ_GLOBAL};
120 : use crate::pgdatadir_mapping::{CalculateLogicalSizeError, MetricsUpdate};
121 : use crate::tenant::config::TenantConfOpt;
122 : use pageserver_api::reltag::RelTag;
123 : use pageserver_api::shard::ShardIndex;
124 :
125 : use postgres_connection::PgConnectionConfig;
126 : use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
127 : use utils::{
128 : completion,
129 : generation::Generation,
130 : id::TimelineId,
131 : lsn::{AtomicLsn, Lsn, RecordLsn},
132 : seqwait::SeqWait,
133 : simple_rcu::{Rcu, RcuReadGuard},
134 : };
135 :
136 : use crate::task_mgr;
137 : use crate::task_mgr::TaskKind;
138 : use crate::tenant::gc_result::GcResult;
139 : use crate::ZERO_PAGE;
140 : use pageserver_api::key::Key;
141 :
142 : use self::delete::DeleteTimelineFlow;
143 : pub(super) use self::eviction_task::EvictionTaskTenantState;
144 : use self::eviction_task::EvictionTaskTimelineState;
145 : use self::layer_manager::LayerManager;
146 : use self::logical_size::LogicalSize;
147 : use self::walreceiver::{WalReceiver, WalReceiverConf};
148 :
149 : use super::{
150 : config::TenantConf, storage_layer::LayerVisibilityHint, upload_queue::NotInitialized,
151 : MaybeOffloaded,
152 : };
153 : use super::{
154 : debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf, HeatMapTimeline,
155 : };
156 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
157 : use super::{
158 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
159 : storage_layer::ReadableLayer,
160 : };
161 : use super::{secondary::heatmap::HeatMapLayer, GcError};
162 :
163 : #[cfg(test)]
164 : use pageserver_api::value::Value;
165 :
166 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
167 : pub(crate) enum FlushLoopState {
168 : NotStarted,
169 : Running {
170 : #[cfg(test)]
171 : expect_initdb_optimization: bool,
172 : #[cfg(test)]
173 : initdb_optimization_count: usize,
174 : },
175 : Exited,
176 : }
177 :
178 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
179 : pub enum ImageLayerCreationMode {
180 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
181 : Try,
182 : /// Force creating the image layers if possible. For now, no image layers will be created
183 : /// for metadata keys. Used in compaction code path with force flag enabled.
184 : Force,
185 : /// Initial ingestion of the data, and no data should be dropped in this function. This
186 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
187 : /// code path.
188 : Initial,
189 : }
190 :
191 : #[derive(Clone, Debug, Default)]
192 : pub enum LastImageLayerCreationStatus {
193 : Incomplete {
194 : /// The last key of the partition (exclusive) that was processed in the last
195 : /// image layer creation attempt. We will continue from this key in the next
196 : /// attempt.
197 : last_key: Key,
198 : },
199 : Complete,
200 : #[default]
201 : Initial,
202 : }
203 :
204 : impl std::fmt::Display for ImageLayerCreationMode {
205 1180 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
206 1180 : write!(f, "{:?}", self)
207 1180 : }
208 : }
209 :
210 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
211 : /// Can be removed after all refactors are done.
212 92 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
213 92 : drop(rlock)
214 92 : }
215 :
216 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
217 : /// Can be removed after all refactors are done.
218 1272 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
219 1272 : drop(rlock)
220 1272 : }
221 :
222 : /// The outward-facing resources required to build a Timeline
223 : pub struct TimelineResources {
224 : pub remote_client: RemoteTimelineClient,
225 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
226 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
227 : pub l0_compaction_trigger: Arc<Notify>,
228 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
229 : }
230 :
231 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
232 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
233 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
234 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
235 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
236 : pub(crate) struct RelSizeCache {
237 : pub(crate) complete_as_of: Lsn,
238 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
239 : }
240 :
241 : pub struct Timeline {
242 : pub(crate) conf: &'static PageServerConf,
243 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
244 :
245 : myself: Weak<Self>,
246 :
247 : pub(crate) tenant_shard_id: TenantShardId,
248 : pub timeline_id: TimelineId,
249 :
250 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
251 : /// Never changes for the lifetime of this [`Timeline`] object.
252 : ///
253 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
254 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
255 : pub(crate) generation: Generation,
256 :
257 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
258 : /// to shards, and is constant through the lifetime of this Timeline.
259 : shard_identity: ShardIdentity,
260 :
261 : pub pg_version: u32,
262 :
263 : /// The tuple has two elements.
264 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
265 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
266 : ///
267 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
268 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
269 : ///
270 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
271 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
272 : /// `PersistentLayerDesc`'s.
273 : ///
274 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
275 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
276 : /// runtime, e.g., during page reconstruction.
277 : ///
278 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
279 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
280 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
281 :
282 : last_freeze_at: AtomicLsn,
283 : // Atomic would be more appropriate here.
284 : last_freeze_ts: RwLock<Instant>,
285 :
286 : pub(crate) standby_horizon: AtomicLsn,
287 :
288 : // WAL redo manager. `None` only for broken tenants.
289 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
290 :
291 : /// Remote storage client.
292 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
293 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
294 :
295 : // What page versions do we hold in the repository? If we get a
296 : // request > last_record_lsn, we need to wait until we receive all
297 : // the WAL up to the request. The SeqWait provides functions for
298 : // that. TODO: If we get a request for an old LSN, such that the
299 : // versions have already been garbage collected away, we should
300 : // throw an error, but we don't track that currently.
301 : //
302 : // last_record_lsn.load().last points to the end of last processed WAL record.
303 : //
304 : // We also remember the starting point of the previous record in
305 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
306 : // first WAL record when the node is started up. But here, we just
307 : // keep track of it.
308 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
309 :
310 : // All WAL records have been processed and stored durably on files on
311 : // local disk, up to this LSN. On crash and restart, we need to re-process
312 : // the WAL starting from this point.
313 : //
314 : // Some later WAL records might have been processed and also flushed to disk
315 : // already, so don't be surprised to see some, but there's no guarantee on
316 : // them yet.
317 : disk_consistent_lsn: AtomicLsn,
318 :
319 : // Parent timeline that this timeline was branched from, and the LSN
320 : // of the branch point.
321 : ancestor_timeline: Option<Arc<Timeline>>,
322 : ancestor_lsn: Lsn,
323 :
324 : pub(super) metrics: TimelineMetrics,
325 :
326 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
327 : // in `crate::page_service` writes these metrics.
328 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
329 :
330 : directory_metrics_inited: [AtomicBool; DirectoryKind::KINDS_NUM],
331 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
332 :
333 : /// Ensures layers aren't frozen by checkpointer between
334 : /// [`Timeline::get_layer_for_write`] and layer reads.
335 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
336 : /// Must always be acquired before the layer map/individual layer lock
337 : /// to avoid deadlock.
338 : ///
339 : /// The state is cleared upon freezing.
340 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
341 :
342 : /// Used to avoid multiple `flush_loop` tasks running
343 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
344 :
345 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
346 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
347 : /// The flush cycle counter is sent back on the layer_flush_done channel when
348 : /// the flush finishes. You can use that to wait for the flush to finish.
349 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
350 : /// read by whoever sends an update
351 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
352 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
353 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
354 :
355 : // The LSN at which we have executed GC: whereas [`Self::gc_info`] records the LSN at which
356 : // we _intend_ to GC (i.e. the PITR cutoff), this LSN records where we actually last did it.
357 : // Because PITR interval is mutable, it's possible for this LSN to be earlier or later than
358 : // the planned GC cutoff.
359 : pub applied_gc_cutoff_lsn: Rcu<Lsn>,
360 :
361 : pub(crate) gc_compaction_layer_update_lock: tokio::sync::RwLock<()>,
362 :
363 : // List of child timelines and their branch points. This is needed to avoid
364 : // garbage collecting data that is still needed by the child timelines.
365 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
366 :
367 : pub(crate) last_image_layer_creation_status: ArcSwap<LastImageLayerCreationStatus>,
368 :
369 : // It may change across major versions so for simplicity
370 : // keep it after running initdb for a timeline.
371 : // It is needed in checks when we want to error on some operations
372 : // when they are requested for pre-initdb lsn.
373 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
374 : // though let's keep them both for better error visibility.
375 : pub initdb_lsn: Lsn,
376 :
377 : /// The repartitioning result. Allows a single writer and multiple readers.
378 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
379 :
380 : /// Configuration: how often should the partitioning be recalculated.
381 : repartition_threshold: u64,
382 :
383 : last_image_layer_creation_check_at: AtomicLsn,
384 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
385 :
386 : /// Current logical size of the "datadir", at the last LSN.
387 : current_logical_size: LogicalSize,
388 :
389 : /// Information about the last processed message by the WAL receiver,
390 : /// or None if WAL receiver has not received anything for this timeline
391 : /// yet.
392 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
393 : pub walreceiver: Mutex<Option<WalReceiver>>,
394 :
395 : /// Relation size cache
396 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
397 :
398 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
399 :
400 : state: watch::Sender<TimelineState>,
401 :
402 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
403 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
404 : pub delete_progress: TimelineDeleteProgress,
405 :
406 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
407 :
408 : /// Load or creation time information about the disk_consistent_lsn and when the loading
409 : /// happened. Used for consumption metrics.
410 : pub(crate) loaded_at: (Lsn, SystemTime),
411 :
412 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
413 : pub(crate) gate: Gate,
414 :
415 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
416 : /// to the timeline should drop out when this token fires.
417 : pub(crate) cancel: CancellationToken,
418 :
419 : /// Make sure we only have one running compaction at a time in tests.
420 : ///
421 : /// Must only be taken in two places:
422 : /// - [`Timeline::compact`] (this file)
423 : /// - [`delete::delete_local_timeline_directory`]
424 : ///
425 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
426 : compaction_lock: tokio::sync::Mutex<()>,
427 :
428 : /// If true, the last compaction failed.
429 : compaction_failed: AtomicBool,
430 :
431 : /// Notifies the tenant compaction loop that there is pending L0 compaction work.
432 : l0_compaction_trigger: Arc<Notify>,
433 :
434 : /// Make sure we only have one running gc at a time.
435 : ///
436 : /// Must only be taken in two places:
437 : /// - [`Timeline::gc`] (this file)
438 : /// - [`delete::delete_local_timeline_directory`]
439 : ///
440 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
441 : gc_lock: tokio::sync::Mutex<()>,
442 :
443 : /// Cloned from [`super::Tenant::pagestream_throttle`] on construction.
444 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
445 :
446 : /// Size estimator for aux file v2
447 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
448 :
449 : /// Some test cases directly place keys into the timeline without actually modifying the directory
450 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
451 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
452 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
453 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
454 : #[cfg(test)]
455 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
456 :
457 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
458 :
459 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
460 :
461 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
462 :
463 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
464 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
465 :
466 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
467 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
468 :
469 : previous_heatmap: ArcSwapOption<PreviousHeatmap>,
470 : }
471 :
472 : pub(crate) enum PreviousHeatmap {
473 : Active {
474 : heatmap: HeatMapTimeline,
475 : read_at: std::time::Instant,
476 : },
477 : Obsolete,
478 : }
479 :
480 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
481 :
482 : pub struct WalReceiverInfo {
483 : pub wal_source_connconf: PgConnectionConfig,
484 : pub last_received_msg_lsn: Lsn,
485 : pub last_received_msg_ts: u128,
486 : }
487 :
488 : /// Information about how much history needs to be retained, needed by
489 : /// Garbage Collection.
490 : #[derive(Default)]
491 : pub(crate) struct GcInfo {
492 : /// Specific LSNs that are needed.
493 : ///
494 : /// Currently, this includes all points where child branches have
495 : /// been forked off from. In the future, could also include
496 : /// explicit user-defined snapshot points.
497 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
498 :
499 : /// The cutoff coordinates, which are combined by selecting the minimum.
500 : pub(crate) cutoffs: GcCutoffs,
501 :
502 : /// Leases granted to particular LSNs.
503 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
504 :
505 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
506 : pub(crate) within_ancestor_pitr: bool,
507 : }
508 :
509 : impl GcInfo {
510 600 : pub(crate) fn min_cutoff(&self) -> Lsn {
511 600 : self.cutoffs.select_min()
512 600 : }
513 :
514 464 : pub(super) fn insert_child(
515 464 : &mut self,
516 464 : child_id: TimelineId,
517 464 : child_lsn: Lsn,
518 464 : is_offloaded: MaybeOffloaded,
519 464 : ) {
520 464 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
521 464 : self.retain_lsns.sort_by_key(|i| i.0);
522 464 : }
523 :
524 8 : pub(super) fn remove_child_maybe_offloaded(
525 8 : &mut self,
526 8 : child_id: TimelineId,
527 8 : maybe_offloaded: MaybeOffloaded,
528 8 : ) -> bool {
529 8 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
530 8 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
531 8 : let mut removed = false;
532 12 : self.retain_lsns.retain(|i| {
533 12 : if removed {
534 4 : return true;
535 8 : }
536 8 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
537 8 : removed |= remove;
538 8 : !remove
539 12 : });
540 8 : removed
541 8 : }
542 :
543 8 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
544 8 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
545 8 : }
546 :
547 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
548 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
549 0 : }
550 464 : pub(crate) fn lsn_covered_by_lease(&self, lsn: Lsn) -> bool {
551 464 : self.leases.contains_key(&lsn)
552 464 : }
553 : }
554 :
555 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
556 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
557 : /// between time-based and space-based retention for observability and consumption metrics purposes.
558 : #[derive(Debug, Clone)]
559 : pub(crate) struct GcCutoffs {
560 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
561 : /// history we must keep to retain a specified number of bytes of WAL.
562 : pub(crate) space: Lsn,
563 :
564 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
565 : /// history we must keep to enable reading back at least the PITR interval duration.
566 : pub(crate) time: Lsn,
567 : }
568 :
569 : impl Default for GcCutoffs {
570 896 : fn default() -> Self {
571 896 : Self {
572 896 : space: Lsn::INVALID,
573 896 : time: Lsn::INVALID,
574 896 : }
575 896 : }
576 : }
577 :
578 : impl GcCutoffs {
579 600 : fn select_min(&self) -> Lsn {
580 600 : std::cmp::min(self.space, self.time)
581 600 : }
582 : }
583 :
584 : pub(crate) struct TimelineVisitOutcome {
585 : completed_keyspace: KeySpace,
586 : image_covered_keyspace: KeySpace,
587 : }
588 :
589 : /// An error happened in a get() operation.
590 : #[derive(thiserror::Error, Debug)]
591 : pub(crate) enum PageReconstructError {
592 : #[error(transparent)]
593 : Other(anyhow::Error),
594 :
595 : #[error("Ancestor LSN wait error: {0}")]
596 : AncestorLsnTimeout(WaitLsnError),
597 :
598 : #[error("timeline shutting down")]
599 : Cancelled,
600 :
601 : /// An error happened replaying WAL records
602 : #[error(transparent)]
603 : WalRedo(anyhow::Error),
604 :
605 : #[error("{0}")]
606 : MissingKey(MissingKeyError),
607 : }
608 :
609 : impl From<anyhow::Error> for PageReconstructError {
610 0 : fn from(value: anyhow::Error) -> Self {
611 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
612 0 : match value.downcast::<PageReconstructError>() {
613 0 : Ok(pre) => pre,
614 0 : Err(other) => PageReconstructError::Other(other),
615 : }
616 0 : }
617 : }
618 :
619 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
620 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
621 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
622 0 : }
623 : }
624 :
625 : impl From<layer_manager::Shutdown> for PageReconstructError {
626 0 : fn from(_: layer_manager::Shutdown) -> Self {
627 0 : PageReconstructError::Cancelled
628 0 : }
629 : }
630 :
631 : impl GetVectoredError {
632 : #[cfg(test)]
633 12 : pub(crate) fn is_missing_key_error(&self) -> bool {
634 12 : matches!(self, Self::MissingKey(_))
635 12 : }
636 : }
637 :
638 : impl From<layer_manager::Shutdown> for GetVectoredError {
639 0 : fn from(_: layer_manager::Shutdown) -> Self {
640 0 : GetVectoredError::Cancelled
641 0 : }
642 : }
643 :
644 : /// A layer identifier when used in the [`ReadPath`] structure. This enum is for observability purposes
645 : /// only and not used by the "real read path".
646 : pub enum ReadPathLayerId {
647 : PersistentLayer(PersistentLayerKey),
648 : InMemoryLayer(Range<Lsn>),
649 : }
650 :
651 : impl std::fmt::Display for ReadPathLayerId {
652 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
653 0 : match self {
654 0 : ReadPathLayerId::PersistentLayer(key) => write!(f, "{}", key),
655 0 : ReadPathLayerId::InMemoryLayer(range) => {
656 0 : write!(f, "in-mem {}..{}", range.start, range.end)
657 : }
658 : }
659 0 : }
660 : }
661 : pub struct ReadPath {
662 : keyspace: KeySpace,
663 : lsn: Lsn,
664 : path: Vec<(ReadPathLayerId, KeySpace, Range<Lsn>)>,
665 : }
666 :
667 : impl ReadPath {
668 1255230 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
669 1255230 : Self {
670 1255230 : keyspace,
671 1255230 : lsn,
672 1255230 : path: Vec::new(),
673 1255230 : }
674 1255230 : }
675 :
676 1768237 : pub fn record_layer_visit(
677 1768237 : &mut self,
678 1768237 : layer_to_read: &ReadableLayer,
679 1768237 : keyspace_to_read: &KeySpace,
680 1768237 : lsn_range: &Range<Lsn>,
681 1768237 : ) {
682 1768237 : let id = match layer_to_read {
683 554959 : ReadableLayer::PersistentLayer(layer) => {
684 554959 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
685 : }
686 1213278 : ReadableLayer::InMemoryLayer(layer) => {
687 1213278 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
688 : }
689 : };
690 1768237 : self.path
691 1768237 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
692 1768237 : }
693 : }
694 :
695 : impl std::fmt::Display for ReadPath {
696 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
697 0 : writeln!(f, "Read path for {} at lsn {}:", self.keyspace, self.lsn)?;
698 0 : for (idx, (layer_id, keyspace, lsn_range)) in self.path.iter().enumerate() {
699 0 : writeln!(
700 0 : f,
701 0 : "{}: {} {}..{} {}",
702 0 : idx, layer_id, lsn_range.start, lsn_range.end, keyspace
703 0 : )?;
704 : }
705 0 : Ok(())
706 0 : }
707 : }
708 :
709 : #[derive(thiserror::Error)]
710 : pub struct MissingKeyError {
711 : key: Key,
712 : shard: ShardNumber,
713 : cont_lsn: Lsn,
714 : request_lsn: Lsn,
715 : ancestor_lsn: Option<Lsn>,
716 : /// Debug information about the read path if there's an error
717 : read_path: Option<ReadPath>,
718 : backtrace: Option<std::backtrace::Backtrace>,
719 : }
720 :
721 : impl std::fmt::Debug for MissingKeyError {
722 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
723 0 : write!(f, "{}", self)
724 0 : }
725 : }
726 :
727 : impl std::fmt::Display for MissingKeyError {
728 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
729 0 : write!(
730 0 : f,
731 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
732 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
733 0 : )?;
734 :
735 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
736 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
737 0 : }
738 :
739 0 : if let Some(ref read_path) = self.read_path {
740 0 : write!(f, "\n{}", read_path)?;
741 0 : }
742 :
743 0 : if let Some(ref backtrace) = self.backtrace {
744 0 : write!(f, "\n{}", backtrace)?;
745 0 : }
746 :
747 0 : Ok(())
748 0 : }
749 : }
750 :
751 : impl PageReconstructError {
752 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
753 0 : pub(crate) fn is_stopping(&self) -> bool {
754 : use PageReconstructError::*;
755 0 : match self {
756 0 : Cancelled => true,
757 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
758 : }
759 0 : }
760 : }
761 :
762 : #[derive(thiserror::Error, Debug)]
763 : pub(crate) enum CreateImageLayersError {
764 : #[error("timeline shutting down")]
765 : Cancelled,
766 :
767 : #[error("read failed")]
768 : GetVectoredError(#[source] GetVectoredError),
769 :
770 : #[error("reconstruction failed")]
771 : PageReconstructError(#[source] PageReconstructError),
772 :
773 : #[error(transparent)]
774 : Other(#[from] anyhow::Error),
775 : }
776 :
777 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
778 0 : fn from(_: layer_manager::Shutdown) -> Self {
779 0 : CreateImageLayersError::Cancelled
780 0 : }
781 : }
782 :
783 : #[derive(thiserror::Error, Debug, Clone)]
784 : pub(crate) enum FlushLayerError {
785 : /// Timeline cancellation token was cancelled
786 : #[error("timeline shutting down")]
787 : Cancelled,
788 :
789 : /// We tried to flush a layer while the Timeline is in an unexpected state
790 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
791 : NotRunning(FlushLoopState),
792 :
793 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
794 : // loop via a watch channel, where we can only borrow it.
795 : #[error("create image layers (shared)")]
796 : CreateImageLayersError(Arc<CreateImageLayersError>),
797 :
798 : #[error("other (shared)")]
799 : Other(#[from] Arc<anyhow::Error>),
800 : }
801 :
802 : impl FlushLayerError {
803 : // When crossing from generic anyhow errors to this error type, we explicitly check
804 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
805 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
806 0 : let cancelled = timeline.cancel.is_cancelled()
807 : // The upload queue might have been shut down before the official cancellation of the timeline.
808 0 : || err
809 0 : .downcast_ref::<NotInitialized>()
810 0 : .map(NotInitialized::is_stopping)
811 0 : .unwrap_or_default();
812 0 : if cancelled {
813 0 : Self::Cancelled
814 : } else {
815 0 : Self::Other(Arc::new(err))
816 : }
817 0 : }
818 : }
819 :
820 : impl From<layer_manager::Shutdown> for FlushLayerError {
821 0 : fn from(_: layer_manager::Shutdown) -> Self {
822 0 : FlushLayerError::Cancelled
823 0 : }
824 : }
825 :
826 : #[derive(thiserror::Error, Debug)]
827 : pub(crate) enum GetVectoredError {
828 : #[error("timeline shutting down")]
829 : Cancelled,
830 :
831 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
832 : Oversized(u64),
833 :
834 : #[error("requested at invalid LSN: {0}")]
835 : InvalidLsn(Lsn),
836 :
837 : #[error("requested key not found: {0}")]
838 : MissingKey(MissingKeyError),
839 :
840 : #[error("ancestry walk")]
841 : GetReadyAncestorError(#[source] GetReadyAncestorError),
842 :
843 : #[error(transparent)]
844 : Other(#[from] anyhow::Error),
845 : }
846 :
847 : impl From<GetReadyAncestorError> for GetVectoredError {
848 4 : fn from(value: GetReadyAncestorError) -> Self {
849 : use GetReadyAncestorError::*;
850 4 : match value {
851 0 : Cancelled => GetVectoredError::Cancelled,
852 : AncestorLsnTimeout(_) | BadState { .. } => {
853 4 : GetVectoredError::GetReadyAncestorError(value)
854 : }
855 : }
856 4 : }
857 : }
858 :
859 : #[derive(thiserror::Error, Debug)]
860 : pub(crate) enum GetReadyAncestorError {
861 : #[error("ancestor LSN wait error")]
862 : AncestorLsnTimeout(#[from] WaitLsnError),
863 :
864 : #[error("bad state on timeline {timeline_id}: {state:?}")]
865 : BadState {
866 : timeline_id: TimelineId,
867 : state: TimelineState,
868 : },
869 :
870 : #[error("cancelled")]
871 : Cancelled,
872 : }
873 :
874 : #[derive(Clone, Copy)]
875 : pub enum LogicalSizeCalculationCause {
876 : Initial,
877 : ConsumptionMetricsSyntheticSize,
878 : EvictionTaskImitation,
879 : TenantSizeHandler,
880 : }
881 :
882 : pub enum GetLogicalSizePriority {
883 : User,
884 : Background,
885 : }
886 :
887 0 : #[derive(Debug, enumset::EnumSetType)]
888 : pub(crate) enum CompactFlags {
889 : ForceRepartition,
890 : ForceImageLayerCreation,
891 : ForceL0Compaction,
892 : OnlyL0Compaction,
893 : EnhancedGcBottomMostCompaction,
894 : DryRun,
895 : /// Disables compaction yielding e.g. due to high L0 count. This is set e.g. when requesting
896 : /// compaction via HTTP API.
897 : NoYield,
898 : }
899 :
900 : #[serde_with::serde_as]
901 0 : #[derive(Debug, Clone, serde::Deserialize)]
902 : pub(crate) struct CompactRequest {
903 : pub compact_key_range: Option<CompactKeyRange>,
904 : pub compact_lsn_range: Option<CompactLsnRange>,
905 : /// Whether the compaction job should be scheduled.
906 : #[serde(default)]
907 : pub scheduled: bool,
908 : /// Whether the compaction job should be split across key ranges.
909 : #[serde(default)]
910 : pub sub_compaction: bool,
911 : /// Max job size for each subcompaction job.
912 : pub sub_compaction_max_job_size_mb: Option<u64>,
913 : }
914 :
915 : #[derive(Debug, Clone, Default)]
916 : pub(crate) struct CompactOptions {
917 : pub flags: EnumSet<CompactFlags>,
918 : /// If set, the compaction will only compact the key range specified by this option.
919 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
920 : pub compact_key_range: Option<CompactKeyRange>,
921 : /// If set, the compaction will only compact the LSN within this value.
922 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
923 : pub compact_lsn_range: Option<CompactLsnRange>,
924 : /// Enable sub-compaction (split compaction job across key ranges).
925 : /// This option is only used by GC compaction.
926 : pub sub_compaction: bool,
927 : /// Set job size for the GC compaction.
928 : /// This option is only used by GC compaction.
929 : pub sub_compaction_max_job_size_mb: Option<u64>,
930 : }
931 :
932 : impl std::fmt::Debug for Timeline {
933 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
934 0 : write!(f, "Timeline<{}>", self.timeline_id)
935 0 : }
936 : }
937 :
938 : #[derive(thiserror::Error, Debug)]
939 : pub(crate) enum WaitLsnError {
940 : // Called on a timeline which is shutting down
941 : #[error("Shutdown")]
942 : Shutdown,
943 :
944 : // Called on an timeline not in active state or shutting down
945 : #[error("Bad timeline state: {0:?}")]
946 : BadState(TimelineState),
947 :
948 : // Timeout expired while waiting for LSN to catch up with goal.
949 : #[error("{0}")]
950 : Timeout(String),
951 : }
952 :
953 : // The impls below achieve cancellation mapping for errors.
954 : // Perhaps there's a way of achieving this with less cruft.
955 :
956 : impl From<CreateImageLayersError> for CompactionError {
957 0 : fn from(e: CreateImageLayersError) -> Self {
958 0 : match e {
959 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
960 0 : CreateImageLayersError::Other(e) => {
961 0 : CompactionError::Other(e.context("create image layers"))
962 : }
963 0 : _ => CompactionError::Other(e.into()),
964 : }
965 0 : }
966 : }
967 :
968 : impl From<CreateImageLayersError> for FlushLayerError {
969 0 : fn from(e: CreateImageLayersError) -> Self {
970 0 : match e {
971 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
972 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
973 : }
974 0 : }
975 : }
976 :
977 : impl From<PageReconstructError> for CreateImageLayersError {
978 0 : fn from(e: PageReconstructError) -> Self {
979 0 : match e {
980 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
981 0 : _ => CreateImageLayersError::PageReconstructError(e),
982 : }
983 0 : }
984 : }
985 :
986 : impl From<GetVectoredError> for CreateImageLayersError {
987 0 : fn from(e: GetVectoredError) -> Self {
988 0 : match e {
989 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
990 0 : _ => CreateImageLayersError::GetVectoredError(e),
991 : }
992 0 : }
993 : }
994 :
995 : impl From<GetVectoredError> for PageReconstructError {
996 12 : fn from(e: GetVectoredError) -> Self {
997 12 : match e {
998 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
999 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
1000 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
1001 8 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
1002 4 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
1003 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
1004 : }
1005 12 : }
1006 : }
1007 :
1008 : impl From<GetReadyAncestorError> for PageReconstructError {
1009 4 : fn from(e: GetReadyAncestorError) -> Self {
1010 : use GetReadyAncestorError::*;
1011 4 : match e {
1012 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
1013 4 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
1014 0 : Cancelled => PageReconstructError::Cancelled,
1015 : }
1016 4 : }
1017 : }
1018 :
1019 : pub(crate) enum WaitLsnTimeout {
1020 : Custom(Duration),
1021 : // Use the [`PageServerConf::wait_lsn_timeout`] default
1022 : Default,
1023 : }
1024 :
1025 : pub(crate) enum WaitLsnWaiter<'a> {
1026 : Timeline(&'a Timeline),
1027 : Tenant,
1028 : PageService,
1029 : HttpEndpoint,
1030 : }
1031 :
1032 : /// Argument to [`Timeline::shutdown`].
1033 : #[derive(Debug, Clone, Copy)]
1034 : pub(crate) enum ShutdownMode {
1035 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
1036 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
1037 : ///
1038 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1039 : /// the call to [`Timeline::shutdown`].
1040 : FreezeAndFlush,
1041 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1042 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1043 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1044 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1045 : Reload,
1046 : /// Shut down immediately, without waiting for any open layers to flush.
1047 : Hard,
1048 : }
1049 :
1050 : enum ImageLayerCreationOutcome {
1051 : /// We generated an image layer
1052 : Generated {
1053 : unfinished_image_layer: ImageLayerWriter,
1054 : },
1055 : /// The key range is empty
1056 : Empty,
1057 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1058 : /// the image layer creation.
