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 enumset::EnumSet;
22 : use fail::fail_point;
23 : use handle::ShardTimelineId;
24 : use offload::OffloadError;
25 : use once_cell::sync::Lazy;
26 : use pageserver_api::models::PageTraceEvent;
27 : use pageserver_api::{
28 : config::tenant_conf_defaults::DEFAULT_COMPACTION_THRESHOLD,
29 : key::{
30 : KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
31 : SPARSE_RANGE,
32 : },
33 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
34 : models::{
35 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
36 : DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy,
37 : InMemoryLayerInfo, LayerMapInfo, LsnLease, TimelineState,
38 : },
39 : reltag::BlockNumber,
40 : shard::{ShardIdentity, ShardNumber, TenantShardId},
41 : };
42 : use rand::Rng;
43 : use remote_storage::DownloadError;
44 : use serde_with::serde_as;
45 : use storage_broker::BrokerClientChannel;
46 : use tokio::sync::mpsc::Sender;
47 : use tokio::{
48 : runtime::Handle,
49 : sync::{oneshot, watch},
50 : };
51 : use tokio_util::sync::CancellationToken;
52 : use tracing::*;
53 : use utils::{
54 : fs_ext,
55 : guard_arc_swap::GuardArcSwap,
56 : pausable_failpoint,
57 : postgres_client::PostgresClientProtocol,
58 : sync::gate::{Gate, GateGuard},
59 : };
60 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
61 :
62 : use std::sync::atomic::Ordering as AtomicOrdering;
63 : use std::sync::{Arc, Mutex, RwLock, Weak};
64 : use std::time::{Duration, Instant, SystemTime};
65 : use std::{
66 : array,
67 : collections::{BTreeMap, HashMap, HashSet},
68 : sync::atomic::AtomicU64,
69 : };
70 : use std::{cmp::min, ops::ControlFlow};
71 : use std::{
72 : collections::btree_map::Entry,
73 : ops::{Deref, Range},
74 : };
75 : use std::{pin::pin, sync::OnceLock};
76 :
77 : use crate::{
78 : aux_file::AuxFileSizeEstimator,
79 : tenant::{
80 : config::AttachmentMode,
81 : layer_map::{LayerMap, SearchResult},
82 : metadata::TimelineMetadata,
83 : storage_layer::{inmemory_layer::IndexEntry, PersistentLayerDesc},
84 : },
85 : walingest::WalLagCooldown,
86 : walredo,
87 : };
88 : use crate::{
89 : context::{DownloadBehavior, RequestContext},
90 : disk_usage_eviction_task::DiskUsageEvictionInfo,
91 : pgdatadir_mapping::CollectKeySpaceError,
92 : };
93 : use crate::{
94 : disk_usage_eviction_task::finite_f32,
95 : tenant::storage_layer::{
96 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
97 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
98 : ValuesReconstructState,
99 : },
100 : };
101 : use crate::{
102 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
103 : };
104 : use crate::{
105 : l0_flush::{self, L0FlushGlobalState},
106 : metrics::GetKind,
107 : };
108 : use crate::{
109 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
110 : };
111 : use crate::{
112 : pgdatadir_mapping::DirectoryKind,
113 : virtual_file::{MaybeFatalIo, VirtualFile},
114 : };
115 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
116 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
117 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
118 :
119 : use crate::config::PageServerConf;
120 : use crate::keyspace::{KeyPartitioning, KeySpace};
121 : use crate::metrics::TimelineMetrics;
122 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
123 : use crate::tenant::config::TenantConfOpt;
124 : use pageserver_api::reltag::RelTag;
125 : use pageserver_api::shard::ShardIndex;
126 :
127 : use postgres_connection::PgConnectionConfig;
128 : use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
129 : use utils::{
130 : completion,
131 : generation::Generation,
132 : id::TimelineId,
133 : lsn::{AtomicLsn, Lsn, RecordLsn},
134 : seqwait::SeqWait,
135 : simple_rcu::{Rcu, RcuReadGuard},
136 : };
137 :
138 : use crate::task_mgr;
139 : use crate::task_mgr::TaskKind;
140 : use crate::tenant::gc_result::GcResult;
141 : use crate::ZERO_PAGE;
142 : use pageserver_api::key::Key;
143 :
144 : use self::delete::DeleteTimelineFlow;
145 : pub(super) use self::eviction_task::EvictionTaskTenantState;
146 : use self::eviction_task::EvictionTaskTimelineState;
147 : use self::layer_manager::LayerManager;
148 : use self::logical_size::LogicalSize;
149 : use self::walreceiver::{WalReceiver, WalReceiverConf};
150 :
151 : use super::{
152 : config::TenantConf, storage_layer::LayerVisibilityHint, upload_queue::NotInitialized,
153 : MaybeOffloaded,
154 : };
155 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
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::{
162 : secondary::heatmap::{HeatMapLayer, HeatMapTimeline},
163 : GcError,
164 : };
165 :
166 : #[cfg(test)]
167 : use pageserver_api::value::Value;
168 :
169 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
170 : pub(crate) enum FlushLoopState {
171 : NotStarted,
172 : Running {
173 : #[cfg(test)]
174 : expect_initdb_optimization: bool,
175 : #[cfg(test)]
176 : initdb_optimization_count: usize,
177 : },
178 : Exited,
179 : }
180 :
181 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
182 : pub enum ImageLayerCreationMode {
183 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
184 : Try,
185 : /// Force creating the image layers if possible. For now, no image layers will be created
186 : /// for metadata keys. Used in compaction code path with force flag enabled.
187 : Force,
188 : /// Initial ingestion of the data, and no data should be dropped in this function. This
189 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
190 : /// code path.
191 : Initial,
192 : }
193 :
194 : impl std::fmt::Display for ImageLayerCreationMode {
195 568 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
196 568 : write!(f, "{:?}", self)
197 568 : }
198 : }
199 :
200 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
201 : /// Can be removed after all refactors are done.
202 28 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
203 28 : drop(rlock)
204 28 : }
205 :
206 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
207 : /// Can be removed after all refactors are done.
208 596 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
209 596 : drop(rlock)
210 596 : }
211 :
212 : /// The outward-facing resources required to build a Timeline
213 : pub struct TimelineResources {
214 : pub remote_client: RemoteTimelineClient,
215 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
216 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
217 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
218 : }
219 :
220 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
221 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
222 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
223 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
224 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
225 : pub(crate) struct RelSizeCache {
226 : pub(crate) complete_as_of: Lsn,
227 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
228 : }
229 :
230 : pub struct Timeline {
231 : pub(crate) conf: &'static PageServerConf,
232 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
233 :
234 : myself: Weak<Self>,
235 :
236 : pub(crate) tenant_shard_id: TenantShardId,
237 : pub timeline_id: TimelineId,
238 :
239 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
240 : /// Never changes for the lifetime of this [`Timeline`] object.
241 : ///
242 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
243 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
244 : pub(crate) generation: Generation,
245 :
246 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
247 : /// to shards, and is constant through the lifetime of this Timeline.
248 : shard_identity: ShardIdentity,
249 :
250 : pub pg_version: u32,
251 :
252 : /// The tuple has two elements.
253 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
254 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
255 : ///
256 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
257 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
258 : ///
259 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
260 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
261 : /// `PersistentLayerDesc`'s.
262 : ///
263 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
264 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
265 : /// runtime, e.g., during page reconstruction.
266 : ///
267 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
268 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
269 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
270 :
271 : last_freeze_at: AtomicLsn,
272 : // Atomic would be more appropriate here.
273 : last_freeze_ts: RwLock<Instant>,
274 :
275 : pub(crate) standby_horizon: AtomicLsn,
276 :
277 : // WAL redo manager. `None` only for broken tenants.
278 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
279 :
280 : /// Remote storage client.
281 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
282 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
283 :
284 : // What page versions do we hold in the repository? If we get a
285 : // request > last_record_lsn, we need to wait until we receive all
286 : // the WAL up to the request. The SeqWait provides functions for
287 : // that. TODO: If we get a request for an old LSN, such that the
288 : // versions have already been garbage collected away, we should
289 : // throw an error, but we don't track that currently.
290 : //
291 : // last_record_lsn.load().last points to the end of last processed WAL record.
292 : //
293 : // We also remember the starting point of the previous record in
294 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
295 : // first WAL record when the node is started up. But here, we just
296 : // keep track of it.
297 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
298 :
299 : // All WAL records have been processed and stored durably on files on
300 : // local disk, up to this LSN. On crash and restart, we need to re-process
301 : // the WAL starting from this point.
302 : //
303 : // Some later WAL records might have been processed and also flushed to disk
304 : // already, so don't be surprised to see some, but there's no guarantee on
305 : // them yet.
306 : disk_consistent_lsn: AtomicLsn,
307 :
308 : // Parent timeline that this timeline was branched from, and the LSN
309 : // of the branch point.
310 : ancestor_timeline: Option<Arc<Timeline>>,
311 : ancestor_lsn: Lsn,
312 :
313 : pub(super) metrics: TimelineMetrics,
314 :
315 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
316 : // in `crate::page_service` writes these metrics.
317 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
318 :
319 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
320 :
321 : /// Ensures layers aren't frozen by checkpointer between
322 : /// [`Timeline::get_layer_for_write`] and layer reads.
323 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
324 : /// Must always be acquired before the layer map/individual layer lock
325 : /// to avoid deadlock.
326 : ///
327 : /// The state is cleared upon freezing.
328 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
329 :
330 : /// Used to avoid multiple `flush_loop` tasks running
331 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
332 :
333 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
334 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
335 : /// The flush cycle counter is sent back on the layer_flush_done channel when
336 : /// the flush finishes. You can use that to wait for the flush to finish.
337 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
338 : /// read by whoever sends an update
339 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
340 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
341 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
342 :
343 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
344 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
345 :
346 : // List of child timelines and their branch points. This is needed to avoid
347 : // garbage collecting data that is still needed by the child timelines.
348 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
349 :
350 : // It may change across major versions so for simplicity
351 : // keep it after running initdb for a timeline.
352 : // It is needed in checks when we want to error on some operations
353 : // when they are requested for pre-initdb lsn.
354 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
355 : // though let's keep them both for better error visibility.
356 : pub initdb_lsn: Lsn,
357 :
358 : /// The repartitioning result. Allows a single writer and multiple readers.
359 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
360 :
361 : /// Configuration: how often should the partitioning be recalculated.
362 : repartition_threshold: u64,
363 :
364 : last_image_layer_creation_check_at: AtomicLsn,
365 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
366 :
367 : /// Current logical size of the "datadir", at the last LSN.
368 : current_logical_size: LogicalSize,
369 :
370 : /// Information about the last processed message by the WAL receiver,
371 : /// or None if WAL receiver has not received anything for this timeline
372 : /// yet.
373 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
374 : pub walreceiver: Mutex<Option<WalReceiver>>,
375 :
376 : /// Relation size cache
377 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
378 :
379 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
380 :
381 : state: watch::Sender<TimelineState>,
382 :
383 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
384 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
385 : pub delete_progress: TimelineDeleteProgress,
386 :
387 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
388 :
389 : /// Load or creation time information about the disk_consistent_lsn and when the loading
390 : /// happened. Used for consumption metrics.
391 : pub(crate) loaded_at: (Lsn, SystemTime),
392 :
393 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
394 : pub(crate) gate: Gate,
395 :
396 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
397 : /// to the timeline should drop out when this token fires.
398 : pub(crate) cancel: CancellationToken,
399 :
400 : /// Make sure we only have one running compaction at a time in tests.
401 : ///
402 : /// Must only be taken in two places:
403 : /// - [`Timeline::compact`] (this file)
404 : /// - [`delete::delete_local_timeline_directory`]
405 : ///
406 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
407 : compaction_lock: tokio::sync::Mutex<()>,
408 :
409 : /// Make sure we only have one running gc at a time.
410 : ///
411 : /// Must only be taken in two places:
412 : /// - [`Timeline::gc`] (this file)
413 : /// - [`delete::delete_local_timeline_directory`]
414 : ///
415 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
416 : gc_lock: tokio::sync::Mutex<()>,
417 :
418 : /// Cloned from [`super::Tenant::pagestream_throttle`] on construction.
419 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
420 :
421 : /// Size estimator for aux file v2
422 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
423 :
424 : /// Some test cases directly place keys into the timeline without actually modifying the directory
425 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
426 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
427 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
428 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
429 : #[cfg(test)]
430 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
431 :
432 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
433 :
434 : pub(crate) handles: handle::PerTimelineState<crate::page_service::TenantManagerTypes>,
435 :
436 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
437 :
438 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
439 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
440 :
441 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
442 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
443 : }
444 :
445 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
446 :
447 : pub struct WalReceiverInfo {
448 : pub wal_source_connconf: PgConnectionConfig,
449 : pub last_received_msg_lsn: Lsn,
450 : pub last_received_msg_ts: u128,
451 : }
452 :
453 : /// Information about how much history needs to be retained, needed by
454 : /// Garbage Collection.
455 : #[derive(Default)]
456 : pub(crate) struct GcInfo {
457 : /// Specific LSNs that are needed.
458 : ///
459 : /// Currently, this includes all points where child branches have
460 : /// been forked off from. In the future, could also include
461 : /// explicit user-defined snapshot points.
462 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
463 :
464 : /// The cutoff coordinates, which are combined by selecting the minimum.
465 : pub(crate) cutoffs: GcCutoffs,
466 :
467 : /// Leases granted to particular LSNs.
468 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
469 :
470 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
471 : pub(crate) within_ancestor_pitr: bool,
472 : }
473 :
474 : impl GcInfo {
475 282 : pub(crate) fn min_cutoff(&self) -> Lsn {
476 282 : self.cutoffs.select_min()
477 282 : }
478 :
479 232 : pub(super) fn insert_child(
480 232 : &mut self,
481 232 : child_id: TimelineId,
482 232 : child_lsn: Lsn,
483 232 : is_offloaded: MaybeOffloaded,
484 232 : ) {
485 232 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
486 232 : self.retain_lsns.sort_by_key(|i| i.0);
487 232 : }
488 :
489 4 : pub(super) fn remove_child_maybe_offloaded(
490 4 : &mut self,
491 4 : child_id: TimelineId,
492 4 : maybe_offloaded: MaybeOffloaded,
493 4 : ) -> bool {
494 4 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
495 4 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
496 4 : let mut removed = false;
497 6 : self.retain_lsns.retain(|i| {
498 6 : if removed {
499 2 : return true;
500 4 : }
501 4 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
502 4 : removed |= remove;
503 4 : !remove
504 6 : });
505 4 : removed
506 4 : }
507 :
508 4 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
509 4 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
510 4 : }
511 :
512 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
513 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
514 0 : }
515 : }
516 :
517 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
518 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
519 : /// between time-based and space-based retention for observability and consumption metrics purposes.
520 : #[derive(Debug, Clone)]
521 : pub(crate) struct GcCutoffs {
522 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
523 : /// history we must keep to retain a specified number of bytes of WAL.
524 : pub(crate) space: Lsn,
525 :
526 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
527 : /// history we must keep to enable reading back at least the PITR interval duration.
528 : pub(crate) time: Lsn,
529 : }
530 :
531 : impl Default for GcCutoffs {
532 446 : fn default() -> Self {
533 446 : Self {
534 446 : space: Lsn::INVALID,
535 446 : time: Lsn::INVALID,
536 446 : }
537 446 : }
538 : }
539 :
540 : impl GcCutoffs {
541 282 : fn select_min(&self) -> Lsn {
542 282 : std::cmp::min(self.space, self.time)
543 282 : }
544 : }
545 :
546 : pub(crate) struct TimelineVisitOutcome {
547 : completed_keyspace: KeySpace,
548 : image_covered_keyspace: KeySpace,
549 : }
550 :
551 : /// An error happened in a get() operation.
552 : #[derive(thiserror::Error, Debug)]
553 : pub(crate) enum PageReconstructError {
554 : #[error(transparent)]
555 : Other(anyhow::Error),
556 :
557 : #[error("Ancestor LSN wait error: {0}")]
558 : AncestorLsnTimeout(WaitLsnError),
559 :
560 : #[error("timeline shutting down")]
561 : Cancelled,
562 :
563 : /// An error happened replaying WAL records
564 : #[error(transparent)]
565 : WalRedo(anyhow::Error),
566 :
567 : #[error("{0}")]
568 : MissingKey(MissingKeyError),
569 : }
570 :
571 : impl From<anyhow::Error> for PageReconstructError {
572 0 : fn from(value: anyhow::Error) -> Self {
573 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
574 0 : match value.downcast::<PageReconstructError>() {
575 0 : Ok(pre) => pre,
576 0 : Err(other) => PageReconstructError::Other(other),
577 : }
578 0 : }
579 : }
580 :
581 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
582 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
583 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
584 0 : }
585 : }
586 :
587 : impl From<layer_manager::Shutdown> for PageReconstructError {
588 0 : fn from(_: layer_manager::Shutdown) -> Self {
589 0 : PageReconstructError::Cancelled
590 0 : }
591 : }
592 :
593 : impl GetVectoredError {
594 : #[cfg(test)]
595 6 : pub(crate) fn is_missing_key_error(&self) -> bool {
596 6 : matches!(self, Self::MissingKey(_))
597 6 : }
598 : }
599 :
600 : impl From<layer_manager::Shutdown> for GetVectoredError {
601 0 : fn from(_: layer_manager::Shutdown) -> Self {
602 0 : GetVectoredError::Cancelled
603 0 : }
604 : }
605 :
606 : #[derive(thiserror::Error)]
607 : pub struct MissingKeyError {
608 : key: Key,
609 : shard: ShardNumber,
610 : cont_lsn: Lsn,
611 : request_lsn: Lsn,
612 : ancestor_lsn: Option<Lsn>,
613 : backtrace: Option<std::backtrace::Backtrace>,
614 : }
615 :
616 : impl std::fmt::Debug for MissingKeyError {
617 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
618 0 : write!(f, "{}", self)
619 0 : }
620 : }
621 :
622 : impl std::fmt::Display for MissingKeyError {
623 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
624 0 : write!(
625 0 : f,
626 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
627 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
628 0 : )?;
629 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
630 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
631 0 : }
632 :
633 0 : if let Some(ref backtrace) = self.backtrace {
634 0 : write!(f, "\n{}", backtrace)?;
635 0 : }
636 :
637 0 : Ok(())
638 0 : }
639 : }
640 :
641 : impl PageReconstructError {
642 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
643 0 : pub(crate) fn is_stopping(&self) -> bool {
644 : use PageReconstructError::*;
645 0 : match self {
646 0 : Cancelled => true,
647 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
648 : }
649 0 : }
650 : }
651 :
652 : #[derive(thiserror::Error, Debug)]
653 : pub(crate) enum CreateImageLayersError {
654 : #[error("timeline shutting down")]
655 : Cancelled,
656 :
657 : #[error("read failed")]
658 : GetVectoredError(#[source] GetVectoredError),
659 :
660 : #[error("reconstruction failed")]
661 : PageReconstructError(#[source] PageReconstructError),
662 :
663 : #[error(transparent)]
664 : Other(#[from] anyhow::Error),
665 : }
666 :
667 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
668 0 : fn from(_: layer_manager::Shutdown) -> Self {
669 0 : CreateImageLayersError::Cancelled
670 0 : }
671 : }
672 :
673 : #[derive(thiserror::Error, Debug, Clone)]
674 : pub(crate) enum FlushLayerError {
675 : /// Timeline cancellation token was cancelled
676 : #[error("timeline shutting down")]
677 : Cancelled,
678 :
679 : /// We tried to flush a layer while the Timeline is in an unexpected state
680 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
681 : NotRunning(FlushLoopState),
682 :
683 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
684 : // loop via a watch channel, where we can only borrow it.
685 : #[error("create image layers (shared)")]
686 : CreateImageLayersError(Arc<CreateImageLayersError>),
687 :
688 : #[error("other (shared)")]
689 : Other(#[from] Arc<anyhow::Error>),
690 : }
691 :
692 : impl FlushLayerError {
693 : // When crossing from generic anyhow errors to this error type, we explicitly check
694 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
695 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
696 0 : let cancelled = timeline.cancel.is_cancelled()
697 : // The upload queue might have been shut down before the official cancellation of the timeline.
698 0 : || err
699 0 : .downcast_ref::<NotInitialized>()
700 0 : .map(NotInitialized::is_stopping)
701 0 : .unwrap_or_default();
702 0 : if cancelled {
703 0 : Self::Cancelled
704 : } else {
705 0 : Self::Other(Arc::new(err))
706 : }
707 0 : }
708 : }
709 :
710 : impl From<layer_manager::Shutdown> for FlushLayerError {
711 0 : fn from(_: layer_manager::Shutdown) -> Self {
712 0 : FlushLayerError::Cancelled
713 0 : }
714 : }
715 :
716 : #[derive(thiserror::Error, Debug)]
717 : pub(crate) enum GetVectoredError {
718 : #[error("timeline shutting down")]
719 : Cancelled,
720 :
721 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
722 : Oversized(u64),
723 :
724 : #[error("requested at invalid LSN: {0}")]
725 : InvalidLsn(Lsn),
726 :
727 : #[error("requested key not found: {0}")]
728 : MissingKey(MissingKeyError),
729 :
730 : #[error("ancestry walk")]
731 : GetReadyAncestorError(#[source] GetReadyAncestorError),
732 :
733 : #[error(transparent)]
734 : Other(#[from] anyhow::Error),
735 : }
736 :
737 : impl From<GetReadyAncestorError> for GetVectoredError {
738 2 : fn from(value: GetReadyAncestorError) -> Self {
739 : use GetReadyAncestorError::*;
740 2 : match value {
741 0 : Cancelled => GetVectoredError::Cancelled,
742 : AncestorLsnTimeout(_) | BadState { .. } => {
743 2 : GetVectoredError::GetReadyAncestorError(value)
744 : }
745 : }
746 2 : }
747 : }
748 :
749 : #[derive(thiserror::Error, Debug)]
750 : pub(crate) enum GetReadyAncestorError {
751 : #[error("ancestor LSN wait error")]
752 : AncestorLsnTimeout(#[from] WaitLsnError),
753 :
754 : #[error("bad state on timeline {timeline_id}: {state:?}")]
755 : BadState {
756 : timeline_id: TimelineId,
757 : state: TimelineState,
758 : },
759 :
760 : #[error("cancelled")]
761 : Cancelled,
762 : }
763 :
764 : #[derive(Clone, Copy)]
765 : pub enum LogicalSizeCalculationCause {
766 : Initial,
767 : ConsumptionMetricsSyntheticSize,
768 : EvictionTaskImitation,
769 : TenantSizeHandler,
770 : }
771 :
772 : pub enum GetLogicalSizePriority {
773 : User,
774 : Background,
775 : }
776 :
777 0 : #[derive(Debug, enumset::EnumSetType)]
778 : pub(crate) enum CompactFlags {
779 : ForceRepartition,
780 : ForceImageLayerCreation,
781 : ForceL0Compaction,
782 : EnhancedGcBottomMostCompaction,
783 : DryRun,
784 : }
785 :
786 : #[serde_with::serde_as]
787 0 : #[derive(Debug, Clone, serde::Deserialize)]
788 : pub(crate) struct CompactRequest {
789 : pub compact_key_range: Option<CompactKeyRange>,
790 : pub compact_lsn_range: Option<CompactLsnRange>,
791 : /// Whether the compaction job should be scheduled.
792 : #[serde(default)]
793 : pub scheduled: bool,
794 : /// Whether the compaction job should be split across key ranges.
795 : #[serde(default)]
796 : pub sub_compaction: bool,
797 : /// Max job size for each subcompaction job.
798 : pub sub_compaction_max_job_size_mb: Option<u64>,
799 : }
800 :
801 : #[derive(Debug, Clone, Default)]
802 : pub(crate) struct CompactOptions {
803 : pub flags: EnumSet<CompactFlags>,
804 : /// If set, the compaction will only compact the key range specified by this option.
805 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
806 : pub compact_key_range: Option<CompactKeyRange>,
807 : /// If set, the compaction will only compact the LSN within this value.
808 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
809 : pub compact_lsn_range: Option<CompactLsnRange>,
810 : /// Enable sub-compaction (split compaction job across key ranges).
811 : /// This option is only used by GC compaction.
812 : pub sub_compaction: bool,
813 : /// Set job size for the GC compaction.
814 : /// This option is only used by GC compaction.
815 : pub sub_compaction_max_job_size_mb: Option<u64>,
816 : }
817 :
818 : impl std::fmt::Debug for Timeline {
819 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
820 0 : write!(f, "Timeline<{}>", self.timeline_id)
821 0 : }
822 : }
823 :
824 : #[derive(thiserror::Error, Debug)]
825 : pub(crate) enum WaitLsnError {
826 : // Called on a timeline which is shutting down
827 : #[error("Shutdown")]
828 : Shutdown,
829 :
830 : // Called on an timeline not in active state or shutting down
831 : #[error("Bad timeline state: {0:?}")]
832 : BadState(TimelineState),
833 :
834 : // Timeout expired while waiting for LSN to catch up with goal.
835 : #[error("{0}")]
836 : Timeout(String),
837 : }
838 :
839 : // The impls below achieve cancellation mapping for errors.
840 : // Perhaps there's a way of achieving this with less cruft.
