Line data Source code
1 : pub(crate) mod analysis;
2 : pub(crate) mod compaction;
3 : pub mod delete;
4 : pub(crate) mod detach_ancestor;
5 : mod eviction_task;
6 : pub(crate) mod handle;
7 : mod init;
8 : pub mod layer_manager;
9 : pub(crate) mod logical_size;
10 : pub mod offload;
11 : pub mod span;
12 : pub mod uninit;
13 : mod walreceiver;
14 :
15 : use anyhow::{anyhow, bail, ensure, Context, Result};
16 : use arc_swap::ArcSwap;
17 : use bytes::Bytes;
18 : use camino::Utf8Path;
19 : use chrono::{DateTime, Utc};
20 : use enumset::EnumSet;
21 : use fail::fail_point;
22 : use handle::ShardTimelineId;
23 : use once_cell::sync::Lazy;
24 : use pageserver_api::{
25 : key::{
26 : CompactKey, KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX,
27 : NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE,
28 : },
29 : keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning},
30 : models::{
31 : CompactionAlgorithm, CompactionAlgorithmSettings, DownloadRemoteLayersTaskInfo,
32 : DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy, InMemoryLayerInfo, LayerMapInfo,
33 : LsnLease, TimelineState,
34 : },
35 : reltag::BlockNumber,
36 : shard::{ShardIdentity, ShardNumber, TenantShardId},
37 : };
38 : use rand::Rng;
39 : use serde_with::serde_as;
40 : use storage_broker::BrokerClientChannel;
41 : use tokio::{
42 : runtime::Handle,
43 : sync::{oneshot, watch},
44 : };
45 : use tokio_util::sync::CancellationToken;
46 : use tracing::*;
47 : use utils::{
48 : fs_ext, pausable_failpoint,
49 : sync::gate::{Gate, GateGuard},
50 : };
51 :
52 : use std::sync::atomic::Ordering as AtomicOrdering;
53 : use std::sync::{Arc, Mutex, RwLock, Weak};
54 : use std::time::{Duration, Instant, SystemTime};
55 : use std::{
56 : array,
57 : collections::{BTreeMap, HashMap, HashSet},
58 : sync::atomic::AtomicU64,
59 : };
60 : use std::{cmp::min, ops::ControlFlow};
61 : use std::{
62 : collections::btree_map::Entry,
63 : ops::{Deref, Range},
64 : };
65 : use std::{pin::pin, sync::OnceLock};
66 :
67 : use crate::{
68 : aux_file::AuxFileSizeEstimator,
69 : tenant::{
70 : config::AttachmentMode,
71 : layer_map::{LayerMap, SearchResult},
72 : metadata::TimelineMetadata,
73 : storage_layer::{inmemory_layer::IndexEntry, PersistentLayerDesc},
74 : },
75 : walingest::WalLagCooldown,
76 : walredo,
77 : };
78 : use crate::{
79 : context::{DownloadBehavior, RequestContext},
80 : disk_usage_eviction_task::DiskUsageEvictionInfo,
81 : pgdatadir_mapping::CollectKeySpaceError,
82 : };
83 : use crate::{
84 : disk_usage_eviction_task::finite_f32,
85 : tenant::storage_layer::{
86 : AsLayerDesc, DeltaLayerWriter, EvictionError, ImageLayerWriter, InMemoryLayer, Layer,
87 : LayerAccessStatsReset, LayerName, ResidentLayer, ValueReconstructState,
88 : ValuesReconstructState,
89 : },
90 : };
91 : use crate::{
92 : disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
93 : };
94 : use crate::{
95 : l0_flush::{self, L0FlushGlobalState},
96 : metrics::GetKind,
97 : };
98 : use crate::{
99 : metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
100 : };
101 : use crate::{
102 : pgdatadir_mapping::DirectoryKind,
103 : virtual_file::{MaybeFatalIo, VirtualFile},
104 : };
105 : use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
106 : use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
107 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
108 :
109 : use crate::config::PageServerConf;
110 : use crate::keyspace::{KeyPartitioning, KeySpace};
111 : use crate::metrics::TimelineMetrics;
112 : use crate::pgdatadir_mapping::CalculateLogicalSizeError;
113 : use crate::tenant::config::TenantConfOpt;
114 : use pageserver_api::reltag::RelTag;
115 : use pageserver_api::shard::ShardIndex;
116 :
117 : use postgres_connection::PgConnectionConfig;
118 : use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
119 : use utils::{
120 : completion,
121 : generation::Generation,
122 : id::TimelineId,
123 : lsn::{AtomicLsn, Lsn, RecordLsn},
124 : seqwait::SeqWait,
125 : simple_rcu::{Rcu, RcuReadGuard},
126 : };