1059 : Skip,
1060 : }
1061 :
1062 : /// Public interface functions
1063 : impl Timeline {
1064 : /// Get the LSN where this branch was created
1065 776 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1066 776 : self.ancestor_lsn
1067 776 : }
1068 :
1069 : /// Get the ancestor's timeline id
1070 24 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1071 24 : self.ancestor_timeline
1072 24 : .as_ref()
1073 24 : .map(|ancestor| ancestor.timeline_id)
1074 24 : }
1075 :
1076 : /// Get the ancestor timeline
1077 4 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1078 4 : self.ancestor_timeline.as_ref()
1079 4 : }
1080 :
1081 : /// Get the bytes written since the PITR cutoff on this branch, and
1082 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1083 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1084 0 : let gc_info = self.gc_info.read().unwrap();
1085 0 : let history = self
1086 0 : .get_last_record_lsn()
1087 0 : .checked_sub(gc_info.cutoffs.time)
1088 0 : .unwrap_or(Lsn(0))
1089 0 : .0;
1090 0 : (history, gc_info.within_ancestor_pitr)
1091 0 : }
1092 :
1093 : /// Read timeline's GC cutoff: this is the LSN at which GC has started to happen
1094 1707588 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1095 1707588 : self.applied_gc_cutoff_lsn.read()
1096 1707588 : }
1097 :
1098 : /// Read timeline's planned GC cutoff: this is the logical end of history that users
1099 : /// are allowed to read (based on configured PITR), even if physically we have more history.
1100 0 : pub(crate) fn get_gc_cutoff_lsn(&self) -> Lsn {
1101 0 : self.gc_info.read().unwrap().cutoffs.time
1102 0 : }
1103 :
1104 : /// Look up given page version.
1105 : ///
1106 : /// If a remote layer file is needed, it is downloaded as part of this
1107 : /// call.
1108 : ///
1109 : /// This method enforces [`Self::pagestream_throttle`] internally.
1110 : ///
1111 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1112 : /// abstraction above this needs to store suitable metadata to track what
1113 : /// data exists with what keys, in separate metadata entries. If a
1114 : /// non-existent key is requested, we may incorrectly return a value from
1115 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1116 : /// non-existing key.
1117 : ///
1118 : /// # Cancel-Safety
1119 : ///
1120 : /// This method is cancellation-safe.
1121 : #[inline(always)]
1122 1215166 : pub(crate) async fn get(
1123 1215166 : &self,
1124 1215166 : key: Key,
1125 1215166 : lsn: Lsn,
1126 1215166 : ctx: &RequestContext,
1127 1215166 : ) -> Result<Bytes, PageReconstructError> {
1128 1215166 : if !lsn.is_valid() {
1129 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1130 1215166 : }
1131 1215166 :
1132 1215166 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1133 1215166 : // already checked the key against the shard_identity when looking up the Timeline from
1134 1215166 : // page_service.
1135 1215166 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1136 :
1137 1215166 : let keyspace = KeySpace {
1138 1215166 : ranges: vec![key..key.next()],
1139 1215166 : };
1140 1215166 :
1141 1215166 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1142 :
1143 1215166 : let vectored_res = self
1144 1215166 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
1145 1215166 : .await;
1146 :
1147 1215166 : let key_value = vectored_res?.pop_first();
1148 1215154 : match key_value {
1149 1215130 : Some((got_key, value)) => {
1150 1215130 : if got_key != key {
1151 0 : error!(
1152 0 : "Expected {}, but singular vectored get returned {}",
1153 : key, got_key
1154 : );
1155 0 : Err(PageReconstructError::Other(anyhow!(
1156 0 : "Singular vectored get returned wrong key"
1157 0 : )))
1158 : } else {
1159 1215130 : value
1160 : }
1161 : }
1162 24 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1163 24 : key,
1164 24 : shard: self.shard_identity.get_shard_number(&key),
1165 24 : cont_lsn: Lsn(0),
1166 24 : request_lsn: lsn,
1167 24 : ancestor_lsn: None,
1168 24 : backtrace: None,
1169 24 : read_path: None,
1170 24 : })),
1171 : }
1172 1215166 : }
1173 :
1174 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1175 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1176 :
1177 : /// Look up multiple page versions at a given LSN
1178 : ///
1179 : /// This naive implementation will be replaced with a more efficient one
1180 : /// which actually vectorizes the read path.
1181 39332 : pub(crate) async fn get_vectored(
1182 39332 : &self,
1183 39332 : keyspace: KeySpace,
1184 39332 : lsn: Lsn,
1185 39332 : io_concurrency: super::storage_layer::IoConcurrency,
1186 39332 : ctx: &RequestContext,
1187 39332 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1188 39332 : if !lsn.is_valid() {
1189 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1190 39332 : }
1191 39332 :
1192 39332 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1193 39332 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1194 0 : return Err(GetVectoredError::Oversized(key_count));
1195 39332 : }
1196 :
1197 78664 : for range in &keyspace.ranges {
1198 39332 : let mut key = range.start;
1199 79128 : while key != range.end {
1200 39796 : assert!(!self.shard_identity.is_key_disposable(&key));
1201 39796 : key = key.next();
1202 : }
1203 : }
1204 :
1205 39332 : trace!(
1206 0 : "get vectored request for {:?}@{} from task kind {:?}",
1207 0 : keyspace,
1208 0 : lsn,
1209 0 : ctx.task_kind(),
1210 : );
1211 :
1212 39332 : let start = crate::metrics::GET_VECTORED_LATENCY
1213 39332 : .for_task_kind(ctx.task_kind())
1214 39332 : .map(|metric| (metric, Instant::now()));
1215 :
1216 39332 : let res = self
1217 39332 : .get_vectored_impl(
1218 39332 : keyspace.clone(),
1219 39332 : lsn,
1220 39332 : &mut ValuesReconstructState::new(io_concurrency),
1221 39332 : ctx,
1222 39332 : )
1223 39332 : .await;
1224 :
1225 39332 : if let Some((metric, start)) = start {
1226 0 : let elapsed = start.elapsed();
1227 0 : metric.observe(elapsed.as_secs_f64());
1228 39332 : }
1229 :
1230 39332 : res
1231 39332 : }
1232 :
1233 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1234 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1235 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1236 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1237 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1238 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1239 : /// the scan operation will not cause OOM in the future.
1240 24 : pub(crate) async fn scan(
1241 24 : &self,
1242 24 : keyspace: KeySpace,
1243 24 : lsn: Lsn,
1244 24 : ctx: &RequestContext,
1245 24 : io_concurrency: super::storage_layer::IoConcurrency,
1246 24 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1247 24 : if !lsn.is_valid() {
1248 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1249 24 : }
1250 24 :
1251 24 : trace!(
1252 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1253 0 : keyspace,
1254 0 : lsn,
1255 0 : ctx.task_kind()
1256 : );
1257 :
1258 : // We should generalize this into Keyspace::contains in the future.
1259 48 : for range in &keyspace.ranges {
1260 24 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1261 24 : || range.end.field1 > METADATA_KEY_END_PREFIX
1262 : {
1263 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1264 0 : "only metadata keyspace can be scanned"
1265 0 : )));
1266 24 : }
1267 : }
1268 :
1269 24 : let start = crate::metrics::SCAN_LATENCY
1270 24 : .for_task_kind(ctx.task_kind())
1271 24 : .map(ScanLatencyOngoingRecording::start_recording);
1272 :
1273 24 : let vectored_res = self
1274 24 : .get_vectored_impl(
1275 24 : keyspace.clone(),
1276 24 : lsn,
1277 24 : &mut ValuesReconstructState::new(io_concurrency),
1278 24 : ctx,
1279 24 : )
1280 24 : .await;
1281 :
1282 24 : if let Some(recording) = start {
1283 0 : recording.observe();
1284 24 : }
1285 :
1286 24 : vectored_res
1287 24 : }
1288 :
1289 1255230 : pub(super) async fn get_vectored_impl(
1290 1255230 : &self,
1291 1255230 : keyspace: KeySpace,
1292 1255230 : lsn: Lsn,
1293 1255230 : reconstruct_state: &mut ValuesReconstructState,
1294 1255230 : ctx: &RequestContext,
1295 1255230 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1296 1255230 : let read_path = if self.conf.enable_read_path_debugging {
1297 1255230 : Some(ReadPath::new(keyspace.clone(), lsn))
1298 : } else {
1299 0 : None
1300 : };
1301 1255230 : reconstruct_state.read_path = read_path;
1302 :
1303 1255230 : let traversal_res: Result<(), _> = self
1304 1255230 : .get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1305 1255230 : .await;
1306 1255230 : if let Err(err) = traversal_res {
1307 : // Wait for all the spawned IOs to complete.
1308 : // See comments on `spawn_io` inside `storage_layer` for more details.
1309 32 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1310 32 : .into_values()
1311 32 : .map(|state| state.collect_pending_ios())
1312 32 : .collect::<FuturesUnordered<_>>();
1313 32 : while collect_futs.next().await.is_some() {}
1314 32 : return Err(err);
1315 1255198 : };
1316 1255198 :
1317 1255198 : let layers_visited = reconstruct_state.get_layers_visited();
1318 1255198 :
1319 1255198 : let futs = FuturesUnordered::new();
1320 1327862 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1321 1327862 : futs.push({
1322 1327862 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1323 1327862 : async move {
1324 1327862 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1325 :
1326 1327862 : let converted = match state.collect_pending_ios().await {
1327 1327862 : Ok(ok) => ok,
1328 0 : Err(err) => {
1329 0 : return (key, Err(err));
1330 : }
1331 : };
1332 1327862 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1333 1327862 :
1334 1327862 : // The walredo module expects the records to be descending in terms of Lsn.
1335 1327862 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1336 1327862 : debug_assert!(
1337 1327862 : converted
1338 1327862 : .records
1339 1327862 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1340 0 : "{converted:?}"
1341 : );
1342 :
1343 : (
1344 1327862 : key,
1345 1327862 : walredo_self.reconstruct_value(key, lsn, converted).await,
1346 : )
1347 1327862 : }
1348 1327862 : });
1349 1327862 : }
1350 :
1351 1255198 : let results = futs
1352 1255198 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1353 1255198 : .await;
1354 :
1355 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1356 : // when they're missing. Instead they are omitted from the resulting btree
1357 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1358 : // to avoid infinite results.
1359 1255198 : if !results.is_empty() {
1360 : // Record the total number of layers visited towards each key in the batch. While some
1361 : // layers may not intersect with a given read, and the cost of layer visits are
1362 : // amortized across the batch, each visited layer contributes directly to the observed
1363 : // latency for every read in the batch, which is what we care about.
1364 1254710 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1365 0 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1366 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1367 0 : LOG_PACER.lock().unwrap().call(|| {
1368 0 : let num_keys = keyspace.total_raw_size();
1369 0 : let num_pages = results.len();
1370 0 : tracing::info!(
1371 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1372 0 : lsn = %lsn,
1373 0 : "Vectored read for {keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1374 : );
1375 0 : });
1376 1254710 : }
1377 :
1378 2582572 : for _ in &results {
1379 1327862 : self.metrics.layers_per_read.observe(layers_visited as f64);
1380 1327862 : LAYERS_PER_READ_GLOBAL.observe(layers_visited as f64);
1381 1327862 : }
1382 488 : }
1383 :
1384 1255198 : Ok(results)
1385 1255230 : }
1386 :
1387 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1388 548220 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1389 548220 : self.last_record_lsn.load().last
1390 548220 : }
1391 :
1392 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1393 0 : self.last_record_lsn.load().prev
1394 0 : }
1395 :
1396 : /// Atomically get both last and prev.
1397 456 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1398 456 : self.last_record_lsn.load()
1399 456 : }
1400 :
1401 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1402 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1403 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1404 0 : self.last_record_lsn.status_receiver()
1405 0 : }
1406 :
1407 1656 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1408 1656 : self.disk_consistent_lsn.load()
1409 1656 : }
1410 :
1411 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1412 : /// not validated with control plane yet.
1413 : /// See [`Self::get_remote_consistent_lsn_visible`].
1414 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1415 0 : self.remote_client.remote_consistent_lsn_projected()
1416 0 : }
1417 :
1418 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1419 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1420 : /// generation validation in the deletion queue.
1421 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1422 0 : self.remote_client.remote_consistent_lsn_visible()
1423 0 : }
1424 :
1425 : /// The sum of the file size of all historic layers in the layer map.
1426 : /// This method makes no distinction between local and remote layers.
1427 : /// Hence, the result **does not represent local filesystem usage**.
1428 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1429 0 : let guard = self.layers.read().await;
1430 0 : guard.layer_size_sum()
1431 0 : }
1432 :
1433 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1434 0 : self.metrics.resident_physical_size_get()
1435 0 : }
1436 :
1437 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1438 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1439 0 : }
1440 :
1441 : ///
1442 : /// Wait until WAL has been received and processed up to this LSN.
1443 : ///
1444 : /// You should call this before any of the other get_* or list_* functions. Calling
1445 : /// those functions with an LSN that has been processed yet is an error.
1446 : ///
1447 451754 : pub(crate) async fn wait_lsn(
1448 451754 : &self,
1449 451754 : lsn: Lsn,
1450 451754 : who_is_waiting: WaitLsnWaiter<'_>,
1451 451754 : timeout: WaitLsnTimeout,
1452 451754 : ctx: &RequestContext, /* Prepare for use by cancellation */
1453 451754 : ) -> Result<(), WaitLsnError> {
1454 451754 : let state = self.current_state();
1455 451754 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1456 0 : return Err(WaitLsnError::Shutdown);
1457 451754 : } else if !matches!(state, TimelineState::Active) {
1458 0 : return Err(WaitLsnError::BadState(state));
1459 451754 : }
1460 451754 :
1461 451754 : if cfg!(debug_assertions) {
1462 451754 : match ctx.task_kind() {
1463 : TaskKind::WalReceiverManager
1464 : | TaskKind::WalReceiverConnectionHandler
1465 : | TaskKind::WalReceiverConnectionPoller => {
1466 0 : let is_myself = match who_is_waiting {
1467 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1468 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService | WaitLsnWaiter::HttpEndpoint => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1469 : };
1470 0 : if is_myself {
1471 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1472 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1473 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1474 0 : }
1475 0 : } else {
1476 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1477 0 : // our walreceiver task can make progress independent of theirs
1478 0 : }
1479 : }
1480 451754 : _ => {}
1481 : }
1482 0 : }
1483 :
1484 451754 : let timeout = match timeout {
1485 0 : WaitLsnTimeout::Custom(t) => t,
1486 451754 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1487 : };
1488 :
1489 451754 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1490 451754 :
1491 451754 : match self.last_record_lsn.wait_for_timeout(lsn, timeout).await {
1492 451754 : Ok(()) => Ok(()),
1493 0 : Err(e) => {
1494 : use utils::seqwait::SeqWaitError::*;
1495 0 : match e {
1496 0 : Shutdown => Err(WaitLsnError::Shutdown),
1497 : Timeout => {
1498 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1499 0 : drop(_timer);
1500 0 : let walreceiver_status = self.walreceiver_status();
1501 0 : Err(WaitLsnError::Timeout(format!(
1502 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1503 0 : lsn,
1504 0 : self.get_last_record_lsn(),
1505 0 : self.get_disk_consistent_lsn(),
1506 0 : walreceiver_status,
1507 0 : )))
1508 : }
1509 : }
1510 : }
1511 : }
1512 451754 : }
1513 :
1514 0 : pub(crate) fn walreceiver_status(&self) -> String {
1515 0 : match &*self.walreceiver.lock().unwrap() {
1516 0 : None => "stopping or stopped".to_string(),
1517 0 : Some(walreceiver) => match walreceiver.status() {
1518 0 : Some(status) => status.to_human_readable_string(),
1519 0 : None => "Not active".to_string(),
1520 : },
1521 : }
1522 0 : }
1523 :
1524 : /// Check that it is valid to request operations with that lsn.
1525 464 : pub(crate) fn check_lsn_is_in_scope(
1526 464 : &self,
1527 464 : lsn: Lsn,
1528 464 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1529 464 : ) -> anyhow::Result<()> {
1530 464 : ensure!(
1531 464 : lsn >= **latest_gc_cutoff_lsn,
1532 8 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1533 8 : lsn,
1534 8 : **latest_gc_cutoff_lsn,
1535 : );
1536 456 : Ok(())
1537 464 : }
1538 :
1539 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1540 20 : pub(crate) fn init_lsn_lease(
1541 20 : &self,
1542 20 : lsn: Lsn,
1543 20 : length: Duration,
1544 20 : ctx: &RequestContext,
1545 20 : ) -> anyhow::Result<LsnLease> {
1546 20 : self.make_lsn_lease(lsn, length, true, ctx)
1547 20 : }
1548 :
1549 : /// Renews a lease at a particular LSN. The requested LSN is not validated against the `latest_gc_cutoff_lsn` when we are in the grace period.
1550 8 : pub(crate) fn renew_lsn_lease(
1551 8 : &self,
1552 8 : lsn: Lsn,
1553 8 : length: Duration,
1554 8 : ctx: &RequestContext,
1555 8 : ) -> anyhow::Result<LsnLease> {
1556 8 : self.make_lsn_lease(lsn, length, false, ctx)
1557 8 : }
1558 :
1559 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1560 : ///
1561 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1562 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1563 : ///
1564 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1565 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1566 28 : fn make_lsn_lease(
1567 28 : &self,
1568 28 : lsn: Lsn,
1569 28 : length: Duration,
1570 28 : init: bool,
1571 28 : _ctx: &RequestContext,
1572 28 : ) -> anyhow::Result<LsnLease> {
1573 24 : let lease = {
1574 : // Normalize the requested LSN to be aligned, and move to the first record
1575 : // if it points to the beginning of the page (header).
1576 28 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1577 28 :
1578 28 : let mut gc_info = self.gc_info.write().unwrap();
1579 28 : let planned_cutoff = gc_info.min_cutoff();
1580 28 :
1581 28 : let valid_until = SystemTime::now() + length;
1582 28 :
1583 28 : let entry = gc_info.leases.entry(lsn);
1584 28 :
1585 28 : match entry {
1586 12 : Entry::Occupied(mut occupied) => {
1587 12 : let existing_lease = occupied.get_mut();
1588 12 : if valid_until > existing_lease.valid_until {
1589 4 : existing_lease.valid_until = valid_until;
1590 4 : let dt: DateTime<Utc> = valid_until.into();
1591 4 : info!("lease extended to {}", dt);
1592 : } else {
1593 8 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1594 8 : info!("existing lease covers greater length, valid until {}", dt);
1595 : }
1596 :
1597 12 : existing_lease.clone()
1598 : }
1599 16 : Entry::Vacant(vacant) => {
1600 : // Reject already GC-ed LSN if we are in AttachedSingle and
1601 : // not blocked by the lsn lease deadline.
1602 16 : let validate = {
1603 16 : let conf = self.tenant_conf.load();
1604 16 : conf.location.attach_mode == AttachmentMode::Single
1605 16 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1606 : };
1607 :
1608 16 : if init || validate {
1609 16 : let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
1610 16 : if lsn < *latest_gc_cutoff_lsn {
1611 4 : bail!("tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1612 12 : }
1613 12 : if lsn < planned_cutoff {
1614 0 : bail!("tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}", lsn, planned_cutoff);
1615 12 : }
1616 0 : }
1617 :
1618 12 : let dt: DateTime<Utc> = valid_until.into();
1619 12 : info!("lease created, valid until {}", dt);
1620 12 : vacant.insert(LsnLease { valid_until }).clone()
1621 : }
1622 : }
1623 : };
1624 :
1625 24 : Ok(lease)
1626 28 : }
1627 :
1628 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1629 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1630 : #[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))]
1631 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1632 : self.freeze0().await
1633 : }
1634 :
1635 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1636 : #[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))]
1637 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1638 : self.freeze_and_flush0().await
1639 : }
1640 :
1641 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1642 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1643 2244 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1644 2244 : let mut g = self.write_lock.lock().await;
1645 2244 : let to_lsn = self.get_last_record_lsn();
1646 2244 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1647 2244 : }
1648 :
1649 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1650 : // polluting the span hierarchy.
1651 2244 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1652 2244 : let token = self.freeze0().await?;
1653 2244 : self.wait_flush_completion(token).await
1654 2244 : }
1655 :
1656 : // Check if an open ephemeral layer should be closed: this provides
1657 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1658 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1659 : // ephemeral layer bytes has been breached.
1660 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1661 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1662 : // If the write lock is held, there is an active wal receiver: rolling open layers
1663 : // is their responsibility while they hold this lock.
1664 0 : return;
1665 : };
1666 :
1667 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1668 : // time, and this was missed.
1669 : // if write_guard.is_none() { return; }
1670 :
1671 0 : let Ok(layers_guard) = self.layers.try_read() else {
1672 : // Don't block if the layer lock is busy
1673 0 : return;
1674 : };
1675 :
1676 0 : let Ok(lm) = layers_guard.layer_map() else {
1677 0 : return;
1678 : };
1679 :
1680 0 : let Some(open_layer) = &lm.open_layer else {
1681 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1682 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1683 : // that didn't result in writes to this shard.
1684 :
1685 : // Must not hold the layers lock while waiting for a flush.
1686 0 : drop(layers_guard);
1687 0 :
1688 0 : let last_record_lsn = self.get_last_record_lsn();
1689 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1690 0 : if last_record_lsn > disk_consistent_lsn {
1691 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1692 : // we are a sharded tenant and have skipped some WAL
1693 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1694 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1695 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1696 : // without any data ingested (yet) doesn't write a remote index as soon as it
1697 : // sees its LSN advance: we only do this if we've been layer-less
1698 : // for some time.
1699 0 : tracing::debug!(
1700 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1701 : disk_consistent_lsn,
1702 : last_record_lsn
1703 : );
1704 :
1705 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1706 : // We know there is no open layer, so we can request freezing without actually
1707 : // freezing anything. This is true even if we have dropped the layers_guard, we
1708 : // still hold the write_guard.
1709 0 : let _ = async {
1710 0 : let token = self
1711 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1712 0 : .await?;
1713 0 : self.wait_flush_completion(token).await
1714 0 : }
1715 0 : .await;
1716 0 : }
1717 0 : }
1718 :
1719 0 : return;
1720 : };
1721 :
1722 0 : let Some(current_size) = open_layer.try_len() else {
1723 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1724 : // read lock to get size should always succeed.
1725 0 : tracing::warn!("Lock conflict while reading size of open layer");
1726 0 : return;
1727 : };
1728 :
1729 0 : let current_lsn = self.get_last_record_lsn();
1730 :
1731 0 : let checkpoint_distance_override = open_layer.tick().await;
1732 :
1733 0 : if let Some(size_override) = checkpoint_distance_override {
1734 0 : if current_size > size_override {
1735 : // This is not harmful, but it only happens in relatively rare cases where
1736 : // time-based checkpoints are not happening fast enough to keep the amount of
1737 : // ephemeral data within configured limits. It's a sign of stress on the system.
1738 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1739 0 : }
1740 0 : }
1741 :
1742 0 : let checkpoint_distance =
1743 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1744 0 :
1745 0 : if self.should_roll(
1746 0 : current_size,
1747 0 : current_size,
1748 0 : checkpoint_distance,
1749 0 : self.get_last_record_lsn(),
1750 0 : self.last_freeze_at.load(),
1751 0 : open_layer.get_opened_at(),
1752 0 : ) {
1753 0 : match open_layer.info() {
1754 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1755 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1756 0 : // happens asynchronously in the background.
1757 0 : tracing::debug!(
1758 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1759 : );
1760 : }
1761 : InMemoryLayerInfo::Open { .. } => {
1762 : // Upgrade to a write lock and freeze the layer
1763 0 : drop(layers_guard);
1764 0 : let res = self
1765 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1766 0 : .await;
1767 :
1768 0 : if let Err(e) = res {
1769 0 : tracing::info!(
1770 0 : "failed to flush frozen layer after background freeze: {e:#}"
1771 : );
1772 0 : }
1773 : }
1774 : }
1775 0 : }
1776 0 : }
1777 :
1778 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1779 : ///
1780 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1781 : /// child timelines that are not offloaded yet.
1782 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1783 0 : if self.remote_client.is_archived() != Some(true) {
1784 0 : return (false, "the timeline is not archived");
1785 0 : }
1786 0 : if !self.remote_client.no_pending_work() {
1787 : // if the remote client is still processing some work, we can't offload
1788 0 : return (false, "the upload queue is not drained yet");
1789 0 : }
1790 0 :
1791 0 : (true, "ok")
1792 0 : }
1793 :
1794 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1795 : /// compaction tasks.
1796 768 : pub(crate) async fn compact(
1797 768 : self: &Arc<Self>,
1798 768 : cancel: &CancellationToken,
1799 768 : flags: EnumSet<CompactFlags>,
1800 768 : ctx: &RequestContext,
1801 768 : ) -> Result<CompactionOutcome, CompactionError> {
1802 768 : self.compact_with_options(
1803 768 : cancel,
1804 768 : CompactOptions {
1805 768 : flags,
1806 768 : compact_key_range: None,
1807 768 : compact_lsn_range: None,
1808 768 : sub_compaction: false,
1809 768 : sub_compaction_max_job_size_mb: None,
1810 768 : },
1811 768 : ctx,
1812 768 : )
1813 768 : .await
1814 768 : }
1815 :
1816 : /// Outermost timeline compaction operation; downloads needed layers.
1817 : ///
1818 : /// NB: the cancellation token is usually from a background task, but can also come from a
1819 : /// request task.
1820 768 : pub(crate) async fn compact_with_options(
1821 768 : self: &Arc<Self>,
1822 768 : cancel: &CancellationToken,
1823 768 : options: CompactOptions,
1824 768 : ctx: &RequestContext,
1825 768 : ) -> Result<CompactionOutcome, CompactionError> {
1826 768 : // Acquire the compaction lock and task semaphore.
1827 768 : //
1828 768 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
1829 768 : // out by other background tasks (including image compaction). We request this via
1830 768 : // `BackgroundLoopKind::L0Compaction`.
1831 768 : //
1832 768 : // If this is a regular compaction pass, and L0-only compaction is enabled in the config,
1833 768 : // then we should yield for immediate L0 compaction if necessary while we're waiting for the
1834 768 : // background task semaphore. There's no point yielding otherwise, since we'd just end up
1835 768 : // right back here.
1836 768 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
1837 768 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
1838 0 : true => BackgroundLoopKind::L0Compaction,
1839 768 : false => BackgroundLoopKind::Compaction,
1840 : };
1841 768 : let yield_for_l0 = !is_l0_only
1842 768 : && self.get_compaction_l0_first()
1843 0 : && !options.flags.contains(CompactFlags::NoYield);
1844 :
1845 768 : let acquire = async move {
1846 768 : let guard = self.compaction_lock.lock().await;
1847 768 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
1848 768 : (guard, permit)
1849 768 : };
1850 :
1851 768 : let (_guard, _permit) = tokio::select! {
1852 768 : (guard, permit) = acquire => (guard, permit),
1853 768 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
1854 0 : return Ok(CompactionOutcome::YieldForL0);
1855 : }
1856 768 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
1857 768 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
1858 : };
1859 :
1860 768 : let last_record_lsn = self.get_last_record_lsn();
1861 768 :
1862 768 : // Last record Lsn could be zero in case the timeline was just created
1863 768 : if !last_record_lsn.is_valid() {
1864 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1865 0 : return Ok(CompactionOutcome::Skipped);
1866 768 : }
1867 :
1868 768 : let result = match self.get_compaction_algorithm_settings().kind {
1869 : CompactionAlgorithm::Tiered => {
1870 0 : self.compact_tiered(cancel, ctx).await?;
1871 0 : Ok(CompactionOutcome::Done)
1872 : }
1873 768 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1874 : };
1875 :
1876 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
1877 0 : match result {
1878 768 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
1879 : Err(CompactionError::Other(_)) => {
1880 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
1881 : }
1882 : // Don't change the current value on offload failure or shutdown. We don't want to
1883 : // abruptly stall nor resume L0 flushes in these cases.
1884 0 : Err(CompactionError::Offload(_)) => {}
1885 0 : Err(CompactionError::ShuttingDown) => {}
1886 : };
1887 :
1888 768 : result
1889 768 : }
1890 :
1891 : /// Mutate the timeline with a [`TimelineWriter`].
1892 10266384 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1893 10266384 : TimelineWriter {
1894 10266384 : tl: self,
1895 10266384 : write_guard: self.write_lock.lock().await,
1896 : }
1897 10266384 : }
1898 :
1899 0 : pub(crate) fn activate(
1900 0 : self: &Arc<Self>,
1901 0 : parent: Arc<crate::tenant::Tenant>,
1902 0 : broker_client: BrokerClientChannel,
1903 0 : background_jobs_can_start: Option<&completion::Barrier>,
1904 0 : ctx: &RequestContext,
1905 0 : ) {
1906 0 : if self.tenant_shard_id.is_shard_zero() {
1907 0 : // Logical size is only maintained accurately on shard zero.
1908 0 : self.spawn_initial_logical_size_computation_task(ctx);
1909 0 : }
1910 0 : self.launch_wal_receiver(ctx, broker_client);
1911 0 : self.set_state(TimelineState::Active);
1912 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1913 0 : }
1914 :
1915 : /// After this function returns, there are no timeline-scoped tasks are left running.
1916 : ///
1917 : /// The preferred pattern for is:
1918 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1919 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1920 : /// go the extra mile and keep track of JoinHandles
1921 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1922 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1923 : ///
1924 : /// For legacy reasons, we still have multiple tasks spawned using
1925 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1926 : /// We refer to these as "timeline-scoped task_mgr tasks".
1927 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1928 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1929 : /// or [`task_mgr::shutdown_watcher`].
1930 : /// We want to gradually convert the code base away from these.