841 :
842 : impl From<CreateImageLayersError> for CompactionError {
843 0 : fn from(e: CreateImageLayersError) -> Self {
844 0 : match e {
845 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
846 0 : CreateImageLayersError::Other(e) => {
847 0 : CompactionError::Other(e.context("create image layers"))
848 : }
849 0 : _ => CompactionError::Other(e.into()),
850 : }
851 0 : }
852 : }
853 :
854 : impl From<CreateImageLayersError> for FlushLayerError {
855 0 : fn from(e: CreateImageLayersError) -> Self {
856 0 : match e {
857 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
858 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
859 : }
860 0 : }
861 : }
862 :
863 : impl From<PageReconstructError> for CreateImageLayersError {
864 0 : fn from(e: PageReconstructError) -> Self {
865 0 : match e {
866 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
867 0 : _ => CreateImageLayersError::PageReconstructError(e),
868 : }
869 0 : }
870 : }
871 :
872 : impl From<GetVectoredError> for CreateImageLayersError {
873 0 : fn from(e: GetVectoredError) -> Self {
874 0 : match e {
875 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
876 0 : _ => CreateImageLayersError::GetVectoredError(e),
877 : }
878 0 : }
879 : }
880 :
881 : impl From<GetVectoredError> for PageReconstructError {
882 6 : fn from(e: GetVectoredError) -> Self {
883 6 : match e {
884 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
885 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
886 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
887 4 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
888 2 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
889 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
890 : }
891 6 : }
892 : }
893 :
894 : impl From<GetReadyAncestorError> for PageReconstructError {
895 2 : fn from(e: GetReadyAncestorError) -> Self {
896 : use GetReadyAncestorError::*;
897 2 : match e {
898 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
899 2 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
900 0 : Cancelled => PageReconstructError::Cancelled,
901 : }
902 2 : }
903 : }
904 :
905 : pub(crate) enum WaitLsnWaiter<'a> {
906 : Timeline(&'a Timeline),
907 : Tenant,
908 : PageService,
909 : }
910 :
911 : /// Argument to [`Timeline::shutdown`].
912 : #[derive(Debug, Clone, Copy)]
913 : pub(crate) enum ShutdownMode {
914 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
915 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
916 : ///
917 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
918 : /// the call to [`Timeline::shutdown`].
919 : FreezeAndFlush,
920 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
921 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
922 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
923 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
924 : Reload,
925 : /// Shut down immediately, without waiting for any open layers to flush.
926 : Hard,
927 : }
928 :
929 : struct ImageLayerCreationOutcome {
930 : image: Option<ResidentLayer>,
931 : next_start_key: Key,
932 : }
933 :
934 : /// Public interface functions
935 : impl Timeline {
936 : /// Get the LSN where this branch was created
937 4 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
938 4 : self.ancestor_lsn
939 4 : }
940 :
941 : /// Get the ancestor's timeline id
942 12 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
943 12 : self.ancestor_timeline
944 12 : .as_ref()
945 12 : .map(|ancestor| ancestor.timeline_id)
946 12 : }
947 :
948 : /// Get the ancestor timeline
949 2 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
950 2 : self.ancestor_timeline.as_ref()
951 2 : }
952 :
953 : /// Get the bytes written since the PITR cutoff on this branch, and
954 : /// whether this branch's ancestor_lsn is within its parent's PITR.
955 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
956 0 : let gc_info = self.gc_info.read().unwrap();
957 0 : let history = self
958 0 : .get_last_record_lsn()
959 0 : .checked_sub(gc_info.cutoffs.time)
960 0 : .unwrap_or(Lsn(0))
961 0 : .0;
962 0 : (history, gc_info.within_ancestor_pitr)
963 0 : }
964 :
965 : /// Lock and get timeline's GC cutoff
966 302 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
967 302 : self.latest_gc_cutoff_lsn.read()
968 302 : }
969 :
970 : /// Look up given page version.
971 : ///
972 : /// If a remote layer file is needed, it is downloaded as part of this
973 : /// call.
974 : ///
975 : /// This method enforces [`Self::pagestream_throttle`] internally.
976 : ///
977 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
978 : /// abstraction above this needs to store suitable metadata to track what
979 : /// data exists with what keys, in separate metadata entries. If a
980 : /// non-existent key is requested, we may incorrectly return a value from
981 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
982 : /// non-existing key.
983 : ///
984 : /// # Cancel-Safety
985 : ///
986 : /// This method is cancellation-safe.
987 : #[inline(always)]
988 607565 : pub(crate) async fn get(
989 607565 : &self,
990 607565 : key: Key,
991 607565 : lsn: Lsn,
992 607565 : ctx: &RequestContext,
993 607565 : ) -> Result<Bytes, PageReconstructError> {
994 607565 : if !lsn.is_valid() {
995 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
996 607565 : }
997 607565 :
998 607565 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
999 607565 : // already checked the key against the shard_identity when looking up the Timeline from
1000 607565 : // page_service.
1001 607565 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1002 :
1003 607565 : let keyspace = KeySpace {
1004 607565 : ranges: vec![key..key.next()],
1005 607565 : };
1006 607565 :
1007 607565 : // Initialise the reconstruct state for the key with the cache
1008 607565 : // entry returned above.
1009 607565 : let mut reconstruct_state = ValuesReconstructState::new();
1010 :
1011 607565 : let vectored_res = self
1012 607565 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
1013 607565 : .await;
1014 :
1015 607565 : let key_value = vectored_res?.pop_first();
1016 607559 : match key_value {
1017 607547 : Some((got_key, value)) => {
1018 607547 : if got_key != key {
1019 0 : error!(
1020 0 : "Expected {}, but singular vectored get returned {}",
1021 : key, got_key
1022 : );
1023 0 : Err(PageReconstructError::Other(anyhow!(
1024 0 : "Singular vectored get returned wrong key"
1025 0 : )))
1026 : } else {
1027 607547 : value
1028 : }
1029 : }
1030 12 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1031 12 : key,
1032 12 : shard: self.shard_identity.get_shard_number(&key),
1033 12 : cont_lsn: Lsn(0),
1034 12 : request_lsn: lsn,
1035 12 : ancestor_lsn: None,
1036 12 : backtrace: None,
1037 12 : })),
1038 : }
1039 607565 : }
1040 :
1041 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1042 : pub(crate) const VEC_GET_LAYERS_VISITED_WARN_THRESH: f64 = 512.0;
1043 :
1044 : /// Look up multiple page versions at a given LSN
1045 : ///
1046 : /// This naive implementation will be replaced with a more efficient one
1047 : /// which actually vectorizes the read path.
1048 19678 : pub(crate) async fn get_vectored(
1049 19678 : &self,
1050 19678 : keyspace: KeySpace,
1051 19678 : lsn: Lsn,
1052 19678 : ctx: &RequestContext,
1053 19678 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1054 19678 : if !lsn.is_valid() {
1055 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1056 19678 : }
1057 19678 :
1058 19678 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1059 19678 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1060 0 : return Err(GetVectoredError::Oversized(key_count));
1061 19678 : }
1062 :
1063 39356 : for range in &keyspace.ranges {
1064 19678 : let mut key = range.start;
1065 39590 : while key != range.end {
1066 19912 : assert!(!self.shard_identity.is_key_disposable(&key));
1067 19912 : key = key.next();
1068 : }
1069 : }
1070 :
1071 19678 : trace!(
1072 0 : "get vectored request for {:?}@{} from task kind {:?}",
1073 0 : keyspace,
1074 0 : lsn,
1075 0 : ctx.task_kind(),
1076 : );
1077 :
1078 19678 : let start = crate::metrics::GET_VECTORED_LATENCY
1079 19678 : .for_task_kind(ctx.task_kind())
1080 19678 : .map(|metric| (metric, Instant::now()));
1081 :
1082 19678 : let res = self
1083 19678 : .get_vectored_impl(
1084 19678 : keyspace.clone(),
1085 19678 : lsn,
1086 19678 : &mut ValuesReconstructState::new(),
1087 19678 : ctx,
1088 19678 : )
1089 19678 : .await;
1090 :
1091 19678 : if let Some((metric, start)) = start {
1092 0 : let elapsed = start.elapsed();
1093 0 : metric.observe(elapsed.as_secs_f64());
1094 19678 : }
1095 :
1096 19678 : res
1097 19678 : }
1098 :
1099 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1100 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1101 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1102 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1103 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1104 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1105 : /// the scan operation will not cause OOM in the future.
1106 12 : pub(crate) async fn scan(
1107 12 : &self,
1108 12 : keyspace: KeySpace,
1109 12 : lsn: Lsn,
1110 12 : ctx: &RequestContext,
1111 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1112 12 : if !lsn.is_valid() {
1113 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1114 12 : }
1115 12 :
1116 12 : trace!(
1117 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1118 0 : keyspace,
1119 0 : lsn,
1120 0 : ctx.task_kind()
1121 : );
1122 :
1123 : // We should generalize this into Keyspace::contains in the future.
1124 24 : for range in &keyspace.ranges {
1125 12 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1126 12 : || range.end.field1 > METADATA_KEY_END_PREFIX
1127 : {
1128 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1129 0 : "only metadata keyspace can be scanned"
1130 0 : )));
1131 12 : }
1132 : }
1133 :
1134 12 : let start = crate::metrics::SCAN_LATENCY
1135 12 : .for_task_kind(ctx.task_kind())
1136 12 : .map(ScanLatencyOngoingRecording::start_recording);
1137 :
1138 12 : let vectored_res = self
1139 12 : .get_vectored_impl(
1140 12 : keyspace.clone(),
1141 12 : lsn,
1142 12 : &mut ValuesReconstructState::default(),
1143 12 : ctx,
1144 12 : )
1145 12 : .await;
1146 :
1147 12 : if let Some(recording) = start {
1148 0 : recording.observe();
1149 12 : }
1150 :
1151 12 : vectored_res
1152 12 : }
1153 :
1154 627609 : pub(super) async fn get_vectored_impl(
1155 627609 : &self,
1156 627609 : keyspace: KeySpace,
1157 627609 : lsn: Lsn,
1158 627609 : reconstruct_state: &mut ValuesReconstructState,
1159 627609 : ctx: &RequestContext,
1160 627609 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1161 627609 : let get_kind = if keyspace.total_raw_size() == 1 {
1162 627113 : GetKind::Singular
1163 : } else {
1164 496 : GetKind::Vectored
1165 : };
1166 :
1167 627609 : let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
1168 627609 : .for_get_kind(get_kind)
1169 627609 : .start_timer();
1170 627609 : self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1171 627609 : .await?;
1172 627593 : get_data_timer.stop_and_record();
1173 627593 :
1174 627593 : let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
1175 627593 : .for_get_kind(get_kind)
1176 627593 : .start_timer();
1177 627593 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
1178 627593 : let layers_visited = reconstruct_state.get_layers_visited();
1179 :
1180 667927 : for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
1181 667927 : match res {
1182 0 : Err(err) => {
1183 0 : results.insert(key, Err(err));
1184 0 : }
1185 667927 : Ok(state) => {
1186 667927 : let state = ValueReconstructState::from(state);
1187 :
1188 667927 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
1189 667927 : results.insert(key, reconstruct_res);
1190 : }
1191 : }
1192 : }
1193 627593 : reconstruct_timer.stop_and_record();
1194 627593 :
1195 627593 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1196 627593 : // when they're missing. Instead they are omitted from the resulting btree
1197 627593 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1198 627593 : // to avoid infinite results.
1199 627593 : if !results.is_empty() {
1200 627353 : let avg = layers_visited as f64 / results.len() as f64;
1201 627353 : if avg >= Self::VEC_GET_LAYERS_VISITED_WARN_THRESH {
1202 0 : use utils::rate_limit::RateLimit;
1203 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1204 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1205 0 : let mut rate_limit = LOGGED.lock().unwrap();
1206 0 : rate_limit.call(|| {
1207 0 : tracing::info!(
1208 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1209 0 : lsn = %lsn,
1210 0 : "Vectored read for {} visited {} layers on average per key and {} in total. {}/{} pages were returned",
1211 0 : keyspace, avg, layers_visited, results.len(), keyspace.total_raw_size());
1212 0 : });
1213 627353 : }
1214 :
1215 : // Note that this is an approximation. Tracking the exact number of layers visited
1216 : // per key requires virtually unbounded memory usage and is inefficient
1217 : // (i.e. segment tree tracking each range queried from a layer)
1218 627353 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED.observe(avg);
1219 240 : }
1220 :
1221 627593 : Ok(results)
1222 627609 : }
1223 :
1224 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1225 274422 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1226 274422 : self.last_record_lsn.load().last
1227 274422 : }
1228 :
1229 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1230 0 : self.last_record_lsn.load().prev
1231 0 : }
1232 :
1233 : /// Atomically get both last and prev.
1234 228 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1235 228 : self.last_record_lsn.load()
1236 228 : }
1237 :
1238 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1239 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1240 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1241 0 : self.last_record_lsn.status_receiver()
1242 0 : }
1243 :
1244 442 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1245 442 : self.disk_consistent_lsn.load()
1246 442 : }
1247 :
1248 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1249 : /// not validated with control plane yet.
1250 : /// See [`Self::get_remote_consistent_lsn_visible`].
1251 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1252 0 : self.remote_client.remote_consistent_lsn_projected()
1253 0 : }
1254 :
1255 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1256 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1257 : /// generation validation in the deletion queue.
1258 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1259 0 : self.remote_client.remote_consistent_lsn_visible()
1260 0 : }
1261 :
1262 : /// The sum of the file size of all historic layers in the layer map.
1263 : /// This method makes no distinction between local and remote layers.
1264 : /// Hence, the result **does not represent local filesystem usage**.
1265 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1266 0 : let guard = self.layers.read().await;
1267 0 : guard.layer_size_sum()
1268 0 : }
1269 :
1270 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1271 0 : self.metrics.resident_physical_size_get()
1272 0 : }
1273 :
1274 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1275 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1276 0 : }
1277 :
1278 : ///
1279 : /// Wait until WAL has been received and processed up to this LSN.
1280 : ///
1281 : /// You should call this before any of the other get_* or list_* functions. Calling
1282 : /// those functions with an LSN that has been processed yet is an error.
1283 : ///
1284 224743 : pub(crate) async fn wait_lsn(
1285 224743 : &self,
1286 224743 : lsn: Lsn,
1287 224743 : who_is_waiting: WaitLsnWaiter<'_>,
1288 224743 : ctx: &RequestContext, /* Prepare for use by cancellation */
1289 224743 : ) -> Result<(), WaitLsnError> {
1290 224743 : let state = self.current_state();
1291 224743 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1292 0 : return Err(WaitLsnError::Shutdown);
1293 224743 : } else if !matches!(state, TimelineState::Active) {
1294 0 : return Err(WaitLsnError::BadState(state));
1295 224743 : }
1296 224743 :
1297 224743 : if cfg!(debug_assertions) {
1298 224743 : match ctx.task_kind() {
1299 : TaskKind::WalReceiverManager
1300 : | TaskKind::WalReceiverConnectionHandler
1301 : | TaskKind::WalReceiverConnectionPoller => {
1302 0 : let is_myself = match who_is_waiting {
1303 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1304 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1305 : };
1306 0 : if is_myself {
1307 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1308 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1309 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1310 0 : }
1311 0 : } else {
1312 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1313 0 : // our walreceiver task can make progress independent of theirs
1314 0 : }
1315 : }
1316 224743 : _ => {}
1317 : }
1318 0 : }
1319 :
1320 224743 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1321 224743 :
1322 224743 : match self
1323 224743 : .last_record_lsn
1324 224743 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1325 224743 : .await
1326 : {
1327 224743 : Ok(()) => Ok(()),
1328 0 : Err(e) => {
1329 : use utils::seqwait::SeqWaitError::*;
1330 0 : match e {
1331 0 : Shutdown => Err(WaitLsnError::Shutdown),
1332 : Timeout => {
1333 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1334 0 : drop(_timer);
1335 0 : let walreceiver_status = self.walreceiver_status();
1336 0 : Err(WaitLsnError::Timeout(format!(
1337 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1338 0 : lsn,
1339 0 : self.get_last_record_lsn(),
1340 0 : self.get_disk_consistent_lsn(),
1341 0 : walreceiver_status,
1342 0 : )))
1343 : }
1344 : }
1345 : }
1346 : }
1347 224743 : }
1348 :
1349 0 : pub(crate) fn walreceiver_status(&self) -> String {
1350 0 : match &*self.walreceiver.lock().unwrap() {
1351 0 : None => "stopping or stopped".to_string(),
1352 0 : Some(walreceiver) => match walreceiver.status() {
1353 0 : Some(status) => status.to_human_readable_string(),
1354 0 : None => "Not active".to_string(),
1355 : },
1356 : }
1357 0 : }
1358 :
1359 : /// Check that it is valid to request operations with that lsn.
1360 232 : pub(crate) fn check_lsn_is_in_scope(
1361 232 : &self,
1362 232 : lsn: Lsn,
1363 232 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1364 232 : ) -> anyhow::Result<()> {
1365 232 : ensure!(
1366 232 : lsn >= **latest_gc_cutoff_lsn,
1367 4 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1368 4 : lsn,
1369 4 : **latest_gc_cutoff_lsn,
1370 : );
1371 228 : Ok(())
1372 232 : }
1373 :
1374 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1375 10 : pub(crate) fn init_lsn_lease(
1376 10 : &self,
1377 10 : lsn: Lsn,
1378 10 : length: Duration,
1379 10 : ctx: &RequestContext,
1380 10 : ) -> anyhow::Result<LsnLease> {
1381 10 : self.make_lsn_lease(lsn, length, true, ctx)
1382 10 : }
1383 :
1384 : /// 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.
1385 4 : pub(crate) fn renew_lsn_lease(
1386 4 : &self,
1387 4 : lsn: Lsn,
1388 4 : length: Duration,
1389 4 : ctx: &RequestContext,
1390 4 : ) -> anyhow::Result<LsnLease> {
1391 4 : self.make_lsn_lease(lsn, length, false, ctx)
1392 4 : }
1393 :
1394 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1395 : ///
1396 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1397 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1398 : ///
1399 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1400 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1401 14 : fn make_lsn_lease(
1402 14 : &self,
1403 14 : lsn: Lsn,
1404 14 : length: Duration,
1405 14 : init: bool,
1406 14 : _ctx: &RequestContext,
1407 14 : ) -> anyhow::Result<LsnLease> {
1408 12 : let lease = {
1409 : // Normalize the requested LSN to be aligned, and move to the first record
1410 : // if it points to the beginning of the page (header).
1411 14 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1412 14 :
1413 14 : let mut gc_info = self.gc_info.write().unwrap();
1414 14 :
1415 14 : let valid_until = SystemTime::now() + length;
1416 14 :
1417 14 : let entry = gc_info.leases.entry(lsn);
1418 14 :
1419 14 : match entry {
1420 6 : Entry::Occupied(mut occupied) => {
1421 6 : let existing_lease = occupied.get_mut();
1422 6 : if valid_until > existing_lease.valid_until {
1423 2 : existing_lease.valid_until = valid_until;
1424 2 : let dt: DateTime<Utc> = valid_until.into();
1425 2 : info!("lease extended to {}", dt);
1426 : } else {
1427 4 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1428 4 : info!("existing lease covers greater length, valid until {}", dt);
1429 : }
1430 :
1431 6 : existing_lease.clone()
1432 : }
1433 8 : Entry::Vacant(vacant) => {
1434 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
1435 : // not blocked by the lsn lease deadline.
1436 8 : let validate = {
1437 8 : let conf = self.tenant_conf.load();
1438 8 : conf.location.attach_mode == AttachmentMode::Single
1439 8 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1440 : };
1441 :
1442 8 : if init || validate {
1443 8 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1444 8 : if lsn < *latest_gc_cutoff_lsn {
1445 2 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1446 6 : }
1447 0 : }
1448 :
1449 6 : let dt: DateTime<Utc> = valid_until.into();
1450 6 : info!("lease created, valid until {}", dt);
1451 6 : vacant.insert(LsnLease { valid_until }).clone()
1452 : }
1453 : }
1454 : };
1455 :
1456 12 : Ok(lease)
1457 14 : }
1458 :
1459 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1460 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1461 : #[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))]
1462 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1463 : self.freeze0().await
1464 : }
1465 :
1466 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1467 : #[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))]
1468 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1469 : self.freeze_and_flush0().await
1470 : }
1471 :
1472 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1473 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1474 1120 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1475 1120 : let mut g = self.write_lock.lock().await;
1476 1120 : let to_lsn = self.get_last_record_lsn();
1477 1120 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1478 1120 : }
1479 :
1480 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1481 : // polluting the span hierarchy.
1482 1120 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1483 1120 : let token = self.freeze0().await?;
1484 1120 : self.wait_flush_completion(token).await
1485 1120 : }
1486 :
1487 : // Check if an open ephemeral layer should be closed: this provides
1488 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1489 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1490 : // ephemeral layer bytes has been breached.
1491 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1492 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1493 : // If the write lock is held, there is an active wal receiver: rolling open layers
1494 : // is their responsibility while they hold this lock.
1495 0 : return;
1496 : };
1497 :
1498 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1499 : // time, and this was missed.
1500 : // if write_guard.is_none() { return; }
1501 :
1502 0 : let Ok(layers_guard) = self.layers.try_read() else {
1503 : // Don't block if the layer lock is busy
1504 0 : return;
1505 : };
1506 :
1507 0 : let Ok(lm) = layers_guard.layer_map() else {
1508 0 : return;
1509 : };
1510 :
1511 0 : let Some(open_layer) = &lm.open_layer else {
1512 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1513 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1514 : // that didn't result in writes to this shard.
1515 :
1516 : // Must not hold the layers lock while waiting for a flush.
1517 0 : drop(layers_guard);
1518 0 :
1519 0 : let last_record_lsn = self.get_last_record_lsn();
1520 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1521 0 : if last_record_lsn > disk_consistent_lsn {
1522 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1523 : // we are a sharded tenant and have skipped some WAL
1524 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1525 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1526 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1527 : // without any data ingested (yet) doesn't write a remote index as soon as it
1528 : // sees its LSN advance: we only do this if we've been layer-less
1529 : // for some time.
1530 0 : tracing::debug!(
1531 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1532 : disk_consistent_lsn,
1533 : last_record_lsn
1534 : );
1535 :
1536 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1537 : // We know there is no open layer, so we can request freezing without actually
1538 : // freezing anything. This is true even if we have dropped the layers_guard, we
1539 : // still hold the write_guard.
1540 0 : let _ = async {
1541 0 : let token = self
1542 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1543 0 : .await?;
1544 0 : self.wait_flush_completion(token).await
1545 0 : }
1546 0 : .await;
1547 0 : }
1548 0 : }
1549 :
1550 0 : return;
1551 : };
1552 :
1553 0 : let Some(current_size) = open_layer.try_len() else {
1554 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1555 : // read lock to get size should always succeed.
1556 0 : tracing::warn!("Lock conflict while reading size of open layer");
1557 0 : return;
1558 : };
1559 :
1560 0 : let current_lsn = self.get_last_record_lsn();
1561 :
1562 0 : let checkpoint_distance_override = open_layer.tick().await;
1563 :
1564 0 : if let Some(size_override) = checkpoint_distance_override {
1565 0 : if current_size > size_override {
1566 : // This is not harmful, but it only happens in relatively rare cases where
1567 : // time-based checkpoints are not happening fast enough to keep the amount of
1568 : // ephemeral data within configured limits. It's a sign of stress on the system.
1569 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1570 0 : }
1571 0 : }
1572 :
1573 0 : let checkpoint_distance =
1574 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1575 0 :
1576 0 : if self.should_roll(
1577 0 : current_size,
1578 0 : current_size,
1579 0 : checkpoint_distance,
1580 0 : self.get_last_record_lsn(),
1581 0 : self.last_freeze_at.load(),
1582 0 : open_layer.get_opened_at(),
1583 0 : ) {
1584 0 : match open_layer.info() {
1585 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1586 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1587 0 : // happens asynchronously in the background.
1588 0 : tracing::debug!(
1589 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1590 : );
1591 : }
1592 : InMemoryLayerInfo::Open { .. } => {
1593 : // Upgrade to a write lock and freeze the layer
1594 0 : drop(layers_guard);
1595 0 : let res = self
1596 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1597 0 : .await;
1598 :
1599 0 : if let Err(e) = res {
1600 0 : tracing::info!(
1601 0 : "failed to flush frozen layer after background freeze: {e:#}"
1602 : );
1603 0 : }
1604 : }
1605 : }
1606 0 : }
1607 0 : }
1608 :
1609 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1610 : ///
1611 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1612 : /// child timelines that are not offloaded yet.
1613 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1614 0 : if self.remote_client.is_archived() != Some(true) {
1615 0 : return (false, "the timeline is not archived");
1616 0 : }
1617 0 : if !self.remote_client.no_pending_work() {
1618 : // if the remote client is still processing some work, we can't offload
1619 0 : return (false, "the upload queue is not drained yet");
1620 0 : }
1621 0 :
1622 0 : (true, "ok")
1623 0 : }
1624 :
1625 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1626 : /// compaction tasks.
1627 364 : pub(crate) async fn compact(
1628 364 : self: &Arc<Self>,
1629 364 : cancel: &CancellationToken,
1630 364 : flags: EnumSet<CompactFlags>,
1631 364 : ctx: &RequestContext,
1632 364 : ) -> Result<bool, CompactionError> {
1633 364 : self.compact_with_options(
1634 364 : cancel,
1635 364 : CompactOptions {
1636 364 : flags,
1637 364 : compact_key_range: None,
1638 364 : compact_lsn_range: None,
1639 364 : sub_compaction: false,
1640 364 : sub_compaction_max_job_size_mb: None,
1641 364 : },
1642 364 : ctx,
1643 364 : )
1644 364 : .await
1645 364 : }
1646 :
1647 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1648 : /// compaction tasks.
1649 364 : pub(crate) async fn compact_with_options(
1650 364 : self: &Arc<Self>,
1651 364 : cancel: &CancellationToken,
1652 364 : options: CompactOptions,
1653 364 : ctx: &RequestContext,
1654 364 : ) -> Result<bool, CompactionError> {
1655 364 : // most likely the cancellation token is from background task, but in tests it could be the
1656 364 : // request task as well.
1657 364 :
1658 364 : let prepare = async move {
1659 364 : let guard = self.compaction_lock.lock().await;
1660 :
1661 364 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1662 364 : BackgroundLoopKind::Compaction,
1663 364 : ctx,
1664 364 : )
1665 364 : .await;
1666 :
1667 364 : (guard, permit)
1668 364 : };
1669 :
1670 : // this wait probably never needs any "long time spent" logging, because we already nag if
1671 : // compaction task goes over it's period (20s) which is quite often in production.
1672 364 : let (_guard, _permit) = tokio::select! {
1673 364 : tuple = prepare => { tuple },
1674 364 : _ = self.cancel.cancelled() => return Ok(false),
1675 364 : _ = cancel.cancelled() => return Ok(false),
1676 : };
1677 :
1678 364 : let last_record_lsn = self.get_last_record_lsn();
1679 364 :
1680 364 : // Last record Lsn could be zero in case the timeline was just created
1681 364 : if !last_record_lsn.is_valid() {
1682 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1683 0 : return Ok(false);
1684 364 : }
1685 364 :
1686 364 : match self.get_compaction_algorithm_settings().kind {
1687 : CompactionAlgorithm::Tiered => {
1688 0 : self.compact_tiered(cancel, ctx).await?;
1689 0 : Ok(false)
1690 : }
1691 364 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1692 : }
1693 364 : }
1694 :
1695 : /// Mutate the timeline with a [`TimelineWriter`].
1696 5133190 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1697 5133190 : TimelineWriter {
1698 5133190 : tl: self,
1699 5133190 : write_guard: self.write_lock.lock().await,
1700 : }
1701 5133190 : }
1702 :
1703 0 : pub(crate) fn activate(
1704 0 : self: &Arc<Self>,
1705 0 : parent: Arc<crate::tenant::Tenant>,
1706 0 : broker_client: BrokerClientChannel,
1707 0 : background_jobs_can_start: Option<&completion::Barrier>,
1708 0 : ctx: &RequestContext,
1709 0 : ) {
1710 0 : if self.tenant_shard_id.is_shard_zero() {
1711 0 : // Logical size is only maintained accurately on shard zero.
1712 0 : self.spawn_initial_logical_size_computation_task(ctx);
1713 0 : }
1714 0 : self.launch_wal_receiver(ctx, broker_client);
1715 0 : self.set_state(TimelineState::Active);
1716 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1717 0 : }
1718 :
1719 : /// After this function returns, there are no timeline-scoped tasks are left running.
1720 : ///
1721 : /// The preferred pattern for is:
1722 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1723 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1724 : /// go the extra mile and keep track of JoinHandles
1725 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1726 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1727 : ///
1728 : /// For legacy reasons, we still have multiple tasks spawned using
1729 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1730 : /// We refer to these as "timeline-scoped task_mgr tasks".
1731 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1732 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1733 : /// or [`task_mgr::shutdown_watcher`].
1734 : /// We want to gradually convert the code base away from these.
1735 : ///
1736 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1737 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1738 : /// ones that aren't mentioned here):
1739 : /// - [`TaskKind::TimelineDeletionWorker`]
1740 : /// - NB: also used for tenant deletion
1741 : /// - [`TaskKind::RemoteUploadTask`]`
1742 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1743 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1744 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1745 : /// - [`TaskKind::Eviction`]
1746 : /// - [`TaskKind::LayerFlushTask`]
1747 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1748 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1749 10 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1750 10 : debug_assert_current_span_has_tenant_and_timeline_id();
1751 10 :
1752 10 : // Regardless of whether we're going to try_freeze_and_flush
1753 10 : // or not, stop ingesting any more data. Walreceiver only provides
1754 10 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1755 10 : // So, only after the self.gate.close() below will we know for sure that
1756 10 : // no walreceiver tasks are left.
1757 10 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1758 10 : // data during the call to `self.freeze_and_flush()` below.
1759 10 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1760 10 : // which is what we'd need to properly model early shutdown of the walreceiver
1761 10 : // task sub-tree before the other Timeline task sub-trees.
1762 10 : let walreceiver = self.walreceiver.lock().unwrap().take();
1763 10 : tracing::debug!(
1764 0 : is_some = walreceiver.is_some(),
1765 0 : "Waiting for WalReceiverManager..."
1766 : );
1767 10 : if let Some(walreceiver) = walreceiver {
1768 0 : walreceiver.cancel();
1769 10 : }
1770 : // ... and inform any waiters for newer LSNs that there won't be any.
1771 10 : self.last_record_lsn.shutdown();
1772 10 :
1773 10 : if let ShutdownMode::FreezeAndFlush = mode {
1774 6 : if let Some((open, frozen)) = self
1775 6 : .layers
1776 6 : .read()
1777 6 : .await
1778 6 : .layer_map()
1779 6 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1780 6 : .ok()
1781 6 : .filter(|(open, frozen)| *open || *frozen > 0)
1782 : {
1783 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1784 6 : } else {
1785 6 : // this is double-shutdown, ignore it
1786 6 : }
1787 :
1788 : // we shut down walreceiver above, so, we won't add anything more
1789 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1790 : // to reach the disk & upload queue, then shut the upload queue and
1791 : // wait for it to drain.
1792 6 : match self.freeze_and_flush().await {
1793 : Ok(_) => {
1794 : // drain the upload queue
1795 : // if we did not wait for completion here, it might be our shutdown process
1796 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1797 : // be spawned.
1798 : //
1799 : // what is problematic is the shutting down of RemoteTimelineClient, because
1800 : // obviously it does not make sense to stop while we wait for it, but what
1801 : // about corner cases like s3 suddenly hanging up?
1802 6 : self.remote_client.shutdown().await;
1803 : }
1804 : Err(FlushLayerError::Cancelled) => {
1805 : // this is likely the second shutdown, ignore silently.
1806 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1807 0 : debug_assert!(self.cancel.is_cancelled());
1808 : }
1809 0 : Err(e) => {
1810 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1811 0 : // we have some extra WAL replay to do next time the timeline starts.
1812 0 : warn!("failed to freeze and flush: {e:#}");
1813 : }
1814 : }
1815 :
1816 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
1817 : // we also do a final check here to ensure that the queue is empty.
1818 6 : if !self.remote_client.no_pending_work() {
1819 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1820 6 : }
1821 4 : }
1822 :
1823 10 : if let ShutdownMode::Reload = mode {
1824 : // drain the upload queue
1825 2 : self.remote_client.shutdown().await;
1826 2 : if !self.remote_client.no_pending_work() {
1827 0 : warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
1828 2 : }
1829 8 : }
1830 :
1831 : // Signal any subscribers to our cancellation token to drop out
1832 10 : tracing::debug!("Cancelling CancellationToken");
1833 10 : self.cancel.cancel();
1834 10 :
1835 10 : // Ensure Prevent new page service requests from starting.
1836 10 : self.handles.shutdown();
1837 10 :
1838 10 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1839 10 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1840 10 : self.remote_client.stop();
1841 10 :
1842 10 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1843 10 : // to shut down the upload queue tasks.
1844 10 : // TODO: fix that, task management should be encapsulated inside remote_client.
1845 10 : task_mgr::shutdown_tasks(
1846 10 : Some(TaskKind::RemoteUploadTask),
1847 10 : Some(self.tenant_shard_id),
1848 10 : Some(self.timeline_id),
1849 10 : )
1850 10 : .await;
1851 :
1852 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
1853 10 : tracing::debug!("Waiting for tasks...");
1854 10 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1855 :
1856 : {
1857 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
1858 : // open.
1859 10 : let mut write_guard = self.write_lock.lock().await;
1860 10 : self.layers.write().await.shutdown(&mut write_guard);
1861 10 : }
1862 10 :
1863 10 : // Finally wait until any gate-holders are complete.
1864 10 : //
1865 10 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1866 10 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1867 10 : self.gate.close().await;
1868 :
1869 10 : self.metrics.shutdown();
1870 10 : }
1871 :
1872 448 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1873 448 : match (self.current_state(), new_state) {
1874 448 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1875 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1876 : }
1877 0 : (st, TimelineState::Loading) => {
1878 0 : error!("ignoring transition from {st:?} into Loading state");
1879 : }
1880 0 : (TimelineState::Broken { .. }, new_state) => {
1881 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1882 : }
1883 : (TimelineState::Stopping, TimelineState::Active) => {
1884 0 : error!("Not activating a Stopping timeline");
1885 : }
1886 446 : (_, new_state) => {
1887 446 : self.state.send_replace(new_state);
1888 446 : }
1889 : }
1890 448 : }
1891 :
1892 2 : pub(crate) fn set_broken(&self, reason: String) {
1893 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1894 2 : let broken_state = TimelineState::Broken {
1895 2 : reason,
1896 2 : backtrace: backtrace_str,
1897 2 : };
1898 2 : self.set_state(broken_state);
1899 2 :
1900 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1901 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1902 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1903 2 : self.cancel.cancel();
1904 2 : }
1905 :
1906 225781 : pub(crate) fn current_state(&self) -> TimelineState {
1907 225781 : self.state.borrow().clone()
1908 225781 : }
1909 :
1910 6 : pub(crate) fn is_broken(&self) -> bool {
1911 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1912 6 : }
1913 :
1914 222 : pub(crate) fn is_active(&self) -> bool {
1915 222 : self.current_state() == TimelineState::Active
1916 222 : }
1917 :
1918 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
1919 0 : self.remote_client.is_archived()
1920 0 : }
1921 :
1922 368 : pub(crate) fn is_stopping(&self) -> bool {
1923 368 : self.current_state() == TimelineState::Stopping
1924 368 : }
1925 :
1926 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1927 0 : self.state.subscribe()
1928 0 : }
1929 :
1930 224745 : pub(crate) async fn wait_to_become_active(
1931 224745 : &self,
1932 224745 : _ctx: &RequestContext, // Prepare for use by cancellation
1933 224745 : ) -> Result<(), TimelineState> {
1934 224745 : let mut receiver = self.state.subscribe();
1935 : loop {
1936 224745 : let current_state = receiver.borrow().clone();
1937 224745 : match current_state {
1938 : TimelineState::Loading => {
1939 0 : receiver
1940 0 : .changed()
1941 0 : .await
1942 0 : .expect("holding a reference to self");
1943 : }
1944 : TimelineState::Active { .. } => {
1945 224743 : return Ok(());
1946 : }
1947 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1948 : // There's no chance the timeline can transition back into ::Active
1949 2 : return Err(current_state);
1950 : }
1951 : }
1952 : }
1953 224745 : }
1954 :
1955 0 : pub(crate) async fn layer_map_info(
1956 0 : &self,
1957 0 : reset: LayerAccessStatsReset,
1958 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
1959 0 : let guard = self.layers.read().await;
1960 0 : let layer_map = guard.layer_map()?;
1961 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
1962 0 : if let Some(open_layer) = &layer_map.open_layer {
1963 0 : in_memory_layers.push(open_layer.info());
1964 0 : }
1965 0 : for frozen_layer in &layer_map.frozen_layers {
1966 0 : in_memory_layers.push(frozen_layer.info());
1967 0 : }
1968 :
1969 0 : let historic_layers = layer_map
1970 0 : .iter_historic_layers()
1971 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
1972 0 : .collect();
1973 0 :
1974 0 : Ok(LayerMapInfo {
1975 0 : in_memory_layers,
1976 0 : historic_layers,
1977 0 : })
1978 0 : }
1979 :
1980 : #[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))]
1981 : pub(crate) async fn download_layer(
1982 : &self,
1983 : layer_file_name: &LayerName,
1984 : ) -> anyhow::Result<Option<bool>> {
1985 : let Some(layer) = self.find_layer(layer_file_name).await? else {
1986 : return Ok(None);
1987 : };
1988 :
1989 : layer.download().await?;
1990 :
1991 : Ok(Some(true))
1992 : }
1993 :
1994 : /// Evict just one layer.
1995 : ///
1996 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
1997 0 : pub(crate) async fn evict_layer(
1998 0 : &self,
1999 0 : layer_file_name: &LayerName,
2000 0 : ) -> anyhow::Result<Option<bool>> {
2001 0 : let _gate = self
2002 0 : .gate
2003 0 : .enter()
2004 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2005 :
2006 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2007 0 : return Ok(None);
2008 : };
2009 :
2010 : // curl has this by default
2011 0 : let timeout = std::time::Duration::from_secs(120);
2012 0 :
2013 0 : match local_layer.evict_and_wait(timeout).await {
2014 0 : Ok(()) => Ok(Some(true)),
2015 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2016 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2017 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2018 : }
2019 0 : }
2020 :
2021 4803010 : fn should_roll(
2022 4803010 : &self,
2023 4803010 : layer_size: u64,
2024 4803010 : projected_layer_size: u64,
2025 4803010 : checkpoint_distance: u64,
2026 4803010 : projected_lsn: Lsn,
2027 4803010 : last_freeze_at: Lsn,
2028 4803010 : opened_at: Instant,
2029 4803010 : ) -> bool {
2030 4803010 : let distance = projected_lsn.widening_sub(last_freeze_at);
2031 4803010 :
2032 4803010 : // Rolling the open layer can be triggered by:
2033 4803010 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2034 4803010 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2035 4803010 : // account for how writes are distributed across shards: we expect each node to consume
2036 4803010 : // 1/count of the LSN on average.
2037 4803010 : // 2. The size of the currently open layer.
2038 4803010 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2039 4803010 : // up and suspend activity.
2040 4803010 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2041 0 : info!(
2042 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2043 : projected_lsn, layer_size, distance
2044 : );
2045 :
2046 0 : true
2047 4803010 : } else if projected_layer_size >= checkpoint_distance {
2048 : // NB: this check is relied upon by:
2049 80 : let _ = IndexEntry::validate_checkpoint_distance;
2050 80 : info!(
2051 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2052 : projected_lsn, layer_size, projected_layer_size
2053 : );
2054 :
2055 80 : true
2056 4802930 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2057 0 : info!(
2058 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2059 0 : projected_lsn,
2060 0 : layer_size,
2061 0 : opened_at.elapsed()
2062 : );
2063 :
2064 0 : true
2065 : } else {
2066 4802930 : false
2067 : }
2068 4803010 : }
2069 : }
2070 :
2071 : /// Number of times we will compute partition within a checkpoint distance.
2072 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2073 :
2074 : // Private functions
2075 : impl Timeline {
2076 12 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2077 12 : let tenant_conf = self.tenant_conf.load();
2078 12 : tenant_conf
2079 12 : .tenant_conf
2080 12 : .lsn_lease_length
2081 12 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2082 12 : }
2083 :
2084 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2085 0 : let tenant_conf = self.tenant_conf.load();
2086 0 : tenant_conf
2087 0 : .tenant_conf
2088 0 : .lsn_lease_length_for_ts
2089 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2090 0 : }
2091 :
2092 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2093 0 : let tenant_conf = self.tenant_conf.load();
2094 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2095 0 : }
2096 :
2097 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2098 0 : let tenant_conf = self.tenant_conf.load();
2099 0 : tenant_conf
2100 0 : .tenant_conf
2101 0 : .lazy_slru_download
2102 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2103 0 : }
2104 :
2105 4804496 : fn get_checkpoint_distance(&self) -> u64 {
2106 4804496 : let tenant_conf = self.tenant_conf.load();
2107 4804496 : tenant_conf
2108 4804496 : .tenant_conf
2109 4804496 : .checkpoint_distance
2110 4804496 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2111 4804496 : }
2112 :
2113 4802930 : fn get_checkpoint_timeout(&self) -> Duration {
2114 4802930 : let tenant_conf = self.tenant_conf.load();
2115 4802930 : tenant_conf
2116 4802930 : .tenant_conf
2117 4802930 : .checkpoint_timeout
2118 4802930 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2119 4802930 : }
2120 :
2121 662 : fn get_compaction_target_size(&self) -> u64 {
2122 662 : let tenant_conf = self.tenant_conf.load();
2123 662 : tenant_conf
2124 662 : .tenant_conf
2125 662 : .compaction_target_size
2126 662 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2127 662 : }
2128 :
2129 1564 : fn get_compaction_threshold(&self) -> usize {
2130 1564 : let tenant_conf = self.tenant_conf.load();
2131 1564 : tenant_conf
2132 1564 : .tenant_conf
2133 1564 : .compaction_threshold
2134 1564 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2135 1564 : }
2136 :
2137 14 : fn get_image_creation_threshold(&self) -> usize {
2138 14 : let tenant_conf = self.tenant_conf.load();
2139 14 : tenant_conf
2140 14 : .tenant_conf
2141 14 : .image_creation_threshold
2142 14 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2143 14 : }
2144 :
2145 364 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2146 364 : let tenant_conf = &self.tenant_conf.load();
2147 364 : tenant_conf
2148 364 : .tenant_conf
2149 364 : .compaction_algorithm
2150 364 : .as_ref()
2151 364 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2152 364 : .clone()
2153 364 : }
2154 :
2155 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2156 0 : let tenant_conf = self.tenant_conf.load();
2157 0 : tenant_conf
2158 0 : .tenant_conf
2159 0 : .eviction_policy
2160 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2161 0 : }
2162 :
2163 446 : fn get_evictions_low_residence_duration_metric_threshold(
2164 446 : tenant_conf: &TenantConfOpt,
2165 446 : default_tenant_conf: &TenantConf,
2166 446 : ) -> Duration {
2167 446 : tenant_conf
2168 446 : .evictions_low_residence_duration_metric_threshold
2169 446 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2170 446 : }
2171 :
2172 568 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2173 568 : let tenant_conf = self.tenant_conf.load();
2174 568 : tenant_conf
2175 568 : .tenant_conf
2176 568 : .image_layer_creation_check_threshold
2177 568 : .unwrap_or(
2178 568 : self.conf
2179 568 : .default_tenant_conf
2180 568 : .image_layer_creation_check_threshold,
2181 568 : )
2182 568 : }
2183 :
2184 : /// Resolve the effective WAL receiver protocol to use for this tenant.
2185 : ///
2186 : /// Priority order is:
2187 : /// 1. Tenant config override
2188 : /// 2. Default value for tenant config override
2189 : /// 3. Pageserver config override
2190 : /// 4. Pageserver config default
2191 0 : pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
2192 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2193 0 : tenant_conf
2194 0 : .wal_receiver_protocol_override
2195 0 : .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
2196 0 : .unwrap_or(self.conf.wal_receiver_protocol)
2197 0 : }
2198 :
2199 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2200 0 : // NB: Most tenant conf options are read by background loops, so,
2201 0 : // changes will automatically be picked up.
2202 0 :
2203 0 : // The threshold is embedded in the metric. So, we need to update it.
2204 0 : {
2205 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2206 0 : &new_conf.tenant_conf,
2207 0 : &self.conf.default_tenant_conf,
2208 0 : );
2209 0 :
2210 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2211 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2212 0 :
2213 0 : let timeline_id_str = self.timeline_id.to_string();
2214 0 :
2215 0 : self.remote_client.update_config(&new_conf.location);
2216 0 :
2217 0 : self.metrics
2218 0 : .evictions_with_low_residence_duration
2219 0 : .write()
2220 0 : .unwrap()
2221 0 : .change_threshold(
2222 0 : &tenant_id_str,
2223 0 : &shard_id_str,
2224 0 : &timeline_id_str,
2225 0 : new_threshold,
2226 0 : );
2227 0 : }
2228 0 : }
2229 :
2230 : /// Open a Timeline handle.
2231 : ///
2232 : /// Loads the metadata for the timeline into memory, but not the layer map.
2233 : #[allow(clippy::too_many_arguments)]
2234 446 : pub(super) fn new(
2235 446 : conf: &'static PageServerConf,
2236 446 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2237 446 : metadata: &TimelineMetadata,
2238 446 : ancestor: Option<Arc<Timeline>>,
2239 446 : timeline_id: TimelineId,
2240 446 : tenant_shard_id: TenantShardId,
2241 446 : generation: Generation,
2242 446 : shard_identity: ShardIdentity,
2243 446 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2244 446 : resources: TimelineResources,
2245 446 : pg_version: u32,
2246 446 : state: TimelineState,
2247 446 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2248 446 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2249 446 : cancel: CancellationToken,
2250 446 : ) -> Arc<Self> {
2251 446 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2252 446 : let (state, _) = watch::channel(state);
2253 446 :
2254 446 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2255 446 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2256 446 :
2257 446 : let evictions_low_residence_duration_metric_threshold = {
2258 446 : let loaded_tenant_conf = tenant_conf.load();
2259 446 : Self::get_evictions_low_residence_duration_metric_threshold(
2260 446 : &loaded_tenant_conf.tenant_conf,
2261 446 : &conf.default_tenant_conf,
2262 446 : )
2263 : };
2264 :
2265 446 : if let Some(ancestor) = &ancestor {
2266 230 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2267 230 : // If we construct an explicit timeline object, it's obviously not offloaded
2268 230 : let is_offloaded = MaybeOffloaded::No;
2269 230 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2270 230 : }
2271 :
2272 446 : Arc::new_cyclic(|myself| {
2273 446 : let metrics = TimelineMetrics::new(
2274 446 : &tenant_shard_id,
2275 446 : &timeline_id,
2276 446 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2277 446 : "mtime",
2278 446 : evictions_low_residence_duration_metric_threshold,
2279 446 : ),
2280 446 : );
2281 446 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2282 :
2283 446 : let mut result = Timeline {
2284 446 : conf,
2285 446 : tenant_conf,
2286 446 : myself: myself.clone(),
2287 446 : timeline_id,
2288 446 : tenant_shard_id,
2289 446 : generation,
2290 446 : shard_identity,
2291 446 : pg_version,
2292 446 : layers: Default::default(),
2293 446 :
2294 446 : walredo_mgr,
2295 446 : walreceiver: Mutex::new(None),
2296 446 :
2297 446 : remote_client: Arc::new(resources.remote_client),
2298 446 :
2299 446 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2300 446 : last_record_lsn: SeqWait::new(RecordLsn {
2301 446 : last: disk_consistent_lsn,
2302 446 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2303 446 : }),
2304 446 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2305 446 :
2306 446 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2307 446 : last_freeze_ts: RwLock::new(Instant::now()),
2308 446 :
2309 446 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2310 446 :
2311 446 : ancestor_timeline: ancestor,
2312 446 : ancestor_lsn: metadata.ancestor_lsn(),
2313 446 :
2314 446 : metrics,
2315 446 :
2316 446 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2317 446 : &tenant_shard_id,
2318 446 : &timeline_id,
2319 446 : resources.pagestream_throttle_metrics,
2320 446 : ),
2321 446 :
2322 3122 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2323 446 :
2324 446 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2325 446 :
2326 446 : layer_flush_start_tx,
2327 446 : layer_flush_done_tx,
2328 446 :
2329 446 : write_lock: tokio::sync::Mutex::new(None),
2330 446 :
2331 446 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2332 446 :
2333 446 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2334 446 : initdb_lsn: metadata.initdb_lsn(),
2335 446 :
2336 446 : current_logical_size: if disk_consistent_lsn.is_valid() {
2337 : // we're creating timeline data with some layer files existing locally,
2338 : // need to recalculate timeline's logical size based on data in the layers.
2339 234 : LogicalSize::deferred_initial(disk_consistent_lsn)
2340 : } else {
2341 : // we're creating timeline data without any layers existing locally,
2342 : // initial logical size is 0.
2343 212 : LogicalSize::empty_initial()
2344 : },
2345 :
2346 446 : partitioning: GuardArcSwap::new((
2347 446 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2348 446 : Lsn(0),
2349 446 : )),
2350 446 : repartition_threshold: 0,
2351 446 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2352 446 : last_image_layer_creation_check_instant: Mutex::new(None),
2353 446 :
2354 446 : last_received_wal: Mutex::new(None),
2355 446 : rel_size_cache: RwLock::new(RelSizeCache {
2356 446 : complete_as_of: disk_consistent_lsn,
2357 446 : map: HashMap::new(),
2358 446 : }),
2359 446 :
2360 446 : download_all_remote_layers_task_info: RwLock::new(None),
2361 446 :
2362 446 : state,
2363 446 :
2364 446 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2365 446 : EvictionTaskTimelineState::default(),
2366 446 : ),
2367 446 : delete_progress: TimelineDeleteProgress::default(),
2368 446 :
2369 446 : cancel,
2370 446 : gate: Gate::default(),
2371 446 :
2372 446 : compaction_lock: tokio::sync::Mutex::default(),
2373 446 : gc_lock: tokio::sync::Mutex::default(),
2374 446 :
2375 446 : standby_horizon: AtomicLsn::new(0),
2376 446 :
2377 446 : pagestream_throttle: resources.pagestream_throttle,
2378 446 :
2379 446 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2380 446 :
2381 446 : #[cfg(test)]
2382 446 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2383 446 :
2384 446 : l0_flush_global_state: resources.l0_flush_global_state,
2385 446 :
2386 446 : handles: Default::default(),
2387 446 :
2388 446 : attach_wal_lag_cooldown,
2389 446 :
2390 446 : create_idempotency,
2391 446 :
2392 446 : page_trace: Default::default(),
2393 446 : };
2394 446 :
2395 446 : result.repartition_threshold =
2396 446 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2397 446 :
2398 446 : result
2399 446 : .metrics
2400 446 : .last_record_lsn_gauge
2401 446 : .set(disk_consistent_lsn.0 as i64);
2402 446 : result
2403 446 : })
2404 446 : }
2405 :
2406 644 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2407 644 : let Ok(guard) = self.gate.enter() else {
2408 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2409 0 : return;
2410 : };
2411 644 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2412 644 : match *flush_loop_state {
2413 440 : FlushLoopState::NotStarted => (),
2414 : FlushLoopState::Running { .. } => {
2415 204 : info!(
2416 0 : "skipping attempt to start flush_loop twice {}/{}",
2417 0 : self.tenant_shard_id, self.timeline_id
2418 : );
2419 204 : return;
2420 : }
2421 : FlushLoopState::Exited => {
2422 0 : warn!(
2423 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2424 0 : self.tenant_shard_id, self.timeline_id
2425 : );
2426 0 : return;
2427 : }
2428 : }
2429 :
2430 440 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2431 440 : let self_clone = Arc::clone(self);
2432 440 :
2433 440 : debug!("spawning flush loop");
2434 440 : *flush_loop_state = FlushLoopState::Running {
2435 440 : #[cfg(test)]
2436 440 : expect_initdb_optimization: false,
2437 440 : #[cfg(test)]
2438 440 : initdb_optimization_count: 0,
2439 440 : };
2440 440 : task_mgr::spawn(
2441 440 : task_mgr::BACKGROUND_RUNTIME.handle(),
2442 440 : task_mgr::TaskKind::LayerFlushTask,
2443 440 : self.tenant_shard_id,
2444 440 : Some(self.timeline_id),
2445 440 : "layer flush task",
2446 440 : async move {
2447 440 : let _guard = guard;
2448 440 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2449 440 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2450 10 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2451 10 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2452 10 : *flush_loop_state = FlushLoopState::Exited;
2453 10 : Ok(())
2454 10 : }
2455 440 : .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))
2456 : );
2457 644 : }
2458 :
2459 : /// Creates and starts the wal receiver.