1931 : ///
1932 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1933 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1934 : /// ones that aren't mentioned here):
1935 : /// - [`TaskKind::TimelineDeletionWorker`]
1936 : /// - NB: also used for tenant deletion
1937 : /// - [`TaskKind::RemoteUploadTask`]`
1938 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1939 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1940 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1941 : /// - [`TaskKind::Eviction`]
1942 : /// - [`TaskKind::LayerFlushTask`]
1943 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1944 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1945 20 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1946 20 : debug_assert_current_span_has_tenant_and_timeline_id();
1947 20 :
1948 20 : // Regardless of whether we're going to try_freeze_and_flush
1949 20 : // or not, stop ingesting any more data. Walreceiver only provides
1950 20 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1951 20 : // So, only after the self.gate.close() below will we know for sure that
1952 20 : // no walreceiver tasks are left.
1953 20 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1954 20 : // data during the call to `self.freeze_and_flush()` below.
1955 20 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1956 20 : // which is what we'd need to properly model early shutdown of the walreceiver
1957 20 : // task sub-tree before the other Timeline task sub-trees.
1958 20 : let walreceiver = self.walreceiver.lock().unwrap().take();
1959 20 : tracing::debug!(
1960 0 : is_some = walreceiver.is_some(),
1961 0 : "Waiting for WalReceiverManager..."
1962 : );
1963 20 : if let Some(walreceiver) = walreceiver {
1964 0 : walreceiver.cancel();
1965 20 : }
1966 : // ... and inform any waiters for newer LSNs that there won't be any.
1967 20 : self.last_record_lsn.shutdown();
1968 20 :
1969 20 : if let ShutdownMode::FreezeAndFlush = mode {
1970 12 : let do_flush = if let Some((open, frozen)) = self
1971 12 : .layers
1972 12 : .read()
1973 12 : .await
1974 12 : .layer_map()
1975 12 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1976 12 : .ok()
1977 12 : .filter(|(open, frozen)| *open || *frozen > 0)
1978 : {
1979 0 : if self.remote_client.is_archived() == Some(true) {
1980 : // No point flushing on shutdown for an archived timeline: it is not important
1981 : // to have it nice and fresh after our restart, and trying to flush here might
1982 : // race with trying to offload it (which also stops the flush loop)
1983 0 : false
1984 : } else {
1985 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1986 0 : true
1987 : }
1988 : } else {
1989 : // this is double-shutdown, it'll be a no-op
1990 12 : true
1991 : };
1992 :
1993 : // we shut down walreceiver above, so, we won't add anything more
1994 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1995 : // to reach the disk & upload queue, then shut the upload queue and
1996 : // wait for it to drain.
1997 12 : if do_flush {
1998 12 : match self.freeze_and_flush().await {
1999 : Ok(_) => {
2000 : // drain the upload queue
2001 : // if we did not wait for completion here, it might be our shutdown process
2002 : // didn't wait for remote uploads to complete at all, as new tasks can forever
2003 : // be spawned.
2004 : //
2005 : // what is problematic is the shutting down of RemoteTimelineClient, because
2006 : // obviously it does not make sense to stop while we wait for it, but what
2007 : // about corner cases like s3 suddenly hanging up?
2008 12 : self.remote_client.shutdown().await;
2009 : }
2010 : Err(FlushLayerError::Cancelled) => {
2011 : // this is likely the second shutdown, ignore silently.
2012 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2013 0 : debug_assert!(self.cancel.is_cancelled());
2014 : }
2015 0 : Err(e) => {
2016 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
2017 0 : // we have some extra WAL replay to do next time the timeline starts.
2018 0 : warn!("failed to freeze and flush: {e:#}");
2019 : }
2020 : }
2021 :
2022 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2023 : // we also do a final check here to ensure that the queue is empty.
2024 12 : if !self.remote_client.no_pending_work() {
2025 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
2026 12 : }
2027 0 : }
2028 8 : }
2029 :
2030 20 : if let ShutdownMode::Reload = mode {
2031 : // drain the upload queue
2032 4 : self.remote_client.shutdown().await;
2033 4 : if !self.remote_client.no_pending_work() {
2034 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
2035 4 : }
2036 16 : }
2037 :
2038 : // Signal any subscribers to our cancellation token to drop out
2039 20 : tracing::debug!("Cancelling CancellationToken");
2040 20 : self.cancel.cancel();
2041 20 :
2042 20 : // Ensure Prevent new page service requests from starting.
2043 20 : self.handles.shutdown();
2044 20 :
2045 20 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2046 20 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2047 20 : self.remote_client.stop();
2048 20 :
2049 20 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2050 20 : // to shut down the upload queue tasks.
2051 20 : // TODO: fix that, task management should be encapsulated inside remote_client.
2052 20 : task_mgr::shutdown_tasks(
2053 20 : Some(TaskKind::RemoteUploadTask),
2054 20 : Some(self.tenant_shard_id),
2055 20 : Some(self.timeline_id),
2056 20 : )
2057 20 : .await;
2058 :
2059 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2060 20 : tracing::debug!("Waiting for tasks...");
2061 20 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2062 :
2063 : {
2064 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2065 : // open.
2066 20 : let mut write_guard = self.write_lock.lock().await;
2067 20 : self.layers.write().await.shutdown(&mut write_guard);
2068 20 : }
2069 20 :
2070 20 : // Finally wait until any gate-holders are complete.
2071 20 : //
2072 20 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2073 20 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2074 20 : self.gate.close().await;
2075 :
2076 20 : self.metrics.shutdown();
2077 20 : }
2078 :
2079 900 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2080 900 : match (self.current_state(), new_state) {
2081 900 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2082 4 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2083 : }
2084 0 : (st, TimelineState::Loading) => {
2085 0 : error!("ignoring transition from {st:?} into Loading state");
2086 : }
2087 0 : (TimelineState::Broken { .. }, new_state) => {
2088 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2089 : }
2090 : (TimelineState::Stopping, TimelineState::Active) => {
2091 0 : error!("Not activating a Stopping timeline");
2092 : }
2093 896 : (_, new_state) => {
2094 896 : self.state.send_replace(new_state);
2095 896 : }
2096 : }
2097 900 : }
2098 :
2099 4 : pub(crate) fn set_broken(&self, reason: String) {
2100 4 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2101 4 : let broken_state = TimelineState::Broken {
2102 4 : reason,
2103 4 : backtrace: backtrace_str,
2104 4 : };
2105 4 : self.set_state(broken_state);
2106 4 :
2107 4 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2108 4 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2109 4 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2110 4 : self.cancel.cancel();
2111 4 : }
2112 :
2113 453902 : pub(crate) fn current_state(&self) -> TimelineState {
2114 453902 : self.state.borrow().clone()
2115 453902 : }
2116 :
2117 12 : pub(crate) fn is_broken(&self) -> bool {
2118 12 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2119 12 : }
2120 :
2121 472 : pub(crate) fn is_active(&self) -> bool {
2122 472 : self.current_state() == TimelineState::Active
2123 472 : }
2124 :
2125 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
2126 0 : self.remote_client.is_archived()
2127 0 : }
2128 :
2129 776 : pub(crate) fn is_stopping(&self) -> bool {
2130 776 : self.current_state() == TimelineState::Stopping
2131 776 : }
2132 :
2133 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2134 0 : self.state.subscribe()
2135 0 : }
2136 :
2137 451758 : pub(crate) async fn wait_to_become_active(
2138 451758 : &self,
2139 451758 : _ctx: &RequestContext, // Prepare for use by cancellation
2140 451758 : ) -> Result<(), TimelineState> {
2141 451758 : let mut receiver = self.state.subscribe();
2142 : loop {
2143 451758 : let current_state = receiver.borrow().clone();
2144 451758 : match current_state {
2145 : TimelineState::Loading => {
2146 0 : receiver
2147 0 : .changed()
2148 0 : .await
2149 0 : .expect("holding a reference to self");
2150 : }
2151 : TimelineState::Active { .. } => {
2152 451754 : return Ok(());
2153 : }
2154 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2155 : // There's no chance the timeline can transition back into ::Active
2156 4 : return Err(current_state);
2157 : }
2158 : }
2159 : }
2160 451758 : }
2161 :
2162 0 : pub(crate) async fn layer_map_info(
2163 0 : &self,
2164 0 : reset: LayerAccessStatsReset,
2165 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2166 0 : let guard = self.layers.read().await;
2167 0 : let layer_map = guard.layer_map()?;
2168 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2169 0 : if let Some(open_layer) = &layer_map.open_layer {
2170 0 : in_memory_layers.push(open_layer.info());
2171 0 : }
2172 0 : for frozen_layer in &layer_map.frozen_layers {
2173 0 : in_memory_layers.push(frozen_layer.info());
2174 0 : }
2175 :
2176 0 : let historic_layers = layer_map
2177 0 : .iter_historic_layers()
2178 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2179 0 : .collect();
2180 0 :
2181 0 : Ok(LayerMapInfo {
2182 0 : in_memory_layers,
2183 0 : historic_layers,
2184 0 : })
2185 0 : }
2186 :
2187 : #[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))]
2188 : pub(crate) async fn download_layer(
2189 : &self,
2190 : layer_file_name: &LayerName,
2191 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2192 : let Some(layer) = self
2193 : .find_layer(layer_file_name)
2194 : .await
2195 0 : .map_err(|e| match e {
2196 0 : layer_manager::Shutdown => {
2197 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2198 0 : }
2199 0 : })?
2200 : else {
2201 : return Ok(None);
2202 : };
2203 :
2204 : layer.download().await?;
2205 :
2206 : Ok(Some(true))
2207 : }
2208 :
2209 : /// Evict just one layer.
2210 : ///
2211 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2212 0 : pub(crate) async fn evict_layer(
2213 0 : &self,
2214 0 : layer_file_name: &LayerName,
2215 0 : ) -> anyhow::Result<Option<bool>> {
2216 0 : let _gate = self
2217 0 : .gate
2218 0 : .enter()
2219 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2220 :
2221 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2222 0 : return Ok(None);
2223 : };
2224 :
2225 : // curl has this by default
2226 0 : let timeout = std::time::Duration::from_secs(120);
2227 0 :
2228 0 : match local_layer.evict_and_wait(timeout).await {
2229 0 : Ok(()) => Ok(Some(true)),
2230 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2231 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2232 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2233 : }
2234 0 : }
2235 :
2236 9606020 : fn should_roll(
2237 9606020 : &self,
2238 9606020 : layer_size: u64,
2239 9606020 : projected_layer_size: u64,
2240 9606020 : checkpoint_distance: u64,
2241 9606020 : projected_lsn: Lsn,
2242 9606020 : last_freeze_at: Lsn,
2243 9606020 : opened_at: Instant,
2244 9606020 : ) -> bool {
2245 9606020 : let distance = projected_lsn.widening_sub(last_freeze_at);
2246 9606020 :
2247 9606020 : // Rolling the open layer can be triggered by:
2248 9606020 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2249 9606020 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2250 9606020 : // account for how writes are distributed across shards: we expect each node to consume
2251 9606020 : // 1/count of the LSN on average.
2252 9606020 : // 2. The size of the currently open layer.
2253 9606020 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2254 9606020 : // up and suspend activity.
2255 9606020 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2256 0 : info!(
2257 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2258 : projected_lsn, layer_size, distance
2259 : );
2260 :
2261 0 : true
2262 9606020 : } else if projected_layer_size >= checkpoint_distance {
2263 : // NB: this check is relied upon by:
2264 160 : let _ = IndexEntry::validate_checkpoint_distance;
2265 160 : info!(
2266 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2267 : projected_lsn, layer_size, projected_layer_size
2268 : );
2269 :
2270 160 : true
2271 9605860 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2272 0 : info!(
2273 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2274 0 : projected_lsn,
2275 0 : layer_size,
2276 0 : opened_at.elapsed()
2277 : );
2278 :
2279 0 : true
2280 : } else {
2281 9605860 : false
2282 : }
2283 9606020 : }
2284 : }
2285 :
2286 : /// Number of times we will compute partition within a checkpoint distance.
2287 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2288 :
2289 : // Private functions
2290 : impl Timeline {
2291 24 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2292 24 : let tenant_conf = self.tenant_conf.load();
2293 24 : tenant_conf
2294 24 : .tenant_conf
2295 24 : .lsn_lease_length
2296 24 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2297 24 : }
2298 :
2299 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2300 0 : let tenant_conf = self.tenant_conf.load();
2301 0 : tenant_conf
2302 0 : .tenant_conf
2303 0 : .lsn_lease_length_for_ts
2304 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2305 0 : }
2306 :
2307 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2308 0 : let tenant_conf = self.tenant_conf.load();
2309 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2310 0 : }
2311 :
2312 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2313 0 : let tenant_conf = self.tenant_conf.load();
2314 0 : tenant_conf
2315 0 : .tenant_conf
2316 0 : .lazy_slru_download
2317 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2318 0 : }
2319 :
2320 9609116 : fn get_checkpoint_distance(&self) -> u64 {
2321 9609116 : let tenant_conf = self.tenant_conf.load();
2322 9609116 : tenant_conf
2323 9609116 : .tenant_conf
2324 9609116 : .checkpoint_distance
2325 9609116 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2326 9609116 : }
2327 :
2328 9605860 : fn get_checkpoint_timeout(&self) -> Duration {
2329 9605860 : let tenant_conf = self.tenant_conf.load();
2330 9605860 : tenant_conf
2331 9605860 : .tenant_conf
2332 9605860 : .checkpoint_timeout
2333 9605860 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2334 9605860 : }
2335 :
2336 5016 : fn get_compaction_period(&self) -> Duration {
2337 5016 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2338 5016 : tenant_conf
2339 5016 : .compaction_period
2340 5016 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2341 5016 : }
2342 :
2343 1368 : fn get_compaction_target_size(&self) -> u64 {
2344 1368 : let tenant_conf = self.tenant_conf.load();
2345 1368 : tenant_conf
2346 1368 : .tenant_conf
2347 1368 : .compaction_target_size
2348 1368 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2349 1368 : }
2350 :
2351 3208 : fn get_compaction_threshold(&self) -> usize {
2352 3208 : let tenant_conf = self.tenant_conf.load();
2353 3208 : tenant_conf
2354 3208 : .tenant_conf
2355 3208 : .compaction_threshold
2356 3208 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2357 3208 : }
2358 :
2359 3892 : pub(crate) fn get_rel_size_v2_enabled(&self) -> bool {
2360 3892 : let tenant_conf = self.tenant_conf.load();
2361 3892 : tenant_conf
2362 3892 : .tenant_conf
2363 3892 : .rel_size_v2_enabled
2364 3892 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
2365 3892 : }
2366 :
2367 92 : fn get_compaction_upper_limit(&self) -> usize {
2368 92 : let tenant_conf = self.tenant_conf.load();
2369 92 : tenant_conf
2370 92 : .tenant_conf
2371 92 : .compaction_upper_limit
2372 92 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2373 92 : }
2374 :
2375 768 : pub fn get_compaction_l0_first(&self) -> bool {
2376 768 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2377 768 : tenant_conf
2378 768 : .compaction_l0_first
2379 768 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2380 768 : }
2381 :
2382 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2383 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2384 0 : tenant_conf
2385 0 : .compaction_l0_semaphore
2386 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2387 0 : }
2388 :
2389 2508 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2390 : // Disable L0 flushes by default. This and compaction needs further tuning.
2391 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 0; // TODO: default to e.g. 3
2392 :
2393 : // If compaction is disabled, don't delay.
2394 2508 : if self.get_compaction_period() == Duration::ZERO {
2395 2508 : return None;
2396 0 : }
2397 0 :
2398 0 : let compaction_threshold = self.get_compaction_threshold();
2399 0 : let tenant_conf = self.tenant_conf.load();
2400 0 : let l0_flush_delay_threshold = tenant_conf
2401 0 : .tenant_conf
2402 0 : .l0_flush_delay_threshold
2403 0 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2404 0 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2405 0 :
2406 0 : // 0 disables backpressure.
2407 0 : if l0_flush_delay_threshold == 0 {
2408 0 : return None;
2409 0 : }
2410 0 :
2411 0 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2412 0 : // backpressure flushes below this.
2413 0 : // TODO: the tenant config should have validation to prevent this instead.
2414 0 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2415 0 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2416 2508 : }
2417 :
2418 2508 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2419 : // Disable L0 stalls by default. In ingest benchmarks, we see image compaction take >10
2420 : // minutes, blocking L0 compaction, and we can't stall L0 flushes for that long.
2421 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2422 :
2423 : // If compaction is disabled, don't stall.
2424 2508 : if self.get_compaction_period() == Duration::ZERO {
2425 2508 : return None;
2426 0 : }
2427 0 :
2428 0 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2429 0 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2430 0 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2431 0 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2432 0 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2433 0 : return None;
2434 0 : }
2435 0 :
2436 0 : let compaction_threshold = self.get_compaction_threshold();
2437 0 : let tenant_conf = self.tenant_conf.load();
2438 0 : let l0_flush_stall_threshold = tenant_conf
2439 0 : .tenant_conf
2440 0 : .l0_flush_stall_threshold
2441 0 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2442 0 :
2443 0 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2444 0 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2445 0 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2446 0 : if cfg!(feature = "testing")
2447 0 : && compaction_threshold == 1
2448 0 : && l0_flush_stall_threshold.is_none()
2449 : {
2450 0 : return None;
2451 0 : }
2452 0 :
2453 0 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2454 0 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2455 0 :
2456 0 : // 0 disables backpressure.
2457 0 : if l0_flush_stall_threshold == 0 {
2458 0 : return None;
2459 0 : }
2460 0 :
2461 0 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2462 0 : // backpressure flushes below this.
2463 0 : // TODO: the tenant config should have validation to prevent this instead.
2464 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2465 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2466 2508 : }
2467 :
2468 2348 : fn get_l0_flush_wait_upload(&self) -> bool {
2469 2348 : let tenant_conf = self.tenant_conf.load();
2470 2348 : tenant_conf
2471 2348 : .tenant_conf
2472 2348 : .l0_flush_wait_upload
2473 2348 : .unwrap_or(self.conf.default_tenant_conf.l0_flush_wait_upload)
2474 2348 : }
2475 :
2476 0 : fn get_image_creation_threshold(&self) -> usize {
2477 0 : let tenant_conf = self.tenant_conf.load();
2478 0 : tenant_conf
2479 0 : .tenant_conf
2480 0 : .image_creation_threshold
2481 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2482 0 : }
2483 :
2484 768 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2485 768 : let tenant_conf = &self.tenant_conf.load();
2486 768 : tenant_conf
2487 768 : .tenant_conf
2488 768 : .compaction_algorithm
2489 768 : .as_ref()
2490 768 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2491 768 : .clone()
2492 768 : }
2493 :
2494 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2495 0 : let tenant_conf = self.tenant_conf.load();
2496 0 : tenant_conf
2497 0 : .tenant_conf
2498 0 : .eviction_policy
2499 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2500 0 : }
2501 :
2502 896 : fn get_evictions_low_residence_duration_metric_threshold(
2503 896 : tenant_conf: &TenantConfOpt,
2504 896 : default_tenant_conf: &TenantConf,
2505 896 : ) -> Duration {
2506 896 : tenant_conf
2507 896 : .evictions_low_residence_duration_metric_threshold
2508 896 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2509 896 : }
2510 :
2511 1180 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2512 1180 : let tenant_conf = self.tenant_conf.load();
2513 1180 : tenant_conf
2514 1180 : .tenant_conf
2515 1180 : .image_layer_creation_check_threshold
2516 1180 : .unwrap_or(
2517 1180 : self.conf
2518 1180 : .default_tenant_conf
2519 1180 : .image_layer_creation_check_threshold,
2520 1180 : )
2521 1180 : }
2522 :
2523 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
2524 0 : let tenant_conf = self.tenant_conf.load();
2525 0 : tenant_conf
2526 0 : .tenant_conf
2527 0 : .image_creation_preempt_threshold
2528 0 : .unwrap_or(
2529 0 : self.conf
2530 0 : .default_tenant_conf
2531 0 : .image_creation_preempt_threshold,
2532 0 : )
2533 0 : }
2534 :
2535 : /// Resolve the effective WAL receiver protocol to use for this tenant.
2536 : ///
2537 : /// Priority order is:
2538 : /// 1. Tenant config override
2539 : /// 2. Default value for tenant config override
2540 : /// 3. Pageserver config override
2541 : /// 4. Pageserver config default
2542 0 : pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
2543 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2544 0 : tenant_conf
2545 0 : .wal_receiver_protocol_override
2546 0 : .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
2547 0 : .unwrap_or(self.conf.wal_receiver_protocol)
2548 0 : }
2549 :
2550 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2551 0 : // NB: Most tenant conf options are read by background loops, so,
2552 0 : // changes will automatically be picked up.
2553 0 :
2554 0 : // The threshold is embedded in the metric. So, we need to update it.
2555 0 : {
2556 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2557 0 : &new_conf.tenant_conf,
2558 0 : &self.conf.default_tenant_conf,
2559 0 : );
2560 0 :
2561 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2562 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2563 0 :
2564 0 : let timeline_id_str = self.timeline_id.to_string();
2565 0 :
2566 0 : self.remote_client.update_config(&new_conf.location);
2567 0 :
2568 0 : self.metrics
2569 0 : .evictions_with_low_residence_duration
2570 0 : .write()
2571 0 : .unwrap()
2572 0 : .change_threshold(
2573 0 : &tenant_id_str,
2574 0 : &shard_id_str,
2575 0 : &timeline_id_str,
2576 0 : new_threshold,
2577 0 : );
2578 0 : }
2579 0 : }
2580 :
2581 : /// Open a Timeline handle.
2582 : ///
2583 : /// Loads the metadata for the timeline into memory, but not the layer map.
2584 : #[allow(clippy::too_many_arguments)]
2585 896 : pub(super) fn new(
2586 896 : conf: &'static PageServerConf,
2587 896 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2588 896 : metadata: &TimelineMetadata,
2589 896 : previous_heatmap: Option<PreviousHeatmap>,
2590 896 : ancestor: Option<Arc<Timeline>>,
2591 896 : timeline_id: TimelineId,
2592 896 : tenant_shard_id: TenantShardId,
2593 896 : generation: Generation,
2594 896 : shard_identity: ShardIdentity,
2595 896 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2596 896 : resources: TimelineResources,
2597 896 : pg_version: u32,
2598 896 : state: TimelineState,
2599 896 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2600 896 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2601 896 : cancel: CancellationToken,
2602 896 : ) -> Arc<Self> {
2603 896 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2604 896 : let (state, _) = watch::channel(state);
2605 896 :
2606 896 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2607 896 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2608 896 :
2609 896 : let evictions_low_residence_duration_metric_threshold = {
2610 896 : let loaded_tenant_conf = tenant_conf.load();
2611 896 : Self::get_evictions_low_residence_duration_metric_threshold(
2612 896 : &loaded_tenant_conf.tenant_conf,
2613 896 : &conf.default_tenant_conf,
2614 896 : )
2615 : };
2616 :
2617 896 : if let Some(ancestor) = &ancestor {
2618 460 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2619 460 : // If we construct an explicit timeline object, it's obviously not offloaded
2620 460 : let is_offloaded = MaybeOffloaded::No;
2621 460 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2622 460 : }
2623 :
2624 896 : Arc::new_cyclic(|myself| {
2625 896 : let metrics = TimelineMetrics::new(
2626 896 : &tenant_shard_id,
2627 896 : &timeline_id,
2628 896 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2629 896 : "mtime",
2630 896 : evictions_low_residence_duration_metric_threshold,
2631 896 : ),
2632 896 : );
2633 896 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2634 :
2635 896 : let mut result = Timeline {
2636 896 : conf,
2637 896 : tenant_conf,
2638 896 : myself: myself.clone(),
2639 896 : timeline_id,
2640 896 : tenant_shard_id,
2641 896 : generation,
2642 896 : shard_identity,
2643 896 : pg_version,
2644 896 : layers: Default::default(),
2645 896 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
2646 896 :
2647 896 : walredo_mgr,
2648 896 : walreceiver: Mutex::new(None),
2649 896 :
2650 896 : remote_client: Arc::new(resources.remote_client),
2651 896 :
2652 896 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2653 896 : last_record_lsn: SeqWait::new(RecordLsn {
2654 896 : last: disk_consistent_lsn,
2655 896 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2656 896 : }),
2657 896 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2658 896 :
2659 896 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2660 896 : last_freeze_ts: RwLock::new(Instant::now()),
2661 896 :
2662 896 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2663 896 :
2664 896 : ancestor_timeline: ancestor,
2665 896 : ancestor_lsn: metadata.ancestor_lsn(),
2666 896 :
2667 896 : metrics,
2668 896 :
2669 896 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2670 896 : &tenant_shard_id,
2671 896 : &timeline_id,
2672 896 : resources.pagestream_throttle_metrics,
2673 896 : ),
2674 896 :
2675 7168 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2676 7168 : directory_metrics_inited: array::from_fn(|_| AtomicBool::new(false)),
2677 896 :
2678 896 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2679 896 :
2680 896 : layer_flush_start_tx,
2681 896 : layer_flush_done_tx,
2682 896 :
2683 896 : write_lock: tokio::sync::Mutex::new(None),
2684 896 :
2685 896 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2686 896 :
2687 896 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
2688 896 : LastImageLayerCreationStatus::default(),
2689 896 : )),
2690 896 :
2691 896 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2692 896 : initdb_lsn: metadata.initdb_lsn(),
2693 896 :
2694 896 : current_logical_size: if disk_consistent_lsn.is_valid() {
2695 : // we're creating timeline data with some layer files existing locally,
2696 : // need to recalculate timeline's logical size based on data in the layers.
2697 468 : LogicalSize::deferred_initial(disk_consistent_lsn)
2698 : } else {
2699 : // we're creating timeline data without any layers existing locally,
2700 : // initial logical size is 0.
2701 428 : LogicalSize::empty_initial()
2702 : },
2703 :
2704 896 : partitioning: GuardArcSwap::new((
2705 896 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2706 896 : Lsn(0),
2707 896 : )),
2708 896 : repartition_threshold: 0,
2709 896 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2710 896 : last_image_layer_creation_check_instant: Mutex::new(None),
2711 896 :
2712 896 : last_received_wal: Mutex::new(None),
2713 896 : rel_size_cache: RwLock::new(RelSizeCache {
2714 896 : complete_as_of: disk_consistent_lsn,
2715 896 : map: HashMap::new(),
2716 896 : }),
2717 896 :
2718 896 : download_all_remote_layers_task_info: RwLock::new(None),
2719 896 :
2720 896 : state,
2721 896 :
2722 896 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2723 896 : EvictionTaskTimelineState::default(),
2724 896 : ),
2725 896 : delete_progress: TimelineDeleteProgress::default(),
2726 896 :
2727 896 : cancel,
2728 896 : gate: Gate::default(),
2729 896 :
2730 896 : compaction_lock: tokio::sync::Mutex::default(),
2731 896 : compaction_failed: AtomicBool::default(),
2732 896 : l0_compaction_trigger: resources.l0_compaction_trigger,
2733 896 : gc_lock: tokio::sync::Mutex::default(),
2734 896 :
2735 896 : standby_horizon: AtomicLsn::new(0),
2736 896 :
2737 896 : pagestream_throttle: resources.pagestream_throttle,
2738 896 :
2739 896 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2740 896 :
2741 896 : #[cfg(test)]
2742 896 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2743 896 :
2744 896 : l0_flush_global_state: resources.l0_flush_global_state,
2745 896 :
2746 896 : handles: Default::default(),
2747 896 :
2748 896 : attach_wal_lag_cooldown,
2749 896 :
2750 896 : create_idempotency,
2751 896 :
2752 896 : page_trace: Default::default(),
2753 896 :
2754 896 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
2755 896 : };
2756 896 :
2757 896 : result.repartition_threshold =
2758 896 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2759 896 :
2760 896 : result
2761 896 : .metrics
2762 896 : .last_record_lsn_gauge
2763 896 : .set(disk_consistent_lsn.0 as i64);
2764 896 : result
2765 896 : })
2766 896 : }
2767 :
2768 1296 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2769 1296 : let Ok(guard) = self.gate.enter() else {
2770 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2771 0 : return;
2772 : };
2773 1296 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2774 1296 : match *flush_loop_state {
2775 884 : FlushLoopState::NotStarted => (),
2776 : FlushLoopState::Running { .. } => {
2777 412 : info!(
2778 0 : "skipping attempt to start flush_loop twice {}/{}",
2779 0 : self.tenant_shard_id, self.timeline_id
2780 : );
2781 412 : return;
2782 : }
2783 : FlushLoopState::Exited => {
2784 0 : info!(
2785 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2786 0 : self.tenant_shard_id, self.timeline_id
2787 : );
2788 0 : return;
2789 : }
2790 : }
2791 :
2792 884 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2793 884 : let self_clone = Arc::clone(self);
2794 884 :
2795 884 : debug!("spawning flush loop");
2796 884 : *flush_loop_state = FlushLoopState::Running {
2797 884 : #[cfg(test)]
2798 884 : expect_initdb_optimization: false,
2799 884 : #[cfg(test)]
2800 884 : initdb_optimization_count: 0,
2801 884 : };
2802 884 : task_mgr::spawn(
2803 884 : task_mgr::BACKGROUND_RUNTIME.handle(),
2804 884 : task_mgr::TaskKind::LayerFlushTask,
2805 884 : self.tenant_shard_id,
2806 884 : Some(self.timeline_id),
2807 884 : "layer flush task",
2808 884 : async move {
2809 884 : let _guard = guard;
2810 884 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2811 884 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2812 20 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2813 20 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2814 20 : *flush_loop_state = FlushLoopState::Exited;
2815 20 : Ok(())
2816 20 : }
2817 884 : .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))
2818 : );
2819 1296 : }
2820 :
2821 : /// Creates and starts the wal receiver.
2822 : ///
2823 : /// This function is expected to be called at most once per Timeline's lifecycle
2824 : /// when the timeline is activated.