2460 : ///
2461 : /// This function is expected to be called at most once per Timeline's lifecycle
2462 : /// when the timeline is activated.
2463 0 : fn launch_wal_receiver(
2464 0 : self: &Arc<Self>,
2465 0 : ctx: &RequestContext,
2466 0 : broker_client: BrokerClientChannel,
2467 0 : ) {
2468 0 : info!(
2469 0 : "launching WAL receiver for timeline {} of tenant {}",
2470 0 : self.timeline_id, self.tenant_shard_id
2471 : );
2472 :
2473 0 : let tenant_conf = self.tenant_conf.load();
2474 0 : let wal_connect_timeout = tenant_conf
2475 0 : .tenant_conf
2476 0 : .walreceiver_connect_timeout
2477 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2478 0 : let lagging_wal_timeout = tenant_conf
2479 0 : .tenant_conf
2480 0 : .lagging_wal_timeout
2481 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2482 0 : let max_lsn_wal_lag = tenant_conf
2483 0 : .tenant_conf
2484 0 : .max_lsn_wal_lag
2485 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2486 0 :
2487 0 : let mut guard = self.walreceiver.lock().unwrap();
2488 0 : assert!(
2489 0 : guard.is_none(),
2490 0 : "multiple launches / re-launches of WAL receiver are not supported"
2491 : );
2492 0 : *guard = Some(WalReceiver::start(
2493 0 : Arc::clone(self),
2494 0 : WalReceiverConf {
2495 0 : protocol: self.resolve_wal_receiver_protocol(),
2496 0 : wal_connect_timeout,
2497 0 : lagging_wal_timeout,
2498 0 : max_lsn_wal_lag,
2499 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2500 0 : availability_zone: self.conf.availability_zone.clone(),
2501 0 : ingest_batch_size: self.conf.ingest_batch_size,
2502 0 : },
2503 0 : broker_client,
2504 0 : ctx,
2505 0 : ));
2506 0 : }
2507 :
2508 : /// Initialize with an empty layer map. Used when creating a new timeline.
2509 440 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2510 440 : let mut layers = self.layers.try_write().expect(
2511 440 : "in the context where we call this function, no other task has access to the object",
2512 440 : );
2513 440 : layers
2514 440 : .open_mut()
2515 440 : .expect("in this context the LayerManager must still be open")
2516 440 : .initialize_empty(Lsn(start_lsn.0));
2517 440 : }
2518 :
2519 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2520 : /// files.
2521 6 : pub(super) async fn load_layer_map(
2522 6 : &self,
2523 6 : disk_consistent_lsn: Lsn,
2524 6 : index_part: IndexPart,
2525 6 : ) -> anyhow::Result<()> {
2526 : use init::{Decision::*, Discovered, DismissedLayer};
2527 : use LayerName::*;
2528 :
2529 6 : let mut guard = self.layers.write().await;
2530 :
2531 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2532 6 :
2533 6 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2534 6 : // structs representing all files on disk
2535 6 : let timeline_path = self
2536 6 : .conf
2537 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2538 6 : let conf = self.conf;
2539 6 : let span = tracing::Span::current();
2540 6 :
2541 6 : // Copy to move into the task we're about to spawn
2542 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2543 :
2544 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2545 6 : move || {
2546 6 : let _g = span.entered();
2547 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2548 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2549 6 : let mut unrecognized_files = Vec::new();
2550 6 :
2551 6 : let mut path = timeline_path;
2552 :
2553 22 : for discovered in discovered {
2554 16 : let (name, kind) = match discovered {
2555 16 : Discovered::Layer(layer_file_name, local_metadata) => {
2556 16 : discovered_layers.push((layer_file_name, local_metadata));
2557 16 : continue;
2558 : }
2559 0 : Discovered::IgnoredBackup(path) => {
2560 0 : std::fs::remove_file(path)
2561 0 : .or_else(fs_ext::ignore_not_found)
2562 0 : .fatal_err("Removing .old file");
2563 0 : continue;
2564 : }
2565 0 : Discovered::Unknown(file_name) => {
2566 0 : // we will later error if there are any
2567 0 : unrecognized_files.push(file_name);
2568 0 : continue;
2569 : }
2570 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2571 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2572 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2573 : };
2574 0 : path.push(Utf8Path::new(&name));
2575 0 : init::cleanup(&path, kind)?;
2576 0 : path.pop();
2577 : }
2578 :
2579 6 : if !unrecognized_files.is_empty() {
2580 : // assume that if there are any there are many many.
2581 0 : let n = unrecognized_files.len();
2582 0 : let first = &unrecognized_files[..n.min(10)];
2583 0 : anyhow::bail!(
2584 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2585 0 : );
2586 6 : }
2587 6 :
2588 6 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2589 6 :
2590 6 : let mut loaded_layers = Vec::new();
2591 6 : let mut needs_cleanup = Vec::new();
2592 6 : let mut total_physical_size = 0;
2593 :
2594 22 : for (name, decision) in decided {
2595 16 : let decision = match decision {
2596 16 : Ok(decision) => decision,
2597 0 : Err(DismissedLayer::Future { local }) => {
2598 0 : if let Some(local) = local {
2599 0 : init::cleanup_future_layer(
2600 0 : &local.local_path,
2601 0 : &name,
2602 0 : disk_consistent_lsn,
2603 0 : )?;
2604 0 : }
2605 0 : needs_cleanup.push(name);
2606 0 : continue;
2607 : }
2608 0 : Err(DismissedLayer::LocalOnly(local)) => {
2609 0 : init::cleanup_local_only_file(&name, &local)?;
2610 : // this file never existed remotely, we will have to do rework
2611 0 : continue;
2612 : }
2613 0 : Err(DismissedLayer::BadMetadata(local)) => {
2614 0 : init::cleanup_local_file_for_remote(&local)?;
2615 : // this file never existed remotely, we will have to do rework
2616 0 : continue;
2617 : }
2618 : };
2619 :
2620 16 : match &name {
2621 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2622 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2623 : }
2624 :
2625 16 : tracing::debug!(layer=%name, ?decision, "applied");
2626 :
2627 16 : let layer = match decision {
2628 16 : Resident { local, remote } => {
2629 16 : total_physical_size += local.file_size;
2630 16 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2631 16 : .drop_eviction_guard()
2632 : }
2633 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2634 : };
2635 :
2636 16 : loaded_layers.push(layer);
2637 : }
2638 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2639 6 : }
2640 6 : })
2641 6 : .await
2642 6 : .map_err(anyhow::Error::new)
2643 6 : .and_then(|x| x)?;
2644 :
2645 6 : let num_layers = loaded_layers.len();
2646 6 :
2647 6 : guard
2648 6 : .open_mut()
2649 6 : .expect("layermanager must be open during init")
2650 6 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2651 6 :
2652 6 : self.remote_client
2653 6 : .schedule_layer_file_deletion(&needs_cleanup)?;
2654 6 : self.remote_client
2655 6 : .schedule_index_upload_for_file_changes()?;
2656 : // This barrier orders above DELETEs before any later operations.
2657 : // This is critical because code executing after the barrier might
2658 : // create again objects with the same key that we just scheduled for deletion.
2659 : // For example, if we just scheduled deletion of an image layer "from the future",
2660 : // later compaction might run again and re-create the same image layer.
2661 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2662 : // "same" here means same key range and LSN.
2663 : //
2664 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2665 : // the upload queue may execute the PUT first, then the DELETE.
2666 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2667 : //
2668 : // 1. a future image layer is created and uploaded
2669 : // 2. ps restart
2670 : // 3. the future layer from (1) is deleted during load layer map
2671 : // 4. image layer is re-created and uploaded
2672 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2673 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2674 : //
2675 : // See https://github.com/neondatabase/neon/issues/5878
2676 : //
2677 : // NB: generation numbers naturally protect against this because they disambiguate
2678 : // (1) and (4)
2679 : // TODO: this is basically a no-op now, should we remove it?
2680 6 : self.remote_client.schedule_barrier()?;
2681 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2682 : // on retry.
2683 :
2684 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2685 6 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2686 6 : self.update_layer_visibility().await?;
2687 :
2688 6 : info!(
2689 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2690 : num_layers, disk_consistent_lsn, total_physical_size
2691 : );
2692 :
2693 6 : timer.stop_and_record();
2694 6 : Ok(())
2695 6 : }
2696 :
2697 : /// Retrieve current logical size of the timeline.
2698 : ///
2699 : /// The size could be lagging behind the actual number, in case
2700 : /// the initial size calculation has not been run (gets triggered on the first size access).
2701 : ///
2702 : /// return size and boolean flag that shows if the size is exact
2703 0 : pub(crate) fn get_current_logical_size(
2704 0 : self: &Arc<Self>,
2705 0 : priority: GetLogicalSizePriority,
2706 0 : ctx: &RequestContext,
2707 0 : ) -> logical_size::CurrentLogicalSize {
2708 0 : if !self.tenant_shard_id.is_shard_zero() {
2709 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2710 : // when HTTP API is serving a GET for timeline zero, return zero
2711 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2712 0 : }
2713 0 :
2714 0 : let current_size = self.current_logical_size.current_size();
2715 0 : debug!("Current size: {current_size:?}");
2716 :
2717 0 : match (current_size.accuracy(), priority) {
2718 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2719 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2720 0 : // background task will eventually deliver an exact value, we're in no rush
2721 0 : }
2722 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2723 : // background task is not ready, but user is asking for it now;
2724 : // => make the background task skip the line
2725 : // (The alternative would be to calculate the size here, but,
2726 : // it can actually take a long time if the user has a lot of rels.
2727 : // And we'll inevitable need it again; So, let the background task do the work.)
2728 0 : match self
2729 0 : .current_logical_size
2730 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2731 0 : .get()
2732 : {
2733 0 : Some(cancel) => cancel.cancel(),
2734 : None => {
2735 0 : match self.current_state() {
2736 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2737 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2738 0 : // Don't make noise.
2739 0 : }
2740 : TimelineState::Loading => {
2741 : // Import does not return an activated timeline.
2742 0 : info!("discarding priority boost for logical size calculation because timeline is not yet active");
2743 : }
2744 : TimelineState::Active => {
2745 : // activation should be setting the once cell
2746 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2747 0 : debug_assert!(false);
2748 : }
2749 : }
2750 : }
2751 : }
2752 : }
2753 : }
2754 :
2755 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2756 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2757 0 : let first = self
2758 0 : .current_logical_size
2759 0 : .did_return_approximate_to_walreceiver
2760 0 : .compare_exchange(
2761 0 : false,
2762 0 : true,
2763 0 : AtomicOrdering::Relaxed,
2764 0 : AtomicOrdering::Relaxed,
2765 0 : )
2766 0 : .is_ok();
2767 0 : if first {
2768 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2769 0 : }
2770 0 : }
2771 0 : }
2772 :
2773 0 : current_size
2774 0 : }
2775 :
2776 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2777 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2778 : // nothing to do for freshly created timelines;
2779 0 : assert_eq!(
2780 0 : self.current_logical_size.current_size().accuracy(),
2781 0 : logical_size::Accuracy::Exact,
2782 0 : );
2783 0 : self.current_logical_size.initialized.add_permits(1);
2784 0 : return;
2785 : };
2786 :
2787 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2788 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2789 0 : self.current_logical_size
2790 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2791 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2792 0 :
2793 0 : let self_clone = Arc::clone(self);
2794 0 : let background_ctx = ctx.detached_child(
2795 0 : TaskKind::InitialLogicalSizeCalculation,
2796 0 : DownloadBehavior::Download,
2797 0 : );
2798 0 : task_mgr::spawn(
2799 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2800 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2801 0 : self.tenant_shard_id,
2802 0 : Some(self.timeline_id),
2803 0 : "initial size calculation",
2804 : // NB: don't log errors here, task_mgr will do that.
2805 0 : async move {
2806 0 : let cancel = task_mgr::shutdown_token();
2807 0 : self_clone
2808 0 : .initial_logical_size_calculation_task(
2809 0 : initial_part_end,
2810 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2811 0 : cancel,
2812 0 : background_ctx,
2813 0 : )
2814 0 : .await;
2815 0 : Ok(())
2816 0 : }
2817 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)),
2818 : );
2819 0 : }
2820 :
2821 0 : async fn initial_logical_size_calculation_task(
2822 0 : self: Arc<Self>,
2823 0 : initial_part_end: Lsn,
2824 0 : skip_concurrency_limiter: CancellationToken,
2825 0 : cancel: CancellationToken,
2826 0 : background_ctx: RequestContext,
2827 0 : ) {
2828 0 : scopeguard::defer! {
2829 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2830 0 : self.current_logical_size.initialized.add_permits(1);
2831 0 : }
2832 0 :
2833 0 : let try_once = |attempt: usize| {
2834 0 : let background_ctx = &background_ctx;
2835 0 : let self_ref = &self;
2836 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2837 0 : async move {
2838 0 : let cancel = task_mgr::shutdown_token();
2839 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2840 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2841 0 : background_ctx,
2842 0 : );
2843 :
2844 : use crate::metrics::initial_logical_size::StartCircumstances;
2845 0 : let (_maybe_permit, circumstances) = tokio::select! {
2846 0 : permit = wait_for_permit => {
2847 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2848 : }
2849 0 : _ = self_ref.cancel.cancelled() => {
2850 0 : return Err(CalculateLogicalSizeError::Cancelled);
2851 : }
2852 0 : _ = cancel.cancelled() => {
2853 0 : return Err(CalculateLogicalSizeError::Cancelled);
2854 : },
2855 0 : () = skip_concurrency_limiter.cancelled() => {
2856 : // Some action that is part of a end user interaction requested logical size
2857 : // => break out of the rate limit
2858 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2859 : // but then again what happens if they cancel; also, we should just be using
2860 : // one runtime across the entire process, so, let's leave this for now.
2861 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2862 : }
2863 : };
2864 :
2865 0 : let metrics_guard = if attempt == 1 {
2866 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2867 : } else {
2868 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2869 : };
2870 :
2871 0 : let calculated_size = self_ref
2872 0 : .logical_size_calculation_task(
2873 0 : initial_part_end,
2874 0 : LogicalSizeCalculationCause::Initial,
2875 0 : background_ctx,
2876 0 : )
2877 0 : .await?;
2878 :
2879 0 : self_ref
2880 0 : .trigger_aux_file_size_computation(initial_part_end, background_ctx)
2881 0 : .await?;
2882 :
2883 : // TODO: add aux file size to logical size
2884 :
2885 0 : Ok((calculated_size, metrics_guard))
2886 0 : }
2887 0 : };
2888 :
2889 0 : let retrying = async {
2890 0 : let mut attempt = 0;
2891 : loop {
2892 0 : attempt += 1;
2893 0 :
2894 0 : match try_once(attempt).await {
2895 0 : Ok(res) => return ControlFlow::Continue(res),
2896 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
2897 : Err(
2898 0 : e @ (CalculateLogicalSizeError::Decode(_)
2899 0 : | CalculateLogicalSizeError::PageRead(_)),
2900 0 : ) => {
2901 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2902 : // exponential back-off doesn't make sense at these long intervals;
2903 : // use fixed retry interval with generous jitter instead
2904 0 : let sleep_duration = Duration::from_secs(
2905 0 : u64::try_from(
2906 0 : // 1hour base
2907 0 : (60_i64 * 60_i64)
2908 0 : // 10min jitter
2909 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2910 0 : )
2911 0 : .expect("10min < 1hour"),
2912 0 : );
2913 0 : tokio::time::sleep(sleep_duration).await;
2914 : }
2915 : }
2916 : }
2917 0 : };
2918 :
2919 0 : let (calculated_size, metrics_guard) = tokio::select! {
2920 0 : res = retrying => {
2921 0 : match res {
2922 0 : ControlFlow::Continue(calculated_size) => calculated_size,
2923 0 : ControlFlow::Break(()) => return,
2924 : }
2925 : }
2926 0 : _ = cancel.cancelled() => {
2927 0 : return;
2928 : }
2929 : };
2930 :
2931 : // we cannot query current_logical_size.current_size() to know the current
2932 : // *negative* value, only truncated to u64.
2933 0 : let added = self
2934 0 : .current_logical_size
2935 0 : .size_added_after_initial
2936 0 : .load(AtomicOrdering::Relaxed);
2937 0 :
2938 0 : let sum = calculated_size.saturating_add_signed(added);
2939 0 :
2940 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2941 0 : self.metrics.current_logical_size_gauge.set(sum);
2942 0 :
2943 0 : self.current_logical_size
2944 0 : .initial_logical_size
2945 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
2946 0 : .ok()
2947 0 : .expect("only this task sets it");
2948 0 : }
2949 :
2950 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
2951 0 : self: &Arc<Self>,
2952 0 : lsn: Lsn,
2953 0 : cause: LogicalSizeCalculationCause,
2954 0 : ctx: RequestContext,
2955 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
2956 0 : let (sender, receiver) = oneshot::channel();
2957 0 : let self_clone = Arc::clone(self);
2958 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
2959 0 : // we should stop the size calculation work and return an error.
2960 0 : // That would require restructuring this function's API to
2961 0 : // return the result directly, instead of a Receiver for the result.
2962 0 : let ctx = ctx.detached_child(
2963 0 : TaskKind::OndemandLogicalSizeCalculation,
2964 0 : DownloadBehavior::Download,
2965 0 : );
2966 0 : task_mgr::spawn(
2967 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2968 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
2969 0 : self.tenant_shard_id,
2970 0 : Some(self.timeline_id),
2971 0 : "ondemand logical size calculation",
2972 0 : async move {
2973 0 : let res = self_clone
2974 0 : .logical_size_calculation_task(lsn, cause, &ctx)
2975 0 : .await;
2976 0 : let _ = sender.send(res).ok();
2977 0 : Ok(()) // Receiver is responsible for handling errors
2978 0 : }
2979 0 : .in_current_span(),
2980 0 : );
2981 0 : receiver
2982 0 : }
2983 :
2984 : /// # Cancel-Safety
2985 : ///
2986 : /// This method is cancellation-safe.
2987 : #[instrument(skip_all)]
2988 : async fn logical_size_calculation_task(
2989 : self: &Arc<Self>,
2990 : lsn: Lsn,
2991 : cause: LogicalSizeCalculationCause,
2992 : ctx: &RequestContext,
2993 : ) -> Result<u64, CalculateLogicalSizeError> {
2994 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
2995 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
2996 : // accurate relation sizes, and they do not emit consumption metrics.
2997 : debug_assert!(self.tenant_shard_id.is_shard_zero());
2998 :
2999 : let guard = self
3000 : .gate
3001 : .enter()
3002 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3003 :
3004 : let self_calculation = Arc::clone(self);
3005 :
3006 0 : let mut calculation = pin!(async {
3007 0 : let ctx = ctx.attached_child();
3008 0 : self_calculation
3009 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
3010 0 : .await
3011 0 : });
3012 :
3013 : tokio::select! {
3014 : res = &mut calculation => { res }
3015 : _ = self.cancel.cancelled() => {
3016 : debug!("cancelling logical size calculation for timeline shutdown");
3017 : calculation.await
3018 : }
3019 : }
3020 : }
3021 :
3022 : /// Calculate the logical size of the database at the latest LSN.
3023 : ///
3024 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3025 : /// especially if we need to download remote layers.
3026 : ///
3027 : /// # Cancel-Safety
3028 : ///
3029 : /// This method is cancellation-safe.
3030 0 : async fn calculate_logical_size(
3031 0 : &self,
3032 0 : up_to_lsn: Lsn,
3033 0 : cause: LogicalSizeCalculationCause,
3034 0 : _guard: &GateGuard,
3035 0 : ctx: &RequestContext,
3036 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3037 0 : info!(
3038 0 : "Calculating logical size for timeline {} at {}",
3039 : self.timeline_id, up_to_lsn
3040 : );
3041 :
3042 0 : pausable_failpoint!("timeline-calculate-logical-size-pause");
3043 :
3044 : // See if we've already done the work for initial size calculation.
3045 : // This is a short-cut for timelines that are mostly unused.
3046 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3047 0 : return Ok(size);
3048 0 : }
3049 0 : let storage_time_metrics = match cause {
3050 : LogicalSizeCalculationCause::Initial
3051 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3052 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3053 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3054 0 : &self.metrics.imitate_logical_size_histo
3055 : }
3056 : };
3057 0 : let timer = storage_time_metrics.start_timer();
3058 0 : let logical_size = self
3059 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3060 0 : .await?;
3061 0 : debug!("calculated logical size: {logical_size}");
3062 0 : timer.stop_and_record();
3063 0 : Ok(logical_size)
3064 0 : }
3065 :
3066 : /// Update current logical size, adding `delta' to the old value.
3067 270570 : fn update_current_logical_size(&self, delta: i64) {
3068 270570 : let logical_size = &self.current_logical_size;
3069 270570 : logical_size.increment_size(delta);
3070 270570 :
3071 270570 : // Also set the value in the prometheus gauge. Note that
3072 270570 : // there is a race condition here: if this is is called by two
3073 270570 : // threads concurrently, the prometheus gauge might be set to
3074 270570 : // one value while current_logical_size is set to the
3075 270570 : // other.
3076 270570 : match logical_size.current_size() {
3077 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
3078 270570 : .metrics
3079 270570 : .current_logical_size_gauge
3080 270570 : .set(new_current_size.into()),
3081 0 : CurrentLogicalSize::Approximate(_) => {
3082 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3083 0 : // forth between the initial size calculation task.
3084 0 : }
3085 : }
3086 270570 : }
3087 :
3088 2956 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
3089 2956 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3090 2956 : let aux_metric =
3091 2956 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3092 2956 :
3093 2956 : let sum_of_entries = self
3094 2956 : .directory_metrics
3095 2956 : .iter()
3096 20692 : .map(|v| v.load(AtomicOrdering::Relaxed))
3097 2956 : .sum();
3098 : // Set a high general threshold and a lower threshold for the auxiliary files,
3099 : // as we can have large numbers of relations in the db directory.
3100 : const SUM_THRESHOLD: u64 = 5000;
3101 : const AUX_THRESHOLD: u64 = 1000;
3102 2956 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3103 0 : self.metrics
3104 0 : .directory_entries_count_gauge
3105 0 : .set(sum_of_entries);
3106 2956 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3107 0 : metric.set(sum_of_entries);
3108 2956 : }
3109 2956 : }
3110 :
3111 0 : async fn find_layer(
3112 0 : &self,
3113 0 : layer_name: &LayerName,
3114 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3115 0 : let guard = self.layers.read().await;
3116 0 : let layer = guard
3117 0 : .layer_map()?
3118 0 : .iter_historic_layers()
3119 0 : .find(|l| &l.layer_name() == layer_name)
3120 0 : .map(|found| guard.get_from_desc(&found));
3121 0 : Ok(layer)
3122 0 : }
3123 :
3124 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3125 : /// indicating which layers are currently on-disk on the primary.
3126 : ///
3127 : /// None is returned if the Timeline is in a state where uploading a heatmap
3128 : /// doesn't make sense, such as shutting down or initializing. The caller
3129 : /// should treat this as a cue to simply skip doing any heatmap uploading
3130 : /// for this timeline.
3131 2 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3132 2 : if !self.is_active() {
3133 0 : return None;
3134 2 : }
3135 :
3136 2 : let guard = self.layers.read().await;
3137 :
3138 10 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3139 10 : match layer.visibility() {
3140 : LayerVisibilityHint::Visible => {
3141 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3142 8 : let last_activity_ts = layer.latest_activity();
3143 8 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
3144 : }
3145 : LayerVisibilityHint::Covered => {
3146 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3147 2 : None
3148 : }
3149 : }
3150 10 : });
3151 2 :
3152 2 : let mut layers = resident.collect::<Vec<_>>();
3153 2 :
3154 2 : // Sort layers in order of which to download first. For a large set of layers to download, we
3155 2 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3156 2 : // or hours later:
3157 2 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3158 2 : // only exist for a few minutes before being compacted into L1s.
3159 2 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3160 2 : // the layer is likely to be covered by an image layer during compaction.
3161 18 : layers.sort_by_key(|(desc, _meta, _atime)| {
3162 18 : std::cmp::Reverse((
3163 18 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3164 18 : desc.lsn_range.end,
3165 18 : ))
3166 18 : });
3167 2 :
3168 2 : let layers = layers
3169 2 : .into_iter()
3170 8 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3171 2 : .collect();
3172 2 :
3173 2 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3174 2 : }
3175 :
3176 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3177 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3178 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3179 0 : // branchpoint in the value in IndexPart::lineage
3180 0 : self.ancestor_lsn == lsn
3181 0 : || (self.ancestor_lsn == Lsn::INVALID
3182 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3183 0 : }
3184 : }
3185 :
3186 : impl Timeline {
3187 : #[allow(clippy::doc_lazy_continuation)]
3188 : /// Get the data needed to reconstruct all keys in the provided keyspace
3189 : ///
3190 : /// The algorithm is as follows:
3191 : /// 1. While some keys are still not done and there's a timeline to visit:
3192 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3193 : /// 2.1: Build the fringe for the current keyspace
3194 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3195 : /// intersects
3196 : /// 2.3. Pop the timeline from the fringe
3197 : /// 2.4. If the fringe is empty, go back to 1
3198 627609 : async fn get_vectored_reconstruct_data(
3199 627609 : &self,
3200 627609 : mut keyspace: KeySpace,
3201 627609 : request_lsn: Lsn,
3202 627609 : reconstruct_state: &mut ValuesReconstructState,
3203 627609 : ctx: &RequestContext,
3204 627609 : ) -> Result<(), GetVectoredError> {
3205 627609 : let mut timeline_owned: Arc<Timeline>;
3206 627609 : let mut timeline = self;
3207 627609 :
3208 627609 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3209 :
3210 627607 : let missing_keyspace = loop {
3211 852352 : if self.cancel.is_cancelled() {
3212 0 : return Err(GetVectoredError::Cancelled);
3213 852352 : }
3214 :
3215 : let TimelineVisitOutcome {
3216 852352 : completed_keyspace: completed,
3217 852352 : image_covered_keyspace,
3218 852352 : } = Self::get_vectored_reconstruct_data_timeline(
3219 852352 : timeline,
3220 852352 : keyspace.clone(),
3221 852352 : cont_lsn,
3222 852352 : reconstruct_state,
3223 852352 : &self.cancel,
3224 852352 : ctx,
3225 852352 : )
3226 852352 : .await?;
3227 :
3228 852352 : keyspace.remove_overlapping_with(&completed);
3229 852352 :
3230 852352 : // Do not descend into the ancestor timeline for aux files.
3231 852352 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3232 852352 : // stalling compaction.
3233 852352 : keyspace.remove_overlapping_with(&KeySpace {
3234 852352 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
3235 852352 : });
3236 852352 :
3237 852352 : // Keyspace is fully retrieved
3238 852352 : if keyspace.is_empty() {
3239 627579 : break None;
3240 224773 : }
3241 :
3242 224773 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3243 : // Not fully retrieved but no ancestor timeline.
3244 28 : break Some(keyspace);
3245 : };
3246 :
3247 : // Now we see if there are keys covered by the image layer but does not exist in the
3248 : // image layer, which means that the key does not exist.
3249 :
3250 : // The block below will stop the vectored search if any of the keys encountered an image layer
3251 : // which did not contain a snapshot for said key. Since we have already removed all completed
3252 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3253 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3254 : // and stop the search as a result of that.
3255 224745 : let mut removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3256 224745 : // Do not fire missing key error for sparse keys.
3257 224745 : removed.remove_overlapping_with(&KeySpace {
3258 224745 : ranges: vec![SPARSE_RANGE],
3259 224745 : });
3260 224745 : if !removed.is_empty() {
3261 0 : break Some(removed);
3262 224745 : }
3263 224745 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3264 224745 : // keyspace.
3265 224745 :
3266 224745 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3267 224745 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3268 224745 : timeline_owned = timeline
3269 224745 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3270 224745 : .await?;
3271 224743 : timeline = &*timeline_owned;
3272 : };
3273 :
3274 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
3275 627607 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
3276 28 : let mut missing_keyspace = missing_keyspace;
3277 28 : missing_keyspace.remove_overlapping_with(&KeySpace {
3278 28 : ranges: vec![SPARSE_RANGE],
3279 28 : });
3280 28 : if missing_keyspace.is_empty() {
3281 14 : None
3282 : } else {
3283 14 : Some(missing_keyspace)
3284 : }
3285 : } else {
3286 627579 : None
3287 : };
3288 :
3289 627607 : if let Some(missing_keyspace) = missing_keyspace {
3290 14 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3291 14 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3292 14 : shard: self
3293 14 : .shard_identity
3294 14 : .get_shard_number(&missing_keyspace.start().unwrap()),
3295 14 : cont_lsn,
3296 14 : request_lsn,
3297 14 : ancestor_lsn: Some(timeline.ancestor_lsn),
3298 14 : backtrace: None,
3299 14 : }));
3300 627593 : }
3301 627593 :
3302 627593 : Ok(())
3303 627609 : }
3304 :
3305 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3306 : ///
3307 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3308 : /// the current keyspace. The current keyspace of the search at any given timeline
3309 : /// is the original keyspace minus all the keys that have been completed minus
3310 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3311 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3312 : ///
3313 : /// This is basically a depth-first search visitor implementation where a vertex
3314 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3315 : ///
3316 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3317 : /// and get all the required reconstruct data from the layer in one go.
3318 : ///
3319 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3320 : /// decides how to deal with these two keyspaces.
3321 852352 : async fn get_vectored_reconstruct_data_timeline(
3322 852352 : timeline: &Timeline,
3323 852352 : keyspace: KeySpace,
3324 852352 : mut cont_lsn: Lsn,
3325 852352 : reconstruct_state: &mut ValuesReconstructState,
3326 852352 : cancel: &CancellationToken,
3327 852352 : ctx: &RequestContext,
3328 852352 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3329 852352 : let mut unmapped_keyspace = keyspace.clone();
3330 852352 : let mut fringe = LayerFringe::new();
3331 852352 :
3332 852352 : let mut completed_keyspace = KeySpace::default();
3333 852352 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3334 :
3335 : loop {
3336 1699054 : if cancel.is_cancelled() {
3337 0 : return Err(GetVectoredError::Cancelled);
3338 1699054 : }
3339 1699054 :
3340 1699054 : let (keys_done_last_step, keys_with_image_coverage) =
3341 1699054 : reconstruct_state.consume_done_keys();
3342 1699054 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3343 1699054 : completed_keyspace.merge(&keys_done_last_step);
3344 1699054 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3345 22546 : unmapped_keyspace
3346 22546 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3347 22546 : image_covered_keyspace.add_range(keys_with_image_coverage);
3348 1676508 : }
3349 :
3350 : // Do not descent any further if the last layer we visited
3351 : // completed all keys in the keyspace it inspected. This is not
3352 : // required for correctness, but avoids visiting extra layers
3353 : // which turns out to be a perf bottleneck in some cases.
3354 1699054 : if !unmapped_keyspace.is_empty() {
3355 1073717 : let guard = timeline.layers.read().await;
3356 1073717 : let layers = guard.layer_map()?;
3357 :
3358 1073717 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3359 912481 : let start_lsn = l.get_lsn_range().start;
3360 912481 : cont_lsn > start_lsn
3361 1073717 : });
3362 1073717 :
3363 1073717 : match in_memory_layer {
3364 606677 : Some(l) => {
3365 606677 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3366 606677 : fringe.update(
3367 606677 : ReadableLayer::InMemoryLayer(l),
3368 606677 : unmapped_keyspace.clone(),
3369 606677 : lsn_range,
3370 606677 : );
3371 606677 : }
3372 : None => {
3373 467064 : for range in unmapped_keyspace.ranges.iter() {
3374 467064 : let results = layers.range_search(range.clone(), cont_lsn);
3375 467064 :
3376 467064 : results
3377 467064 : .found
3378 467064 : .into_iter()
3379 467064 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3380 240039 : (
3381 240039 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3382 240039 : keyspace_accum.to_keyspace(),
3383 240039 : lsn_floor..cont_lsn,
3384 240039 : )
3385 467064 : })
3386 467064 : .for_each(|(layer, keyspace, lsn_range)| {
3387 240039 : fringe.update(layer, keyspace, lsn_range)
3388 467064 : });
3389 467064 : }
3390 : }
3391 : }
3392 :
3393 : // It's safe to drop the layer map lock after planning the next round of reads.
3394 : // The fringe keeps readable handles for the layers which are safe to read even
3395 : // if layers were compacted or flushed.
3396 : //
3397 : // The more interesting consideration is: "Why is the read algorithm still correct
3398 : // if the layer map changes while it is operating?". Doing a vectored read on a
3399 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3400 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3401 : // range at *a particular point in time*. It is fine for the answer to be different
3402 : // at two different time points.
3403 1073717 : drop(guard);
3404 625337 : }
3405 :
3406 1699054 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3407 846702 : let next_cont_lsn = lsn_range.start;
3408 846702 : layer_to_read
3409 846702 : .get_values_reconstruct_data(
3410 846702 : keyspace_to_read.clone(),
3411 846702 : lsn_range,
3412 846702 : reconstruct_state,
3413 846702 : ctx,
3414 846702 : )
3415 846702 : .await?;
3416 :
3417 846702 : unmapped_keyspace = keyspace_to_read;
3418 846702 : cont_lsn = next_cont_lsn;
3419 846702 :
3420 846702 : reconstruct_state.on_layer_visited(&layer_to_read);
3421 : } else {
3422 852352 : break;
3423 852352 : }
3424 852352 : }
3425 852352 :
3426 852352 : Ok(TimelineVisitOutcome {
3427 852352 : completed_keyspace,
3428 852352 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3429 852352 : })
3430 852352 : }
3431 :
3432 224745 : async fn get_ready_ancestor_timeline(
3433 224745 : &self,
3434 224745 : ancestor: &Arc<Timeline>,
3435 224745 : ctx: &RequestContext,
3436 224745 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3437 224745 : // It's possible that the ancestor timeline isn't active yet, or
3438 224745 : // is active but hasn't yet caught up to the branch point. Wait
3439 224745 : // for it.
3440 224745 : //
3441 224745 : // This cannot happen while the pageserver is running normally,
3442 224745 : // because you cannot create a branch from a point that isn't
3443 224745 : // present in the pageserver yet. However, we don't wait for the
3444 224745 : // branch point to be uploaded to cloud storage before creating
3445 224745 : // a branch. I.e., the branch LSN need not be remote consistent
3446 224745 : // for the branching operation to succeed.
3447 224745 : //
3448 224745 : // Hence, if we try to load a tenant in such a state where
3449 224745 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3450 224745 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3451 224745 : // then we will need to wait for the ancestor timeline to
3452 224745 : // re-stream WAL up to branch_lsn before we access it.
3453 224745 : //
3454 224745 : // How can a tenant get in such a state?
3455 224745 : // - ungraceful pageserver process exit
3456 224745 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3457 224745 : //
3458 224745 : // NB: this could be avoided by requiring
3459 224745 : // branch_lsn >= remote_consistent_lsn
3460 224745 : // during branch creation.
3461 224745 : match ancestor.wait_to_become_active(ctx).await {
3462 224743 : Ok(()) => {}
3463 : Err(TimelineState::Stopping) => {
3464 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3465 0 : return Err(GetReadyAncestorError::Cancelled);
3466 : }
3467 2 : Err(state) => {
3468 2 : return Err(GetReadyAncestorError::BadState {
3469 2 : timeline_id: ancestor.timeline_id,
3470 2 : state,
3471 2 : });
3472 : }
3473 : }
3474 224743 : ancestor
3475 224743 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3476 224743 : .await
3477 224743 : .map_err(|e| match e {
3478 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3479 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3480 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3481 0 : timeline_id: ancestor.timeline_id,
3482 0 : state,
3483 0 : },
3484 224743 : })?;
3485 :
3486 224743 : Ok(ancestor.clone())
3487 224745 : }
3488 :
3489 297156 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3490 297156 : &self.shard_identity
3491 297156 : }
3492 :
3493 : #[inline(always)]
3494 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3495 0 : ShardTimelineId {
3496 0 : shard_index: ShardIndex {
3497 0 : shard_number: self.shard_identity.number,
3498 0 : shard_count: self.shard_identity.count,
3499 0 : },
3500 0 : timeline_id: self.timeline_id,
3501 0 : }
3502 0 : }
3503 :
3504 : /// Returns a non-frozen open in-memory layer for ingestion.
3505 : ///
3506 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3507 : /// this function without holding the mutex.
3508 1296 : async fn get_layer_for_write(
3509 1296 : &self,
3510 1296 : lsn: Lsn,
3511 1296 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3512 1296 : ctx: &RequestContext,
3513 1296 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3514 1296 : let mut guard = self.layers.write().await;
3515 :
3516 1296 : let last_record_lsn = self.get_last_record_lsn();
3517 1296 : ensure!(
3518 1296 : lsn > last_record_lsn,
3519 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3520 : lsn,
3521 : last_record_lsn,
3522 : );
3523 :
3524 1296 : let layer = guard
3525 1296 : .open_mut()?
3526 1296 : .get_layer_for_write(
3527 1296 : lsn,
3528 1296 : self.conf,
3529 1296 : self.timeline_id,
3530 1296 : self.tenant_shard_id,
3531 1296 : &self.gate,
3532 1296 : ctx,
3533 1296 : )
3534 1296 : .await?;
3535 1296 : Ok(layer)
3536 1296 : }
3537 :
3538 5279092 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3539 5279092 : assert!(new_lsn.is_aligned());
3540 :
3541 5279092 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
3542 5279092 : self.last_record_lsn.advance(new_lsn);
3543 5279092 : }
3544 :
3545 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3546 : ///
3547 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3548 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3549 1200 : async fn freeze_inmem_layer_at(
3550 1200 : &self,
3551 1200 : at: Lsn,
3552 1200 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3553 1200 : ) -> Result<u64, FlushLayerError> {
3554 1200 : let frozen = {
3555 1200 : let mut guard = self.layers.write().await;
3556 1200 : guard
3557 1200 : .open_mut()?
3558 1200 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock)
3559 1200 : .await
3560 : };
3561 :
3562 1200 : if frozen {
3563 1172 : let now = Instant::now();
3564 1172 : *(self.last_freeze_ts.write().unwrap()) = now;
3565 1172 : }
3566 :
3567 : // Increment the flush cycle counter and wake up the flush task.
3568 : // Remember the new value, so that when we listen for the flush
3569 : // to finish, we know when the flush that we initiated has
3570 : // finished, instead of some other flush that was started earlier.
3571 1200 : let mut my_flush_request = 0;
3572 1200 :
3573 1200 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3574 1200 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3575 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3576 1200 : }
3577 1200 :
3578 1200 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3579 1200 : my_flush_request = *counter + 1;
3580 1200 : *counter = my_flush_request;
3581 1200 : *lsn = std::cmp::max(at, *lsn);
3582 1200 : });
3583 1200 :
3584 1200 : assert_ne!(my_flush_request, 0);
3585 :
3586 1200 : Ok(my_flush_request)
3587 1200 : }
3588 :
3589 : /// Layer flusher task's main loop.
3590 440 : async fn flush_loop(
3591 440 : self: &Arc<Self>,
3592 440 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3593 440 : ctx: &RequestContext,
3594 440 : ) {
3595 440 : info!("started flush loop");
3596 : loop {
3597 1598 : tokio::select! {
3598 1598 : _ = self.cancel.cancelled() => {
3599 10 : info!("shutting down layer flush task due to Timeline::cancel");
3600 10 : break;
3601 : },
3602 1598 : _ = layer_flush_start_rx.changed() => {}
3603 1158 : }
3604 1158 : trace!("waking up");
3605 1158 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3606 1158 :
3607 1158 : // The highest LSN to which we flushed in the loop over frozen layers
3608 1158 : let mut flushed_to_lsn = Lsn(0);
3609 :
3610 1158 : let result = loop {
3611 2330 : if self.cancel.is_cancelled() {
3612 0 : info!("dropping out of flush loop for timeline shutdown");
3613 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3614 : // anyone waiting on that will respect self.cancel as well: they will stop
3615 : // waiting at the same time we as drop out of this loop.
3616 0 : return;
3617 2330 : }
3618 2330 :
3619 2330 : let timer = self.metrics.flush_time_histo.start_timer();
3620 :
3621 : let num_frozen_layers;
3622 : let frozen_layer_total_size;
3623 2330 : let layer_to_flush = {
3624 2330 : let guard = self.layers.read().await;
3625 2330 : let Ok(lm) = guard.layer_map() else {
3626 0 : info!("dropping out of flush loop for timeline shutdown");
3627 0 : return;
3628 : };
3629 2330 : num_frozen_layers = lm.frozen_layers.len();
3630 2330 : frozen_layer_total_size = lm
3631 2330 : .frozen_layers
3632 2330 : .iter()
3633 2330 : .map(|l| l.estimated_in_mem_size())
3634 2330 : .sum::<u64>();
3635 2330 : lm.frozen_layers.front().cloned()
3636 : // drop 'layers' lock to allow concurrent reads and writes
3637 : };
3638 2330 : let Some(layer_to_flush) = layer_to_flush else {
3639 1158 : break Ok(());
3640 : };
3641 1172 : if num_frozen_layers
3642 1172 : > std::cmp::max(
3643 1172 : self.get_compaction_threshold(),
3644 1172 : DEFAULT_COMPACTION_THRESHOLD,
3645 1172 : )
3646 0 : && frozen_layer_total_size >= /* 128 MB */ 128000000
3647 : {
3648 0 : tracing::warn!(
3649 0 : "too many frozen layers: {num_frozen_layers} layers with estimated in-mem size of {frozen_layer_total_size} bytes",
3650 : );
3651 1172 : }
3652 1172 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3653 1172 : Ok(this_layer_to_lsn) => {
3654 1172 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3655 1172 : }
3656 : Err(FlushLayerError::Cancelled) => {
3657 0 : info!("dropping out of flush loop for timeline shutdown");
3658 0 : return;
3659 : }
3660 0 : err @ Err(
3661 0 : FlushLayerError::NotRunning(_)
3662 0 : | FlushLayerError::Other(_)
3663 0 : | FlushLayerError::CreateImageLayersError(_),
3664 0 : ) => {
3665 0 : error!("could not flush frozen layer: {err:?}");
3666 0 : break err.map(|_| ());
3667 : }
3668 : }
3669 1172 : timer.stop_and_record();
3670 : };
3671 :
3672 : // Unsharded tenants should never advance their LSN beyond the end of the
3673 : // highest layer they write: such gaps between layer data and the frozen LSN
3674 : // are only legal on sharded tenants.
3675 1158 : debug_assert!(
3676 1158 : self.shard_identity.count.count() > 1
3677 1158 : || flushed_to_lsn >= frozen_to_lsn
3678 65 : || !flushed_to_lsn.is_valid()
3679 : );
3680 :
3681 1158 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3682 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3683 : // to us via layer_flush_start_rx, then advance it here.
3684 : //
3685 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3686 : // never encounter a gap in the wal.
3687 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3688 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3689 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3690 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3691 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3692 0 : }
3693 0 : }
3694 1158 : }
3695 :
3696 : // Notify any listeners that we're done
3697 1158 : let _ = self
3698 1158 : .layer_flush_done_tx
3699 1158 : .send_replace((flush_counter, result));
3700 : }
3701 10 : }
3702 :
3703 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
3704 1120 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3705 1120 : let mut rx = self.layer_flush_done_tx.subscribe();
3706 : loop {
3707 : {
3708 2240 : let (last_result_counter, last_result) = &*rx.borrow();
3709 2240 : if *last_result_counter >= request {
3710 1120 : if let Err(err) = last_result {
3711 : // We already logged the original error in
3712 : // flush_loop. We cannot propagate it to the caller
3713 : // here, because it might not be Cloneable
3714 0 : return Err(err.clone());
3715 : } else {
3716 1120 : return Ok(());
3717 : }
3718 1120 : }
3719 1120 : }
3720 1120 : trace!("waiting for flush to complete");
3721 1120 : tokio::select! {
3722 1120 : rx_e = rx.changed() => {
3723 1120 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3724 : },
3725 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3726 : // the notification from [`flush_loop`] that it completed.
3727 1120 : _ = self.cancel.cancelled() => {
3728 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3729 0 : return Ok(())
3730 : }
3731 : };
3732 1120 : trace!("done")
3733 : }
3734 1120 : }
3735 :
3736 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3737 : ///
3738 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3739 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3740 : async fn flush_frozen_layer(
3741 : self: &Arc<Self>,
3742 : frozen_layer: Arc<InMemoryLayer>,
3743 : ctx: &RequestContext,
3744 : ) -> Result<Lsn, FlushLayerError> {
3745 : debug_assert_current_span_has_tenant_and_timeline_id();
3746 :
3747 : // As a special case, when we have just imported an image into the repository,
3748 : // instead of writing out a L0 delta layer, we directly write out image layer
3749 : // files instead. This is possible as long as *all* the data imported into the
3750 : // repository have the same LSN.
3751 : let lsn_range = frozen_layer.get_lsn_range();
3752 :
3753 : // Whether to directly create image layers for this flush, or flush them as delta layers
3754 : let create_image_layer =
3755 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3756 :
3757 : #[cfg(test)]
3758 : {
3759 : match &mut *self.flush_loop_state.lock().unwrap() {
3760 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3761 : panic!("flush loop not running")
3762 : }
3763 : FlushLoopState::Running {
3764 : expect_initdb_optimization,
3765 : initdb_optimization_count,
3766 : ..
3767 : } => {
3768 : if create_image_layer {
3769 : *initdb_optimization_count += 1;
3770 : } else {
3771 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3772 : }
3773 : }
3774 : }
3775 : }
3776 :
3777 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3778 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3779 : // require downloading anything during initial import.
3780 : let ((rel_partition, metadata_partition), _lsn) = self
3781 : .repartition(
3782 : self.initdb_lsn,
3783 : self.get_compaction_target_size(),
3784 : EnumSet::empty(),
3785 : ctx,
3786 : )
3787 : .await
3788 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
3789 :
3790 : if self.cancel.is_cancelled() {
3791 : return Err(FlushLayerError::Cancelled);
3792 : }
3793 :
3794 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
3795 : // So that the key ranges don't overlap.
3796 : let mut partitions = KeyPartitioning::default();
3797 : partitions.parts.extend(rel_partition.parts);
3798 : if !metadata_partition.parts.is_empty() {
3799 : assert_eq!(
3800 : metadata_partition.parts.len(),
3801 : 1,
3802 : "currently sparse keyspace should only contain a single metadata keyspace"
3803 : );
3804 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
3805 : // every single key within the keyspace, and therefore, it's safe to force converting it
3806 : // into a dense keyspace before calling this function.
3807 : partitions
3808 : .parts
3809 : .extend(metadata_partition.into_dense().parts);
3810 : }
3811 :
3812 : let mut layers_to_upload = Vec::new();
3813 : layers_to_upload.extend(
3814 : self.create_image_layers(
3815 : &partitions,
3816 : self.initdb_lsn,
3817 : ImageLayerCreationMode::Initial,
3818 : ctx,
3819 : )
3820 : .await?,
3821 : );
3822 :
3823 : (layers_to_upload, None)
3824 : } else {
3825 : // Normal case, write out a L0 delta layer file.
3826 : // `create_delta_layer` will not modify the layer map.
3827 : // We will remove frozen layer and add delta layer in one atomic operation later.
3828 : let Some(layer) = self
3829 : .create_delta_layer(&frozen_layer, None, ctx)
3830 : .await
3831 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
3832 : else {
3833 : panic!("delta layer cannot be empty if no filter is applied");
3834 : };
3835 : (
3836 : // FIXME: even though we have a single image and single delta layer assumption
3837 : // we push them to vec
3838 : vec![layer.clone()],
3839 : Some(layer),
3840 : )
3841 : };
3842 :
3843 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
3844 :
3845 : if self.cancel.is_cancelled() {
3846 : return Err(FlushLayerError::Cancelled);
3847 : }
3848 :
3849 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
3850 :
3851 : // The new on-disk layers are now in the layer map. We can remove the
3852 : // in-memory layer from the map now. The flushed layer is stored in
3853 : // the mapping in `create_delta_layer`.
3854 : {
3855 : let mut guard = self.layers.write().await;
3856 :
3857 : guard.open_mut()?.finish_flush_l0_layer(
3858 : delta_layer_to_add.as_ref(),
3859 : &frozen_layer,
3860 : &self.metrics,
3861 : );
3862 :
3863 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
3864 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
3865 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
3866 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
3867 : }
3868 : // release lock on 'layers'
3869 : };
3870 :
3871 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
3872 : // This makes us refuse ingest until the new layers have been persisted to the remote
3873 : let start = Instant::now();
3874 : self.remote_client
3875 : .wait_completion()
3876 : .await
3877 0 : .map_err(|e| match e {
3878 : WaitCompletionError::UploadQueueShutDownOrStopped
3879 : | WaitCompletionError::NotInitialized(
3880 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
3881 0 : ) => FlushLayerError::Cancelled,
3882 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
3883 0 : FlushLayerError::Other(anyhow!(e).into())
3884 : }
3885 0 : })?;
3886 : let duration = start.elapsed().as_secs_f64();
3887 : self.metrics.flush_wait_upload_time_gauge_add(duration);
3888 :
3889 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
3890 : // a compaction can delete the file and then it won't be available for uploads any more.
3891 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
3892 : // race situation.
3893 : // See https://github.com/neondatabase/neon/issues/4526
3894 : pausable_failpoint!("flush-frozen-pausable");
3895 :
3896 : // This failpoint is used by another test case `test_pageserver_recovery`.
3897 : fail_point!("flush-frozen-exit");
3898 :
3899 : Ok(Lsn(lsn_range.end.0 - 1))
3900 : }
3901 :
3902 : /// Return true if the value changed
3903 : ///
3904 : /// This function must only be used from the layer flush task.
3905 1172 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
3906 1172 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
3907 1172 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
3908 :
3909 1172 : self.metrics
3910 1172 : .disk_consistent_lsn_gauge
3911 1172 : .set(new_value.0 as i64);
3912 1172 : new_value != old_value
3913 1172 : }
3914 :
3915 : /// Update metadata file
3916 1222 : fn schedule_uploads(
3917 1222 : &self,
3918 1222 : disk_consistent_lsn: Lsn,
3919 1222 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
3920 1222 : ) -> anyhow::Result<()> {
3921 1222 : // We can only save a valid 'prev_record_lsn' value on disk if we
3922 1222 : // flushed *all* in-memory changes to disk. We only track
3923 1222 : // 'prev_record_lsn' in memory for the latest processed record, so we
3924 1222 : // don't remember what the correct value that corresponds to some old
3925 1222 : // LSN is. But if we flush everything, then the value corresponding
3926 1222 : // current 'last_record_lsn' is correct and we can store it on disk.