2825 0 : fn launch_wal_receiver(
2826 0 : self: &Arc<Self>,
2827 0 : ctx: &RequestContext,
2828 0 : broker_client: BrokerClientChannel,
2829 0 : ) {
2830 0 : info!(
2831 0 : "launching WAL receiver for timeline {} of tenant {}",
2832 0 : self.timeline_id, self.tenant_shard_id
2833 : );
2834 :
2835 0 : let tenant_conf = self.tenant_conf.load();
2836 0 : let wal_connect_timeout = tenant_conf
2837 0 : .tenant_conf
2838 0 : .walreceiver_connect_timeout
2839 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2840 0 : let lagging_wal_timeout = tenant_conf
2841 0 : .tenant_conf
2842 0 : .lagging_wal_timeout
2843 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2844 0 : let max_lsn_wal_lag = tenant_conf
2845 0 : .tenant_conf
2846 0 : .max_lsn_wal_lag
2847 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2848 0 :
2849 0 : let mut guard = self.walreceiver.lock().unwrap();
2850 0 : assert!(
2851 0 : guard.is_none(),
2852 0 : "multiple launches / re-launches of WAL receiver are not supported"
2853 : );
2854 0 : *guard = Some(WalReceiver::start(
2855 0 : Arc::clone(self),
2856 0 : WalReceiverConf {
2857 0 : protocol: self.resolve_wal_receiver_protocol(),
2858 0 : wal_connect_timeout,
2859 0 : lagging_wal_timeout,
2860 0 : max_lsn_wal_lag,
2861 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2862 0 : availability_zone: self.conf.availability_zone.clone(),
2863 0 : ingest_batch_size: self.conf.ingest_batch_size,
2864 0 : },
2865 0 : broker_client,
2866 0 : ctx,
2867 0 : ));
2868 0 : }
2869 :
2870 : /// Initialize with an empty layer map. Used when creating a new timeline.
2871 884 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2872 884 : let mut layers = self.layers.try_write().expect(
2873 884 : "in the context where we call this function, no other task has access to the object",
2874 884 : );
2875 884 : layers
2876 884 : .open_mut()
2877 884 : .expect("in this context the LayerManager must still be open")
2878 884 : .initialize_empty(Lsn(start_lsn.0));
2879 884 : }
2880 :
2881 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2882 : /// files.
2883 12 : pub(super) async fn load_layer_map(
2884 12 : &self,
2885 12 : disk_consistent_lsn: Lsn,
2886 12 : index_part: IndexPart,
2887 12 : ) -> anyhow::Result<()> {
2888 : use init::{Decision::*, Discovered, DismissedLayer};
2889 : use LayerName::*;
2890 :
2891 12 : let mut guard = self.layers.write().await;
2892 :
2893 12 : let timer = self.metrics.load_layer_map_histo.start_timer();
2894 12 :
2895 12 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2896 12 : // structs representing all files on disk
2897 12 : let timeline_path = self
2898 12 : .conf
2899 12 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2900 12 : let conf = self.conf;
2901 12 : let span = tracing::Span::current();
2902 12 :
2903 12 : // Copy to move into the task we're about to spawn
2904 12 : let this = self.myself.upgrade().expect("&self method holds the arc");
2905 :
2906 12 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2907 12 : move || {
2908 12 : let _g = span.entered();
2909 12 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2910 12 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2911 12 : let mut unrecognized_files = Vec::new();
2912 12 :
2913 12 : let mut path = timeline_path;
2914 :
2915 44 : for discovered in discovered {
2916 32 : let (name, kind) = match discovered {
2917 32 : Discovered::Layer(layer_file_name, local_metadata) => {
2918 32 : discovered_layers.push((layer_file_name, local_metadata));
2919 32 : continue;
2920 : }
2921 0 : Discovered::IgnoredBackup(path) => {
2922 0 : std::fs::remove_file(path)
2923 0 : .or_else(fs_ext::ignore_not_found)
2924 0 : .fatal_err("Removing .old file");
2925 0 : continue;
2926 : }
2927 0 : Discovered::Unknown(file_name) => {
2928 0 : // we will later error if there are any
2929 0 : unrecognized_files.push(file_name);
2930 0 : continue;
2931 : }
2932 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2933 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2934 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2935 : };
2936 0 : path.push(Utf8Path::new(&name));
2937 0 : init::cleanup(&path, kind)?;
2938 0 : path.pop();
2939 : }
2940 :
2941 12 : if !unrecognized_files.is_empty() {
2942 : // assume that if there are any there are many many.
2943 0 : let n = unrecognized_files.len();
2944 0 : let first = &unrecognized_files[..n.min(10)];
2945 0 : anyhow::bail!(
2946 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2947 0 : );
2948 12 : }
2949 12 :
2950 12 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2951 12 :
2952 12 : let mut loaded_layers = Vec::new();
2953 12 : let mut needs_cleanup = Vec::new();
2954 12 : let mut total_physical_size = 0;
2955 :
2956 44 : for (name, decision) in decided {
2957 32 : let decision = match decision {
2958 32 : Ok(decision) => decision,
2959 0 : Err(DismissedLayer::Future { local }) => {
2960 0 : if let Some(local) = local {
2961 0 : init::cleanup_future_layer(
2962 0 : &local.local_path,
2963 0 : &name,
2964 0 : disk_consistent_lsn,
2965 0 : )?;
2966 0 : }
2967 0 : needs_cleanup.push(name);
2968 0 : continue;
2969 : }
2970 0 : Err(DismissedLayer::LocalOnly(local)) => {
2971 0 : init::cleanup_local_only_file(&name, &local)?;
2972 : // this file never existed remotely, we will have to do rework
2973 0 : continue;
2974 : }
2975 0 : Err(DismissedLayer::BadMetadata(local)) => {
2976 0 : init::cleanup_local_file_for_remote(&local)?;
2977 : // this file never existed remotely, we will have to do rework
2978 0 : continue;
2979 : }
2980 : };
2981 :
2982 32 : match &name {
2983 24 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2984 8 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2985 : }
2986 :
2987 32 : tracing::debug!(layer=%name, ?decision, "applied");
2988 :
2989 32 : let layer = match decision {
2990 32 : Resident { local, remote } => {
2991 32 : total_physical_size += local.file_size;
2992 32 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2993 32 : .drop_eviction_guard()
2994 : }
2995 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2996 : };
2997 :
2998 32 : loaded_layers.push(layer);
2999 : }
3000 12 : Ok((loaded_layers, needs_cleanup, total_physical_size))
3001 12 : }
3002 12 : })
3003 12 : .await
3004 12 : .map_err(anyhow::Error::new)
3005 12 : .and_then(|x| x)?;
3006 :
3007 12 : let num_layers = loaded_layers.len();
3008 12 :
3009 12 : guard
3010 12 : .open_mut()
3011 12 : .expect("layermanager must be open during init")
3012 12 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3013 12 :
3014 12 : self.remote_client
3015 12 : .schedule_layer_file_deletion(&needs_cleanup)?;
3016 12 : self.remote_client
3017 12 : .schedule_index_upload_for_file_changes()?;
3018 : // This barrier orders above DELETEs before any later operations.
3019 : // This is critical because code executing after the barrier might
3020 : // create again objects with the same key that we just scheduled for deletion.
3021 : // For example, if we just scheduled deletion of an image layer "from the future",
3022 : // later compaction might run again and re-create the same image layer.
3023 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3024 : // "same" here means same key range and LSN.
3025 : //
3026 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3027 : // the upload queue may execute the PUT first, then the DELETE.
3028 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3029 : //
3030 : // 1. a future image layer is created and uploaded
3031 : // 2. ps restart
3032 : // 3. the future layer from (1) is deleted during load layer map
3033 : // 4. image layer is re-created and uploaded
3034 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3035 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3036 : //
3037 : // See https://github.com/neondatabase/neon/issues/5878
3038 : //
3039 : // NB: generation numbers naturally protect against this because they disambiguate
3040 : // (1) and (4)
3041 : // TODO: this is basically a no-op now, should we remove it?
3042 12 : self.remote_client.schedule_barrier()?;
3043 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
3044 : // on retry.
3045 :
3046 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
3047 12 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
3048 12 : self.update_layer_visibility().await?;
3049 :
3050 12 : info!(
3051 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3052 : num_layers, disk_consistent_lsn, total_physical_size
3053 : );
3054 :
3055 12 : timer.stop_and_record();
3056 12 : Ok(())
3057 12 : }
3058 :
3059 : /// Retrieve current logical size of the timeline.
3060 : ///
3061 : /// The size could be lagging behind the actual number, in case
3062 : /// the initial size calculation has not been run (gets triggered on the first size access).
3063 : ///
3064 : /// return size and boolean flag that shows if the size is exact
3065 0 : pub(crate) fn get_current_logical_size(
3066 0 : self: &Arc<Self>,
3067 0 : priority: GetLogicalSizePriority,
3068 0 : ctx: &RequestContext,
3069 0 : ) -> logical_size::CurrentLogicalSize {
3070 0 : if !self.tenant_shard_id.is_shard_zero() {
3071 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3072 : // when HTTP API is serving a GET for timeline zero, return zero
3073 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3074 0 : }
3075 0 :
3076 0 : let current_size = self.current_logical_size.current_size();
3077 0 : debug!("Current size: {current_size:?}");
3078 :
3079 0 : match (current_size.accuracy(), priority) {
3080 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3081 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3082 0 : // background task will eventually deliver an exact value, we're in no rush
3083 0 : }
3084 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3085 : // background task is not ready, but user is asking for it now;
3086 : // => make the background task skip the line
3087 : // (The alternative would be to calculate the size here, but,
3088 : // it can actually take a long time if the user has a lot of rels.
3089 : // And we'll inevitable need it again; So, let the background task do the work.)
3090 0 : match self
3091 0 : .current_logical_size
3092 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3093 0 : .get()
3094 : {
3095 0 : Some(cancel) => cancel.cancel(),
3096 : None => {
3097 0 : match self.current_state() {
3098 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3099 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3100 0 : // Don't make noise.
3101 0 : }
3102 : TimelineState::Loading => {
3103 : // Import does not return an activated timeline.
3104 0 : info!("discarding priority boost for logical size calculation because timeline is not yet active");
3105 : }
3106 : TimelineState::Active => {
3107 : // activation should be setting the once cell
3108 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
3109 0 : debug_assert!(false);
3110 : }
3111 : }
3112 : }
3113 : }
3114 : }
3115 : }
3116 :
3117 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3118 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3119 0 : let first = self
3120 0 : .current_logical_size
3121 0 : .did_return_approximate_to_walreceiver
3122 0 : .compare_exchange(
3123 0 : false,
3124 0 : true,
3125 0 : AtomicOrdering::Relaxed,
3126 0 : AtomicOrdering::Relaxed,
3127 0 : )
3128 0 : .is_ok();
3129 0 : if first {
3130 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3131 0 : }
3132 0 : }
3133 0 : }
3134 :
3135 0 : current_size
3136 0 : }
3137 :
3138 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3139 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3140 : // nothing to do for freshly created timelines;
3141 0 : assert_eq!(
3142 0 : self.current_logical_size.current_size().accuracy(),
3143 0 : logical_size::Accuracy::Exact,
3144 0 : );
3145 0 : self.current_logical_size.initialized.add_permits(1);
3146 0 : return;
3147 : };
3148 :
3149 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3150 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3151 0 : self.current_logical_size
3152 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3153 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3154 0 :
3155 0 : let self_clone = Arc::clone(self);
3156 0 : let background_ctx = ctx.detached_child(
3157 0 : TaskKind::InitialLogicalSizeCalculation,
3158 0 : DownloadBehavior::Download,
3159 0 : );
3160 0 : task_mgr::spawn(
3161 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3162 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3163 0 : self.tenant_shard_id,
3164 0 : Some(self.timeline_id),
3165 0 : "initial size calculation",
3166 : // NB: don't log errors here, task_mgr will do that.
3167 0 : async move {
3168 0 : self_clone
3169 0 : .initial_logical_size_calculation_task(
3170 0 : initial_part_end,
3171 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3172 0 : background_ctx,
3173 0 : )
3174 0 : .await;
3175 0 : Ok(())
3176 0 : }
3177 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)),
3178 : );
3179 0 : }
3180 :
3181 : /// # Cancellation
3182 : ///
3183 : /// This method is sensitive to `Timeline::cancel`.
3184 : ///
3185 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3186 : ///
3187 : /// # Cancel-Safety
3188 : ///
3189 : /// It does Timeline IO, hence this should be polled to completion because
3190 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3191 : /// to reason about.
3192 0 : async fn initial_logical_size_calculation_task(
3193 0 : self: Arc<Self>,
3194 0 : initial_part_end: Lsn,
3195 0 : skip_concurrency_limiter: CancellationToken,
3196 0 : background_ctx: RequestContext,
3197 0 : ) {
3198 0 : scopeguard::defer! {
3199 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3200 0 : self.current_logical_size.initialized.add_permits(1);
3201 0 : }
3202 0 :
3203 0 : let try_once = |attempt: usize| {
3204 0 : let background_ctx = &background_ctx;
3205 0 : let self_ref = &self;
3206 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3207 0 : async move {
3208 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3209 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3210 0 : background_ctx,
3211 0 : );
3212 :
3213 : use crate::metrics::initial_logical_size::StartCircumstances;
3214 0 : let (_maybe_permit, circumstances) = tokio::select! {
3215 0 : permit = wait_for_permit => {
3216 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3217 : }
3218 0 : _ = self_ref.cancel.cancelled() => {
3219 0 : return Err(CalculateLogicalSizeError::Cancelled);
3220 : }
3221 0 : () = skip_concurrency_limiter.cancelled() => {
3222 : // Some action that is part of a end user interaction requested logical size
3223 : // => break out of the rate limit
3224 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3225 : // but then again what happens if they cancel; also, we should just be using
3226 : // one runtime across the entire process, so, let's leave this for now.
3227 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3228 : }
3229 : };
3230 :
3231 0 : let metrics_guard = if attempt == 1 {
3232 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3233 : } else {
3234 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3235 : };
3236 :
3237 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3238 0 : self_ref.conf,
3239 0 : self_ref
3240 0 : .gate
3241 0 : .enter()
3242 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3243 : );
3244 :
3245 0 : let calculated_size = self_ref
3246 0 : .logical_size_calculation_task(
3247 0 : initial_part_end,
3248 0 : LogicalSizeCalculationCause::Initial,
3249 0 : background_ctx,
3250 0 : )
3251 0 : .await?;
3252 :
3253 0 : self_ref
3254 0 : .trigger_aux_file_size_computation(
3255 0 : initial_part_end,
3256 0 : background_ctx,
3257 0 : io_concurrency,
3258 0 : )
3259 0 : .await?;
3260 :
3261 : // TODO: add aux file size to logical size
3262 :
3263 0 : Ok((calculated_size, metrics_guard))
3264 0 : }
3265 0 : };
3266 :
3267 0 : let retrying = async {
3268 0 : let mut attempt = 0;
3269 : loop {
3270 0 : attempt += 1;
3271 0 :
3272 0 : match try_once(attempt).await {
3273 0 : Ok(res) => return ControlFlow::Continue(res),
3274 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3275 : Err(
3276 0 : e @ (CalculateLogicalSizeError::Decode(_)
3277 0 : | CalculateLogicalSizeError::PageRead(_)),
3278 0 : ) => {
3279 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3280 : // exponential back-off doesn't make sense at these long intervals;
3281 : // use fixed retry interval with generous jitter instead
3282 0 : let sleep_duration = Duration::from_secs(
3283 0 : u64::try_from(
3284 0 : // 1hour base
3285 0 : (60_i64 * 60_i64)
3286 0 : // 10min jitter
3287 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3288 0 : )
3289 0 : .expect("10min < 1hour"),
3290 0 : );
3291 0 : tokio::select! {
3292 0 : _ = tokio::time::sleep(sleep_duration) => {}
3293 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3294 : }
3295 : }
3296 : }
3297 : }
3298 0 : };
3299 :
3300 0 : let (calculated_size, metrics_guard) = match retrying.await {
3301 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3302 0 : ControlFlow::Break(()) => return,
3303 : };
3304 :
3305 : // we cannot query current_logical_size.current_size() to know the current
3306 : // *negative* value, only truncated to u64.
3307 0 : let added = self
3308 0 : .current_logical_size
3309 0 : .size_added_after_initial
3310 0 : .load(AtomicOrdering::Relaxed);
3311 0 :
3312 0 : let sum = calculated_size.saturating_add_signed(added);
3313 0 :
3314 0 : // set the gauge value before it can be set in `update_current_logical_size`.
3315 0 : self.metrics.current_logical_size_gauge.set(sum);
3316 0 :
3317 0 : self.current_logical_size
3318 0 : .initial_logical_size
3319 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3320 0 : .ok()
3321 0 : .expect("only this task sets it");
3322 0 : }
3323 :
3324 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3325 0 : self: &Arc<Self>,
3326 0 : lsn: Lsn,
3327 0 : cause: LogicalSizeCalculationCause,
3328 0 : ctx: RequestContext,
3329 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3330 0 : let (sender, receiver) = oneshot::channel();
3331 0 : let self_clone = Arc::clone(self);
3332 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3333 0 : // we should stop the size calculation work and return an error.
3334 0 : // That would require restructuring this function's API to
3335 0 : // return the result directly, instead of a Receiver for the result.
3336 0 : let ctx = ctx.detached_child(
3337 0 : TaskKind::OndemandLogicalSizeCalculation,
3338 0 : DownloadBehavior::Download,
3339 0 : );
3340 0 : task_mgr::spawn(
3341 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3342 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3343 0 : self.tenant_shard_id,
3344 0 : Some(self.timeline_id),
3345 0 : "ondemand logical size calculation",
3346 0 : async move {
3347 0 : let res = self_clone
3348 0 : .logical_size_calculation_task(lsn, cause, &ctx)
3349 0 : .await;
3350 0 : let _ = sender.send(res).ok();
3351 0 : Ok(()) // Receiver is responsible for handling errors
3352 0 : }
3353 0 : .in_current_span(),
3354 0 : );
3355 0 : receiver
3356 0 : }
3357 :
3358 : #[instrument(skip_all)]
3359 : async fn logical_size_calculation_task(
3360 : self: &Arc<Self>,
3361 : lsn: Lsn,
3362 : cause: LogicalSizeCalculationCause,
3363 : ctx: &RequestContext,
3364 : ) -> Result<u64, CalculateLogicalSizeError> {
3365 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3366 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3367 : // accurate relation sizes, and they do not emit consumption metrics.
3368 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3369 :
3370 : let guard = self
3371 : .gate
3372 : .enter()
3373 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3374 :
3375 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3376 : }
3377 :
3378 : /// Calculate the logical size of the database at the latest LSN.
3379 : ///
3380 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3381 : /// especially if we need to download remote layers.
3382 0 : async fn calculate_logical_size(
3383 0 : &self,
3384 0 : up_to_lsn: Lsn,
3385 0 : cause: LogicalSizeCalculationCause,
3386 0 : _guard: &GateGuard,
3387 0 : ctx: &RequestContext,
3388 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3389 0 : info!(
3390 0 : "Calculating logical size for timeline {} at {}",
3391 : self.timeline_id, up_to_lsn
3392 : );
3393 :
3394 0 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
3395 : {
3396 0 : return Err(CalculateLogicalSizeError::Cancelled);
3397 0 : }
3398 :
3399 : // See if we've already done the work for initial size calculation.
3400 : // This is a short-cut for timelines that are mostly unused.
3401 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3402 0 : return Ok(size);
3403 0 : }
3404 0 : let storage_time_metrics = match cause {
3405 : LogicalSizeCalculationCause::Initial
3406 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3407 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3408 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3409 0 : &self.metrics.imitate_logical_size_histo
3410 : }
3411 : };
3412 0 : let timer = storage_time_metrics.start_timer();
3413 0 : let logical_size = self
3414 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3415 0 : .await?;
3416 0 : debug!("calculated logical size: {logical_size}");
3417 0 : timer.stop_and_record();
3418 0 : Ok(logical_size)
3419 0 : }
3420 :
3421 : /// Update current logical size, adding `delta' to the old value.
3422 541140 : fn update_current_logical_size(&self, delta: i64) {
3423 541140 : let logical_size = &self.current_logical_size;
3424 541140 : logical_size.increment_size(delta);
3425 541140 :
3426 541140 : // Also set the value in the prometheus gauge. Note that
3427 541140 : // there is a race condition here: if this is is called by two
3428 541140 : // threads concurrently, the prometheus gauge might be set to
3429 541140 : // one value while current_logical_size is set to the
3430 541140 : // other.
3431 541140 : match logical_size.current_size() {
3432 541140 : CurrentLogicalSize::Exact(ref new_current_size) => self
3433 541140 : .metrics
3434 541140 : .current_logical_size_gauge
3435 541140 : .set(new_current_size.into()),
3436 0 : CurrentLogicalSize::Approximate(_) => {
3437 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3438 0 : // forth between the initial size calculation task.
3439 0 : }
3440 : }
3441 541140 : }
3442 :
3443 5948 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: MetricsUpdate) {
3444 5948 : // TODO: this directory metrics is not correct -- we could have multiple reldirs in the system
3445 5948 : // for each of the database, but we only store one value, and therefore each pgdirmodification
3446 5948 : // would overwrite the previous value if they modify different databases.
3447 5948 :
3448 5948 : match count {
3449 2104 : MetricsUpdate::Set(count) => {
3450 2104 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3451 2104 : self.directory_metrics_inited[kind.offset()].store(true, AtomicOrdering::Relaxed);
3452 2104 : }
3453 3840 : MetricsUpdate::Add(count) => {
3454 3840 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3455 3840 : // it's fine.
3456 3840 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3457 3840 : // The metrics has been initialized with `MetricsUpdate::Set` before, so we can add/sub
3458 3840 : // the value reliably.
3459 3840 : self.directory_metrics[kind.offset()].fetch_add(count, AtomicOrdering::Relaxed);
3460 3840 : }
3461 : // Otherwise, ignore this update
3462 : }
3463 4 : MetricsUpdate::Sub(count) => {
3464 4 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3465 4 : // it's fine.
3466 4 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3467 4 : // The metrics has been initialized with `MetricsUpdate::Set` before.
3468 4 : // The operation could overflow so we need to normalize the value.
3469 4 : let prev_val =
3470 4 : self.directory_metrics[kind.offset()].load(AtomicOrdering::Relaxed);
3471 4 : let res = prev_val.saturating_sub(count);
3472 4 : self.directory_metrics[kind.offset()].store(res, AtomicOrdering::Relaxed);
3473 4 : }
3474 : // Otherwise, ignore this update
3475 : }
3476 : };
3477 :
3478 : // TODO: remove this, there's no place in the code that updates this aux metrics.
3479 5948 : let aux_metric =
3480 5948 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3481 5948 :
3482 5948 : let sum_of_entries = self
3483 5948 : .directory_metrics
3484 5948 : .iter()
3485 47584 : .map(|v| v.load(AtomicOrdering::Relaxed))
3486 5948 : .sum();
3487 : // Set a high general threshold and a lower threshold for the auxiliary files,
3488 : // as we can have large numbers of relations in the db directory.
3489 : const SUM_THRESHOLD: u64 = 5000;
3490 : const AUX_THRESHOLD: u64 = 1000;
3491 5948 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3492 0 : self.metrics
3493 0 : .directory_entries_count_gauge
3494 0 : .set(sum_of_entries);
3495 5948 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3496 0 : metric.set(sum_of_entries);
3497 5948 : }
3498 5948 : }
3499 :
3500 0 : async fn find_layer(
3501 0 : &self,
3502 0 : layer_name: &LayerName,
3503 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3504 0 : let guard = self.layers.read().await;
3505 0 : let layer = guard
3506 0 : .layer_map()?
3507 0 : .iter_historic_layers()
3508 0 : .find(|l| &l.layer_name() == layer_name)
3509 0 : .map(|found| guard.get_from_desc(&found));
3510 0 : Ok(layer)
3511 0 : }
3512 :
3513 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3514 : /// indicating which layers are currently on-disk on the primary.
3515 : ///
3516 : /// None is returned if the Timeline is in a state where uploading a heatmap
3517 : /// doesn't make sense, such as shutting down or initializing. The caller
3518 : /// should treat this as a cue to simply skip doing any heatmap uploading
3519 : /// for this timeline.
3520 32 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3521 32 : if !self.is_active() {
3522 0 : return None;
3523 32 : }
3524 :
3525 32 : let guard = self.layers.read().await;
3526 :
3527 : // Firstly, if there's any heatmap left over from when this location
3528 : // was a secondary, take that into account. Keep layers that are:
3529 : // * present in the layer map
3530 : // * visible
3531 : // * non-resident
3532 : // * not evicted since we read the heatmap
3533 : //
3534 : // Without this, a new cold, attached location would clobber the previous
3535 : // heatamp.
3536 32 : let previous_heatmap = self.previous_heatmap.load();
3537 32 : let visible_non_resident = match previous_heatmap.as_deref() {
3538 24 : Some(PreviousHeatmap::Active { heatmap, read_at }) => {
3539 92 : Some(heatmap.layers.iter().filter_map(|hl| {
3540 92 : let desc: PersistentLayerDesc = hl.name.clone().into();
3541 92 : let layer = guard.try_get_from_key(&desc.key())?;
3542 :
3543 92 : if layer.visibility() == LayerVisibilityHint::Covered {
3544 0 : return None;
3545 92 : }
3546 92 :
3547 92 : if layer.is_likely_resident() {
3548 40 : return None;
3549 52 : }
3550 52 :
3551 52 : if layer.last_evicted_at().happened_after(*read_at) {
3552 12 : return None;
3553 40 : }
3554 40 :
3555 40 : Some((desc, hl.metadata.clone(), hl.access_time))
3556 92 : }))
3557 : }
3558 0 : Some(PreviousHeatmap::Obsolete) => None,
3559 8 : None => None,
3560 : };
3561 :
3562 : // Secondly, all currently visible, resident layers are included.
3563 72 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3564 72 : match layer.visibility() {
3565 : LayerVisibilityHint::Visible => {
3566 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3567 68 : let last_activity_ts = layer.latest_activity();
3568 68 : Some((
3569 68 : layer.layer_desc().clone(),
3570 68 : layer.metadata(),
3571 68 : last_activity_ts,
3572 68 : ))
3573 : }
3574 : LayerVisibilityHint::Covered => {
3575 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3576 4 : None
3577 : }
3578 : }
3579 72 : });
3580 :
3581 32 : let mut layers = match visible_non_resident {
3582 24 : Some(non_resident) => {
3583 24 : let mut non_resident = non_resident.peekable();
3584 24 : if non_resident.peek().is_none() {
3585 8 : self.previous_heatmap
3586 8 : .store(Some(PreviousHeatmap::Obsolete.into()));
3587 16 : }
3588 :
3589 24 : non_resident.chain(resident).collect::<Vec<_>>()
3590 : }
3591 8 : None => resident.collect::<Vec<_>>(),
3592 : };
3593 :
3594 : // Sort layers in order of which to download first. For a large set of layers to download, we
3595 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3596 : // or hours later:
3597 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3598 : // only exist for a few minutes before being compacted into L1s.
3599 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3600 : // the layer is likely to be covered by an image layer during compaction.
3601 238 : layers.sort_by_key(|(desc, _meta, _atime)| {
3602 238 : std::cmp::Reverse((
3603 238 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3604 238 : desc.lsn_range.end,
3605 238 : ))
3606 238 : });
3607 32 :
3608 32 : let layers = layers
3609 32 : .into_iter()
3610 108 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3611 32 : .collect();
3612 32 :
3613 32 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3614 32 : }
3615 :
3616 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3617 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3618 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3619 0 : // branchpoint in the value in IndexPart::lineage
3620 0 : self.ancestor_lsn == lsn
3621 0 : || (self.ancestor_lsn == Lsn::INVALID
3622 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3623 0 : }
3624 : }
3625 :
3626 : impl Timeline {
3627 : #[allow(clippy::doc_lazy_continuation)]
3628 : /// Get the data needed to reconstruct all keys in the provided keyspace
3629 : ///
3630 : /// The algorithm is as follows:
3631 : /// 1. While some keys are still not done and there's a timeline to visit:
3632 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3633 : /// 2.1: Build the fringe for the current keyspace
3634 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3635 : /// intersects
3636 : /// 2.3. Pop the timeline from the fringe
3637 : /// 2.4. If the fringe is empty, go back to 1
3638 1255230 : async fn get_vectored_reconstruct_data(
3639 1255230 : &self,
3640 1255230 : mut keyspace: KeySpace,
3641 1255230 : request_lsn: Lsn,
3642 1255230 : reconstruct_state: &mut ValuesReconstructState,
3643 1255230 : ctx: &RequestContext,
3644 1255230 : ) -> Result<(), GetVectoredError> {
3645 1255230 : let mut timeline_owned: Arc<Timeline>;
3646 1255230 : let mut timeline = self;
3647 1255230 :
3648 1255230 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3649 :
3650 1255226 : let missing_keyspace = loop {
3651 1706984 : if self.cancel.is_cancelled() {
3652 0 : return Err(GetVectoredError::Cancelled);
3653 1706984 : }
3654 :
3655 : let TimelineVisitOutcome {
3656 1706984 : completed_keyspace: completed,
3657 1706984 : image_covered_keyspace,
3658 1706984 : } = Self::get_vectored_reconstruct_data_timeline(
3659 1706984 : timeline,
3660 1706984 : keyspace.clone(),
3661 1706984 : cont_lsn,
3662 1706984 : reconstruct_state,
3663 1706984 : &self.cancel,
3664 1706984 : ctx,
3665 1706984 : )
3666 1706984 : .await?;
3667 :
3668 1706984 : keyspace.remove_overlapping_with(&completed);
3669 1706984 :
3670 1706984 : // Do not descend into the ancestor timeline for aux files.
3671 1706984 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3672 1706984 : // stalling compaction.