3927 1222 : let RecordLsn {
3928 1222 : last: last_record_lsn,
3929 1222 : prev: prev_record_lsn,
3930 1222 : } = self.last_record_lsn.load();
3931 1222 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
3932 1094 : Some(prev_record_lsn)
3933 : } else {
3934 128 : None
3935 : };
3936 :
3937 1222 : let update = crate::tenant::metadata::MetadataUpdate::new(
3938 1222 : disk_consistent_lsn,
3939 1222 : ondisk_prev_record_lsn,
3940 1222 : *self.latest_gc_cutoff_lsn.read(),
3941 1222 : );
3942 1222 :
3943 1222 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
3944 0 : "{}",
3945 0 : x.unwrap()
3946 1222 : ));
3947 :
3948 2406 : for layer in layers_to_upload {
3949 1184 : self.remote_client.schedule_layer_file_upload(layer)?;
3950 : }
3951 1222 : self.remote_client
3952 1222 : .schedule_index_upload_for_metadata_update(&update)?;
3953 :
3954 1222 : Ok(())
3955 1222 : }
3956 :
3957 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
3958 0 : self.remote_client
3959 0 : .preserve_initdb_archive(
3960 0 : &self.tenant_shard_id.tenant_id,
3961 0 : &self.timeline_id,
3962 0 : &self.cancel,
3963 0 : )
3964 0 : .await
3965 0 : }
3966 :
3967 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
3968 : // in layer map immediately. The caller is responsible to put it into the layer map.
3969 968 : async fn create_delta_layer(
3970 968 : self: &Arc<Self>,
3971 968 : frozen_layer: &Arc<InMemoryLayer>,
3972 968 : key_range: Option<Range<Key>>,
3973 968 : ctx: &RequestContext,
3974 968 : ) -> anyhow::Result<Option<ResidentLayer>> {
3975 968 : let self_clone = Arc::clone(self);
3976 968 : let frozen_layer = Arc::clone(frozen_layer);
3977 968 : let ctx = ctx.attached_child();
3978 968 : let work = async move {
3979 968 : let Some((desc, path)) = frozen_layer
3980 968 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
3981 968 : .await?
3982 : else {
3983 0 : return Ok(None);
3984 : };
3985 968 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
3986 :
3987 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
3988 : // We just need to fsync the directory in which these inodes are linked,
3989 : // which we know to be the timeline directory.
3990 : //
3991 : // We use fatal_err() below because the after write_to_disk returns with success,
3992 : // the in-memory state of the filesystem already has the layer file in its final place,
3993 : // and subsequent pageserver code could think it's durable while it really isn't.
3994 968 : let timeline_dir = VirtualFile::open(
3995 968 : &self_clone
3996 968 : .conf
3997 968 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
3998 968 : &ctx,
3999 968 : )
4000 968 : .await
4001 968 : .fatal_err("VirtualFile::open for timeline dir fsync");
4002 968 : timeline_dir
4003 968 : .sync_all()
4004 968 : .await
4005 968 : .fatal_err("VirtualFile::sync_all timeline dir");
4006 968 : anyhow::Ok(Some(new_delta))
4007 968 : };
4008 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
4009 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
4010 : use crate::virtual_file::io_engine::IoEngine;
4011 968 : match crate::virtual_file::io_engine::get() {
4012 0 : IoEngine::NotSet => panic!("io engine not set"),
4013 : IoEngine::StdFs => {
4014 484 : let span = tracing::info_span!("blocking");
4015 484 : tokio::task::spawn_blocking({
4016 484 : move || Handle::current().block_on(work.instrument(span))
4017 484 : })
4018 484 : .await
4019 484 : .context("spawn_blocking")
4020 484 : .and_then(|x| x)
4021 : }
4022 : #[cfg(target_os = "linux")]
4023 484 : IoEngine::TokioEpollUring => work.await,
4024 : }
4025 968 : }
4026 :
4027 568 : async fn repartition(
4028 568 : &self,
4029 568 : lsn: Lsn,
4030 568 : partition_size: u64,
4031 568 : flags: EnumSet<CompactFlags>,
4032 568 : ctx: &RequestContext,
4033 568 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
4034 568 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
4035 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
4036 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
4037 : // and hence before the compaction task starts.
4038 0 : return Err(CompactionError::Other(anyhow!(
4039 0 : "repartition() called concurrently"
4040 0 : )));
4041 : };
4042 568 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
4043 568 : if lsn < *partition_lsn {
4044 0 : return Err(CompactionError::Other(anyhow!(
4045 0 : "repartition() called with LSN going backwards, this should not happen"
4046 0 : )));
4047 568 : }
4048 568 :
4049 568 : let distance = lsn.0 - partition_lsn.0;
4050 568 : if *partition_lsn != Lsn(0)
4051 262 : && distance <= self.repartition_threshold
4052 262 : && !flags.contains(CompactFlags::ForceRepartition)
4053 : {
4054 248 : debug!(
4055 : distance,
4056 : threshold = self.repartition_threshold,
4057 0 : "no repartitioning needed"
4058 : );
4059 248 : return Ok((
4060 248 : (dense_partition.clone(), sparse_partition.clone()),
4061 248 : *partition_lsn,
4062 248 : ));
4063 320 : }
4064 :
4065 320 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4066 320 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4067 320 : let sparse_partitioning = SparseKeyPartitioning {
4068 320 : parts: vec![sparse_ks],
4069 320 : }; // no partitioning for metadata keys for now
4070 320 : let result = ((dense_partitioning, sparse_partitioning), lsn);
4071 320 : guard.write(result.clone());
4072 320 : Ok(result)
4073 568 : }
4074 :
4075 : // Is it time to create a new image layer for the given partition?
4076 14 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4077 14 : let threshold = self.get_image_creation_threshold();
4078 :
4079 14 : let guard = self.layers.read().await;
4080 14 : let Ok(layers) = guard.layer_map() else {
4081 0 : return false;
4082 : };
4083 :
4084 14 : let mut max_deltas = 0;
4085 28 : for part_range in &partition.ranges {
4086 14 : let image_coverage = layers.image_coverage(part_range, lsn);
4087 28 : for (img_range, last_img) in image_coverage {
4088 14 : let img_lsn = if let Some(last_img) = last_img {
4089 0 : last_img.get_lsn_range().end
4090 : } else {
4091 14 : Lsn(0)
4092 : };
4093 : // Let's consider an example:
4094 : //
4095 : // delta layer with LSN range 71-81
4096 : // delta layer with LSN range 81-91
4097 : // delta layer with LSN range 91-101
4098 : // image layer at LSN 100
4099 : //
4100 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4101 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4102 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4103 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4104 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4105 14 : if img_lsn < lsn {
4106 14 : let num_deltas =
4107 14 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4108 14 :
4109 14 : max_deltas = max_deltas.max(num_deltas);
4110 14 : if num_deltas >= threshold {
4111 0 : debug!(
4112 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4113 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4114 : );
4115 0 : return true;
4116 14 : }
4117 0 : }
4118 : }
4119 : }
4120 :
4121 14 : debug!(
4122 : max_deltas,
4123 0 : "none of the partitioned ranges had >= {threshold} deltas"
4124 : );
4125 14 : false
4126 14 : }
4127 :
4128 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4129 : /// so that at most one image layer will be produced from this function.
4130 230 : async fn create_image_layer_for_rel_blocks(
4131 230 : self: &Arc<Self>,
4132 230 : partition: &KeySpace,
4133 230 : mut image_layer_writer: ImageLayerWriter,
4134 230 : lsn: Lsn,
4135 230 : ctx: &RequestContext,
4136 230 : img_range: Range<Key>,
4137 230 : start: Key,
4138 230 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4139 230 : let mut wrote_keys = false;
4140 230 :
4141 230 : let mut key_request_accum = KeySpaceAccum::new();
4142 1522 : for range in &partition.ranges {
4143 1292 : let mut key = range.start;
4144 2800 : while key < range.end {
4145 : // Decide whether to retain this key: usually we do, but sharded tenants may
4146 : // need to drop keys that don't belong to them. If we retain the key, add it
4147 : // to `key_request_accum` for later issuing a vectored get
4148 1508 : if self.shard_identity.is_key_disposable(&key) {
4149 0 : debug!(
4150 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4151 0 : key,
4152 0 : self.shard_identity.get_shard_number(&key)
4153 : );
4154 1508 : } else {
4155 1508 : key_request_accum.add_key(key);
4156 1508 : }
4157 :
4158 1508 : let last_key_in_range = key.next() == range.end;
4159 1508 : key = key.next();
4160 1508 :
4161 1508 : // Maybe flush `key_rest_accum`
4162 1508 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4163 1508 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4164 : {
4165 1292 : let results = self
4166 1292 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
4167 1292 : .await?;
4168 :
4169 1292 : if self.cancel.is_cancelled() {
4170 0 : return Err(CreateImageLayersError::Cancelled);
4171 1292 : }
4172 :
4173 2800 : for (img_key, img) in results {
4174 1508 : let img = match img {
4175 1508 : Ok(img) => img,
4176 0 : Err(err) => {
4177 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4178 0 : // page without losing any actual user data. That seems better
4179 0 : // than failing repeatedly and getting stuck.
4180 0 : //
4181 0 : // We had a bug at one point, where we truncated the FSM and VM
4182 0 : // in the pageserver, but the Postgres didn't know about that
4183 0 : // and continued to generate incremental WAL records for pages
4184 0 : // that didn't exist in the pageserver. Trying to replay those
4185 0 : // WAL records failed to find the previous image of the page.
4186 0 : // This special case allows us to recover from that situation.
4187 0 : // See https://github.com/neondatabase/neon/issues/2601.
4188 0 : //
4189 0 : // Unfortunately we cannot do this for the main fork, or for
4190 0 : // any metadata keys, keys, as that would lead to actual data
4191 0 : // loss.
4192 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4193 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4194 0 : ZERO_PAGE.clone()
4195 : } else {
4196 0 : return Err(CreateImageLayersError::from(err));
4197 : }
4198 : }
4199 : };
4200 :
4201 : // Write all the keys we just read into our new image layer.
4202 1508 : image_layer_writer.put_image(img_key, img, ctx).await?;
4203 1508 : wrote_keys = true;
4204 : }
4205 216 : }
4206 : }
4207 : }
4208 :
4209 230 : if wrote_keys {
4210 : // Normal path: we have written some data into the new image layer for this
4211 : // partition, so flush it to disk.
4212 230 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4213 230 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4214 230 : info!("created image layer for rel {}", image_layer.local_path());
4215 230 : Ok(ImageLayerCreationOutcome {
4216 230 : image: Some(image_layer),
4217 230 : next_start_key: img_range.end,
4218 230 : })
4219 : } else {
4220 : // Special case: the image layer may be empty if this is a sharded tenant and the
4221 : // partition does not cover any keys owned by this shard. In this case, to ensure
4222 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4223 : // layer we write will cover the key range that we just scanned.
4224 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4225 0 : Ok(ImageLayerCreationOutcome {
4226 0 : image: None,
4227 0 : next_start_key: start,
4228 0 : })
4229 : }
4230 230 : }
4231 :
4232 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4233 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4234 : /// would not be too large to fit in a single image layer.
4235 : #[allow(clippy::too_many_arguments)]
4236 220 : async fn create_image_layer_for_metadata_keys(
4237 220 : self: &Arc<Self>,
4238 220 : partition: &KeySpace,
4239 220 : mut image_layer_writer: ImageLayerWriter,
4240 220 : lsn: Lsn,
4241 220 : ctx: &RequestContext,
4242 220 : img_range: Range<Key>,
4243 220 : mode: ImageLayerCreationMode,
4244 220 : start: Key,
4245 220 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4246 220 : // Metadata keys image layer creation.
4247 220 : let mut reconstruct_state = ValuesReconstructState::default();
4248 220 : let begin = Instant::now();
4249 220 : let data = self
4250 220 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4251 220 : .await?;
4252 220 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4253 220 : let mut new_data = BTreeMap::new();
4254 220 : let mut total_kb_retrieved = 0;
4255 220 : let mut total_keys_retrieved = 0;
4256 10232 : for (k, v) in data {
4257 10012 : let v = v?;
4258 10012 : total_kb_retrieved += KEY_SIZE + v.len();
4259 10012 : total_keys_retrieved += 1;
4260 10012 : new_data.insert(k, v);
4261 : }
4262 220 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4263 220 : };
4264 220 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4265 220 : let elapsed = begin.elapsed();
4266 220 :
4267 220 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4268 220 : info!(
4269 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()
4270 : );
4271 :
4272 220 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4273 2 : return Ok(ImageLayerCreationOutcome {
4274 2 : image: None,
4275 2 : next_start_key: img_range.end,
4276 2 : });
4277 218 : }
4278 218 : if self.cancel.is_cancelled() {
4279 0 : return Err(CreateImageLayersError::Cancelled);
4280 218 : }
4281 218 : let mut wrote_any_image = false;
4282 10230 : for (k, v) in data {
4283 10012 : if v.is_empty() {
4284 : // the key has been deleted, it does not need an image
4285 : // in metadata keyspace, an empty image == tombstone
4286 8 : continue;
4287 10004 : }
4288 10004 : wrote_any_image = true;
4289 10004 :
4290 10004 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4291 10004 :
4292 10004 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4293 10004 : // on the normal data path either.
4294 10004 : image_layer_writer.put_image(k, v, ctx).await?;
4295 : }
4296 :
4297 218 : if wrote_any_image {
4298 : // Normal path: we have written some data into the new image layer for this
4299 : // partition, so flush it to disk.
4300 12 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4301 12 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4302 12 : info!(
4303 0 : "created image layer for metadata {}",
4304 0 : image_layer.local_path()
4305 : );
4306 12 : Ok(ImageLayerCreationOutcome {
4307 12 : image: Some(image_layer),
4308 12 : next_start_key: img_range.end,
4309 12 : })
4310 : } else {
4311 : // Special case: the image layer may be empty if this is a sharded tenant and the
4312 : // partition does not cover any keys owned by this shard. In this case, to ensure
4313 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4314 : // layer we write will cover the key range that we just scanned.
4315 206 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4316 206 : Ok(ImageLayerCreationOutcome {
4317 206 : image: None,
4318 206 : next_start_key: start,
4319 206 : })
4320 : }
4321 220 : }
4322 :
4323 : /// Predicate function which indicates whether we should check if new image layers
4324 : /// are required. Since checking if new image layers are required is expensive in
4325 : /// terms of CPU, we only do it in the following cases:
4326 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4327 : /// 2. If enough time has passed since the last check:
4328 : /// 1. For large tenants, we wish to perform the check more often since they
4329 : /// suffer from the lack of image layers
4330 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4331 568 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4332 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4333 :
4334 568 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4335 568 : let distance = lsn
4336 568 : .checked_sub(last_checks_at)
4337 568 : .expect("Attempt to compact with LSN going backwards");
4338 568 : let min_distance =
4339 568 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4340 568 :
4341 568 : let distance_based_decision = distance.0 >= min_distance;
4342 568 :
4343 568 : let mut time_based_decision = false;
4344 568 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4345 568 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4346 466 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4347 : {
4348 0 : self.get_checkpoint_timeout()
4349 : } else {
4350 466 : Duration::from_secs(3600 * 48)
4351 : };
4352 :
4353 466 : time_based_decision = match *last_check_instant {
4354 262 : Some(last_check) => {
4355 262 : let elapsed = last_check.elapsed();
4356 262 : elapsed >= check_required_after
4357 : }
4358 204 : None => true,
4359 : };
4360 102 : }
4361 :
4362 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4363 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4364 : // check.
4365 568 : let decision = distance_based_decision || time_based_decision;
4366 :
4367 568 : if decision {
4368 206 : self.last_image_layer_creation_check_at.store(lsn);
4369 206 : *last_check_instant = Some(Instant::now());
4370 362 : }
4371 :
4372 568 : decision
4373 568 : }
4374 :
4375 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4376 : async fn create_image_layers(
4377 : self: &Arc<Timeline>,
4378 : partitioning: &KeyPartitioning,
4379 : lsn: Lsn,
4380 : mode: ImageLayerCreationMode,
4381 : ctx: &RequestContext,
4382 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4383 : let timer = self.metrics.create_images_time_histo.start_timer();
4384 : let mut image_layers = Vec::new();
4385 :
4386 : // We need to avoid holes between generated image layers.
4387 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4388 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4389 : //
4390 : // How such hole between partitions can appear?
4391 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4392 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4393 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4394 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4395 : let mut start = Key::MIN;
4396 :
4397 : let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
4398 :
4399 : for partition in partitioning.parts.iter() {
4400 : if self.cancel.is_cancelled() {
4401 : return Err(CreateImageLayersError::Cancelled);
4402 : }
4403 :
4404 : let img_range = start..partition.ranges.last().unwrap().end;
4405 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4406 : if compact_metadata {
4407 : for range in &partition.ranges {
4408 : assert!(
4409 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4410 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4411 : "metadata keys must be partitioned separately"
4412 : );
4413 : }
4414 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4415 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4416 : // might mess up with evictions.
4417 : start = img_range.end;
4418 : continue;
4419 : }
4420 : // For initial and force modes, we always generate image layers for metadata keys.
4421 : } else if let ImageLayerCreationMode::Try = mode {
4422 : // check_for_image_layers = false -> skip
4423 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4424 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4425 : start = img_range.end;
4426 : continue;
4427 : }
4428 : }
4429 : if let ImageLayerCreationMode::Force = mode {
4430 : // When forced to create image layers, we might try and create them where they already
4431 : // exist. This mode is only used in tests/debug.
4432 : let layers = self.layers.read().await;
4433 : if layers.contains_key(&PersistentLayerKey {
4434 : key_range: img_range.clone(),
4435 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4436 : is_delta: false,
4437 : }) {
4438 : tracing::info!(
4439 : "Skipping image layer at {lsn} {}..{}, already exists",
4440 : img_range.start,
4441 : img_range.end
4442 : );
4443 : start = img_range.end;
4444 : continue;
4445 : }
4446 : }
4447 :
4448 : let image_layer_writer = ImageLayerWriter::new(
4449 : self.conf,
4450 : self.timeline_id,
4451 : self.tenant_shard_id,
4452 : &img_range,
4453 : lsn,
4454 : ctx,
4455 : )
4456 : .await?;
4457 :
4458 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4459 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4460 0 : "failpoint image-layer-writer-fail-before-finish"
4461 0 : )))
4462 0 : });
4463 :
4464 : if !compact_metadata {
4465 : let ImageLayerCreationOutcome {
4466 : image,
4467 : next_start_key,
4468 : } = self
4469 : .create_image_layer_for_rel_blocks(
4470 : partition,
4471 : image_layer_writer,
4472 : lsn,
4473 : ctx,
4474 : img_range,
4475 : start,
4476 : )
4477 : .await?;
4478 :
4479 : start = next_start_key;
4480 : image_layers.extend(image);
4481 : } else {
4482 : let ImageLayerCreationOutcome {
4483 : image,
4484 : next_start_key,
4485 : } = self
4486 : .create_image_layer_for_metadata_keys(
4487 : partition,
4488 : image_layer_writer,
4489 : lsn,
4490 : ctx,
4491 : img_range,
4492 : mode,
4493 : start,
4494 : )
4495 : .await?;
4496 : start = next_start_key;
4497 : image_layers.extend(image);
4498 : }
4499 : }
4500 :
4501 : let mut guard = self.layers.write().await;
4502 :
4503 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4504 : // now they are being scheduled outside of write lock; current way is inconsistent with
4505 : // compaction lock order.
4506 : guard
4507 : .open_mut()?
4508 : .track_new_image_layers(&image_layers, &self.metrics);
4509 : drop_wlock(guard);
4510 : timer.stop_and_record();
4511 :
4512 : // Creating image layers may have caused some previously visible layers to be covered
4513 : if !image_layers.is_empty() {
4514 : self.update_layer_visibility().await?;
4515 : }
4516 :
4517 : Ok(image_layers)
4518 : }
4519 :
4520 : /// Wait until the background initial logical size calculation is complete, or
4521 : /// this Timeline is shut down. Calling this function will cause the initial
4522 : /// logical size calculation to skip waiting for the background jobs barrier.
4523 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4524 0 : if !self.shard_identity.is_shard_zero() {
4525 : // We don't populate logical size on shard >0: skip waiting for it.
4526 0 : return;
4527 0 : }
4528 0 :
4529 0 : if self.remote_client.is_deleting() {
4530 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4531 0 : return;
4532 0 : }
4533 0 :
4534 0 : if self.current_logical_size.current_size().is_exact() {
4535 : // root timelines are initialized with exact count, but never start the background
4536 : // calculation
4537 0 : return;
4538 0 : }
4539 :
4540 0 : if let Some(await_bg_cancel) = self
4541 0 : .current_logical_size
4542 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4543 0 : .get()
4544 0 : {
4545 0 : await_bg_cancel.cancel();
4546 0 : } else {
4547 : // We should not wait if we were not able to explicitly instruct
4548 : // the logical size cancellation to skip the concurrency limit semaphore.
4549 : // TODO: this is an unexpected case. We should restructure so that it
4550 : // can't happen.
4551 0 : tracing::warn!(
4552 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4553 : );
4554 0 : debug_assert!(false);
4555 : }
4556 :
4557 0 : tokio::select!(
4558 0 : _ = self.current_logical_size.initialized.acquire() => {},
4559 0 : _ = self.cancel.cancelled() => {}
4560 : )
4561 0 : }
4562 :
4563 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4564 : /// Timelines layers up to the ancestor_lsn.
4565 : ///
4566 : /// Requires a timeline that:
4567 : /// - has an ancestor to detach from
4568 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4569 : /// a technical requirement
4570 : ///
4571 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4572 : /// polled again until completion.
4573 : ///
4574 : /// During the operation all timelines sharing the data with this timeline will be reparented
4575 : /// from our ancestor to be branches of this timeline.
4576 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4577 0 : self: &Arc<Timeline>,
4578 0 : tenant: &crate::tenant::Tenant,
4579 0 : options: detach_ancestor::Options,
4580 0 : ctx: &RequestContext,
4581 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
4582 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4583 0 : }
4584 :
4585 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
4586 : /// reparents any reparentable children of previous ancestor.
4587 : ///
4588 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
4589 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
4590 : /// successfully, tenant must be reloaded.
4591 : ///
4592 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
4593 : /// resetting the tenant.
4594 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
4595 0 : self: &Arc<Timeline>,
4596 0 : tenant: &crate::tenant::Tenant,
4597 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4598 0 : ctx: &RequestContext,
4599 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
4600 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
4601 0 : }
4602 :
4603 : /// Final step which unblocks the GC.
4604 : ///
4605 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
4606 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4607 0 : self: &Arc<Timeline>,
4608 0 : tenant: &crate::tenant::Tenant,
4609 0 : attempt: detach_ancestor::Attempt,
4610 0 : ctx: &RequestContext,
4611 0 : ) -> Result<(), detach_ancestor::Error> {
4612 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
4613 0 : }
4614 : }
4615 :
4616 : impl Drop for Timeline {
4617 10 : fn drop(&mut self) {
4618 10 : if let Some(ancestor) = &self.ancestor_timeline {
4619 : // This lock should never be poisoned, but in case it is we do a .map() instead of
4620 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
4621 4 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
4622 4 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
4623 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
4624 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
4625 4 : }
4626 0 : }
4627 6 : }
4628 10 : }
4629 : }
4630 :
4631 : /// Top-level failure to compact.
4632 : #[derive(Debug, thiserror::Error)]
4633 : pub(crate) enum CompactionError {
4634 : #[error("The timeline or pageserver is shutting down")]
4635 : ShuttingDown,
4636 : /// Compaction tried to offload a timeline and failed
4637 : #[error("Failed to offload timeline: {0}")]
4638 : Offload(OffloadError),
4639 : /// Compaction cannot be done right now; page reconstruction and so on.
4640 : #[error(transparent)]
4641 : Other(anyhow::Error),
4642 : }
4643 :
4644 : impl From<OffloadError> for CompactionError {
4645 0 : fn from(e: OffloadError) -> Self {
4646 0 : match e {
4647 0 : OffloadError::Cancelled => Self::ShuttingDown,
4648 0 : _ => Self::Offload(e),
4649 : }
4650 0 : }
4651 : }
4652 :
4653 : impl CompactionError {
4654 0 : pub fn is_cancelled(&self) -> bool {
4655 0 : matches!(self, CompactionError::ShuttingDown)
4656 0 : }
4657 : }
4658 :
4659 : impl From<CollectKeySpaceError> for CompactionError {
4660 0 : fn from(err: CollectKeySpaceError) -> Self {
4661 0 : match err {
4662 : CollectKeySpaceError::Cancelled
4663 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4664 0 : CompactionError::ShuttingDown
4665 : }
4666 0 : e => CompactionError::Other(e.into()),
4667 : }
4668 0 : }
4669 : }
4670 :
4671 : impl From<super::upload_queue::NotInitialized> for CompactionError {
4672 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
4673 0 : match value {
4674 : super::upload_queue::NotInitialized::Uninitialized => {
4675 0 : CompactionError::Other(anyhow::anyhow!(value))
4676 : }
4677 : super::upload_queue::NotInitialized::ShuttingDown
4678 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
4679 : }
4680 0 : }
4681 : }
4682 :
4683 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
4684 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
4685 0 : match e {
4686 : super::storage_layer::layer::DownloadError::TimelineShutdown
4687 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
4688 0 : CompactionError::ShuttingDown
4689 : }
4690 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
4691 : | super::storage_layer::layer::DownloadError::DownloadRequired
4692 : | super::storage_layer::layer::DownloadError::NotFile(_)
4693 : | super::storage_layer::layer::DownloadError::DownloadFailed
4694 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
4695 0 : CompactionError::Other(anyhow::anyhow!(e))
4696 : }
4697 : #[cfg(test)]
4698 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
4699 0 : CompactionError::Other(anyhow::anyhow!(e))
4700 : }
4701 : }
4702 0 : }
4703 : }
4704 :
4705 : impl From<layer_manager::Shutdown> for CompactionError {
4706 0 : fn from(_: layer_manager::Shutdown) -> Self {
4707 0 : CompactionError::ShuttingDown
4708 0 : }
4709 : }
4710 :
4711 : #[serde_as]
4712 196 : #[derive(serde::Serialize)]
4713 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4714 :
4715 : #[derive(Default)]
4716 : enum DurationRecorder {
4717 : #[default]
4718 : NotStarted,
4719 : Recorded(RecordedDuration, tokio::time::Instant),
4720 : }
4721 :
4722 : impl DurationRecorder {
4723 504 : fn till_now(&self) -> DurationRecorder {
4724 504 : match self {
4725 : DurationRecorder::NotStarted => {
4726 0 : panic!("must only call on recorded measurements")
4727 : }
4728 504 : DurationRecorder::Recorded(_, ended) => {
4729 504 : let now = tokio::time::Instant::now();
4730 504 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4731 504 : }
4732 504 : }
4733 504 : }
4734 196 : fn into_recorded(self) -> Option<RecordedDuration> {
4735 196 : match self {
4736 0 : DurationRecorder::NotStarted => None,
4737 196 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4738 : }
4739 196 : }
4740 : }
4741 :
4742 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
4743 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
4744 : /// the layer descriptor requires the user to provide the ranges, which should cover all
4745 : /// keys specified in the `data` field.
4746 : #[cfg(test)]
4747 : #[derive(Clone)]
4748 : pub struct DeltaLayerTestDesc {
4749 : pub lsn_range: Range<Lsn>,
4750 : pub key_range: Range<Key>,
4751 : pub data: Vec<(Key, Lsn, Value)>,
4752 : }
4753 :
4754 : #[cfg(test)]
4755 : impl DeltaLayerTestDesc {
4756 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
4757 2 : Self {
4758 2 : lsn_range,
4759 2 : key_range,
4760 2 : data,
4761 2 : }
4762 2 : }
4763 :
4764 88 : pub fn new_with_inferred_key_range(
4765 88 : lsn_range: Range<Lsn>,
4766 88 : data: Vec<(Key, Lsn, Value)>,
4767 88 : ) -> Self {
4768 220 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
4769 220 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
4770 88 : Self {
4771 88 : key_range: (*key_min)..(key_max.next()),
4772 88 : lsn_range,
4773 88 : data,
4774 88 : }
4775 88 : }
4776 :
4777 10 : pub(crate) fn layer_name(&self) -> LayerName {
4778 10 : LayerName::Delta(super::storage_layer::DeltaLayerName {
4779 10 : key_range: self.key_range.clone(),
4780 10 : lsn_range: self.lsn_range.clone(),
4781 10 : })
4782 10 : }
4783 : }
4784 :
4785 : impl Timeline {
4786 28 : async fn finish_compact_batch(
4787 28 : self: &Arc<Self>,
4788 28 : new_deltas: &[ResidentLayer],
4789 28 : new_images: &[ResidentLayer],
4790 28 : layers_to_remove: &[Layer],
4791 28 : ) -> Result<(), CompactionError> {
4792 28 : let mut guard = tokio::select! {
4793 28 : guard = self.layers.write() => guard,
4794 28 : _ = self.cancel.cancelled() => {
4795 0 : return Err(CompactionError::ShuttingDown);
4796 : }
4797 : };
4798 :
4799 28 : let mut duplicated_layers = HashSet::new();
4800 28 :
4801 28 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
4802 :
4803 336 : for l in new_deltas {
4804 308 : if guard.contains(l.as_ref()) {
4805 : // expected in tests
4806 0 : tracing::error!(layer=%l, "duplicated L1 layer");
4807 :
4808 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
4809 : // `guard` on self.layers. as of writing this, there are no error returns except
4810 : // for compact_level0_phase1 creating an L0, which does not happen in practice
4811 : // because we have not implemented L0 => L0 compaction.
4812 0 : duplicated_layers.insert(l.layer_desc().key());
4813 308 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
4814 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
4815 308 : } else {
4816 308 : insert_layers.push(l.clone());
4817 308 : }
4818 : }
4819 :
4820 : // only remove those inputs which were not outputs
4821 28 : let remove_layers: Vec<Layer> = layers_to_remove
4822 28 : .iter()
4823 402 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
4824 28 : .cloned()
4825 28 : .collect();
4826 28 :
4827 28 : if !new_images.is_empty() {
4828 0 : guard
4829 0 : .open_mut()?
4830 0 : .track_new_image_layers(new_images, &self.metrics);
4831 28 : }
4832 :
4833 28 : guard
4834 28 : .open_mut()?
4835 28 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
4836 28 :
4837 28 : self.remote_client
4838 28 : .schedule_compaction_update(&remove_layers, new_deltas)?;
4839 :
4840 28 : drop_wlock(guard);
4841 28 :
4842 28 : Ok(())
4843 28 : }
4844 :
4845 0 : async fn rewrite_layers(
4846 0 : self: &Arc<Self>,
4847 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
4848 0 : mut drop_layers: Vec<Layer>,
4849 0 : ) -> Result<(), CompactionError> {
4850 0 : let mut guard = self.layers.write().await;
4851 :
4852 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
4853 : // to avoid double-removing, and avoid rewriting something that was removed.
4854 0 : replace_layers.retain(|(l, _)| guard.contains(l));
4855 0 : drop_layers.retain(|l| guard.contains(l));
4856 0 :
4857 0 : guard
4858 0 : .open_mut()?
4859 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
4860 0 :
4861 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
4862 0 :
4863 0 : self.remote_client
4864 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
4865 :
4866 0 : Ok(())
4867 0 : }
4868 :
4869 : /// Schedules the uploads of the given image layers
4870 364 : fn upload_new_image_layers(
4871 364 : self: &Arc<Self>,
4872 364 : new_images: impl IntoIterator<Item = ResidentLayer>,
4873 364 : ) -> Result<(), super::upload_queue::NotInitialized> {
4874 390 : for layer in new_images {
4875 26 : self.remote_client.schedule_layer_file_upload(layer)?;
4876 : }
4877 : // should any new image layer been created, not uploading index_part will
4878 : // result in a mismatch between remote_physical_size and layermap calculated
4879 : // size, which will fail some tests, but should not be an issue otherwise.
4880 364 : self.remote_client
4881 364 : .schedule_index_upload_for_file_changes()?;
4882 364 : Ok(())
4883 364 : }
4884 :
4885 0 : async fn find_gc_time_cutoff(
4886 0 : &self,
4887 0 : now: SystemTime,
4888 0 : pitr: Duration,
4889 0 : cancel: &CancellationToken,
4890 0 : ctx: &RequestContext,
4891 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
4892 0 : debug_assert_current_span_has_tenant_and_timeline_id();
4893 0 : if self.shard_identity.is_shard_zero() {
4894 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
4895 0 : let time_range = if pitr == Duration::ZERO {
4896 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
4897 : } else {
4898 0 : pitr
4899 : };
4900 :
4901 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
4902 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
4903 0 : let timestamp = to_pg_timestamp(time_cutoff);
4904 :
4905 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
4906 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
4907 0 : LsnForTimestamp::Future(lsn) => {
4908 0 : // The timestamp is in the future. That sounds impossible,
4909 0 : // but what it really means is that there hasn't been
4910 0 : // any commits since the cutoff timestamp.
4911 0 : //
4912 0 : // In this case we should use the LSN of the most recent commit,
4913 0 : // which is implicitly the last LSN in the log.
4914 0 : debug!("future({})", lsn);
4915 0 : Some(self.get_last_record_lsn())
4916 : }
4917 0 : LsnForTimestamp::Past(lsn) => {
4918 0 : debug!("past({})", lsn);
4919 0 : None
4920 : }
4921 0 : LsnForTimestamp::NoData(lsn) => {
4922 0 : debug!("nodata({})", lsn);
4923 0 : None
4924 : }
4925 : };
4926 0 : Ok(time_cutoff)
4927 : } else {
4928 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
4929 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
4930 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
4931 : // space cutoff that we would also have respected ourselves.
4932 0 : match self
4933 0 : .remote_client
4934 0 : .download_foreign_index(ShardNumber(0), cancel)
4935 0 : .await
4936 : {
4937 0 : Ok((index_part, index_generation, _index_mtime)) => {
4938 0 : tracing::info!("GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
4939 0 : index_part.metadata.latest_gc_cutoff_lsn());
4940 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
4941 : }
4942 : Err(DownloadError::NotFound) => {
4943 : // This is unexpected, because during timeline creations shard zero persists to remote
4944 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
4945 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
4946 : // state, then shard zero should _eventually_ write an index when it GCs.
4947 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
4948 0 : Ok(None)
4949 : }
4950 0 : Err(e) => {
4951 0 : // TODO: this function should return a different error type than page reconstruct error
4952 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
4953 : }
4954 : }
4955 :
4956 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
4957 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
4958 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
4959 : // new location. This is legal in principle, but problematic in practice because it might result
4960 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
4961 : // because they have GC'd past the branch point.
4962 : }
4963 0 : }
4964 :
4965 : /// Find the Lsns above which layer files need to be retained on
4966 : /// garbage collection.
4967 : ///
4968 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
4969 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
4970 : /// the space-based retention.
4971 : ///
4972 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
4973 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
4974 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
4975 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
4976 : /// in the response as the GC cutoff point for the timeline.
4977 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
4978 : pub(super) async fn find_gc_cutoffs(
4979 : &self,
4980 : now: SystemTime,
4981 : space_cutoff: Lsn,
4982 : pitr: Duration,
4983 : cancel: &CancellationToken,
4984 : ctx: &RequestContext,
4985 : ) -> Result<GcCutoffs, PageReconstructError> {
4986 : let _timer = self
4987 : .metrics
4988 : .find_gc_cutoffs_histo
4989 : .start_timer()
4990 : .record_on_drop();
4991 :
4992 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
4993 :
4994 : if cfg!(test) {
4995 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
4996 : if pitr == Duration::ZERO {
4997 : return Ok(GcCutoffs {
4998 : time: self.get_last_record_lsn(),
4999 : space: space_cutoff,
5000 : });
5001 : }
5002 : }
5003 :
5004 : // Calculate a time-based limit on how much to retain:
5005 : // - if PITR interval is set, then this is our cutoff.
5006 : // - if PITR interval is not set, then we do a lookup
5007 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
5008 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
5009 :
5010 : Ok(match (pitr, time_cutoff) {
5011 : (Duration::ZERO, Some(time_cutoff)) => {
5012 : // PITR is not set. Retain the size-based limit, or the default time retention,
5013 : // whichever requires less data.
5014 : GcCutoffs {
5015 : time: self.get_last_record_lsn(),
5016 : space: std::cmp::max(time_cutoff, space_cutoff),
5017 : }
5018 : }
5019 : (Duration::ZERO, None) => {
5020 : // PITR is not set, and time lookup failed
5021 : GcCutoffs {
5022 : time: self.get_last_record_lsn(),
5023 : space: space_cutoff,
5024 : }
5025 : }
5026 : (_, None) => {
5027 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
5028 : // cannot advance beyond what was already GC'd, and respect space-based retention
5029 : GcCutoffs {
5030 : time: *self.get_latest_gc_cutoff_lsn(),
5031 : space: space_cutoff,
5032 : }
5033 : }
5034 : (_, Some(time_cutoff)) => {
5035 : // PITR interval is set and we looked up timestamp successfully. Ignore
5036 : // size based retention and make time cutoff authoritative
5037 : GcCutoffs {
5038 : time: time_cutoff,
5039 : space: time_cutoff,
5040 : }
5041 : }
5042 : })
5043 : }
5044 :
5045 : /// Garbage collect layer files on a timeline that are no longer needed.
5046 : ///
5047 : /// Currently, we don't make any attempt at removing unneeded page versions
5048 : /// within a layer file. We can only remove the whole file if it's fully
5049 : /// obsolete.
5050 4 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
5051 : // this is most likely the background tasks, but it might be the spawned task from
5052 : // immediate_gc
5053 4 : let _g = tokio::select! {
5054 4 : guard = self.gc_lock.lock() => guard,
5055 4 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
5056 : };
5057 4 : let timer = self.metrics.garbage_collect_histo.start_timer();
5058 4 :
5059 4 : fail_point!("before-timeline-gc");
5060 4 :
5061 4 : // Is the timeline being deleted?
5062 4 : if self.is_stopping() {
5063 0 : return Err(GcError::TimelineCancelled);
5064 4 : }
5065 4 :
5066 4 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
5067 4 : let gc_info = self.gc_info.read().unwrap();
5068 4 :
5069 4 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
5070 4 : let time_cutoff = gc_info.cutoffs.time;
5071 4 : let retain_lsns = gc_info
5072 4 : .retain_lsns
5073 4 : .iter()
5074 4 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
5075 4 : .collect();
5076 4 :
5077 4 : // Gets the maximum LSN that holds the valid lease.
5078 4 : //
5079 4 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
5080 4 : // Here, we do not check for stale leases again.
5081 4 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
5082 4 :
5083 4 : (
5084 4 : space_cutoff,
5085 4 : time_cutoff,
5086 4 : retain_lsns,
5087 4 : max_lsn_with_valid_lease,
5088 4 : )
5089 4 : };
5090 4 :
5091 4 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
5092 4 : let standby_horizon = self.standby_horizon.load();
5093 4 : // Hold GC for the standby, but as a safety guard do it only within some
5094 4 : // reasonable lag.
5095 4 : if standby_horizon != Lsn::INVALID {
5096 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
5097 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
5098 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
5099 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
5100 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
5101 : } else {
5102 0 : warn!(
5103 0 : "standby is lagging for more than {}MB, not holding gc for it",
5104 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
5105 : )
5106 : }
5107 0 : }
5108 4 : }
5109 :
5110 : // Reset standby horizon to ignore it if it is not updated till next GC.
5111 : // It is an easy way to unset it when standby disappears without adding
5112 : // more conf options.
5113 4 : self.standby_horizon.store(Lsn::INVALID);
5114 4 : self.metrics
5115 4 : .standby_horizon_gauge
5116 4 : .set(Lsn::INVALID.0 as i64);
5117 :
5118 4 : let res = self
5119 4 : .gc_timeline(
5120 4 : space_cutoff,
5121 4 : time_cutoff,
5122 4 : retain_lsns,
5123 4 : max_lsn_with_valid_lease,
5124 4 : new_gc_cutoff,
5125 4 : )
5126 4 : .instrument(
5127 4 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5128 : )
5129 4 : .await?;
5130 :
5131 : // only record successes
5132 4 : timer.stop_and_record();
5133 4 :
5134 4 : Ok(res)
5135 4 : }
5136 :
5137 4 : async fn gc_timeline(
5138 4 : &self,
5139 4 : space_cutoff: Lsn,
5140 4 : time_cutoff: Lsn,
5141 4 : retain_lsns: Vec<Lsn>,
5142 4 : max_lsn_with_valid_lease: Option<Lsn>,
5143 4 : new_gc_cutoff: Lsn,
5144 4 : ) -> Result<GcResult, GcError> {
5145 4 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5146 4 :
5147 4 : let now = SystemTime::now();
5148 4 : let mut result: GcResult = GcResult::default();
5149 4 :
5150 4 : // Nothing to GC. Return early.
5151 4 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
5152 4 : if latest_gc_cutoff >= new_gc_cutoff {
5153 0 : info!(
5154 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5155 : );
5156 0 : return Ok(result);
5157 4 : }
5158 :
5159 : // We need to ensure that no one tries to read page versions or create
5160 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
5161 : // for details. This will block until the old value is no longer in use.
5162 : //
5163 : // The GC cutoff should only ever move forwards.
5164 4 : let waitlist = {
5165 4 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
5166 4 : if *write_guard > new_gc_cutoff {
5167 0 : return Err(GcError::BadLsn {
5168 0 : why: format!(
5169 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
5170 0 : *write_guard, new_gc_cutoff
5171 0 : ),
5172 0 : });
5173 4 : }
5174 4 :
5175 4 : write_guard.store_and_unlock(new_gc_cutoff)
5176 4 : };
5177 4 : waitlist.wait().await;
5178 :
5179 4 : info!("GC starting");
5180 :
5181 4 : debug!("retain_lsns: {:?}", retain_lsns);
5182 :
5183 4 : let mut layers_to_remove = Vec::new();
5184 :
5185 : // Scan all layers in the timeline (remote or on-disk).
5186 : //
5187 : // Garbage collect the layer if all conditions are satisfied:
5188 : // 1. it is older than cutoff LSN;
5189 : // 2. it is older than PITR interval;
5190 : // 3. it doesn't need to be retained for 'retain_lsns';
5191 : // 4. it does not need to be kept for LSNs holding valid leases.
5192 : // 5. newer on-disk image layers cover the layer's whole key range
5193 : //
5194 : // TODO holding a write lock is too agressive and avoidable
5195 4 : let mut guard = self.layers.write().await;
5196 4 : let layers = guard.layer_map()?;
5197 24 : 'outer: for l in layers.iter_historic_layers() {
5198 24 : result.layers_total += 1;
5199 24 :
5200 24 : // 1. Is it newer than GC horizon cutoff point?
5201 24 : if l.get_lsn_range().end > space_cutoff {
5202 2 : info!(
5203 0 : "keeping {} because it's newer than space_cutoff {}",
5204 0 : l.layer_name(),
5205 : space_cutoff,
5206 : );
5207 2 : result.layers_needed_by_cutoff += 1;
5208 2 : continue 'outer;
5209 22 : }
5210 22 :
5211 22 : // 2. It is newer than PiTR cutoff point?
5212 22 : if l.get_lsn_range().end > time_cutoff {
5213 0 : info!(
5214 0 : "keeping {} because it's newer than time_cutoff {}",
5215 0 : l.layer_name(),
5216 : time_cutoff,
5217 : );
5218 0 : result.layers_needed_by_pitr += 1;
5219 0 : continue 'outer;
5220 22 : }
5221 :
5222 : // 3. Is it needed by a child branch?
5223 : // NOTE With that we would keep data that
5224 : // might be referenced by child branches forever.
5225 : // We can track this in child timeline GC and delete parent layers when
5226 : // they are no longer needed. This might be complicated with long inheritance chains.
5227 : //
5228 : // TODO Vec is not a great choice for `retain_lsns`
5229 22 : for retain_lsn in &retain_lsns {
5230 : // start_lsn is inclusive
5231 0 : if &l.get_lsn_range().start <= retain_lsn {
5232 0 : info!(
5233 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5234 0 : l.layer_name(),
5235 0 : retain_lsn,
5236 0 : l.is_incremental(),
5237 : );
5238 0 : result.layers_needed_by_branches += 1;
5239 0 : continue 'outer;
5240 0 : }
5241 : }
5242 :
5243 : // 4. Is there a valid lease that requires us to keep this layer?
5244 22 : if let Some(lsn) = &max_lsn_with_valid_lease {
5245 : // keep if layer start <= any of the lease
5246 18 : if &l.get_lsn_range().start <= lsn {
5247 14 : info!(
5248 0 : "keeping {} because there is a valid lease preventing GC at {}",
5249 0 : l.layer_name(),
5250 : lsn,
5251 : );
5252 14 : result.layers_needed_by_leases += 1;
5253 14 : continue 'outer;
5254 4 : }
5255 4 : }
5256 :
5257 : // 5. Is there a later on-disk layer for this relation?
5258 : //
5259 : // The end-LSN is exclusive, while disk_consistent_lsn is
5260 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5261 : // OK for a delta layer to have end LSN 101, but if the end LSN
5262 : // is 102, then it might not have been fully flushed to disk
5263 : // before crash.
5264 : //
5265 : // For example, imagine that the following layers exist:
5266 : //
5267 : // 1000 - image (A)
5268 : // 1000-2000 - delta (B)
5269 : // 2000 - image (C)
5270 : // 2000-3000 - delta (D)
5271 : // 3000 - image (E)
5272 : //
5273 : // If GC horizon is at 2500, we can remove layers A and B, but
5274 : // we cannot remove C, even though it's older than 2500, because
5275 : // the delta layer 2000-3000 depends on it.
5276 8 : if !layers
5277 8 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5278 : {
5279 6 : info!("keeping {} because it is the latest layer", l.layer_name());
5280 6 : result.layers_not_updated += 1;
5281 6 : continue 'outer;
5282 2 : }
5283 2 :
5284 2 : // We didn't find any reason to keep this file, so remove it.
5285 2 : info!(
5286 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5287 0 : l.layer_name(),
5288 0 : l.is_incremental(),
5289 : );
5290 2 : layers_to_remove.push(l);
5291 : }
5292 :
5293 4 : if !layers_to_remove.is_empty() {
5294 : // Persist the new GC cutoff value before we actually remove anything.
5295 : // This unconditionally schedules also an index_part.json update, even though, we will
5296 : // be doing one a bit later with the unlinked gc'd layers.
5297 2 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5298 2 : self.schedule_uploads(disk_consistent_lsn, None)
5299 2 : .map_err(|e| {
5300 0 : if self.cancel.is_cancelled() {
5301 0 : GcError::TimelineCancelled
5302 : } else {
5303 0 : GcError::Remote(e)
5304 : }
5305 2 : })?;
5306 :
5307 2 : let gc_layers = layers_to_remove
5308 2 : .iter()
5309 2 : .map(|x| guard.get_from_desc(x))
5310 2 : .collect::<Vec<Layer>>();
5311 2 :
5312 2 : result.layers_removed = gc_layers.len() as u64;
5313 2 :
5314 2 : self.remote_client.schedule_gc_update(&gc_layers)?;
5315 :
5316 2 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5317 2 :
5318 2 : #[cfg(feature = "testing")]
5319 2 : {
5320 2 : result.doomed_layers = gc_layers;
5321 2 : }
5322 2 : }
5323 :
5324 4 : info!(
5325 0 : "GC completed removing {} layers, cutoff {}",
5326 : result.layers_removed, new_gc_cutoff
5327 : );
5328 :
5329 4 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5330 4 : Ok(result)
5331 4 : }
5332 :
5333 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5334 669967 : async fn reconstruct_value(
5335 669967 : &self,
5336 669967 : key: Key,
5337 669967 : request_lsn: Lsn,
5338 669967 : mut data: ValueReconstructState,
5339 669967 : ) -> Result<Bytes, PageReconstructError> {
5340 669967 : // Perform WAL redo if needed
5341 669967 : data.records.reverse();
5342 669967 :
5343 669967 : // If we have a page image, and no WAL, we're all set
5344 669967 : if data.records.is_empty() {
5345 669149 : if let Some((img_lsn, img)) = &data.img {
5346 669149 : trace!(
5347 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5348 : key,
5349 : img_lsn,
5350 : request_lsn,
5351 : );
5352 669149 : Ok(img.clone())
5353 : } else {
5354 0 : Err(PageReconstructError::from(anyhow!(
5355 0 : "base image for {key} at {request_lsn} not found"
5356 0 : )))
5357 : }
5358 : } else {
5359 : // We need to do WAL redo.
5360 : //
5361 : // If we don't have a base image, then the oldest WAL record better initialize
5362 : // the page
5363 818 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5364 0 : Err(PageReconstructError::from(anyhow!(
5365 0 : "Base image for {} at {} not found, but got {} WAL records",
5366 0 : key,
5367 0 : request_lsn,
5368 0 : data.records.len()
5369 0 : )))
5370 : } else {
5371 818 : if data.img.is_some() {
5372 752 : trace!(
5373 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5374 0 : data.records.len(),
5375 : key,
5376 : request_lsn
5377 : );
5378 : } else {
5379 66 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5380 : };
5381 818 : let res = self
5382 818 : .walredo_mgr
5383 818 : .as_ref()
5384 818 : .context("timeline has no walredo manager")
5385 818 : .map_err(PageReconstructError::WalRedo)?
5386 818 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5387 818 : .await;
5388 818 : let img = match res {
5389 818 : Ok(img) => img,
5390 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5391 0 : Err(walredo::Error::Other(e)) => {
5392 0 : return Err(PageReconstructError::WalRedo(
5393 0 : e.context("reconstruct a page image"),
5394 0 : ))
5395 : }
5396 : };
5397 818 : Ok(img)
5398 : }
5399 : }
5400 669967 : }
5401 :
5402 0 : pub(crate) async fn spawn_download_all_remote_layers(
5403 0 : self: Arc<Self>,
5404 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5405 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5406 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5407 :
5408 : // this is not really needed anymore; it has tests which really check the return value from
5409 : // http api. it would be better not to maintain this anymore.
5410 :
5411 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5412 0 : if let Some(st) = &*status_guard {
5413 0 : match &st.state {
5414 : DownloadRemoteLayersTaskState::Running => {
5415 0 : return Err(st.clone());
5416 : }
5417 : DownloadRemoteLayersTaskState::ShutDown
5418 0 : | DownloadRemoteLayersTaskState::Completed => {
5419 0 : *status_guard = None;
5420 0 : }
5421 : }
5422 0 : }
5423 :
5424 0 : let self_clone = Arc::clone(&self);
5425 0 : let task_id = task_mgr::spawn(
5426 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5427 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5428 0 : self.tenant_shard_id,
5429 0 : Some(self.timeline_id),
5430 0 : "download all remote layers task",
5431 0 : async move {
5432 0 : self_clone.download_all_remote_layers(request).await;
5433 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5434 0 : match &mut *status_guard {
5435 : None => {
5436 0 : warn!("tasks status is supposed to be Some(), since we are running");
5437 : }
5438 0 : Some(st) => {
5439 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5440 0 : if st.task_id != exp_task_id {
5441 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5442 0 : } else {
5443 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5444 0 : }
5445 : }
5446 : };
5447 0 : Ok(())
5448 0 : }
5449 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))
5450 : );
5451 :
5452 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5453 0 : task_id: format!("{task_id}"),
5454 0 : state: DownloadRemoteLayersTaskState::Running,
5455 0 : total_layer_count: 0,
5456 0 : successful_download_count: 0,
5457 0 : failed_download_count: 0,
5458 0 : };
5459 0 : *status_guard = Some(initial_info.clone());
5460 0 :
5461 0 : Ok(initial_info)
5462 0 : }
5463 :
5464 0 : async fn download_all_remote_layers(
5465 0 : self: &Arc<Self>,
5466 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5467 0 : ) {
5468 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5469 :
5470 0 : let remaining = {
5471 0 : let guard = self.layers.read().await;
5472 0 : let Ok(lm) = guard.layer_map() else {
5473 : // technically here we could look into iterating accessible layers, but downloading
5474 : // all layers of a shutdown timeline makes no sense regardless.