3673 1706984 : keyspace.remove_overlapping_with(&KeySpace {
3674 1706984 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
3675 1706984 : });
3676 1706984 :
3677 1706984 : // Keyspace is fully retrieved
3678 1706984 : if keyspace.is_empty() {
3679 1254730 : break None;
3680 452254 : }
3681 :
3682 452254 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3683 : // Not fully retrieved but no ancestor timeline.
3684 496 : break Some(keyspace);
3685 : };
3686 :
3687 : // Now we see if there are keys covered by the image layer but does not exist in the
3688 : // image layer, which means that the key does not exist.
3689 :
3690 : // The block below will stop the vectored search if any of the keys encountered an image layer
3691 : // which did not contain a snapshot for said key. Since we have already removed all completed
3692 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3693 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3694 : // and stop the search as a result of that.
3695 451758 : let mut removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3696 451758 : // Do not fire missing key error and end early for sparse keys. Note that we hava already removed
3697 451758 : // non-inherited keyspaces before, so we can safely do a full `SPARSE_RANGE` remove instead of
3698 451758 : // figuring out what is the inherited key range and do a fine-grained pruning.
3699 451758 : removed.remove_overlapping_with(&KeySpace {
3700 451758 : ranges: vec![SPARSE_RANGE],
3701 451758 : });
3702 451758 : if !removed.is_empty() {
3703 0 : break Some(removed);
3704 451758 : }
3705 451758 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3706 451758 : // keyspace.
3707 451758 :
3708 451758 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3709 451758 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3710 451758 : timeline_owned = timeline
3711 451758 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3712 451758 : .await?;
3713 451754 : timeline = &*timeline_owned;
3714 : };
3715 :
3716 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
3717 1255226 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
3718 496 : let mut missing_keyspace = missing_keyspace;
3719 496 : missing_keyspace.remove_overlapping_with(&KeySpace {
3720 496 : ranges: vec![SPARSE_RANGE],
3721 496 : });
3722 496 : if missing_keyspace.is_empty() {
3723 468 : None
3724 : } else {
3725 28 : Some(missing_keyspace)
3726 : }
3727 : } else {
3728 1254730 : None
3729 : };
3730 :
3731 1255226 : if let Some(missing_keyspace) = missing_keyspace {
3732 28 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3733 28 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3734 28 : shard: self
3735 28 : .shard_identity
3736 28 : .get_shard_number(&missing_keyspace.start().unwrap()),
3737 28 : cont_lsn,
3738 28 : request_lsn,
3739 28 : ancestor_lsn: Some(timeline.ancestor_lsn),
3740 28 : backtrace: None,
3741 28 : read_path: std::mem::take(&mut reconstruct_state.read_path),
3742 28 : }));
3743 1255198 : }
3744 1255198 :
3745 1255198 : Ok(())
3746 1255230 : }
3747 :
3748 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3749 : ///
3750 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3751 : /// the current keyspace. The current keyspace of the search at any given timeline
3752 : /// is the original keyspace minus all the keys that have been completed minus
3753 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3754 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3755 : ///
3756 : /// This is basically a depth-first search visitor implementation where a vertex
3757 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3758 : ///
3759 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3760 : /// and get all the required reconstruct data from the layer in one go.
3761 : ///
3762 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3763 : /// decides how to deal with these two keyspaces.
3764 1706984 : async fn get_vectored_reconstruct_data_timeline(
3765 1706984 : timeline: &Timeline,
3766 1706984 : keyspace: KeySpace,
3767 1706984 : mut cont_lsn: Lsn,
3768 1706984 : reconstruct_state: &mut ValuesReconstructState,
3769 1706984 : cancel: &CancellationToken,
3770 1706984 : ctx: &RequestContext,
3771 1706984 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3772 1706984 : let mut unmapped_keyspace = keyspace.clone();
3773 1706984 : let mut fringe = LayerFringe::new();
3774 1706984 :
3775 1706984 : let mut completed_keyspace = KeySpace::default();
3776 1706984 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3777 1706984 :
3778 1706984 : // Prevent GC from progressing while visiting the current timeline.
3779 1706984 : // If we are GC-ing because a new image layer was added while traversing
3780 1706984 : // the timeline, then it will remove layers that are required for fulfilling
3781 1706984 : // the current get request (read-path cannot "look back" and notice the new
3782 1706984 : // image layer).
3783 1706984 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
3784 :
3785 : // See `compaction::compact_with_gc` for why we need this.
3786 1706984 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
3787 :
3788 : loop {
3789 3475221 : if cancel.is_cancelled() {
3790 0 : return Err(GetVectoredError::Cancelled);
3791 3475221 : }
3792 3475221 :
3793 3475221 : let (keys_done_last_step, keys_with_image_coverage) =
3794 3475221 : reconstruct_state.consume_done_keys();
3795 3475221 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3796 3475221 : completed_keyspace.merge(&keys_done_last_step);
3797 3475221 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3798 4998 : unmapped_keyspace
3799 4998 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3800 4998 : image_covered_keyspace.add_range(keys_with_image_coverage);
3801 3470223 : }
3802 :
3803 : // Do not descent any further if the last layer we visited
3804 : // completed all keys in the keyspace it inspected. This is not
3805 : // required for correctness, but avoids visiting extra layers
3806 : // which turns out to be a perf bottleneck in some cases.
3807 3475221 : if !unmapped_keyspace.is_empty() {
3808 2264581 : let guard = timeline.layers.read().await;
3809 2264581 : let layers = guard.layer_map()?;
3810 :
3811 2264581 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3812 1827764 : let start_lsn = l.get_lsn_range().start;
3813 1827764 : cont_lsn > start_lsn
3814 2264581 : });
3815 2264581 :
3816 2264581 : match in_memory_layer {
3817 1213278 : Some(l) => {
3818 1213278 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3819 1213278 : fringe.update(
3820 1213278 : ReadableLayer::InMemoryLayer(l),
3821 1213278 : unmapped_keyspace.clone(),
3822 1213278 : lsn_range,
3823 1213278 : );
3824 1213278 : }
3825 : None => {
3826 1051807 : for range in unmapped_keyspace.ranges.iter() {
3827 1051807 : let results = layers.range_search(range.clone(), cont_lsn);
3828 1051807 :
3829 1051807 : results
3830 1051807 : .found
3831 1051807 : .into_iter()
3832 1051807 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3833 555039 : (
3834 555039 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3835 555039 : keyspace_accum.to_keyspace(),
3836 555039 : lsn_floor..cont_lsn,
3837 555039 : )
3838 1051807 : })
3839 1051807 : .for_each(|(layer, keyspace, lsn_range)| {
3840 555039 : fringe.update(layer, keyspace, lsn_range)
3841 1051807 : });
3842 1051807 : }
3843 : }
3844 : }
3845 :
3846 : // It's safe to drop the layer map lock after planning the next round of reads.
3847 : // The fringe keeps readable handles for the layers which are safe to read even
3848 : // if layers were compacted or flushed.
3849 : //
3850 : // The more interesting consideration is: "Why is the read algorithm still correct
3851 : // if the layer map changes while it is operating?". Doing a vectored read on a
3852 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3853 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3854 : // range at *a particular point in time*. It is fine for the answer to be different
3855 : // at two different time points.
3856 2264581 : drop(guard);
3857 1210640 : }
3858 :
3859 3475221 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3860 1768237 : if let Some(ref mut read_path) = reconstruct_state.read_path {
3861 1768237 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
3862 1768237 : }
3863 1768237 : let next_cont_lsn = lsn_range.start;
3864 1768237 : layer_to_read
3865 1768237 : .get_values_reconstruct_data(
3866 1768237 : keyspace_to_read.clone(),
3867 1768237 : lsn_range,
3868 1768237 : reconstruct_state,
3869 1768237 : ctx,
3870 1768237 : )
3871 1768237 : .await?;
3872 :
3873 1768237 : unmapped_keyspace = keyspace_to_read;
3874 1768237 : cont_lsn = next_cont_lsn;
3875 1768237 :
3876 1768237 : reconstruct_state.on_layer_visited(&layer_to_read);
3877 : } else {
3878 1706984 : break;
3879 1706984 : }
3880 1706984 : }
3881 1706984 :
3882 1706984 : Ok(TimelineVisitOutcome {
3883 1706984 : completed_keyspace,
3884 1706984 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3885 1706984 : })
3886 1706984 : }
3887 :
3888 451758 : async fn get_ready_ancestor_timeline(
3889 451758 : &self,
3890 451758 : ancestor: &Arc<Timeline>,
3891 451758 : ctx: &RequestContext,
3892 451758 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3893 451758 : // It's possible that the ancestor timeline isn't active yet, or
3894 451758 : // is active but hasn't yet caught up to the branch point. Wait
3895 451758 : // for it.
3896 451758 : //
3897 451758 : // This cannot happen while the pageserver is running normally,
3898 451758 : // because you cannot create a branch from a point that isn't
3899 451758 : // present in the pageserver yet. However, we don't wait for the
3900 451758 : // branch point to be uploaded to cloud storage before creating
3901 451758 : // a branch. I.e., the branch LSN need not be remote consistent
3902 451758 : // for the branching operation to succeed.
3903 451758 : //
3904 451758 : // Hence, if we try to load a tenant in such a state where
3905 451758 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3906 451758 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3907 451758 : // then we will need to wait for the ancestor timeline to
3908 451758 : // re-stream WAL up to branch_lsn before we access it.
3909 451758 : //
3910 451758 : // How can a tenant get in such a state?
3911 451758 : // - ungraceful pageserver process exit
3912 451758 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3913 451758 : //
3914 451758 : // NB: this could be avoided by requiring
3915 451758 : // branch_lsn >= remote_consistent_lsn
3916 451758 : // during branch creation.
3917 451758 : match ancestor.wait_to_become_active(ctx).await {
3918 451754 : Ok(()) => {}
3919 : Err(TimelineState::Stopping) => {
3920 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3921 0 : return Err(GetReadyAncestorError::Cancelled);
3922 : }
3923 4 : Err(state) => {
3924 4 : return Err(GetReadyAncestorError::BadState {
3925 4 : timeline_id: ancestor.timeline_id,
3926 4 : state,
3927 4 : });
3928 : }
3929 : }
3930 451754 : ancestor
3931 451754 : .wait_lsn(
3932 451754 : self.ancestor_lsn,
3933 451754 : WaitLsnWaiter::Timeline(self),
3934 451754 : WaitLsnTimeout::Default,
3935 451754 : ctx,
3936 451754 : )
3937 451754 : .await
3938 451754 : .map_err(|e| match e {
3939 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3940 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3941 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3942 0 : timeline_id: ancestor.timeline_id,
3943 0 : state,
3944 0 : },
3945 451754 : })?;
3946 :
3947 451754 : Ok(ancestor.clone())
3948 451758 : }
3949 :
3950 594312 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3951 594312 : &self.shard_identity
3952 594312 : }
3953 :
3954 : #[inline(always)]
3955 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3956 0 : ShardTimelineId {
3957 0 : shard_index: ShardIndex {
3958 0 : shard_number: self.shard_identity.number,
3959 0 : shard_count: self.shard_identity.count,
3960 0 : },
3961 0 : timeline_id: self.timeline_id,
3962 0 : }
3963 0 : }
3964 :
3965 : /// Returns a non-frozen open in-memory layer for ingestion.
3966 : ///
3967 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3968 : /// this function without holding the mutex.
3969 2596 : async fn get_layer_for_write(
3970 2596 : &self,
3971 2596 : lsn: Lsn,
3972 2596 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3973 2596 : ctx: &RequestContext,
3974 2596 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3975 2596 : let mut guard = self.layers.write().await;
3976 :
3977 2596 : let last_record_lsn = self.get_last_record_lsn();
3978 2596 : ensure!(
3979 2596 : lsn > last_record_lsn,
3980 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3981 : lsn,
3982 : last_record_lsn,
3983 : );
3984 :
3985 2596 : let layer = guard
3986 2596 : .open_mut()?
3987 2596 : .get_layer_for_write(
3988 2596 : lsn,
3989 2596 : self.conf,
3990 2596 : self.timeline_id,
3991 2596 : self.tenant_shard_id,
3992 2596 : &self.gate,
3993 2596 : ctx,
3994 2596 : )
3995 2596 : .await?;
3996 2596 : Ok(layer)
3997 2596 : }
3998 :
3999 10558188 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4000 10558188 : assert!(new_lsn.is_aligned());
4001 :
4002 10558188 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
4003 10558188 : self.last_record_lsn.advance(new_lsn);
4004 10558188 : }
4005 :
4006 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
4007 : ///
4008 : /// Unconditional flush loop notification is given because in sharded cases we will want to
4009 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
4010 2404 : async fn freeze_inmem_layer_at(
4011 2404 : &self,
4012 2404 : at: Lsn,
4013 2404 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4014 2404 : ) -> Result<u64, FlushLayerError> {
4015 2404 : let frozen = {
4016 2404 : let mut guard = self.layers.write().await;
4017 2404 : guard
4018 2404 : .open_mut()?
4019 2404 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
4020 2404 : .await
4021 : };
4022 :
4023 2404 : if frozen {
4024 2348 : let now = Instant::now();
4025 2348 : *(self.last_freeze_ts.write().unwrap()) = now;
4026 2348 : }
4027 :
4028 : // Increment the flush cycle counter and wake up the flush task.
4029 : // Remember the new value, so that when we listen for the flush
4030 : // to finish, we know when the flush that we initiated has
4031 : // finished, instead of some other flush that was started earlier.
4032 2404 : let mut my_flush_request = 0;
4033 2404 :
4034 2404 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
4035 2404 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
4036 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
4037 2404 : }
4038 2404 :
4039 2404 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
4040 2404 : my_flush_request = *counter + 1;
4041 2404 : *counter = my_flush_request;
4042 2404 : *lsn = std::cmp::max(at, *lsn);
4043 2404 : });
4044 2404 :
4045 2404 : assert_ne!(my_flush_request, 0);
4046 :
4047 2404 : Ok(my_flush_request)
4048 2404 : }
4049 :
4050 : /// Layer flusher task's main loop.
4051 884 : async fn flush_loop(
4052 884 : self: &Arc<Self>,
4053 884 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4054 884 : ctx: &RequestContext,
4055 884 : ) {
4056 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4057 884 : let mut watch_l0 = match self.layers.read().await.layer_map() {
4058 884 : Ok(lm) => lm.watch_level0_deltas(),
4059 0 : Err(Shutdown) => return,
4060 : };
4061 :
4062 884 : info!("started flush loop");
4063 : loop {
4064 3215 : tokio::select! {
4065 3215 : _ = self.cancel.cancelled() => {
4066 20 : info!("shutting down layer flush task due to Timeline::cancel");
4067 20 : break;
4068 : },
4069 3215 : _ = layer_flush_start_rx.changed() => {}
4070 2331 : }
4071 2331 : trace!("waking up");
4072 2331 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4073 2331 :
4074 2331 : // The highest LSN to which we flushed in the loop over frozen layers
4075 2331 : let mut flushed_to_lsn = Lsn(0);
4076 :
4077 2331 : let result = loop {
4078 4679 : if self.cancel.is_cancelled() {
4079 0 : info!("dropping out of flush loop for timeline shutdown");
4080 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
4081 : // anyone waiting on that will respect self.cancel as well: they will stop
4082 : // waiting at the same time we as drop out of this loop.
4083 0 : return;
4084 4679 : }
4085 4679 :
4086 4679 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4087 4679 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4088 4679 : if flushed_to_lsn >= frozen_to_lsn {
4089 2275 : break Ok(());
4090 2404 : }
4091 :
4092 : // Fetch the next layer to flush, if any.
4093 2404 : let (layer, l0_count, frozen_count, frozen_size) = {
4094 2404 : let layers = self.layers.read().await;
4095 2404 : let Ok(lm) = layers.layer_map() else {
4096 0 : info!("dropping out of flush loop for timeline shutdown");
4097 0 : return;
4098 : };
4099 2404 : let l0_count = lm.level0_deltas().len();
4100 2404 : let frozen_count = lm.frozen_layers.len();
4101 2404 : let frozen_size: u64 = lm
4102 2404 : .frozen_layers
4103 2404 : .iter()
4104 2422 : .map(|l| l.estimated_in_mem_size())
4105 2404 : .sum();
4106 2404 : let layer = lm.frozen_layers.front().cloned();
4107 2404 : (layer, l0_count, frozen_count, frozen_size)
4108 2404 : // drop 'layers' lock
4109 2404 : };
4110 2404 : let Some(layer) = layer else {
4111 56 : break Ok(());
4112 : };
4113 :
4114 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
4115 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
4116 : //
4117 : // NB: the compaction loop only checks `compaction_threshold` every 20 seconds, so
4118 : // we can end up stalling before compaction even starts. Consider making it more
4119 : // responsive (e.g. via `watch_level0_deltas`).
4120 2348 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
4121 0 : if l0_count >= stall_threshold {
4122 0 : warn!(
4123 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
4124 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4125 : );
4126 0 : let stall_timer = self
4127 0 : .metrics
4128 0 : .flush_delay_histo
4129 0 : .start_timer()
4130 0 : .record_on_drop();
4131 0 : tokio::select! {
4132 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
4133 0 : if let Ok(l0) = result.as_deref() {
4134 0 : let delay = stall_timer.elapsed().as_secs_f64();
4135 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
4136 0 : }
4137 : },
4138 0 : _ = self.cancel.cancelled() => {},
4139 : }
4140 0 : continue; // check again
4141 0 : }
4142 2348 : }
4143 :
4144 : // Flush the layer.
4145 2348 : let flush_timer = self.metrics.flush_time_histo.start_timer();
4146 2348 : match self.flush_frozen_layer(layer, ctx).await {
4147 2348 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
4148 : Err(FlushLayerError::Cancelled) => {
4149 0 : info!("dropping out of flush loop for timeline shutdown");
4150 0 : return;
4151 : }
4152 0 : err @ Err(
4153 0 : FlushLayerError::NotRunning(_)
4154 0 : | FlushLayerError::Other(_)
4155 0 : | FlushLayerError::CreateImageLayersError(_),
4156 0 : ) => {
4157 0 : error!("could not flush frozen layer: {err:?}");
4158 0 : break err.map(|_| ());
4159 : }
4160 : }
4161 2348 : let flush_duration = flush_timer.stop_and_record();
4162 2348 :
4163 2348 : // Notify the tenant compaction loop if L0 compaction is needed.
4164 2348 : let l0_count = *watch_l0.borrow();
4165 2348 : if l0_count >= self.get_compaction_threshold() {
4166 952 : self.l0_compaction_trigger.notify_one();
4167 1396 : }
4168 :
4169 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
4170 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
4171 : // ingestion by having ephemeral layer rolls wait for flushes.
4172 2348 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
4173 0 : if l0_count >= delay_threshold {
4174 0 : let delay = flush_duration.as_secs_f64();
4175 0 : info!(
4176 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
4177 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4178 : );
4179 0 : let _delay_timer = self
4180 0 : .metrics
4181 0 : .flush_delay_histo
4182 0 : .start_timer()
4183 0 : .record_on_drop();
4184 0 : tokio::select! {
4185 0 : _ = tokio::time::sleep(flush_duration) => {},
4186 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
4187 0 : _ = self.cancel.cancelled() => {},
4188 : }
4189 0 : }
4190 2348 : }
4191 : };
4192 :
4193 : // Unsharded tenants should never advance their LSN beyond the end of the
4194 : // highest layer they write: such gaps between layer data and the frozen LSN
4195 : // are only legal on sharded tenants.
4196 2331 : debug_assert!(
4197 2331 : self.shard_identity.count.count() > 1
4198 2331 : || flushed_to_lsn >= frozen_to_lsn
4199 56 : || !flushed_to_lsn.is_valid()
4200 : );
4201 :
4202 2331 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
4203 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
4204 : // to us via layer_flush_start_rx, then advance it here.
4205 : //
4206 : // This path is only taken for tenants with multiple shards: single sharded tenants should
4207 : // never encounter a gap in the wal.
4208 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
4209 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
4210 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
4211 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
4212 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
4213 0 : }
4214 0 : }
4215 2331 : }
4216 :
4217 : // Notify any listeners that we're done
4218 2331 : let _ = self
4219 2331 : .layer_flush_done_tx
4220 2331 : .send_replace((flush_counter, result));
4221 : }
4222 20 : }
4223 :
4224 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
4225 2244 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
4226 2244 : let mut rx = self.layer_flush_done_tx.subscribe();
4227 : loop {
4228 : {
4229 4559 : let (last_result_counter, last_result) = &*rx.borrow();
4230 4559 : if *last_result_counter >= request {
4231 2244 : if let Err(err) = last_result {
4232 : // We already logged the original error in
4233 : // flush_loop. We cannot propagate it to the caller
4234 : // here, because it might not be Cloneable
4235 0 : return Err(err.clone());
4236 : } else {
4237 2244 : return Ok(());
4238 : }
4239 2315 : }
4240 2315 : }
4241 2315 : trace!("waiting for flush to complete");
4242 2315 : tokio::select! {
4243 2315 : rx_e = rx.changed() => {
4244 2315 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
4245 : },
4246 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
4247 : // the notification from [`flush_loop`] that it completed.
4248 2315 : _ = self.cancel.cancelled() => {
4249 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
4250 0 : return Ok(())
4251 : }
4252 : };
4253 2315 : trace!("done")
4254 : }
4255 2244 : }
4256 :
4257 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
4258 : ///
4259 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
4260 : #[instrument(skip_all, fields(layer=%frozen_layer))]
4261 : async fn flush_frozen_layer(
4262 : self: &Arc<Self>,
4263 : frozen_layer: Arc<InMemoryLayer>,
4264 : ctx: &RequestContext,
4265 : ) -> Result<Lsn, FlushLayerError> {
4266 : debug_assert_current_span_has_tenant_and_timeline_id();
4267 :
4268 : // As a special case, when we have just imported an image into the repository,
4269 : // instead of writing out a L0 delta layer, we directly write out image layer
4270 : // files instead. This is possible as long as *all* the data imported into the
4271 : // repository have the same LSN.
4272 : let lsn_range = frozen_layer.get_lsn_range();
4273 :
4274 : // Whether to directly create image layers for this flush, or flush them as delta layers
4275 : let create_image_layer =
4276 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
4277 :
4278 : #[cfg(test)]
4279 : {
4280 : match &mut *self.flush_loop_state.lock().unwrap() {
4281 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
4282 : panic!("flush loop not running")
4283 : }
4284 : FlushLoopState::Running {
4285 : expect_initdb_optimization,
4286 : initdb_optimization_count,
4287 : ..
4288 : } => {
4289 : if create_image_layer {
4290 : *initdb_optimization_count += 1;
4291 : } else {
4292 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
4293 : }
4294 : }
4295 : }
4296 : }
4297 :
4298 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
4299 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
4300 : // require downloading anything during initial import.
4301 : let ((rel_partition, metadata_partition), _lsn) = self
4302 : .repartition(
4303 : self.initdb_lsn,
4304 : self.get_compaction_target_size(),
4305 : EnumSet::empty(),
4306 : ctx,
4307 : )
4308 : .await
4309 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
4310 :
4311 : if self.cancel.is_cancelled() {
4312 : return Err(FlushLayerError::Cancelled);
4313 : }
4314 :
4315 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
4316 : // So that the key ranges don't overlap.
4317 : let mut partitions = KeyPartitioning::default();
4318 : partitions.parts.extend(rel_partition.parts);
4319 : if !metadata_partition.parts.is_empty() {
4320 : assert_eq!(
4321 : metadata_partition.parts.len(),
4322 : 1,
4323 : "currently sparse keyspace should only contain a single metadata keyspace"
4324 : );
4325 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
4326 : // every single key within the keyspace, and therefore, it's safe to force converting it
4327 : // into a dense keyspace before calling this function.
4328 : partitions
4329 : .parts
4330 : .extend(metadata_partition.into_dense().parts);
4331 : }
4332 :
4333 : let mut layers_to_upload = Vec::new();
4334 : let (generated_image_layers, is_complete) = self
4335 : .create_image_layers(
4336 : &partitions,
4337 : self.initdb_lsn,
4338 : ImageLayerCreationMode::Initial,
4339 : ctx,
4340 : LastImageLayerCreationStatus::Initial,
4341 : false, // don't yield for L0, we're flushing L0
4342 : )
4343 : .await?;
4344 : debug_assert!(
4345 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
4346 : "init image generation mode must fully cover the keyspace"
4347 : );
4348 : layers_to_upload.extend(generated_image_layers);
4349 :
4350 : (layers_to_upload, None)
4351 : } else {
4352 : // Normal case, write out a L0 delta layer file.
4353 : // `create_delta_layer` will not modify the layer map.
4354 : // We will remove frozen layer and add delta layer in one atomic operation later.
4355 : let Some(layer) = self
4356 : .create_delta_layer(&frozen_layer, None, ctx)
4357 : .await
4358 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
4359 : else {
4360 : panic!("delta layer cannot be empty if no filter is applied");
4361 : };
4362 : (
4363 : // FIXME: even though we have a single image and single delta layer assumption
4364 : // we push them to vec
4365 : vec![layer.clone()],
4366 : Some(layer),
4367 : )
4368 : };
4369 :
4370 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
4371 :
4372 : if self.cancel.is_cancelled() {
4373 : return Err(FlushLayerError::Cancelled);
4374 : }
4375 :
4376 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
4377 :
4378 : // The new on-disk layers are now in the layer map. We can remove the
4379 : // in-memory layer from the map now. The flushed layer is stored in
4380 : // the mapping in `create_delta_layer`.
4381 : {
4382 : let mut guard = self.layers.write().await;
4383 :
4384 : guard.open_mut()?.finish_flush_l0_layer(
4385 : delta_layer_to_add.as_ref(),
4386 : &frozen_layer,
4387 : &self.metrics,
4388 : );
4389 :
4390 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
4391 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
4392 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
4393 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
4394 : }
4395 : // release lock on 'layers'
4396 : };
4397 :
4398 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
4399 : // This makes us refuse ingest until the new layers have been persisted to the remote
4400 : // TODO: remove this, and rely on l0_flush_{delay,stall}_threshold instead.
4401 : if self.get_l0_flush_wait_upload() {
4402 : let start = Instant::now();
4403 : self.remote_client
4404 : .wait_completion()
4405 : .await
4406 0 : .map_err(|e| match e {
4407 : WaitCompletionError::UploadQueueShutDownOrStopped
4408 : | WaitCompletionError::NotInitialized(
4409 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
4410 0 : ) => FlushLayerError::Cancelled,
4411 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
4412 0 : FlushLayerError::Other(anyhow!(e).into())
4413 : }
4414 0 : })?;
4415 : let duration = start.elapsed().as_secs_f64();
4416 : self.metrics.flush_wait_upload_time_gauge_add(duration);
4417 : }
4418 :
4419 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
4420 : // a compaction can delete the file and then it won't be available for uploads any more.
4421 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
4422 : // race situation.
4423 : // See https://github.com/neondatabase/neon/issues/4526
4424 : pausable_failpoint!("flush-frozen-pausable");
4425 :
4426 : // This failpoint is used by another test case `test_pageserver_recovery`.
4427 : fail_point!("flush-frozen-exit");
4428 :
4429 : Ok(Lsn(lsn_range.end.0 - 1))
4430 : }
4431 :
4432 : /// Return true if the value changed
4433 : ///
4434 : /// This function must only be used from the layer flush task.
4435 2348 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
4436 2348 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
4437 2348 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
4438 :
4439 2348 : self.metrics
4440 2348 : .disk_consistent_lsn_gauge
4441 2348 : .set(new_value.0 as i64);
4442 2348 : new_value != old_value
4443 2348 : }
4444 :
4445 : /// Update metadata file
4446 2448 : fn schedule_uploads(
4447 2448 : &self,
4448 2448 : disk_consistent_lsn: Lsn,
4449 2448 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
4450 2448 : ) -> anyhow::Result<()> {
4451 2448 : // We can only save a valid 'prev_record_lsn' value on disk if we
4452 2448 : // flushed *all* in-memory changes to disk. We only track
4453 2448 : // 'prev_record_lsn' in memory for the latest processed record, so we
4454 2448 : // don't remember what the correct value that corresponds to some old
4455 2448 : // LSN is. But if we flush everything, then the value corresponding
4456 2448 : // current 'last_record_lsn' is correct and we can store it on disk.
4457 2448 : let RecordLsn {
4458 2448 : last: last_record_lsn,
4459 2448 : prev: prev_record_lsn,
4460 2448 : } = self.last_record_lsn.load();
4461 2448 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
4462 2193 : Some(prev_record_lsn)
4463 : } else {
4464 255 : None
4465 : };
4466 :
4467 2448 : let update = crate::tenant::metadata::MetadataUpdate::new(
4468 2448 : disk_consistent_lsn,
4469 2448 : ondisk_prev_record_lsn,
4470 2448 : *self.applied_gc_cutoff_lsn.read(),
4471 2448 : );
4472 2448 :
4473 2448 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
4474 0 : "{}",
4475 0 : x.unwrap()
4476 2448 : ));
4477 :
4478 4820 : for layer in layers_to_upload {
4479 2372 : self.remote_client.schedule_layer_file_upload(layer)?;
4480 : }
4481 2448 : self.remote_client
4482 2448 : .schedule_index_upload_for_metadata_update(&update)?;
4483 :
4484 2448 : Ok(())
4485 2448 : }
4486 :
4487 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
4488 0 : self.remote_client
4489 0 : .preserve_initdb_archive(
4490 0 : &self.tenant_shard_id.tenant_id,
4491 0 : &self.timeline_id,
4492 0 : &self.cancel,
4493 0 : )
4494 0 : .await
4495 0 : }
4496 :
4497 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
4498 : // in layer map immediately. The caller is responsible to put it into the layer map.
4499 1936 : async fn create_delta_layer(
4500 1936 : self: &Arc<Self>,
4501 1936 : frozen_layer: &Arc<InMemoryLayer>,
4502 1936 : key_range: Option<Range<Key>>,
4503 1936 : ctx: &RequestContext,
4504 1936 : ) -> anyhow::Result<Option<ResidentLayer>> {
4505 1936 : let self_clone = Arc::clone(self);
4506 1936 : let frozen_layer = Arc::clone(frozen_layer);
4507 1936 : let ctx = ctx.attached_child();
4508 1936 : let work = async move {
4509 1936 : let Some((desc, path)) = frozen_layer
4510 1936 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
4511 1936 : .await?
4512 : else {
4513 0 : return Ok(None);
4514 : };
4515 1936 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
4516 :
4517 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
4518 : // We just need to fsync the directory in which these inodes are linked,
4519 : // which we know to be the timeline directory.
4520 : //
4521 : // We use fatal_err() below because the after write_to_disk returns with success,
4522 : // the in-memory state of the filesystem already has the layer file in its final place,
4523 : // and subsequent pageserver code could think it's durable while it really isn't.
4524 1936 : let timeline_dir = VirtualFile::open(
4525 1936 : &self_clone
4526 1936 : .conf
4527 1936 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
4528 1936 : &ctx,
4529 1936 : )
4530 1936 : .await
4531 1936 : .fatal_err("VirtualFile::open for timeline dir fsync");
4532 1936 : timeline_dir
4533 1936 : .sync_all()
4534 1936 : .await
4535 1936 : .fatal_err("VirtualFile::sync_all timeline dir");
4536 1936 : anyhow::Ok(Some(new_delta))
4537 1936 : };
4538 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
4539 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
4540 : use crate::virtual_file::io_engine::IoEngine;
4541 1936 : match crate::virtual_file::io_engine::get() {
4542 0 : IoEngine::NotSet => panic!("io engine not set"),
4543 : IoEngine::StdFs => {
4544 968 : let span = tracing::info_span!("blocking");
4545 968 : tokio::task::spawn_blocking({
4546 968 : move || Handle::current().block_on(work.instrument(span))
4547 968 : })
4548 968 : .await
4549 968 : .context("spawn_blocking")
4550 968 : .and_then(|x| x)
4551 : }
4552 : #[cfg(target_os = "linux")]
4553 968 : IoEngine::TokioEpollUring => work.await,
4554 : }
4555 1936 : }
4556 :
4557 1180 : async fn repartition(
4558 1180 : &self,
4559 1180 : lsn: Lsn,
4560 1180 : partition_size: u64,
4561 1180 : flags: EnumSet<CompactFlags>,
4562 1180 : ctx: &RequestContext,
4563 1180 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
4564 1180 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
4565 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
4566 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
4567 : // and hence before the compaction task starts.
4568 0 : return Err(CompactionError::Other(anyhow!(
4569 0 : "repartition() called concurrently"
4570 0 : )));
4571 : };
4572 1180 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
4573 1180 : if lsn < *partition_lsn {
4574 0 : return Err(CompactionError::Other(anyhow!(
4575 0 : "repartition() called with LSN going backwards, this should not happen"
4576 0 : )));
4577 1180 : }
4578 1180 :
4579 1180 : let distance = lsn.0 - partition_lsn.0;
4580 1180 : if *partition_lsn != Lsn(0)
4581 564 : && distance <= self.repartition_threshold
4582 564 : && !flags.contains(CompactFlags::ForceRepartition)
4583 : {
4584 536 : debug!(
4585 : distance,
4586 : threshold = self.repartition_threshold,
4587 0 : "no repartitioning needed"
4588 : );
4589 536 : return Ok((
4590 536 : (dense_partition.clone(), sparse_partition.clone()),
4591 536 : *partition_lsn,
4592 536 : ));
4593 644 : }
4594 :
4595 644 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4596 644 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4597 644 : let sparse_partitioning = SparseKeyPartitioning {
4598 644 : parts: vec![sparse_ks],
4599 644 : }; // no partitioning for metadata keys for now
4600 644 : let result = ((dense_partitioning, sparse_partitioning), lsn);
4601 644 : guard.write(result.clone());
4602 644 : Ok(result)
4603 1180 : }
4604 :
4605 : // Is it time to create a new image layer for the given partition? True if we want to generate.
4606 0 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4607 0 : let threshold = self.get_image_creation_threshold();
4608 :
4609 0 : let guard = self.layers.read().await;
4610 0 : let Ok(layers) = guard.layer_map() else {
4611 0 : return false;
4612 : };
4613 :
4614 0 : let mut max_deltas = 0;
4615 0 : for part_range in &partition.ranges {
4616 0 : let image_coverage = layers.image_coverage(part_range, lsn);
4617 0 : for (img_range, last_img) in image_coverage {
4618 0 : let img_lsn = if let Some(last_img) = last_img {
4619 0 : last_img.get_lsn_range().end
4620 : } else {
4621 0 : Lsn(0)
4622 : };
4623 : // Let's consider an example:
4624 : //
4625 : // delta layer with LSN range 71-81
4626 : // delta layer with LSN range 81-91
4627 : // delta layer with LSN range 91-101
4628 : // image layer at LSN 100
4629 : //
4630 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4631 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4632 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4633 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4634 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4635 0 : if img_lsn < lsn {
4636 0 : let num_deltas =
4637 0 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4638 0 :
4639 0 : max_deltas = max_deltas.max(num_deltas);
4640 0 : if num_deltas >= threshold {
4641 0 : debug!(
4642 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4643 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4644 : );
4645 0 : return true;
4646 0 : }
4647 0 : }
4648 : }
4649 : }
4650 :
4651 0 : debug!(
4652 : max_deltas,
4653 0 : "none of the partitioned ranges had >= {threshold} deltas"
4654 : );
4655 0 : false
4656 0 : }
4657 :
4658 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4659 : /// so that at most one image layer will be produced from this function.
4660 : #[allow(clippy::too_many_arguments)]
4661 456 : async fn create_image_layer_for_rel_blocks(
4662 456 : self: &Arc<Self>,
4663 456 : partition: &KeySpace,
4664 456 : mut image_layer_writer: ImageLayerWriter,
4665 456 : lsn: Lsn,
4666 456 : ctx: &RequestContext,
4667 456 : img_range: Range<Key>,
4668 456 : io_concurrency: IoConcurrency,
4669 456 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4670 456 : let mut wrote_keys = false;
4671 456 :
4672 456 : let mut key_request_accum = KeySpaceAccum::new();
4673 3016 : for range in &partition.ranges {
4674 2560 : let mut key = range.start;
4675 5548 : while key < range.end {
4676 : // Decide whether to retain this key: usually we do, but sharded tenants may
4677 : // need to drop keys that don't belong to them. If we retain the key, add it
4678 : // to `key_request_accum` for later issuing a vectored get
4679 2988 : if self.shard_identity.is_key_disposable(&key) {
4680 0 : debug!(
4681 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4682 0 : key,
4683 0 : self.shard_identity.get_shard_number(&key)
4684 : );
4685 2988 : } else {
4686 2988 : key_request_accum.add_key(key);
4687 2988 : }
4688 :
4689 2988 : let last_key_in_range = key.next() == range.end;
4690 2988 : key = key.next();
4691 2988 :
4692 2988 : // Maybe flush `key_rest_accum`
4693 2988 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4694 2988 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4695 : {
4696 2560 : let results = self
4697 2560 : .get_vectored(
4698 2560 : key_request_accum.consume_keyspace(),
4699 2560 : lsn,
4700 2560 : io_concurrency.clone(),
4701 2560 : ctx,
4702 2560 : )
4703 2560 : .await?;
4704 :
4705 2560 : if self.cancel.is_cancelled() {
4706 0 : return Err(CreateImageLayersError::Cancelled);
4707 2560 : }
4708 :
4709 5548 : for (img_key, img) in results {
4710 2988 : let img = match img {
4711 2988 : Ok(img) => img,
4712 0 : Err(err) => {
4713 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4714 0 : // page without losing any actual user data. That seems better
4715 0 : // than failing repeatedly and getting stuck.
4716 0 : //
4717 0 : // We had a bug at one point, where we truncated the FSM and VM
4718 0 : // in the pageserver, but the Postgres didn't know about that
4719 0 : // and continued to generate incremental WAL records for pages
4720 0 : // that didn't exist in the pageserver. Trying to replay those
4721 0 : // WAL records failed to find the previous image of the page.
4722 0 : // This special case allows us to recover from that situation.
4723 0 : // See https://github.com/neondatabase/neon/issues/2601.
4724 0 : //
4725 0 : // Unfortunately we cannot do this for the main fork, or for
4726 0 : // any metadata keys, keys, as that would lead to actual data
4727 0 : // loss.
4728 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4729 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4730 0 : ZERO_PAGE.clone()
4731 : } else {
4732 0 : return Err(CreateImageLayersError::from(err));
4733 : }
4734 : }
4735 : };
4736 :
4737 : // Write all the keys we just read into our new image layer.
4738 2988 : image_layer_writer.put_image(img_key, img, ctx).await?;
4739 2988 : wrote_keys = true;
4740 : }
4741 428 : }
4742 : }
4743 : }
4744 :
4745 456 : if wrote_keys {
4746 : // Normal path: we have written some data into the new image layer for this
4747 : // partition, so flush it to disk.
4748 456 : info!(
4749 0 : "produced image layer for rel {}",
4750 0 : ImageLayerName {
4751 0 : key_range: img_range.clone(),
4752 0 : lsn
4753 0 : },
4754 : );
4755 456 : Ok(ImageLayerCreationOutcome::Generated {
4756 456 : unfinished_image_layer: image_layer_writer,
4757 456 : })
4758 : } else {
4759 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4760 0 : Ok(ImageLayerCreationOutcome::Empty)
4761 : }
4762 456 : }
4763 :
4764 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4765 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4766 : /// would not be too large to fit in a single image layer.
4767 : ///
4768 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
4769 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
4770 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
4771 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
4772 : #[allow(clippy::too_many_arguments)]
4773 440 : async fn create_image_layer_for_metadata_keys(
4774 440 : self: &Arc<Self>,
4775 440 : partition: &KeySpace,
4776 440 : mut image_layer_writer: ImageLayerWriter,
4777 440 : lsn: Lsn,
4778 440 : ctx: &RequestContext,
4779 440 : img_range: Range<Key>,
4780 440 : mode: ImageLayerCreationMode,
4781 440 : io_concurrency: IoConcurrency,
4782 440 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4783 440 : // Metadata keys image layer creation.
4784 440 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
4785 440 : let begin = Instant::now();
4786 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
4787 : // not contain too many keys, otherwise this takes a lot of memory.
4788 440 : let data = self
4789 440 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4790 440 : .await?;
4791 440 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4792 440 : let mut new_data = BTreeMap::new();
4793 440 : let mut total_kb_retrieved = 0;
4794 440 : let mut total_keys_retrieved = 0;
4795 12464 : for (k, v) in data {
4796 12024 : let v = v?;
4797 12024 : total_kb_retrieved += KEY_SIZE + v.len();
4798 12024 : total_keys_retrieved += 1;
4799 12024 : new_data.insert(k, v);
4800 : }
4801 440 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4802 440 : };
4803 440 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4804 440 : let elapsed = begin.elapsed();
4805 440 :
4806 440 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4807 440 : info!(
4808 0 : "metadata key compaction: trigger_generation={trigger_generation}, delta_files_accessed={delta_files_accessed}, total_kb_retrieved={total_kb_retrieved}, total_keys_retrieved={total_keys_retrieved}, read_time={}s", elapsed.as_secs_f64()
4809 : );
4810 :
4811 440 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4812 0 : return Ok(ImageLayerCreationOutcome::Skip);
4813 440 : }
4814 440 : if self.cancel.is_cancelled() {
4815 0 : return Err(CreateImageLayersError::Cancelled);
4816 440 : }
4817 440 : let mut wrote_any_image = false;
4818 12464 : for (k, v) in data {
4819 12024 : if v.is_empty() {
4820 : // the key has been deleted, it does not need an image
4821 : // in metadata keyspace, an empty image == tombstone
4822 8 : continue;
4823 12016 : }
4824 12016 : wrote_any_image = true;
4825 12016 :
4826 12016 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4827 12016 :
4828 12016 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4829 12016 : // on the normal data path either.
4830 12016 : image_layer_writer.put_image(k, v, ctx).await?;
4831 : }
4832 :
4833 440 : if wrote_any_image {
4834 : // Normal path: we have written some data into the new image layer for this
4835 : // partition, so flush it to disk.
4836 20 : info!(
4837 0 : "created image layer for metadata {}",
4838 0 : ImageLayerName {
4839 0 : key_range: img_range.clone(),
4840 0 : lsn
4841 0 : }
4842 : );
4843 20 : Ok(ImageLayerCreationOutcome::Generated {
4844 20 : unfinished_image_layer: image_layer_writer,
4845 20 : })
4846 : } else {
4847 420 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4848 420 : Ok(ImageLayerCreationOutcome::Empty)
4849 : }
4850 440 : }
4851 :
4852 : /// Predicate function which indicates whether we should check if new image layers
4853 : /// are required. Since checking if new image layers are required is expensive in
4854 : /// terms of CPU, we only do it in the following cases:
4855 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4856 : /// 2. If enough time has passed since the last check:
4857 : /// 1. For large tenants, we wish to perform the check more often since they
4858 : /// suffer from the lack of image layers
4859 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4860 1180 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4861 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4862 :
4863 1180 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4864 1180 : let distance = lsn
4865 1180 : .checked_sub(last_checks_at)
4866 1180 : .expect("Attempt to compact with LSN going backwards");
4867 1180 : let min_distance =
4868 1180 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4869 1180 :
4870 1180 : let distance_based_decision = distance.0 >= min_distance;
4871 1180 :
4872 1180 : let mut time_based_decision = false;
4873 1180 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4874 1180 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4875 976 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4876 : {
4877 0 : self.get_checkpoint_timeout()
4878 : } else {
4879 976 : Duration::from_secs(3600 * 48)
4880 : };
4881 :
4882 976 : time_based_decision = match *last_check_instant {
4883 564 : Some(last_check) => {
4884 564 : let elapsed = last_check.elapsed();
4885 564 : elapsed >= check_required_after
4886 : }
4887 412 : None => true,
4888 : };
4889 204 : }
4890 :
4891 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4892 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4893 : // check.
4894 1180 : let decision = distance_based_decision || time_based_decision;
4895 :
4896 1180 : if decision {
4897 412 : self.last_image_layer_creation_check_at.store(lsn);
4898 412 : *last_check_instant = Some(Instant::now());
4899 768 : }
4900 :
4901 1180 : decision
4902 1180 : }
4903 :
4904 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
4905 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
4906 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4907 : async fn create_image_layers(
4908 : self: &Arc<Timeline>,
4909 : partitioning: &KeyPartitioning,
4910 : lsn: Lsn,
4911 : mode: ImageLayerCreationMode,
4912 : ctx: &RequestContext,
4913 : last_status: LastImageLayerCreationStatus,
4914 : yield_for_l0: bool,
4915 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
4916 : let timer = self.metrics.create_images_time_histo.start_timer();
4917 :
4918 : if partitioning.parts.is_empty() {
4919 : warn!("no partitions to create image layers for");
4920 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
4921 : }
4922 :
4923 : // We need to avoid holes between generated image layers.
4924 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4925 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4926 : //
4927 : // How such hole between partitions can appear?
4928 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4929 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4930 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4931 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4932 : let mut start = Key::MIN;
4933 :
4934 : let check_for_image_layers =
4935 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
4936 : info!(
4937 : "resuming image layer creation: last_status=incomplete, continue from {}",
4938 : last_key
4939 : );
4940 : true
4941 : } else {
4942 : self.should_check_if_image_layers_required(lsn)
4943 : };
4944 :
4945 : let mut batch_image_writer = BatchLayerWriter::new(self.conf).await?;
4946 :
4947 : let mut all_generated = true;
4948 :
4949 : let mut partition_processed = 0;
4950 : let mut total_partitions = partitioning.parts.len();
4951 : let mut last_partition_processed = None;
4952 : let mut partition_parts = partitioning.parts.clone();
4953 :
4954 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
4955 : // We need to skip the partitions that have already been processed.
4956 : let mut found = false;
4957 : for (i, partition) in partition_parts.iter().enumerate() {
4958 : if last_key <= partition.end().unwrap() {
4959 : // ```plain
4960 : // |------|--------|----------|------|
4961 : // ^last_key
4962 : // ^start from this partition
4963 : // ```
4964 : // Why `i+1` instead of `i`?
4965 : // It is possible that the user did some writes after the previous image layer creation attempt so that
4966 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
4967 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
4968 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
4969 : if i + 1 >= total_partitions {
4970 : // In general, this case should not happen -- if last_key is on the last partition, the previous
4971 : // iteration of image layer creation should return a complete status.
4972 : break; // with found=false
4973 : }
4974 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
4975 : total_partitions = partition_parts.len();
4976 : // Update the start key to the partition start.
4977 : start = partition_parts[0].start().unwrap();
4978 : found = true;
4979 : break;
4980 : }
4981 : }
4982 : if !found {
4983 : // Last key is within the last partition, or larger than all partitions.
4984 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
4985 : }
4986 : }
4987 :
4988 : for partition in partition_parts.iter() {
4989 : if self.cancel.is_cancelled() {
4990 : return Err(CreateImageLayersError::Cancelled);
4991 : }
4992 : partition_processed += 1;
4993 : let img_range = start..partition.ranges.last().unwrap().end;
4994 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4995 : if compact_metadata {
4996 : for range in &partition.ranges {
4997 : assert!(
4998 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4999 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
5000 : "metadata keys must be partitioned separately"
5001 : );
5002 : }
5003 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
5004 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
5005 : // might mess up with evictions.
5006 : start = img_range.end;
5007 : continue;
5008 : }
5009 : // For initial and force modes, we always generate image layers for metadata keys.
5010 : } else if let ImageLayerCreationMode::Try = mode {
5011 : // check_for_image_layers = false -> skip
5012 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
5013 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
5014 : start = img_range.end;
5015 : continue;
5016 : }
5017 : }
5018 : if let ImageLayerCreationMode::Force = mode {
5019 : // When forced to create image layers, we might try and create them where they already
5020 : // exist. This mode is only used in tests/debug.
5021 : let layers = self.layers.read().await;
5022 : if layers.contains_key(&PersistentLayerKey {
5023 : key_range: img_range.clone(),
5024 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
5025 : is_delta: false,
5026 : }) {
5027 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
5028 : // in the future.
5029 : tracing::info!(
5030 : "Skipping image layer at {lsn} {}..{}, already exists",
5031 : img_range.start,
5032 : img_range.end
5033 : );
5034 : start = img_range.end;
5035 : continue;
5036 : }
5037 : }
5038 :
5039 : let image_layer_writer = ImageLayerWriter::new(
5040 : self.conf,
5041 : self.timeline_id,
5042 : self.tenant_shard_id,
5043 : &img_range,
5044 : lsn,
5045 : ctx,
5046 : )
5047 : .await?;
5048 :
5049 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
5050 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
5051 0 : "failpoint image-layer-writer-fail-before-finish"
5052 0 : )))
5053 0 : });
5054 :
5055 : let io_concurrency = IoConcurrency::spawn_from_conf(
5056 : self.conf,
5057 : self.gate
5058 : .enter()
5059 0 : .map_err(|_| CreateImageLayersError::Cancelled)?,
5060 : );
5061 :
5062 : let outcome = if !compact_metadata {
5063 : self.create_image_layer_for_rel_blocks(
5064 : partition,
5065 : image_layer_writer,
5066 : lsn,
5067 : ctx,
5068 : img_range.clone(),
5069 : io_concurrency,
5070 : )
5071 : .await?
5072 : } else {
5073 : self.create_image_layer_for_metadata_keys(
5074 : partition,
5075 : image_layer_writer,
5076 : lsn,
5077 : ctx,
5078 : img_range.clone(),
5079 : mode,
5080 : io_concurrency,
5081 : )
5082 : .await?
5083 : };
5084 : match outcome {
5085 : ImageLayerCreationOutcome::Empty => {
5086 : // No data in this partition, so we don't need to create an image layer (for now).
5087 : // The next image layer should cover this key range, so we don't advance the `start`
5088 : // key.
5089 : }
5090 : ImageLayerCreationOutcome::Generated {
5091 : unfinished_image_layer,
5092 : } => {
5093 : batch_image_writer.add_unfinished_image_writer(
5094 : unfinished_image_layer,
5095 : img_range.clone(),
5096 : lsn,
5097 : );
5098 : // The next image layer should be generated right after this one.
5099 : start = img_range.end;
5100 : }
5101 : ImageLayerCreationOutcome::Skip => {
5102 : // We don't need to create an image layer for this partition.
5103 : // The next image layer should NOT cover this range, otherwise
5104 : // the keyspace becomes empty (reads don't go past image layers).
5105 : start = img_range.end;
5106 : }
5107 : }
5108 :
5109 : if let ImageLayerCreationMode::Try = mode {
5110 : // We have at least made some progress
5111 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
5112 : // The `Try` mode is currently only used on the compaction path. We want to avoid
5113 : // image layer generation taking too long time and blocking L0 compaction. So in this
5114 : // mode, we also inspect the current number of L0 layers and skip image layer generation
5115 : // if there are too many of them.
5116 : let num_of_l0_layers = {
5117 : let layers = self.layers.read().await;
5118 : layers.layer_map()?.level0_deltas().len()
5119 : };
5120 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
5121 : * self.get_compaction_threshold();
5122 : if image_preempt_threshold != 0 && num_of_l0_layers >= image_preempt_threshold {
5123 : tracing::info!(
5124 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers {}",
5125 : partition.start().unwrap(), partition.end().unwrap(), num_of_l0_layers
5126 : );
5127 : last_partition_processed = Some(partition.clone());
5128 : all_generated = false;
5129 : break;
5130 : }
5131 : }
5132 : }
5133 : }
5134 :
5135 : let image_layers = batch_image_writer.finish(self, ctx).await?;
5136 :
5137 : let mut guard = self.layers.write().await;
5138 :
5139 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
5140 : // now they are being scheduled outside of write lock; current way is inconsistent with
5141 : // compaction lock order.
5142 : guard
5143 : .open_mut()?
5144 : .track_new_image_layers(&image_layers, &self.metrics);
5145 : drop_wlock(guard);
5146 : let duration = timer.stop_and_record();
5147 :
5148 : // Creating image layers may have caused some previously visible layers to be covered
5149 : if !image_layers.is_empty() {
5150 : self.update_layer_visibility().await?;
5151 : }
5152 :
5153 : let total_layer_size = image_layers
5154 : .iter()
5155 476 : .map(|l| l.metadata().file_size)
5156 : .sum::<u64>();
5157 :
5158 : info!(
5159 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
5160 : image_layers.len(),
5161 : total_layer_size,
5162 : duration.as_secs_f64(),
5163 : partition_processed,
5164 : total_partitions
5165 : );
5166 :
5167 : Ok((
5168 : image_layers,
5169 : if all_generated {
5170 : LastImageLayerCreationStatus::Complete
5171 : } else {
5172 : LastImageLayerCreationStatus::Incomplete {
5173 : last_key: if let Some(last_partition_processed) = last_partition_processed {
5174 : last_partition_processed.end().unwrap_or(Key::MIN)
5175 : } else {
5176 : // This branch should be unreachable, but in case it happens, we can just return the start key.
5177 : Key::MIN
5178 : },
5179 : }
5180 : },
5181 : ))
5182 : }
5183 :
5184 : /// Wait until the background initial logical size calculation is complete, or
5185 : /// this Timeline is shut down. Calling this function will cause the initial
5186 : /// logical size calculation to skip waiting for the background jobs barrier.
5187 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
5188 0 : if !self.shard_identity.is_shard_zero() {
5189 : // We don't populate logical size on shard >0: skip waiting for it.
5190 0 : return;
5191 0 : }
5192 0 :
5193 0 : if self.remote_client.is_deleting() {
5194 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
5195 0 : return;
5196 0 : }
5197 0 :
5198 0 : if self.current_logical_size.current_size().is_exact() {
5199 : // root timelines are initialized with exact count, but never start the background
5200 : // calculation
5201 0 : return;
5202 0 : }
5203 :
5204 0 : if let Some(await_bg_cancel) = self
5205 0 : .current_logical_size
5206 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
5207 0 : .get()
5208 0 : {
5209 0 : await_bg_cancel.cancel();
5210 0 : } else {
5211 : // We should not wait if we were not able to explicitly instruct
5212 : // the logical size cancellation to skip the concurrency limit semaphore.
5213 : // TODO: this is an unexpected case. We should restructure so that it
5214 : // can't happen.
5215 0 : tracing::warn!(
5216 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
5217 : );
5218 0 : debug_assert!(false);
5219 : }
5220 :
5221 0 : tokio::select!(
5222 0 : _ = self.current_logical_size.initialized.acquire() => {},
5223 0 : _ = self.cancel.cancelled() => {}
5224 : )
5225 0 : }
5226 :
5227 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
5228 : /// Timelines layers up to the ancestor_lsn.
5229 : ///
5230 : /// Requires a timeline that:
5231 : /// - has an ancestor to detach from
5232 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
5233 : /// a technical requirement
5234 : ///
5235 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
5236 : /// polled again until completion.
5237 : ///
5238 : /// During the operation all timelines sharing the data with this timeline will be reparented
5239 : /// from our ancestor to be branches of this timeline.
5240 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
5241 0 : self: &Arc<Timeline>,
5242 0 : tenant: &crate::tenant::Tenant,
5243 0 : options: detach_ancestor::Options,
5244 0 : ctx: &RequestContext,
5245 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
5246 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
5247 0 : }
5248 :
5249 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
5250 : /// reparents any reparentable children of previous ancestor.
5251 : ///
5252 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
5253 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
5254 : /// successfully, tenant must be reloaded.
5255 : ///
5256 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
5257 : /// resetting the tenant.
5258 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
5259 0 : self: &Arc<Timeline>,
5260 0 : tenant: &crate::tenant::Tenant,
5261 0 : prepared: detach_ancestor::PreparedTimelineDetach,
5262 0 : ctx: &RequestContext,
5263 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
5264 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
5265 0 : }
5266 :
5267 : /// Final step which unblocks the GC.
5268 : ///
5269 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
5270 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
5271 0 : self: &Arc<Timeline>,
5272 0 : tenant: &crate::tenant::Tenant,
5273 0 : attempt: detach_ancestor::Attempt,
5274 0 : ctx: &RequestContext,
5275 0 : ) -> Result<(), detach_ancestor::Error> {
5276 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
5277 0 : }
5278 : }
5279 :
5280 : impl Drop for Timeline {
5281 20 : fn drop(&mut self) {
5282 20 : if let Some(ancestor) = &self.ancestor_timeline {
5283 : // This lock should never be poisoned, but in case it is we do a .map() instead of
5284 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
5285 8 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
5286 8 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
5287 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
5288 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
5289 8 : }
5290 0 : }
5291 12 : }
5292 20 : info!(
5293 0 : "Timeline {} for tenant {} is being dropped",
5294 : self.timeline_id, self.tenant_shard_id.tenant_id
5295 : );
5296 20 : }
5297 : }
5298 :
5299 : /// Top-level failure to compact.
5300 : #[derive(Debug, thiserror::Error)]
5301 : pub(crate) enum CompactionError {
5302 : #[error("The timeline or pageserver is shutting down")]
5303 : ShuttingDown,
5304 : /// Compaction tried to offload a timeline and failed
5305 : #[error("Failed to offload timeline: {0}")]
5306 : Offload(OffloadError),
5307 : /// Compaction cannot be done right now; page reconstruction and so on.
5308 : #[error(transparent)]
5309 : Other(anyhow::Error),
5310 : }
5311 :
5312 : impl From<OffloadError> for CompactionError {
5313 0 : fn from(e: OffloadError) -> Self {
5314 0 : match e {
5315 0 : OffloadError::Cancelled => Self::ShuttingDown,
5316 0 : _ => Self::Offload(e),
5317 : }
5318 0 : }
5319 : }
5320 :
5321 : impl CompactionError {
5322 0 : pub fn is_cancelled(&self) -> bool {
5323 0 : matches!(self, CompactionError::ShuttingDown)
5324 0 : }
5325 : }
5326 :
5327 : impl From<CollectKeySpaceError> for CompactionError {
5328 0 : fn from(err: CollectKeySpaceError) -> Self {
5329 0 : match err {
5330 : CollectKeySpaceError::Cancelled
5331 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
5332 0 : CompactionError::ShuttingDown
5333 : }
5334 0 : e => CompactionError::Other(e.into()),
5335 : }
5336 0 : }
5337 : }
5338 :
5339 : impl From<super::upload_queue::NotInitialized> for CompactionError {
5340 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
5341 0 : match value {
5342 : super::upload_queue::NotInitialized::Uninitialized => {
5343 0 : CompactionError::Other(anyhow::anyhow!(value))
5344 : }
5345 : super::upload_queue::NotInitialized::ShuttingDown
5346 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
5347 : }
5348 0 : }
5349 : }
5350 :
5351 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
5352 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
5353 0 : match e {
5354 : super::storage_layer::layer::DownloadError::TimelineShutdown
5355 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
5356 0 : CompactionError::ShuttingDown
5357 : }
5358 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
5359 : | super::storage_layer::layer::DownloadError::DownloadRequired
5360 : | super::storage_layer::layer::DownloadError::NotFile(_)
5361 : | super::storage_layer::layer::DownloadError::DownloadFailed
5362 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
5363 0 : CompactionError::Other(anyhow::anyhow!(e))
5364 : }
5365 : #[cfg(test)]
5366 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
5367 0 : CompactionError::Other(anyhow::anyhow!(e))
5368 : }
5369 : }
5370 0 : }
5371 : }
5372 :
5373 : impl From<layer_manager::Shutdown> for CompactionError {
5374 0 : fn from(_: layer_manager::Shutdown) -> Self {
5375 0 : CompactionError::ShuttingDown
5376 0 : }
5377 : }
5378 :
5379 : #[serde_as]
5380 644 : #[derive(serde::Serialize)]
5381 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
5382 :
5383 : #[derive(Default)]
5384 : enum DurationRecorder {
5385 : #[default]
5386 : NotStarted,
5387 : Recorded(RecordedDuration, tokio::time::Instant),
5388 : }
5389 :
5390 : impl DurationRecorder {
5391 1228 : fn till_now(&self) -> DurationRecorder {
5392 1228 : match self {
5393 : DurationRecorder::NotStarted => {
5394 0 : panic!("must only call on recorded measurements")
5395 : }
5396 1228 : DurationRecorder::Recorded(_, ended) => {
5397 1228 : let now = tokio::time::Instant::now();
5398 1228 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
5399 1228 : }
5400 1228 : }
5401 1228 : }
5402 644 : fn into_recorded(self) -> Option<RecordedDuration> {
5403 644 : match self {
5404 0 : DurationRecorder::NotStarted => None,
5405 644 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
5406 : }
5407 644 : }
5408 : }
5409 :
5410 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
5411 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
5412 : /// the layer descriptor requires the user to provide the ranges, which should cover all
5413 : /// keys specified in the `data` field.