5475 0 : tracing::info!("attempted to download all layers of shutdown timeline");
5476 0 : return;
5477 : };
5478 0 : lm.iter_historic_layers()
5479 0 : .map(|desc| guard.get_from_desc(&desc))
5480 0 : .collect::<Vec<_>>()
5481 0 : };
5482 0 : let total_layer_count = remaining.len();
5483 :
5484 : macro_rules! lock_status {
5485 : ($st:ident) => {
5486 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5487 : let st = st
5488 : .as_mut()
5489 : .expect("this function is only called after the task has been spawned");
5490 : assert_eq!(
5491 : st.task_id,
5492 : format!(
5493 : "{}",
5494 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5495 : )
5496 : );
5497 : let $st = st;
5498 : };
5499 : }
5500 :
5501 : {
5502 0 : lock_status!(st);
5503 0 : st.total_layer_count = total_layer_count as u64;
5504 0 : }
5505 0 :
5506 0 : let mut remaining = remaining.into_iter();
5507 0 : let mut have_remaining = true;
5508 0 : let mut js = tokio::task::JoinSet::new();
5509 0 :
5510 0 : let cancel = task_mgr::shutdown_token();
5511 0 :
5512 0 : let limit = request.max_concurrent_downloads;
5513 :
5514 : loop {
5515 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5516 0 : let Some(next) = remaining.next() else {
5517 0 : have_remaining = false;
5518 0 : break;
5519 : };
5520 :
5521 0 : let span = tracing::info_span!("download", layer = %next);
5522 :
5523 0 : js.spawn(
5524 0 : async move {
5525 0 : let res = next.download().await;
5526 0 : (next, res)
5527 0 : }
5528 0 : .instrument(span),
5529 0 : );
5530 0 : }
5531 :
5532 0 : while let Some(res) = js.join_next().await {
5533 0 : match res {
5534 : Ok((_, Ok(_))) => {
5535 0 : lock_status!(st);
5536 0 : st.successful_download_count += 1;
5537 : }
5538 0 : Ok((layer, Err(e))) => {
5539 0 : tracing::error!(%layer, "download failed: {e:#}");
5540 0 : lock_status!(st);
5541 0 : st.failed_download_count += 1;
5542 : }
5543 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5544 0 : Err(je) if je.is_panic() => {
5545 0 : lock_status!(st);
5546 0 : st.failed_download_count += 1;
5547 : }
5548 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5549 : }
5550 : }
5551 :
5552 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5553 0 : break;
5554 0 : }
5555 : }
5556 :
5557 : {
5558 0 : lock_status!(st);
5559 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5560 : }
5561 0 : }
5562 :
5563 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5564 0 : &self,
5565 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5566 0 : self.download_all_remote_layers_task_info
5567 0 : .read()
5568 0 : .unwrap()
5569 0 : .clone()
5570 0 : }
5571 : }
5572 :
5573 : impl Timeline {
5574 : /// Returns non-remote layers for eviction.
5575 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5576 0 : let guard = self.layers.read().await;
5577 0 : let mut max_layer_size: Option<u64> = None;
5578 0 :
5579 0 : let resident_layers = guard
5580 0 : .likely_resident_layers()
5581 0 : .map(|layer| {
5582 0 : let file_size = layer.layer_desc().file_size;
5583 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5584 0 :
5585 0 : let last_activity_ts = layer.latest_activity();
5586 0 :
5587 0 : EvictionCandidate {
5588 0 : layer: layer.to_owned().into(),
5589 0 : last_activity_ts,
5590 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5591 0 : visibility: layer.visibility(),
5592 0 : }
5593 0 : })
5594 0 : .collect();
5595 0 :
5596 0 : DiskUsageEvictionInfo {
5597 0 : max_layer_size,
5598 0 : resident_layers,
5599 0 : }
5600 0 : }
5601 :
5602 1858 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5603 1858 : ShardIndex {
5604 1858 : shard_number: self.tenant_shard_id.shard_number,
5605 1858 : shard_count: self.tenant_shard_id.shard_count,
5606 1858 : }
5607 1858 : }
5608 :
5609 : /// Persistently blocks gc for `Manual` reason.
5610 : ///
5611 : /// Returns true if no such block existed before, false otherwise.
5612 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
5613 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5614 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5615 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
5616 0 : }
5617 :
5618 : /// Persistently unblocks gc for `Manual` reason.
5619 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
5620 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5621 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5622 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
5623 0 : }
5624 :
5625 : #[cfg(test)]
5626 44 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5627 44 : self.last_record_lsn.advance(new_lsn);
5628 44 : }
5629 :
5630 : #[cfg(test)]
5631 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
5632 2 : self.disk_consistent_lsn.store(new_value);
5633 2 : }
5634 :
5635 : /// Force create an image layer and place it into the layer map.
5636 : ///
5637 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5638 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5639 : /// placed into the layer map in one run AND be validated.
5640 : #[cfg(test)]
5641 56 : pub(super) async fn force_create_image_layer(
5642 56 : self: &Arc<Timeline>,
5643 56 : lsn: Lsn,
5644 56 : mut images: Vec<(Key, Bytes)>,
5645 56 : check_start_lsn: Option<Lsn>,
5646 56 : ctx: &RequestContext,
5647 56 : ) -> anyhow::Result<()> {
5648 56 : let last_record_lsn = self.get_last_record_lsn();
5649 56 : assert!(
5650 56 : lsn <= last_record_lsn,
5651 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5652 : );
5653 56 : if let Some(check_start_lsn) = check_start_lsn {
5654 56 : assert!(lsn >= check_start_lsn);
5655 0 : }
5656 174 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5657 56 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5658 56 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5659 56 : let mut image_layer_writer = ImageLayerWriter::new(
5660 56 : self.conf,
5661 56 : self.timeline_id,
5662 56 : self.tenant_shard_id,
5663 56 : &(min_key..end_key),
5664 56 : lsn,
5665 56 : ctx,
5666 56 : )
5667 56 : .await?;
5668 286 : for (key, img) in images {
5669 230 : image_layer_writer.put_image(key, img, ctx).await?;
5670 : }
5671 56 : let (desc, path) = image_layer_writer.finish(ctx).await?;
5672 56 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5673 56 : info!("force created image layer {}", image_layer.local_path());
5674 : {
5675 56 : let mut guard = self.layers.write().await;
5676 56 : guard.open_mut().unwrap().force_insert_layer(image_layer);
5677 56 : }
5678 56 :
5679 56 : Ok(())
5680 56 : }
5681 :
5682 : /// Force create a delta layer and place it into the layer map.
5683 : ///
5684 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5685 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5686 : /// placed into the layer map in one run AND be validated.
5687 : #[cfg(test)]
5688 90 : pub(super) async fn force_create_delta_layer(
5689 90 : self: &Arc<Timeline>,
5690 90 : mut deltas: DeltaLayerTestDesc,
5691 90 : check_start_lsn: Option<Lsn>,
5692 90 : ctx: &RequestContext,
5693 90 : ) -> anyhow::Result<()> {
5694 90 : let last_record_lsn = self.get_last_record_lsn();
5695 90 : deltas
5696 90 : .data
5697 132 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5698 90 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5699 90 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5700 312 : for (_, lsn, _) in &deltas.data {
5701 222 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5702 : }
5703 90 : assert!(
5704 90 : deltas.lsn_range.end <= last_record_lsn,
5705 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
5706 : deltas.lsn_range.end,
5707 : last_record_lsn
5708 : );
5709 90 : if let Some(check_start_lsn) = check_start_lsn {
5710 90 : assert!(deltas.lsn_range.start >= check_start_lsn);
5711 0 : }
5712 90 : let mut delta_layer_writer = DeltaLayerWriter::new(
5713 90 : self.conf,
5714 90 : self.timeline_id,
5715 90 : self.tenant_shard_id,
5716 90 : deltas.key_range.start,
5717 90 : deltas.lsn_range,
5718 90 : ctx,
5719 90 : )
5720 90 : .await?;
5721 312 : for (key, lsn, val) in deltas.data {
5722 222 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5723 : }
5724 90 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
5725 90 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5726 90 : info!("force created delta layer {}", delta_layer.local_path());
5727 : {
5728 90 : let mut guard = self.layers.write().await;
5729 90 : guard.open_mut().unwrap().force_insert_layer(delta_layer);
5730 90 : }
5731 90 :
5732 90 : Ok(())
5733 90 : }
5734 :
5735 : /// Return all keys at the LSN in the image layers
5736 : #[cfg(test)]
5737 6 : pub(crate) async fn inspect_image_layers(
5738 6 : self: &Arc<Timeline>,
5739 6 : lsn: Lsn,
5740 6 : ctx: &RequestContext,
5741 6 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
5742 6 : let mut all_data = Vec::new();
5743 6 : let guard = self.layers.read().await;
5744 34 : for layer in guard.layer_map()?.iter_historic_layers() {
5745 34 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
5746 8 : let layer = guard.get_from_desc(&layer);
5747 8 : let mut reconstruct_data = ValuesReconstructState::default();
5748 8 : layer
5749 8 : .get_values_reconstruct_data(
5750 8 : KeySpace::single(Key::MIN..Key::MAX),
5751 8 : lsn..Lsn(lsn.0 + 1),
5752 8 : &mut reconstruct_data,
5753 8 : ctx,
5754 8 : )
5755 8 : .await?;
5756 74 : for (k, v) in reconstruct_data.keys {
5757 66 : all_data.push((k, v?.img.unwrap().1));
5758 : }
5759 26 : }
5760 : }
5761 6 : all_data.sort();
5762 6 : Ok(all_data)
5763 6 : }
5764 :
5765 : /// Get all historic layer descriptors in the layer map
5766 : #[cfg(test)]
5767 24 : pub(crate) async fn inspect_historic_layers(
5768 24 : self: &Arc<Timeline>,
5769 24 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
5770 24 : let mut layers = Vec::new();
5771 24 : let guard = self.layers.read().await;
5772 114 : for layer in guard.layer_map()?.iter_historic_layers() {
5773 114 : layers.push(layer.key());
5774 114 : }
5775 24 : Ok(layers)
5776 24 : }
5777 :
5778 : #[cfg(test)]
5779 10 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
5780 10 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
5781 10 : keyspace.merge(&ks);
5782 10 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
5783 10 : }
5784 : }
5785 :
5786 : /// Tracking writes ingestion does to a particular in-memory layer.
5787 : ///
5788 : /// Cleared upon freezing a layer.
5789 : pub(crate) struct TimelineWriterState {
5790 : open_layer: Arc<InMemoryLayer>,
5791 : current_size: u64,
5792 : // Previous Lsn which passed through
5793 : prev_lsn: Option<Lsn>,
5794 : // Largest Lsn which passed through the current writer
5795 : max_lsn: Option<Lsn>,
5796 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
5797 : cached_last_freeze_at: Lsn,
5798 : }
5799 :
5800 : impl TimelineWriterState {
5801 1296 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
5802 1296 : Self {
5803 1296 : open_layer,
5804 1296 : current_size,
5805 1296 : prev_lsn: None,
5806 1296 : max_lsn: None,
5807 1296 : cached_last_freeze_at: last_freeze_at,
5808 1296 : }
5809 1296 : }
5810 : }
5811 :
5812 : /// Various functions to mutate the timeline.
5813 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
5814 : // This is probably considered a bad practice in Rust and should be fixed eventually,
5815 : // but will cause large code changes.
5816 : pub(crate) struct TimelineWriter<'a> {
5817 : tl: &'a Timeline,
5818 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
5819 : }
5820 :
5821 : impl Deref for TimelineWriter<'_> {
5822 : type Target = Timeline;
5823 :
5824 9898008 : fn deref(&self) -> &Self::Target {
5825 9898008 : self.tl
5826 9898008 : }
5827 : }
5828 :
5829 : #[derive(PartialEq)]
5830 : enum OpenLayerAction {
5831 : Roll,
5832 : Open,
5833 : None,
5834 : }
5835 :
5836 : impl TimelineWriter<'_> {
5837 4804232 : async fn handle_open_layer_action(
5838 4804232 : &mut self,
5839 4804232 : at: Lsn,
5840 4804232 : action: OpenLayerAction,
5841 4804232 : ctx: &RequestContext,
5842 4804232 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
5843 4804232 : match action {
5844 : OpenLayerAction::Roll => {
5845 80 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
5846 80 : self.roll_layer(freeze_at).await?;
5847 80 : self.open_layer(at, ctx).await?;
5848 : }
5849 1216 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
5850 : OpenLayerAction::None => {
5851 4802936 : assert!(self.write_guard.is_some());
5852 : }
5853 : }
5854 :
5855 4804232 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
5856 4804232 : }
5857 :
5858 1296 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
5859 1296 : let layer = self
5860 1296 : .tl
5861 1296 : .get_layer_for_write(at, &self.write_guard, ctx)
5862 1296 : .await?;
5863 1296 : let initial_size = layer.size().await?;
5864 :
5865 1296 : let last_freeze_at = self.last_freeze_at.load();
5866 1296 : self.write_guard.replace(TimelineWriterState::new(
5867 1296 : layer,
5868 1296 : initial_size,
5869 1296 : last_freeze_at,
5870 1296 : ));
5871 1296 :
5872 1296 : Ok(())
5873 1296 : }
5874 :
5875 80 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
5876 80 : let current_size = self.write_guard.as_ref().unwrap().current_size;
5877 80 :
5878 80 : // self.write_guard will be taken by the freezing
5879 80 : self.tl
5880 80 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
5881 80 : .await?;
5882 :
5883 80 : assert!(self.write_guard.is_none());
5884 :
5885 80 : if current_size >= self.get_checkpoint_distance() * 2 {
5886 0 : warn!("Flushed oversized open layer with size {}", current_size)
5887 80 : }
5888 :
5889 80 : Ok(())
5890 80 : }
5891 :
5892 4804232 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
5893 4804232 : let state = &*self.write_guard;
5894 4804232 : let Some(state) = &state else {
5895 1216 : return OpenLayerAction::Open;
5896 : };
5897 :
5898 : #[cfg(feature = "testing")]
5899 4803016 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
5900 : // this check and assertion are not really needed because
5901 : // LayerManager::try_freeze_in_memory_layer will always clear out the
5902 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
5903 : // is no TimelineWriterState.
5904 0 : assert!(
5905 0 : state.open_layer.end_lsn.get().is_some(),
5906 0 : "our open_layer must be outdated"
5907 : );
5908 :
5909 : // this would be a memory leak waiting to happen because the in-memory layer always has
5910 : // an index
5911 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
5912 4803016 : }
5913 4803016 :
5914 4803016 : if state.prev_lsn == Some(lsn) {
5915 : // Rolling mid LSN is not supported by [downstream code].
5916 : // Hence, only roll at LSN boundaries.
5917 : //
5918 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
5919 6 : return OpenLayerAction::None;
5920 4803010 : }
5921 4803010 :
5922 4803010 : if state.current_size == 0 {
5923 : // Don't roll empty layers
5924 0 : return OpenLayerAction::None;
5925 4803010 : }
5926 4803010 :
5927 4803010 : if self.tl.should_roll(
5928 4803010 : state.current_size,
5929 4803010 : state.current_size + new_value_size,
5930 4803010 : self.get_checkpoint_distance(),
5931 4803010 : lsn,
5932 4803010 : state.cached_last_freeze_at,
5933 4803010 : state.open_layer.get_opened_at(),
5934 4803010 : ) {
5935 80 : OpenLayerAction::Roll
5936 : } else {
5937 4802930 : OpenLayerAction::None
5938 : }
5939 4804232 : }
5940 :
5941 : /// Put a batch of keys at the specified Lsns.
5942 4804230 : pub(crate) async fn put_batch(
5943 4804230 : &mut self,
5944 4804230 : batch: SerializedValueBatch,
5945 4804230 : ctx: &RequestContext,
5946 4804230 : ) -> anyhow::Result<()> {
5947 4804230 : if !batch.has_data() {
5948 0 : return Ok(());
5949 4804230 : }
5950 4804230 :
5951 4804230 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
5952 4804230 : // We don't assert this in release builds, since key ownership policies may change over
5953 4804230 : // time. Stray keys will be removed during compaction.
5954 4804230 : if cfg!(debug_assertions) {
5955 9894896 : for metadata in &batch.metadata {
5956 5090666 : if let ValueMeta::Serialized(metadata) = metadata {
5957 5090666 : let key = Key::from_compact(metadata.key);
5958 5090666 : assert!(
5959 5090666 : self.shard_identity.is_key_local(&key)
5960 0 : || self.shard_identity.is_key_global(&key),
5961 0 : "key {key} does not belong on shard {}",
5962 0 : self.shard_identity.shard_index()
5963 : );
5964 0 : }
5965 : }
5966 0 : }
5967 :
5968 4804230 : let batch_max_lsn = batch.max_lsn;
5969 4804230 : let buf_size: u64 = batch.buffer_size() as u64;
5970 4804230 :
5971 4804230 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
5972 4804230 : let layer = self
5973 4804230 : .handle_open_layer_action(batch_max_lsn, action, ctx)
5974 4804230 : .await?;
5975 :
5976 4804230 : let res = layer.put_batch(batch, ctx).await;
5977 :
5978 4804230 : if res.is_ok() {
5979 4804230 : // Update the current size only when the entire write was ok.
5980 4804230 : // In case of failures, we may have had partial writes which
5981 4804230 : // render the size tracking out of sync. That's ok because
5982 4804230 : // the checkpoint distance should be significantly smaller
5983 4804230 : // than the S3 single shot upload limit of 5GiB.
5984 4804230 : let state = self.write_guard.as_mut().unwrap();
5985 4804230 :
5986 4804230 : state.current_size += buf_size;
5987 4804230 : state.prev_lsn = Some(batch_max_lsn);
5988 4804230 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
5989 4804230 : }
5990 :
5991 4804230 : res
5992 4804230 : }
5993 :
5994 : #[cfg(test)]
5995 : /// Test helper, for tests that would like to poke individual values without composing a batch
5996 4390154 : pub(crate) async fn put(
5997 4390154 : &mut self,
5998 4390154 : key: Key,
5999 4390154 : lsn: Lsn,
6000 4390154 : value: &Value,
6001 4390154 : ctx: &RequestContext,
6002 4390154 : ) -> anyhow::Result<()> {
6003 : use utils::bin_ser::BeSer;
6004 4390154 : if !key.is_valid_key_on_write_path() {
6005 0 : bail!(
6006 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
6007 0 : key
6008 0 : );
6009 4390154 : }
6010 4390154 : let val_ser_size = value.serialized_size().unwrap() as usize;
6011 4390154 : let batch = SerializedValueBatch::from_values(vec![(
6012 4390154 : key.to_compact(),
6013 4390154 : lsn,
6014 4390154 : val_ser_size,
6015 4390154 : value.clone(),
6016 4390154 : )]);
6017 4390154 :
6018 4390154 : self.put_batch(batch, ctx).await
6019 4390154 : }
6020 :
6021 2 : pub(crate) async fn delete_batch(
6022 2 : &mut self,
6023 2 : batch: &[(Range<Key>, Lsn)],
6024 2 : ctx: &RequestContext,
6025 2 : ) -> anyhow::Result<()> {
6026 2 : if let Some((_, lsn)) = batch.first() {
6027 2 : let action = self.get_open_layer_action(*lsn, 0);
6028 2 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
6029 2 : layer.put_tombstones(batch).await?;
6030 0 : }
6031 :
6032 2 : Ok(())
6033 2 : }
6034 :
6035 : /// Track the end of the latest digested WAL record.
6036 : /// Remember the (end of) last valid WAL record remembered in the timeline.
6037 : ///
6038 : /// Call this after you have finished writing all the WAL up to 'lsn'.
6039 : ///
6040 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
6041 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
6042 : /// the latest and can be read.
6043 5279092 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
6044 5279092 : self.tl.finish_write(new_lsn);
6045 5279092 : }
6046 :
6047 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
6048 270570 : self.tl.update_current_logical_size(delta)
6049 270570 : }
6050 : }
6051 :
6052 : // We need TimelineWriter to be send in upcoming conversion of
6053 : // Timeline::layers to tokio::sync::RwLock.
6054 : #[test]
6055 2 : fn is_send() {
6056 2 : fn _assert_send<T: Send>() {}
6057 2 : _assert_send::<TimelineWriter<'_>>();
6058 2 : }
6059 :
6060 : #[cfg(test)]
6061 : mod tests {
6062 : use pageserver_api::key::Key;
6063 : use pageserver_api::value::Value;
6064 : use utils::{id::TimelineId, lsn::Lsn};
6065 :
6066 : use crate::tenant::{
6067 : harness::{test_img, TenantHarness},
6068 : layer_map::LayerMap,
6069 : storage_layer::{Layer, LayerName},
6070 : timeline::{DeltaLayerTestDesc, EvictionError},
6071 : Timeline,
6072 : };
6073 :
6074 : #[tokio::test]
6075 2 : async fn test_heatmap_generation() {
6076 2 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
6077 2 :
6078 2 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6079 2 : Lsn(0x10)..Lsn(0x20),
6080 2 : vec![(
6081 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6082 2 : Lsn(0x11),
6083 2 : Value::Image(test_img("foo")),
6084 2 : )],
6085 2 : );
6086 2 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
6087 2 : Lsn(0x10)..Lsn(0x20),
6088 2 : vec![(
6089 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6090 2 : Lsn(0x11),
6091 2 : Value::Image(test_img("foo")),
6092 2 : )],
6093 2 : );
6094 2 : let l0_delta = DeltaLayerTestDesc::new(
6095 2 : Lsn(0x20)..Lsn(0x30),
6096 2 : Key::from_hex("000000000000000000000000000000000000").unwrap()
6097 2 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
6098 2 : vec![(
6099 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
6100 2 : Lsn(0x25),
6101 2 : Value::Image(test_img("foo")),
6102 2 : )],
6103 2 : );
6104 2 : let delta_layers = vec![
6105 2 : covered_delta.clone(),
6106 2 : visible_delta.clone(),
6107 2 : l0_delta.clone(),
6108 2 : ];
6109 2 :
6110 2 : let image_layer = (
6111 2 : Lsn(0x40),
6112 2 : vec![(
6113 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
6114 2 : test_img("bar"),
6115 2 : )],
6116 2 : );
6117 2 : let image_layers = vec![image_layer];
6118 2 :
6119 2 : let (tenant, ctx) = harness.load().await;
6120 2 : let timeline = tenant
6121 2 : .create_test_timeline_with_layers(
6122 2 : TimelineId::generate(),
6123 2 : Lsn(0x10),
6124 2 : 14,
6125 2 : &ctx,
6126 2 : delta_layers,
6127 2 : image_layers,
6128 2 : Lsn(0x100),
6129 2 : )
6130 2 : .await
6131 2 : .unwrap();
6132 2 :
6133 2 : // Layer visibility is an input to heatmap generation, so refresh it first
6134 2 : timeline.update_layer_visibility().await.unwrap();
6135 2 :
6136 2 : let heatmap = timeline
6137 2 : .generate_heatmap()
6138 2 : .await
6139 2 : .expect("Infallible while timeline is not shut down");
6140 2 :
6141 2 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
6142 2 :
6143 2 : // L0 should come last
6144 2 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
6145 2 :
6146 2 : let mut last_lsn = Lsn::MAX;
6147 10 : for layer in heatmap.layers {
6148 2 : // Covered layer should be omitted
6149 8 : assert!(layer.name != covered_delta.layer_name());
6150 2 :
6151 8 : let layer_lsn = match &layer.name {
6152 4 : LayerName::Delta(d) => d.lsn_range.end,
6153 4 : LayerName::Image(i) => i.lsn,
6154 2 : };
6155 2 :
6156 2 : // Apart from L0s, newest Layers should come first
6157 8 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
6158 6 : assert!(layer_lsn <= last_lsn);
6159 6 : last_lsn = layer_lsn;
6160 2 : }
6161 2 : }
6162 2 : }
6163 :
6164 : #[tokio::test]
6165 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
6166 2 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
6167 2 : .await
6168 2 : .unwrap();
6169 2 :
6170 2 : let (tenant, ctx) = harness.load().await;
6171 2 : let timeline = tenant
6172 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
6173 2 : .await
6174 2 : .unwrap();
6175 2 :
6176 2 : let layer = find_some_layer(&timeline).await;
6177 2 : let layer = layer
6178 2 : .keep_resident()
6179 2 : .await
6180 2 : .expect("no download => no downloading errors")
6181 2 : .drop_eviction_guard();
6182 2 :
6183 2 : let forever = std::time::Duration::from_secs(120);
6184 2 :
6185 2 : let first = layer.evict_and_wait(forever);
6186 2 : let second = layer.evict_and_wait(forever);
6187 2 :
6188 2 : let (first, second) = tokio::join!(first, second);
6189 2 :
6190 2 : let res = layer.keep_resident().await;
6191 2 : assert!(res.is_none(), "{res:?}");
6192 2 :
6193 2 : match (first, second) {
6194 2 : (Ok(()), Ok(())) => {
6195 2 : // because there are no more timeline locks being taken on eviction path, we can
6196 2 : // witness all three outcomes here.
6197 2 : }
6198 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
6199 0 : // if one completes before the other, this is fine just as well.
6200 0 : }
6201 2 : other => unreachable!("unexpected {:?}", other),
6202 2 : }
6203 2 : }
6204 :
6205 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
6206 2 : let layers = timeline.layers.read().await;
6207 2 : let desc = layers
6208 2 : .layer_map()
6209 2 : .unwrap()
6210 2 : .iter_historic_layers()
6211 2 : .next()
6212 2 : .expect("must find one layer to evict");
6213 2 :
6214 2 : layers.get_from_desc(&desc)
6215 2 : }
6216 : }
|