5414 : #[cfg(test)]
5415 : #[derive(Clone)]
5416 : pub struct DeltaLayerTestDesc {
5417 : pub lsn_range: Range<Lsn>,
5418 : pub key_range: Range<Key>,
5419 : pub data: Vec<(Key, Lsn, Value)>,
5420 : }
5421 :
5422 : #[cfg(test)]
5423 : impl DeltaLayerTestDesc {
5424 8 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
5425 8 : Self {
5426 8 : lsn_range,
5427 8 : key_range,
5428 8 : data,
5429 8 : }
5430 8 : }
5431 :
5432 176 : pub fn new_with_inferred_key_range(
5433 176 : lsn_range: Range<Lsn>,
5434 176 : data: Vec<(Key, Lsn, Value)>,
5435 176 : ) -> Self {
5436 440 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
5437 440 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
5438 176 : Self {
5439 176 : key_range: (*key_min)..(key_max.next()),
5440 176 : lsn_range,
5441 176 : data,
5442 176 : }
5443 176 : }
5444 :
5445 20 : pub(crate) fn layer_name(&self) -> LayerName {
5446 20 : LayerName::Delta(super::storage_layer::DeltaLayerName {
5447 20 : key_range: self.key_range.clone(),
5448 20 : lsn_range: self.lsn_range.clone(),
5449 20 : })
5450 20 : }
5451 : }
5452 :
5453 : impl Timeline {
5454 92 : async fn finish_compact_batch(
5455 92 : self: &Arc<Self>,
5456 92 : new_deltas: &[ResidentLayer],
5457 92 : new_images: &[ResidentLayer],
5458 92 : layers_to_remove: &[Layer],
5459 92 : ) -> Result<(), CompactionError> {
5460 92 : let mut guard = tokio::select! {
5461 92 : guard = self.layers.write() => guard,
5462 92 : _ = self.cancel.cancelled() => {
5463 0 : return Err(CompactionError::ShuttingDown);
5464 : }
5465 : };
5466 :
5467 92 : let mut duplicated_layers = HashSet::new();
5468 92 :
5469 92 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
5470 :
5471 744 : for l in new_deltas {
5472 652 : if guard.contains(l.as_ref()) {
5473 : // expected in tests
5474 0 : tracing::error!(layer=%l, "duplicated L1 layer");
5475 :
5476 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
5477 : // `guard` on self.layers. as of writing this, there are no error returns except
5478 : // for compact_level0_phase1 creating an L0, which does not happen in practice
5479 : // because we have not implemented L0 => L0 compaction.
5480 0 : duplicated_layers.insert(l.layer_desc().key());
5481 652 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
5482 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
5483 652 : } else {
5484 652 : insert_layers.push(l.clone());
5485 652 : }
5486 : }
5487 :
5488 : // only remove those inputs which were not outputs
5489 92 : let remove_layers: Vec<Layer> = layers_to_remove
5490 92 : .iter()
5491 804 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
5492 92 : .cloned()
5493 92 : .collect();
5494 92 :
5495 92 : if !new_images.is_empty() {
5496 0 : guard
5497 0 : .open_mut()?
5498 0 : .track_new_image_layers(new_images, &self.metrics);
5499 92 : }
5500 :
5501 92 : guard
5502 92 : .open_mut()?
5503 92 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
5504 92 :
5505 92 : self.remote_client
5506 92 : .schedule_compaction_update(&remove_layers, new_deltas)?;
5507 :
5508 92 : drop_wlock(guard);
5509 92 :
5510 92 : Ok(())
5511 92 : }
5512 :
5513 0 : async fn rewrite_layers(
5514 0 : self: &Arc<Self>,
5515 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
5516 0 : mut drop_layers: Vec<Layer>,
5517 0 : ) -> Result<(), CompactionError> {
5518 0 : let mut guard = self.layers.write().await;
5519 :
5520 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
5521 : // to avoid double-removing, and avoid rewriting something that was removed.
5522 0 : replace_layers.retain(|(l, _)| guard.contains(l));
5523 0 : drop_layers.retain(|l| guard.contains(l));
5524 0 :
5525 0 : guard
5526 0 : .open_mut()?
5527 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
5528 0 :
5529 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
5530 0 :
5531 0 : self.remote_client
5532 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
5533 :
5534 0 : Ok(())
5535 0 : }
5536 :
5537 : /// Schedules the uploads of the given image layers
5538 768 : fn upload_new_image_layers(
5539 768 : self: &Arc<Self>,
5540 768 : new_images: impl IntoIterator<Item = ResidentLayer>,
5541 768 : ) -> Result<(), super::upload_queue::NotInitialized> {
5542 808 : for layer in new_images {
5543 40 : self.remote_client.schedule_layer_file_upload(layer)?;
5544 : }
5545 : // should any new image layer been created, not uploading index_part will
5546 : // result in a mismatch between remote_physical_size and layermap calculated
5547 : // size, which will fail some tests, but should not be an issue otherwise.
5548 768 : self.remote_client
5549 768 : .schedule_index_upload_for_file_changes()?;
5550 768 : Ok(())
5551 768 : }
5552 :
5553 0 : async fn find_gc_time_cutoff(
5554 0 : &self,
5555 0 : now: SystemTime,
5556 0 : pitr: Duration,
5557 0 : cancel: &CancellationToken,
5558 0 : ctx: &RequestContext,
5559 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
5560 0 : debug_assert_current_span_has_tenant_and_timeline_id();
5561 0 : if self.shard_identity.is_shard_zero() {
5562 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
5563 0 : let time_range = if pitr == Duration::ZERO {
5564 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
5565 : } else {
5566 0 : pitr
5567 : };
5568 :
5569 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
5570 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
5571 0 : let timestamp = to_pg_timestamp(time_cutoff);
5572 :
5573 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
5574 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
5575 0 : LsnForTimestamp::Future(lsn) => {
5576 0 : // The timestamp is in the future. That sounds impossible,
5577 0 : // but what it really means is that there hasn't been
5578 0 : // any commits since the cutoff timestamp.
5579 0 : //
5580 0 : // In this case we should use the LSN of the most recent commit,
5581 0 : // which is implicitly the last LSN in the log.
5582 0 : debug!("future({})", lsn);
5583 0 : Some(self.get_last_record_lsn())
5584 : }
5585 0 : LsnForTimestamp::Past(lsn) => {
5586 0 : debug!("past({})", lsn);
5587 0 : None
5588 : }
5589 0 : LsnForTimestamp::NoData(lsn) => {
5590 0 : debug!("nodata({})", lsn);
5591 0 : None
5592 : }
5593 : };
5594 0 : Ok(time_cutoff)
5595 : } else {
5596 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
5597 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
5598 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
5599 : // space cutoff that we would also have respected ourselves.
5600 0 : match self
5601 0 : .remote_client
5602 0 : .download_foreign_index(ShardNumber(0), cancel)
5603 0 : .await
5604 : {
5605 0 : Ok((index_part, index_generation, _index_mtime)) => {
5606 0 : tracing::info!("GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
5607 0 : index_part.metadata.latest_gc_cutoff_lsn());
5608 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
5609 : }
5610 : Err(DownloadError::NotFound) => {
5611 : // This is unexpected, because during timeline creations shard zero persists to remote
5612 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
5613 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
5614 : // state, then shard zero should _eventually_ write an index when it GCs.
5615 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
5616 0 : Ok(None)
5617 : }
5618 0 : Err(e) => {
5619 0 : // TODO: this function should return a different error type than page reconstruct error
5620 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
5621 : }
5622 : }
5623 :
5624 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
5625 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
5626 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
5627 : // new location. This is legal in principle, but problematic in practice because it might result
5628 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
5629 : // because they have GC'd past the branch point.
5630 : }
5631 0 : }
5632 :
5633 : /// Find the Lsns above which layer files need to be retained on
5634 : /// garbage collection.
5635 : ///
5636 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
5637 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
5638 : /// the space-based retention.
5639 : ///
5640 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
5641 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
5642 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
5643 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
5644 : /// in the response as the GC cutoff point for the timeline.
5645 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
5646 : pub(super) async fn find_gc_cutoffs(
5647 : &self,
5648 : now: SystemTime,
5649 : space_cutoff: Lsn,
5650 : pitr: Duration,
5651 : cancel: &CancellationToken,
5652 : ctx: &RequestContext,
5653 : ) -> Result<GcCutoffs, PageReconstructError> {
5654 : let _timer = self
5655 : .metrics
5656 : .find_gc_cutoffs_histo
5657 : .start_timer()
5658 : .record_on_drop();
5659 :
5660 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
5661 :
5662 : if cfg!(test) {
5663 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
5664 : if pitr == Duration::ZERO {
5665 : return Ok(GcCutoffs {
5666 : time: self.get_last_record_lsn(),
5667 : space: space_cutoff,
5668 : });
5669 : }
5670 : }
5671 :
5672 : // Calculate a time-based limit on how much to retain:
5673 : // - if PITR interval is set, then this is our cutoff.
5674 : // - if PITR interval is not set, then we do a lookup
5675 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
5676 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
5677 :
5678 : Ok(match (pitr, time_cutoff) {
5679 : (Duration::ZERO, Some(time_cutoff)) => {
5680 : // PITR is not set. Retain the size-based limit, or the default time retention,
5681 : // whichever requires less data.
5682 : GcCutoffs {
5683 : time: self.get_last_record_lsn(),
5684 : space: std::cmp::max(time_cutoff, space_cutoff),
5685 : }
5686 : }
5687 : (Duration::ZERO, None) => {
5688 : // PITR is not set, and time lookup failed
5689 : GcCutoffs {
5690 : time: self.get_last_record_lsn(),
5691 : space: space_cutoff,
5692 : }
5693 : }
5694 : (_, None) => {
5695 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
5696 : // cannot advance beyond what was already GC'd, and respect space-based retention
5697 : GcCutoffs {
5698 : time: *self.get_applied_gc_cutoff_lsn(),
5699 : space: space_cutoff,
5700 : }
5701 : }
5702 : (_, Some(time_cutoff)) => {
5703 : // PITR interval is set and we looked up timestamp successfully. Ignore
5704 : // size based retention and make time cutoff authoritative
5705 : GcCutoffs {
5706 : time: time_cutoff,
5707 : space: time_cutoff,
5708 : }
5709 : }
5710 : })
5711 : }
5712 :
5713 : /// Garbage collect layer files on a timeline that are no longer needed.
5714 : ///
5715 : /// Currently, we don't make any attempt at removing unneeded page versions
5716 : /// within a layer file. We can only remove the whole file if it's fully
5717 : /// obsolete.
5718 8 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
5719 : // this is most likely the background tasks, but it might be the spawned task from
5720 : // immediate_gc
5721 8 : let _g = tokio::select! {
5722 8 : guard = self.gc_lock.lock() => guard,
5723 8 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
5724 : };
5725 8 : let timer = self.metrics.garbage_collect_histo.start_timer();
5726 8 :
5727 8 : fail_point!("before-timeline-gc");
5728 8 :
5729 8 : // Is the timeline being deleted?
5730 8 : if self.is_stopping() {
5731 0 : return Err(GcError::TimelineCancelled);
5732 8 : }
5733 8 :
5734 8 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
5735 8 : let gc_info = self.gc_info.read().unwrap();
5736 8 :
5737 8 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
5738 8 : let time_cutoff = gc_info.cutoffs.time;
5739 8 : let retain_lsns = gc_info
5740 8 : .retain_lsns
5741 8 : .iter()
5742 8 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
5743 8 : .collect();
5744 8 :
5745 8 : // Gets the maximum LSN that holds the valid lease.
5746 8 : //
5747 8 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
5748 8 : // Here, we do not check for stale leases again.
5749 8 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
5750 8 :
5751 8 : (
5752 8 : space_cutoff,
5753 8 : time_cutoff,
5754 8 : retain_lsns,
5755 8 : max_lsn_with_valid_lease,
5756 8 : )
5757 8 : };
5758 8 :
5759 8 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
5760 8 : let standby_horizon = self.standby_horizon.load();
5761 8 : // Hold GC for the standby, but as a safety guard do it only within some
5762 8 : // reasonable lag.
5763 8 : if standby_horizon != Lsn::INVALID {
5764 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
5765 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
5766 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
5767 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
5768 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
5769 : } else {
5770 0 : warn!(
5771 0 : "standby is lagging for more than {}MB, not holding gc for it",
5772 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
5773 : )
5774 : }
5775 0 : }
5776 8 : }
5777 :
5778 : // Reset standby horizon to ignore it if it is not updated till next GC.
5779 : // It is an easy way to unset it when standby disappears without adding
5780 : // more conf options.
5781 8 : self.standby_horizon.store(Lsn::INVALID);
5782 8 : self.metrics
5783 8 : .standby_horizon_gauge
5784 8 : .set(Lsn::INVALID.0 as i64);
5785 :
5786 8 : let res = self
5787 8 : .gc_timeline(
5788 8 : space_cutoff,
5789 8 : time_cutoff,
5790 8 : retain_lsns,
5791 8 : max_lsn_with_valid_lease,
5792 8 : new_gc_cutoff,
5793 8 : )
5794 8 : .instrument(
5795 8 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5796 : )
5797 8 : .await?;
5798 :
5799 : // only record successes
5800 8 : timer.stop_and_record();
5801 8 :
5802 8 : Ok(res)
5803 8 : }
5804 :
5805 8 : async fn gc_timeline(
5806 8 : &self,
5807 8 : space_cutoff: Lsn,
5808 8 : time_cutoff: Lsn,
5809 8 : retain_lsns: Vec<Lsn>,
5810 8 : max_lsn_with_valid_lease: Option<Lsn>,
5811 8 : new_gc_cutoff: Lsn,
5812 8 : ) -> Result<GcResult, GcError> {
5813 8 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5814 8 :
5815 8 : let now = SystemTime::now();
5816 8 : let mut result: GcResult = GcResult::default();
5817 8 :
5818 8 : // Nothing to GC. Return early.
5819 8 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
5820 8 : if latest_gc_cutoff >= new_gc_cutoff {
5821 0 : info!(
5822 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5823 : );
5824 0 : return Ok(result);
5825 8 : }
5826 :
5827 : // We need to ensure that no one tries to read page versions or create
5828 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5829 : // for details. This will block until the old value is no longer in use.
5830 : //
5831 : // The GC cutoff should only ever move forwards.
5832 8 : let waitlist = {
5833 8 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
5834 8 : if *write_guard > new_gc_cutoff {
5835 0 : return Err(GcError::BadLsn {
5836 0 : why: format!(
5837 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5838 0 : *write_guard, new_gc_cutoff
5839 0 : ),
5840 0 : });
5841 8 : }
5842 8 :
5843 8 : write_guard.store_and_unlock(new_gc_cutoff)
5844 8 : };
5845 8 : waitlist.wait().await;
5846 :
5847 8 : info!("GC starting");
5848 :
5849 8 : debug!("retain_lsns: {:?}", retain_lsns);
5850 :
5851 8 : let mut layers_to_remove = Vec::new();
5852 :
5853 : // Scan all layers in the timeline (remote or on-disk).
5854 : //
5855 : // Garbage collect the layer if all conditions are satisfied:
5856 : // 1. it is older than cutoff LSN;
5857 : // 2. it is older than PITR interval;
5858 : // 3. it doesn't need to be retained for 'retain_lsns';
5859 : // 4. it does not need to be kept for LSNs holding valid leases.
5860 : // 5. newer on-disk image layers cover the layer's whole key range
5861 : //
5862 : // TODO holding a write lock is too agressive and avoidable
5863 8 : let mut guard = self.layers.write().await;
5864 8 : let layers = guard.layer_map()?;
5865 48 : 'outer: for l in layers.iter_historic_layers() {
5866 48 : result.layers_total += 1;
5867 48 :
5868 48 : // 1. Is it newer than GC horizon cutoff point?
5869 48 : if l.get_lsn_range().end > space_cutoff {
5870 4 : info!(
5871 0 : "keeping {} because it's newer than space_cutoff {}",
5872 0 : l.layer_name(),
5873 : space_cutoff,
5874 : );
5875 4 : result.layers_needed_by_cutoff += 1;
5876 4 : continue 'outer;
5877 44 : }
5878 44 :
5879 44 : // 2. It is newer than PiTR cutoff point?
5880 44 : if l.get_lsn_range().end > time_cutoff {
5881 0 : info!(
5882 0 : "keeping {} because it's newer than time_cutoff {}",
5883 0 : l.layer_name(),
5884 : time_cutoff,
5885 : );
5886 0 : result.layers_needed_by_pitr += 1;
5887 0 : continue 'outer;
5888 44 : }
5889 :
5890 : // 3. Is it needed by a child branch?
5891 : // NOTE With that we would keep data that
5892 : // might be referenced by child branches forever.
5893 : // We can track this in child timeline GC and delete parent layers when
5894 : // they are no longer needed. This might be complicated with long inheritance chains.
5895 : //
5896 : // TODO Vec is not a great choice for `retain_lsns`
5897 44 : for retain_lsn in &retain_lsns {
5898 : // start_lsn is inclusive
5899 0 : if &l.get_lsn_range().start <= retain_lsn {
5900 0 : info!(
5901 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5902 0 : l.layer_name(),
5903 0 : retain_lsn,
5904 0 : l.is_incremental(),
5905 : );
5906 0 : result.layers_needed_by_branches += 1;
5907 0 : continue 'outer;
5908 0 : }
5909 : }
5910 :
5911 : // 4. Is there a valid lease that requires us to keep this layer?
5912 44 : if let Some(lsn) = &max_lsn_with_valid_lease {
5913 : // keep if layer start <= any of the lease
5914 36 : if &l.get_lsn_range().start <= lsn {
5915 28 : info!(
5916 0 : "keeping {} because there is a valid lease preventing GC at {}",
5917 0 : l.layer_name(),
5918 : lsn,
5919 : );
5920 28 : result.layers_needed_by_leases += 1;
5921 28 : continue 'outer;
5922 8 : }
5923 8 : }
5924 :
5925 : // 5. Is there a later on-disk layer for this relation?
5926 : //
5927 : // The end-LSN is exclusive, while disk_consistent_lsn is
5928 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5929 : // OK for a delta layer to have end LSN 101, but if the end LSN
5930 : // is 102, then it might not have been fully flushed to disk
5931 : // before crash.
5932 : //
5933 : // For example, imagine that the following layers exist:
5934 : //
5935 : // 1000 - image (A)
5936 : // 1000-2000 - delta (B)
5937 : // 2000 - image (C)
5938 : // 2000-3000 - delta (D)
5939 : // 3000 - image (E)
5940 : //
5941 : // If GC horizon is at 2500, we can remove layers A and B, but
5942 : // we cannot remove C, even though it's older than 2500, because
5943 : // the delta layer 2000-3000 depends on it.
5944 16 : if !layers
5945 16 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5946 : {
5947 12 : info!("keeping {} because it is the latest layer", l.layer_name());
5948 12 : result.layers_not_updated += 1;
5949 12 : continue 'outer;
5950 4 : }
5951 4 :
5952 4 : // We didn't find any reason to keep this file, so remove it.
5953 4 : info!(
5954 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5955 0 : l.layer_name(),
5956 0 : l.is_incremental(),
5957 : );
5958 4 : layers_to_remove.push(l);
5959 : }
5960 :
5961 8 : if !layers_to_remove.is_empty() {
5962 : // Persist the new GC cutoff value before we actually remove anything.
5963 : // This unconditionally schedules also an index_part.json update, even though, we will
5964 : // be doing one a bit later with the unlinked gc'd layers.
5965 4 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5966 4 : self.schedule_uploads(disk_consistent_lsn, None)
5967 4 : .map_err(|e| {
5968 0 : if self.cancel.is_cancelled() {
5969 0 : GcError::TimelineCancelled
5970 : } else {
5971 0 : GcError::Remote(e)
5972 : }
5973 4 : })?;
5974 :
5975 4 : let gc_layers = layers_to_remove
5976 4 : .iter()
5977 4 : .map(|x| guard.get_from_desc(x))
5978 4 : .collect::<Vec<Layer>>();
5979 4 :
5980 4 : result.layers_removed = gc_layers.len() as u64;
5981 4 :
5982 4 : self.remote_client.schedule_gc_update(&gc_layers)?;
5983 :
5984 4 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5985 4 :
5986 4 : #[cfg(feature = "testing")]
5987 4 : {
5988 4 : result.doomed_layers = gc_layers;
5989 4 : }
5990 4 : }
5991 :
5992 8 : info!(
5993 0 : "GC completed removing {} layers, cutoff {}",
5994 : result.layers_removed, new_gc_cutoff
5995 : );
5996 :
5997 8 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5998 8 : Ok(result)
5999 8 : }
6000 :
6001 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
6002 1331942 : async fn reconstruct_value(
6003 1331942 : &self,
6004 1331942 : key: Key,
6005 1331942 : request_lsn: Lsn,
6006 1331942 : mut data: ValueReconstructState,
6007 1331942 : ) -> Result<Bytes, PageReconstructError> {
6008 1331942 : // Perform WAL redo if needed
6009 1331942 : data.records.reverse();
6010 1331942 :
6011 1331942 : // If we have a page image, and no WAL, we're all set
6012 1331942 : if data.records.is_empty() {
6013 1330306 : if let Some((img_lsn, img)) = &data.img {
6014 1330306 : trace!(
6015 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
6016 : key,
6017 : img_lsn,
6018 : request_lsn,
6019 : );
6020 1330306 : Ok(img.clone())
6021 : } else {
6022 0 : Err(PageReconstructError::from(anyhow!(
6023 0 : "base image for {key} at {request_lsn} not found"
6024 0 : )))
6025 : }
6026 : } else {
6027 : // We need to do WAL redo.
6028 : //
6029 : // If we don't have a base image, then the oldest WAL record better initialize
6030 : // the page
6031 1636 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
6032 0 : Err(PageReconstructError::from(anyhow!(
6033 0 : "Base image for {} at {} not found, but got {} WAL records",
6034 0 : key,
6035 0 : request_lsn,
6036 0 : data.records.len()
6037 0 : )))
6038 : } else {
6039 1636 : if data.img.is_some() {
6040 1504 : trace!(
6041 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
6042 0 : data.records.len(),
6043 : key,
6044 : request_lsn
6045 : );
6046 : } else {
6047 132 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
6048 : };
6049 1636 : let res = self
6050 1636 : .walredo_mgr
6051 1636 : .as_ref()
6052 1636 : .context("timeline has no walredo manager")
6053 1636 : .map_err(PageReconstructError::WalRedo)?
6054 1636 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
6055 1636 : .await;
6056 1636 : let img = match res {
6057 1636 : Ok(img) => img,
6058 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
6059 0 : Err(walredo::Error::Other(err)) => {
6060 0 : critical!("walredo failure during page reconstruction: {err:?}");
6061 0 : return Err(PageReconstructError::WalRedo(
6062 0 : err.context("reconstruct a page image"),
6063 0 : ));
6064 : }
6065 : };
6066 1636 : Ok(img)
6067 : }
6068 : }
6069 1331942 : }
6070 :
6071 0 : pub(crate) async fn spawn_download_all_remote_layers(
6072 0 : self: Arc<Self>,
6073 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6074 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
6075 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6076 :
6077 : // this is not really needed anymore; it has tests which really check the return value from
6078 : // http api. it would be better not to maintain this anymore.
6079 :
6080 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
6081 0 : if let Some(st) = &*status_guard {
6082 0 : match &st.state {
6083 : DownloadRemoteLayersTaskState::Running => {
6084 0 : return Err(st.clone());
6085 : }
6086 : DownloadRemoteLayersTaskState::ShutDown
6087 0 : | DownloadRemoteLayersTaskState::Completed => {
6088 0 : *status_guard = None;
6089 0 : }
6090 : }
6091 0 : }
6092 :
6093 0 : let self_clone = Arc::clone(&self);
6094 0 : let task_id = task_mgr::spawn(
6095 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
6096 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
6097 0 : self.tenant_shard_id,
6098 0 : Some(self.timeline_id),
6099 0 : "download all remote layers task",
6100 0 : async move {
6101 0 : self_clone.download_all_remote_layers(request).await;
6102 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
6103 0 : match &mut *status_guard {
6104 : None => {
6105 0 : warn!("tasks status is supposed to be Some(), since we are running");
6106 : }
6107 0 : Some(st) => {
6108 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
6109 0 : if st.task_id != exp_task_id {
6110 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
6111 0 : } else {
6112 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6113 0 : }
6114 : }
6115 : };
6116 0 : Ok(())
6117 0 : }
6118 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))
6119 : );
6120 :
6121 0 : let initial_info = DownloadRemoteLayersTaskInfo {
6122 0 : task_id: format!("{task_id}"),
6123 0 : state: DownloadRemoteLayersTaskState::Running,
6124 0 : total_layer_count: 0,
6125 0 : successful_download_count: 0,
6126 0 : failed_download_count: 0,
6127 0 : };
6128 0 : *status_guard = Some(initial_info.clone());
6129 0 :
6130 0 : Ok(initial_info)
6131 0 : }
6132 :
6133 0 : async fn download_all_remote_layers(
6134 0 : self: &Arc<Self>,
6135 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6136 0 : ) {
6137 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6138 :
6139 0 : let remaining = {
6140 0 : let guard = self.layers.read().await;
6141 0 : let Ok(lm) = guard.layer_map() else {
6142 : // technically here we could look into iterating accessible layers, but downloading
6143 : // all layers of a shutdown timeline makes no sense regardless.
6144 0 : tracing::info!("attempted to download all layers of shutdown timeline");
6145 0 : return;
6146 : };
6147 0 : lm.iter_historic_layers()
6148 0 : .map(|desc| guard.get_from_desc(&desc))
6149 0 : .collect::<Vec<_>>()
6150 0 : };
6151 0 : let total_layer_count = remaining.len();
6152 :
6153 : macro_rules! lock_status {
6154 : ($st:ident) => {
6155 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
6156 : let st = st
6157 : .as_mut()
6158 : .expect("this function is only called after the task has been spawned");
6159 : assert_eq!(
6160 : st.task_id,
6161 : format!(
6162 : "{}",
6163 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
6164 : )
6165 : );
6166 : let $st = st;
6167 : };
6168 : }
6169 :
6170 : {
6171 0 : lock_status!(st);
6172 0 : st.total_layer_count = total_layer_count as u64;
6173 0 : }
6174 0 :
6175 0 : let mut remaining = remaining.into_iter();
6176 0 : let mut have_remaining = true;
6177 0 : let mut js = tokio::task::JoinSet::new();
6178 0 :
6179 0 : let cancel = task_mgr::shutdown_token();
6180 0 :
6181 0 : let limit = request.max_concurrent_downloads;
6182 :
6183 : loop {
6184 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
6185 0 : let Some(next) = remaining.next() else {
6186 0 : have_remaining = false;
6187 0 : break;
6188 : };
6189 :
6190 0 : let span = tracing::info_span!("download", layer = %next);
6191 :
6192 0 : js.spawn(
6193 0 : async move {
6194 0 : let res = next.download().await;
6195 0 : (next, res)
6196 0 : }
6197 0 : .instrument(span),
6198 0 : );
6199 0 : }
6200 :
6201 0 : while let Some(res) = js.join_next().await {
6202 0 : match res {
6203 : Ok((_, Ok(_))) => {
6204 0 : lock_status!(st);
6205 0 : st.successful_download_count += 1;
6206 : }
6207 0 : Ok((layer, Err(e))) => {
6208 0 : tracing::error!(%layer, "download failed: {e:#}");
6209 0 : lock_status!(st);
6210 0 : st.failed_download_count += 1;
6211 : }
6212 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
6213 0 : Err(je) if je.is_panic() => {
6214 0 : lock_status!(st);
6215 0 : st.failed_download_count += 1;
6216 : }
6217 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
6218 : }
6219 : }
6220 :
6221 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
6222 0 : break;
6223 0 : }
6224 : }
6225 :
6226 : {
6227 0 : lock_status!(st);
6228 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6229 : }
6230 0 : }
6231 :
6232 0 : pub(crate) fn get_download_all_remote_layers_task_info(
6233 0 : &self,
6234 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
6235 0 : self.download_all_remote_layers_task_info
6236 0 : .read()
6237 0 : .unwrap()
6238 0 : .clone()
6239 0 : }
6240 : }
6241 :
6242 : impl Timeline {
6243 : /// Returns non-remote layers for eviction.
6244 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
6245 0 : let guard = self.layers.read().await;
6246 0 : let mut max_layer_size: Option<u64> = None;
6247 0 :
6248 0 : let resident_layers = guard
6249 0 : .likely_resident_layers()
6250 0 : .map(|layer| {
6251 0 : let file_size = layer.layer_desc().file_size;
6252 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
6253 0 :
6254 0 : let last_activity_ts = layer.latest_activity();
6255 0 :
6256 0 : EvictionCandidate {
6257 0 : layer: layer.to_owned().into(),
6258 0 : last_activity_ts,
6259 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
6260 0 : visibility: layer.visibility(),
6261 0 : }
6262 0 : })
6263 0 : .collect();
6264 0 :
6265 0 : DiskUsageEvictionInfo {
6266 0 : max_layer_size,
6267 0 : resident_layers,
6268 0 : }
6269 0 : }
6270 :
6271 3756 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
6272 3756 : ShardIndex {
6273 3756 : shard_number: self.tenant_shard_id.shard_number,
6274 3756 : shard_count: self.tenant_shard_id.shard_count,
6275 3756 : }
6276 3756 : }
6277 :
6278 : /// Persistently blocks gc for `Manual` reason.
6279 : ///
6280 : /// Returns true if no such block existed before, false otherwise.
6281 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
6282 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6283 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6284 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
6285 0 : }
6286 :
6287 : /// Persistently unblocks gc for `Manual` reason.
6288 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
6289 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6290 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6291 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
6292 0 : }
6293 :
6294 : #[cfg(test)]
6295 96 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
6296 96 : self.last_record_lsn.advance(new_lsn);
6297 96 : }
6298 :
6299 : #[cfg(test)]
6300 8 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
6301 8 : self.disk_consistent_lsn.store(new_value);
6302 8 : }
6303 :
6304 : /// Force create an image layer and place it into the layer map.
6305 : ///
6306 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
6307 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
6308 : /// placed into the layer map in one run AND be validated.
6309 : #[cfg(test)]
6310 120 : pub(super) async fn force_create_image_layer(
6311 120 : self: &Arc<Timeline>,
6312 120 : lsn: Lsn,
6313 120 : mut images: Vec<(Key, Bytes)>,
6314 120 : check_start_lsn: Option<Lsn>,
6315 120 : ctx: &RequestContext,
6316 120 : ) -> anyhow::Result<()> {
6317 120 : let last_record_lsn = self.get_last_record_lsn();
6318 120 : assert!(
6319 120 : lsn <= last_record_lsn,
6320 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
6321 : );
6322 120 : if let Some(check_start_lsn) = check_start_lsn {
6323 120 : assert!(lsn >= check_start_lsn);
6324 0 : }
6325 348 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
6326 120 : let min_key = *images.first().map(|(k, _)| k).unwrap();
6327 120 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
6328 120 : let mut image_layer_writer = ImageLayerWriter::new(
6329 120 : self.conf,
6330 120 : self.timeline_id,
6331 120 : self.tenant_shard_id,
6332 120 : &(min_key..end_key),
6333 120 : lsn,
6334 120 : ctx,
6335 120 : )
6336 120 : .await?;
6337 588 : for (key, img) in images {
6338 468 : image_layer_writer.put_image(key, img, ctx).await?;
6339 : }
6340 120 : let (desc, path) = image_layer_writer.finish(ctx).await?;
6341 120 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6342 120 : info!("force created image layer {}", image_layer.local_path());
6343 : {
6344 120 : let mut guard = self.layers.write().await;
6345 120 : guard
6346 120 : .open_mut()
6347 120 : .unwrap()
6348 120 : .force_insert_layer(image_layer.clone());
6349 120 : }
6350 120 :
6351 120 : // Update remote_timeline_client state to reflect existence of this layer
6352 120 : self.remote_client
6353 120 : .schedule_layer_file_upload(image_layer)
6354 120 : .unwrap();
6355 120 :
6356 120 : Ok(())
6357 120 : }
6358 :
6359 : /// Force create a delta layer and place it into the layer map.
6360 : ///
6361 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
6362 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
6363 : /// placed into the layer map in one run AND be validated.
6364 : #[cfg(test)]
6365 184 : pub(super) async fn force_create_delta_layer(
6366 184 : self: &Arc<Timeline>,
6367 184 : mut deltas: DeltaLayerTestDesc,
6368 184 : check_start_lsn: Option<Lsn>,
6369 184 : ctx: &RequestContext,
6370 184 : ) -> anyhow::Result<()> {
6371 184 : let last_record_lsn = self.get_last_record_lsn();
6372 184 : deltas
6373 184 : .data
6374 264 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
6375 184 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
6376 184 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
6377 632 : for (_, lsn, _) in &deltas.data {
6378 448 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
6379 : }
6380 184 : assert!(
6381 184 : deltas.lsn_range.end <= last_record_lsn,
6382 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
6383 : deltas.lsn_range.end,
6384 : last_record_lsn
6385 : );
6386 184 : if let Some(check_start_lsn) = check_start_lsn {
6387 184 : assert!(deltas.lsn_range.start >= check_start_lsn);
6388 0 : }
6389 184 : let mut delta_layer_writer = DeltaLayerWriter::new(
6390 184 : self.conf,
6391 184 : self.timeline_id,
6392 184 : self.tenant_shard_id,
6393 184 : deltas.key_range.start,
6394 184 : deltas.lsn_range,
6395 184 : ctx,
6396 184 : )
6397 184 : .await?;
6398 632 : for (key, lsn, val) in deltas.data {
6399 448 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
6400 : }
6401 184 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
6402 184 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6403 184 : info!("force created delta layer {}", delta_layer.local_path());
6404 : {
6405 184 : let mut guard = self.layers.write().await;
6406 184 : guard
6407 184 : .open_mut()
6408 184 : .unwrap()
6409 184 : .force_insert_layer(delta_layer.clone());
6410 184 : }
6411 184 :
6412 184 : // Update remote_timeline_client state to reflect existence of this layer
6413 184 : self.remote_client
6414 184 : .schedule_layer_file_upload(delta_layer)
6415 184 : .unwrap();
6416 184 :
6417 184 : Ok(())
6418 184 : }
6419 :
6420 : /// Return all keys at the LSN in the image layers
6421 : #[cfg(test)]
6422 12 : pub(crate) async fn inspect_image_layers(
6423 12 : self: &Arc<Timeline>,
6424 12 : lsn: Lsn,
6425 12 : ctx: &RequestContext,
6426 12 : io_concurrency: IoConcurrency,
6427 12 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
6428 12 : let mut all_data = Vec::new();
6429 12 : let guard = self.layers.read().await;
6430 68 : for layer in guard.layer_map()?.iter_historic_layers() {
6431 68 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
6432 12 : let layer = guard.get_from_desc(&layer);
6433 12 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
6434 12 : layer
6435 12 : .get_values_reconstruct_data(
6436 12 : KeySpace::single(Key::MIN..Key::MAX),
6437 12 : lsn..Lsn(lsn.0 + 1),
6438 12 : &mut reconstruct_data,
6439 12 : ctx,
6440 12 : )
6441 12 : .await?;
6442 104 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
6443 104 : let v = v.collect_pending_ios().await?;
6444 104 : all_data.push((k, v.img.unwrap().1));
6445 : }
6446 56 : }
6447 : }
6448 12 : all_data.sort();
6449 12 : Ok(all_data)
6450 12 : }
6451 :
6452 : /// Get all historic layer descriptors in the layer map
6453 : #[cfg(test)]
6454 48 : pub(crate) async fn inspect_historic_layers(
6455 48 : self: &Arc<Timeline>,
6456 48 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
6457 48 : let mut layers = Vec::new();
6458 48 : let guard = self.layers.read().await;
6459 228 : for layer in guard.layer_map()?.iter_historic_layers() {
6460 228 : layers.push(layer.key());
6461 228 : }
6462 48 : Ok(layers)
6463 48 : }
6464 :
6465 : #[cfg(test)]
6466 20 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
6467 20 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
6468 20 : keyspace.merge(&ks);
6469 20 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
6470 20 : }
6471 : }
6472 :
6473 : /// Tracking writes ingestion does to a particular in-memory layer.
6474 : ///
6475 : /// Cleared upon freezing a layer.
6476 : pub(crate) struct TimelineWriterState {
6477 : open_layer: Arc<InMemoryLayer>,
6478 : current_size: u64,
6479 : // Previous Lsn which passed through
6480 : prev_lsn: Option<Lsn>,
6481 : // Largest Lsn which passed through the current writer
6482 : max_lsn: Option<Lsn>,
6483 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
6484 : cached_last_freeze_at: Lsn,
6485 : }
6486 :
6487 : impl TimelineWriterState {
6488 2596 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
6489 2596 : Self {
6490 2596 : open_layer,
6491 2596 : current_size,
6492 2596 : prev_lsn: None,
6493 2596 : max_lsn: None,
6494 2596 : cached_last_freeze_at: last_freeze_at,
6495 2596 : }
6496 2596 : }
6497 : }
6498 :
6499 : /// Various functions to mutate the timeline.
6500 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
6501 : // This is probably considered a bad practice in Rust and should be fixed eventually,
6502 : // but will cause large code changes.
6503 : pub(crate) struct TimelineWriter<'a> {
6504 : tl: &'a Timeline,
6505 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
6506 : }
6507 :
6508 : impl Deref for TimelineWriter<'_> {
6509 : type Target = Timeline;
6510 :
6511 19796404 : fn deref(&self) -> &Self::Target {
6512 19796404 : self.tl
6513 19796404 : }
6514 : }
6515 :
6516 : #[derive(PartialEq)]
6517 : enum OpenLayerAction {
6518 : Roll,
6519 : Open,
6520 : None,
6521 : }
6522 :
6523 : impl TimelineWriter<'_> {
6524 9608468 : async fn handle_open_layer_action(
6525 9608468 : &mut self,
6526 9608468 : at: Lsn,
6527 9608468 : action: OpenLayerAction,
6528 9608468 : ctx: &RequestContext,
6529 9608468 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
6530 9608468 : match action {
6531 : OpenLayerAction::Roll => {
6532 160 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
6533 160 : self.roll_layer(freeze_at).await?;
6534 160 : self.open_layer(at, ctx).await?;
6535 : }
6536 2436 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
6537 : OpenLayerAction::None => {
6538 9605872 : assert!(self.write_guard.is_some());
6539 : }
6540 : }
6541 :
6542 9608468 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
6543 9608468 : }
6544 :
6545 2596 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
6546 2596 : let layer = self
6547 2596 : .tl
6548 2596 : .get_layer_for_write(at, &self.write_guard, ctx)
6549 2596 : .await?;
6550 2596 : let initial_size = layer.size().await?;
6551 :
6552 2596 : let last_freeze_at = self.last_freeze_at.load();
6553 2596 : self.write_guard.replace(TimelineWriterState::new(
6554 2596 : layer,
6555 2596 : initial_size,
6556 2596 : last_freeze_at,
6557 2596 : ));
6558 2596 :
6559 2596 : Ok(())
6560 2596 : }
6561 :
6562 160 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
6563 160 : let current_size = self.write_guard.as_ref().unwrap().current_size;
6564 :
6565 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
6566 : // to propagate the backpressure up into WAL ingestion.
6567 160 : let l0_count = self
6568 160 : .tl
6569 160 : .layers
6570 160 : .read()
6571 160 : .await
6572 160 : .layer_map()?
6573 160 : .level0_deltas()
6574 160 : .len();
6575 160 : let wait_thresholds = [
6576 160 : self.get_l0_flush_delay_threshold(),
6577 160 : self.get_l0_flush_stall_threshold(),
6578 160 : ];
6579 160 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
6580 :
6581 : // self.write_guard will be taken by the freezing
6582 160 : let flush_id = self
6583 160 : .tl
6584 160 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
6585 160 : .await?;
6586 :
6587 160 : assert!(self.write_guard.is_none());
6588 :
6589 160 : if let Some(wait_threshold) = wait_threshold {
6590 0 : if l0_count >= wait_threshold {
6591 0 : info!("layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers");
6592 0 : self.tl.wait_flush_completion(flush_id).await?;
6593 0 : }
6594 160 : }
6595 :
6596 160 : if current_size >= self.get_checkpoint_distance() * 2 {
6597 0 : warn!("Flushed oversized open layer with size {}", current_size)
6598 160 : }
6599 :
6600 160 : Ok(())
6601 160 : }
6602 :
6603 9608468 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
6604 9608468 : let state = &*self.write_guard;
6605 9608468 : let Some(state) = &state else {
6606 2436 : return OpenLayerAction::Open;
6607 : };
6608 :
6609 : #[cfg(feature = "testing")]
6610 9606032 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
6611 : // this check and assertion are not really needed because
6612 : // LayerManager::try_freeze_in_memory_layer will always clear out the
6613 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
6614 : // is no TimelineWriterState.
6615 0 : assert!(
6616 0 : state.open_layer.end_lsn.get().is_some(),
6617 0 : "our open_layer must be outdated"
6618 : );
6619 :
6620 : // this would be a memory leak waiting to happen because the in-memory layer always has
6621 : // an index
6622 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
6623 9606032 : }
6624 9606032 :
6625 9606032 : if state.prev_lsn == Some(lsn) {
6626 : // Rolling mid LSN is not supported by [downstream code].
6627 : // Hence, only roll at LSN boundaries.
6628 : //
6629 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
6630 12 : return OpenLayerAction::None;
6631 9606020 : }
6632 9606020 :
6633 9606020 : if state.current_size == 0 {
6634 : // Don't roll empty layers
6635 0 : return OpenLayerAction::None;
6636 9606020 : }
6637 9606020 :
6638 9606020 : if self.tl.should_roll(
6639 9606020 : state.current_size,
6640 9606020 : state.current_size + new_value_size,
6641 9606020 : self.get_checkpoint_distance(),
6642 9606020 : lsn,
6643 9606020 : state.cached_last_freeze_at,
6644 9606020 : state.open_layer.get_opened_at(),
6645 9606020 : ) {
6646 160 : OpenLayerAction::Roll
6647 : } else {
6648 9605860 : OpenLayerAction::None
6649 : }
6650 9608468 : }
6651 :
6652 : /// Put a batch of keys at the specified Lsns.
6653 9608464 : pub(crate) async fn put_batch(
6654 9608464 : &mut self,
6655 9608464 : batch: SerializedValueBatch,
6656 9608464 : ctx: &RequestContext,
6657 9608464 : ) -> anyhow::Result<()> {
6658 9608464 : if !batch.has_data() {
6659 0 : return Ok(());
6660 9608464 : }
6661 9608464 :
6662 9608464 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
6663 9608464 : // We don't assert this in release builds, since key ownership policies may change over
6664 9608464 : // time. Stray keys will be removed during compaction.
6665 9608464 : if cfg!(debug_assertions) {
6666 19789824 : for metadata in &batch.metadata {
6667 10181360 : if let ValueMeta::Serialized(metadata) = metadata {
6668 10181360 : let key = Key::from_compact(metadata.key);
6669 10181360 : assert!(
6670 10181360 : self.shard_identity.is_key_local(&key)
6671 0 : || self.shard_identity.is_key_global(&key),
6672 0 : "key {key} does not belong on shard {}",
6673 0 : self.shard_identity.shard_index()
6674 : );
6675 0 : }
6676 : }
6677 0 : }
6678 :
6679 9608464 : let batch_max_lsn = batch.max_lsn;
6680 9608464 : let buf_size: u64 = batch.buffer_size() as u64;
6681 9608464 :
6682 9608464 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
6683 9608464 : let layer = self
6684 9608464 : .handle_open_layer_action(batch_max_lsn, action, ctx)
6685 9608464 : .await?;
6686 :
6687 9608464 : let res = layer.put_batch(batch, ctx).await;
6688 :
6689 9608464 : if res.is_ok() {
6690 9608464 : // Update the current size only when the entire write was ok.
6691 9608464 : // In case of failures, we may have had partial writes which
6692 9608464 : // render the size tracking out of sync. That's ok because
6693 9608464 : // the checkpoint distance should be significantly smaller
6694 9608464 : // than the S3 single shot upload limit of 5GiB.
6695 9608464 : let state = self.write_guard.as_mut().unwrap();
6696 9608464 :
6697 9608464 : state.current_size += buf_size;
6698 9608464 : state.prev_lsn = Some(batch_max_lsn);
6699 9608464 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
6700 9608464 : }
6701 :
6702 9608464 : res
6703 9608464 : }
6704 :
6705 : #[cfg(test)]
6706 : /// Test helper, for tests that would like to poke individual values without composing a batch
6707 8780308 : pub(crate) async fn put(
6708 8780308 : &mut self,
6709 8780308 : key: Key,
6710 8780308 : lsn: Lsn,
6711 8780308 : value: &Value,
6712 8780308 : ctx: &RequestContext,
6713 8780308 : ) -> anyhow::Result<()> {
6714 : use utils::bin_ser::BeSer;
6715 8780308 : if !key.is_valid_key_on_write_path() {
6716 0 : bail!(
6717 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
6718 0 : key
6719 0 : );
6720 8780308 : }
6721 8780308 : let val_ser_size = value.serialized_size().unwrap() as usize;
6722 8780308 : let batch = SerializedValueBatch::from_values(vec![(
6723 8780308 : key.to_compact(),
6724 8780308 : lsn,
6725 8780308 : val_ser_size,
6726 8780308 : value.clone(),
6727 8780308 : )]);
6728 8780308 :
6729 8780308 : self.put_batch(batch, ctx).await
6730 8780308 : }
6731 :
6732 4 : pub(crate) async fn delete_batch(
6733 4 : &mut self,
6734 4 : batch: &[(Range<Key>, Lsn)],
6735 4 : ctx: &RequestContext,
6736 4 : ) -> anyhow::Result<()> {
6737 4 : if let Some((_, lsn)) = batch.first() {
6738 4 : let action = self.get_open_layer_action(*lsn, 0);
6739 4 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
6740 4 : layer.put_tombstones(batch).await?;
6741 0 : }
6742 :
6743 4 : Ok(())
6744 4 : }
6745 :
6746 : /// Track the end of the latest digested WAL record.
6747 : /// Remember the (end of) last valid WAL record remembered in the timeline.
6748 : ///
6749 : /// Call this after you have finished writing all the WAL up to 'lsn'.
6750 : ///
6751 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
6752 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
6753 : /// the latest and can be read.
6754 10558188 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
6755 10558188 : self.tl.finish_write(new_lsn);
6756 10558188 : }
6757 :
6758 541140 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
6759 541140 : self.tl.update_current_logical_size(delta)
6760 541140 : }
6761 : }
6762 :
6763 : // We need TimelineWriter to be send in upcoming conversion of
6764 : // Timeline::layers to tokio::sync::RwLock.
6765 : #[test]
6766 4 : fn is_send() {
6767 4 : fn _assert_send<T: Send>() {}
6768 4 : _assert_send::<TimelineWriter<'_>>();
6769 4 : }
6770 :
6771 : #[cfg(test)]
6772 : mod tests {
6773 : use std::sync::Arc;
6774 :
6775 : use pageserver_api::key::Key;
6776 : use pageserver_api::value::Value;
6777 : use tracing::Instrument;
6778 : use utils::{id::TimelineId, lsn::Lsn};
6779 :
6780 : use crate::tenant::{
6781 : harness::{test_img, TenantHarness},
6782 : layer_map::LayerMap,
6783 : storage_layer::{Layer, LayerName, LayerVisibilityHint},
6784 : timeline::{DeltaLayerTestDesc, EvictionError},
6785 : PreviousHeatmap, Timeline,
6786 : };
6787 :
6788 : use super::HeatMapTimeline;
6789 :
6790 20 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
6791 20 : assert_eq!(lhs.layers.len(), rhs.layers.len());
6792 20 : let lhs_rhs = lhs.layers.iter().zip(rhs.layers.iter());
6793 100 : for (l, r) in lhs_rhs {
6794 80 : assert_eq!(l.name, r.name);
6795 80 : assert_eq!(l.metadata, r.metadata);
6796 : }
6797 20 : }
6798 :
6799 : #[tokio::test]
6800 4 : async fn test_heatmap_generation() {
6801 4 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
6802 4 :
6803 4 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6804 4 : Lsn(0x10)..Lsn(0x20),
6805 4 : vec![(
6806 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6807 4 : Lsn(0x11),
6808 4 : Value::Image(test_img("foo")),
6809 4 : )],
6810 4 : );
6811 4 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6812 4 : Lsn(0x10)..Lsn(0x20),
6813 4 : vec![(
6814 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6815 4 : Lsn(0x11),
6816 4 : Value::Image(test_img("foo")),
6817 4 : )],
6818 4 : );
6819 4 : let l0_delta = DeltaLayerTestDesc::new(
6820 4 : Lsn(0x20)..Lsn(0x30),
6821 4 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6822 4 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6823 4 : vec![(
6824 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6825 4 : Lsn(0x25),
6826 4 : Value::Image(test_img("foo")),
6827 4 : )],
6828 4 : );
6829 4 : let delta_layers = vec![
6830 4 : covered_delta.clone(),
6831 4 : visible_delta.clone(),
6832 4 : l0_delta.clone(),
6833 4 : ];
6834 4 :
6835 4 : let image_layer = (
6836 4 : Lsn(0x40),
6837 4 : vec![(
6838 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6839 4 : test_img("bar"),
6840 4 : )],
6841 4 : );
6842 4 : let image_layers = vec![image_layer];
6843 4 :
6844 4 : let (tenant, ctx) = harness.load().await;
6845 4 : let timeline = tenant
6846 4 : .create_test_timeline_with_layers(
6847 4 : TimelineId::generate(),
6848 4 : Lsn(0x10),
6849 4 : 14,
6850 4 : &ctx,
6851 4 : delta_layers,
6852 4 : image_layers,
6853 4 : Lsn(0x100),
6854 4 : )
6855 4 : .await
6856 4 : .unwrap();
6857 4 :
6858 4 : // Layer visibility is an input to heatmap generation, so refresh it first
6859 4 : timeline.update_layer_visibility().await.unwrap();
6860 4 :
6861 4 : let heatmap = timeline
6862 4 : .generate_heatmap()
6863 4 : .await
6864 4 : .expect("Infallible while timeline is not shut down");
6865 4 :
6866 4 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
6867 4 :
6868 4 : // L0 should come last
6869 4 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
6870 4 :
6871 4 : let mut last_lsn = Lsn::MAX;
6872 20 : for layer in &heatmap.layers {
6873 4 : // Covered layer should be omitted
6874 16 : assert!(layer.name != covered_delta.layer_name());
6875 4 :
6876 16 : let layer_lsn = match &layer.name {
6877 8 : LayerName::Delta(d) => d.lsn_range.end,
6878 8 : LayerName::Image(i) => i.lsn,
6879 4 : };
6880 4 :
6881 4 : // Apart from L0s, newest Layers should come first
6882 16 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
6883 12 : assert!(layer_lsn <= last_lsn);
6884 12 : last_lsn = layer_lsn;
6885 4 : }
6886 4 : }
6887 4 :
6888 4 : // Evict all the layers and stash the old heatmap in the timeline.
6889 4 : // This simulates a migration to a cold secondary location.
6890 4 :
6891 4 : let guard = timeline.layers.read().await;
6892 4 : let mut all_layers = Vec::new();
6893 4 : let forever = std::time::Duration::from_secs(120);
6894 20 : for layer in guard.likely_resident_layers() {
6895 20 : all_layers.push(layer.clone());
6896 20 : layer.evict_and_wait(forever).await.unwrap();
6897 4 : }
6898 4 : drop(guard);
6899 4 :
6900 4 : timeline
6901 4 : .previous_heatmap
6902 4 : .store(Some(Arc::new(PreviousHeatmap::Active {
6903 4 : heatmap: heatmap.clone(),
6904 4 : read_at: std::time::Instant::now(),
6905 4 : })));
6906 4 :
6907 4 : // Generate a new heatmap and assert that it contains the same layers as the old one.
6908 4 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
6909 4 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
6910 4 :
6911 4 : // Download each layer one by one. Generate the heatmap at each step and check
6912 4 : // that it's stable.
6913 24 : for layer in all_layers {
6914 20 : if layer.visibility() == LayerVisibilityHint::Covered {
6915 4 : continue;
6916 16 : }
6917 16 :
6918 16 : eprintln!("Downloading {layer} and re-generating heatmap");
6919 4 :
6920 16 : let _resident = layer
6921 16 : .download_and_keep_resident()
6922 16 : .instrument(tracing::info_span!(
6923 16 : parent: None,
6924 4 : "download_layer",
6925 4 : tenant_id = %timeline.tenant_shard_id.tenant_id,
6926 0 : shard_id = %timeline.tenant_shard_id.shard_slug(),
6927 0 : timeline_id = %timeline.timeline_id
6928 4 : ))
6929 16 : .await
6930 16 : .unwrap();
6931 4 :
6932 16 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
6933 16 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
6934 4 : }
6935 4 :
6936 4 : // Everything from the post-migration heatmap is now resident.
6937 4 : // Check that we drop it from memory.
6938 4 : assert!(matches!(
6939 4 : timeline.previous_heatmap.load().as_deref(),
6940 4 : Some(PreviousHeatmap::Obsolete)
6941 4 : ));
6942 4 : }
6943 :
6944 : #[tokio::test]
6945 4 : async fn test_previous_heatmap_obsoletion() {
6946 4 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
6947 4 : .await
6948 4 : .unwrap();
6949 4 :
6950 4 : let l0_delta = DeltaLayerTestDesc::new(
6951 4 : Lsn(0x20)..Lsn(0x30),
6952 4 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6953 4 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6954 4 : vec![(
6955 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6956 4 : Lsn(0x25),
6957 4 : Value::Image(test_img("foo")),
6958 4 : )],
6959 4 : );
6960 4 :
6961 4 : let image_layer = (
6962 4 : Lsn(0x40),
6963 4 : vec![(
6964 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6965 4 : test_img("bar"),
6966 4 : )],
6967 4 : );
6968 4 :
6969 4 : let delta_layers = vec![l0_delta];
6970 4 : let image_layers = vec![image_layer];
6971 4 :
6972 4 : let (tenant, ctx) = harness.load().await;
6973 4 : let timeline = tenant
6974 4 : .create_test_timeline_with_layers(
6975 4 : TimelineId::generate(),
6976 4 : Lsn(0x10),
6977 4 : 14,
6978 4 : &ctx,
6979 4 : delta_layers,
6980 4 : image_layers,
6981 4 : Lsn(0x100),
6982 4 : )
6983 4 : .await
6984 4 : .unwrap();
6985 4 :
6986 4 : // Layer visibility is an input to heatmap generation, so refresh it first
6987 4 : timeline.update_layer_visibility().await.unwrap();
6988 4 :
6989 4 : let heatmap = timeline
6990 4 : .generate_heatmap()
6991 4 : .await
6992 4 : .expect("Infallible while timeline is not shut down");
6993 4 :
6994 4 : // Both layers should be in the heatmap
6995 4 : assert!(!heatmap.layers.is_empty());
6996 4 :
6997 4 : // Now simulate a migration.
6998 4 : timeline
6999 4 : .previous_heatmap
7000 4 : .store(Some(Arc::new(PreviousHeatmap::Active {
7001 4 : heatmap: heatmap.clone(),
7002 4 : read_at: std::time::Instant::now(),
7003 4 : })));
7004 4 :
7005 4 : // Evict all the layers in the previous heatmap
7006 4 : let guard = timeline.layers.read().await;
7007 4 : let forever = std::time::Duration::from_secs(120);
7008 12 : for layer in guard.likely_resident_layers() {
7009 12 : layer.evict_and_wait(forever).await.unwrap();
7010 4 : }
7011 4 : drop(guard);
7012 4 :
7013 4 : // Generate a new heatmap and check that the previous heatmap
7014 4 : // has been marked obsolete.
7015 4 : let post_eviction_heatmap = timeline
7016 4 : .generate_heatmap()
7017 4 : .await
7018 4 : .expect("Infallible while timeline is not shut down");
7019 4 :
7020 4 : assert!(post_eviction_heatmap.layers.is_empty());
7021 4 : assert!(matches!(
7022 4 : timeline.previous_heatmap.load().as_deref(),
7023 4 : Some(PreviousHeatmap::Obsolete)
7024 4 : ));
7025 4 : }
7026 :
7027 : #[tokio::test]
7028 4 : async fn two_layer_eviction_attempts_at_the_same_time() {
7029 4 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
7030 4 : .await
7031 4 : .unwrap();
7032 4 :
7033 4 : let (tenant, ctx) = harness.load().await;
7034 4 : let timeline = tenant
7035 4 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
7036 4 : .await
7037 4 : .unwrap();
7038 4 :
7039 4 : let layer = find_some_layer(&timeline).await;
7040 4 : let layer = layer
7041 4 : .keep_resident()
7042 4 : .await
7043 4 : .expect("no download => no downloading errors")
7044 4 : .drop_eviction_guard();
7045 4 :
7046 4 : let forever = std::time::Duration::from_secs(120);
7047 4 :
7048 4 : let first = layer.evict_and_wait(forever);
7049 4 : let second = layer.evict_and_wait(forever);
7050 4 :
7051 4 : let (first, second) = tokio::join!(first, second);
7052 4 :
7053 4 : let res = layer.keep_resident().await;
7054 4 : assert!(res.is_none(), "{res:?}");
7055 4 :
7056 4 : match (first, second) {
7057 4 : (Ok(()), Ok(())) => {
7058 4 : // because there are no more timeline locks being taken on eviction path, we can
7059 4 : // witness all three outcomes here.
7060 4 : }
7061 4 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
7062 0 : // if one completes before the other, this is fine just as well.
7063 0 : }
7064 4 : other => unreachable!("unexpected {:?}", other),
7065 4 : }
7066 4 : }
7067 :
7068 4 : async fn find_some_layer(timeline: &Timeline) -> Layer {
7069 4 : let layers = timeline.layers.read().await;
7070 4 : let desc = layers
7071 4 : .layer_map()
7072 4 : .unwrap()
7073 4 : .iter_historic_layers()
7074 4 : .next()
7075 4 : .expect("must find one layer to evict");
7076 4 :
7077 4 : layers.get_from_desc(&desc)
7078 4 : }
7079 : }
|