127 :
128 : use crate::repository::GcResult;
129 : use crate::repository::{Key, Value};
130 : use crate::task_mgr;
131 : use crate::task_mgr::TaskKind;
132 : use crate::ZERO_PAGE;
133 :
134 : use self::delete::DeleteTimelineFlow;
135 : pub(super) use self::eviction_task::EvictionTaskTenantState;
136 : use self::eviction_task::EvictionTaskTimelineState;
137 : use self::layer_manager::LayerManager;
138 : use self::logical_size::LogicalSize;
139 : use self::walreceiver::{WalReceiver, WalReceiverConf};
140 :
141 : use super::{
142 : config::TenantConf,
143 : storage_layer::{inmemory_layer, LayerVisibilityHint},
144 : upload_queue::NotInitialized,
145 : MaybeOffloaded,
146 : };
147 : use super::{debug_assert_current_span_has_tenant_and_timeline_id, AttachedTenantConf};
148 : use super::{remote_timeline_client::index::IndexPart, storage_layer::LayerFringe};
149 : use super::{
150 : remote_timeline_client::RemoteTimelineClient, remote_timeline_client::WaitCompletionError,
151 : storage_layer::ReadableLayer,
152 : };
153 : use super::{
154 : secondary::heatmap::{HeatMapLayer, HeatMapTimeline},
155 : GcError,
156 : };
157 :
158 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
159 : pub(crate) enum FlushLoopState {
160 : NotStarted,
161 : Running {
162 : #[cfg(test)]
163 : expect_initdb_optimization: bool,
164 : #[cfg(test)]
165 : initdb_optimization_count: usize,
166 : },
167 : Exited,
168 : }
169 :
170 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
171 : pub enum ImageLayerCreationMode {
172 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
173 : Try,
174 : /// Force creating the image layers if possible. For now, no image layers will be created
175 : /// for metadata keys. Used in compaction code path with force flag enabled.
176 : Force,
177 : /// Initial ingestion of the data, and no data should be dropped in this function. This
178 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
179 : /// code path.
180 : Initial,
181 : }
182 :
183 : impl std::fmt::Display for ImageLayerCreationMode {
184 704 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
185 704 : write!(f, "{:?}", self)
186 704 : }
187 : }
188 :
189 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
190 : /// Can be removed after all refactors are done.
191 28 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
192 28 : drop(rlock)
193 28 : }
194 :
195 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
196 : /// Can be removed after all refactors are done.
197 732 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
198 732 : drop(rlock)
199 732 : }
200 :
201 : /// The outward-facing resources required to build a Timeline
202 : pub struct TimelineResources {
203 : pub remote_client: RemoteTimelineClient,
204 : pub timeline_get_throttle:
205 : Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
206 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
207 : }
208 :
209 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
210 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
211 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
212 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
213 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
214 : pub(crate) struct RelSizeCache {
215 : pub(crate) complete_as_of: Lsn,
216 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
217 : }
218 :
219 : pub struct Timeline {
220 : pub(crate) conf: &'static PageServerConf,
221 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
222 :
223 : myself: Weak<Self>,
224 :
225 : pub(crate) tenant_shard_id: TenantShardId,
226 : pub timeline_id: TimelineId,
227 :
228 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
229 : /// Never changes for the lifetime of this [`Timeline`] object.
230 : ///
231 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
232 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
233 : pub(crate) generation: Generation,
234 :
235 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
236 : /// to shards, and is constant through the lifetime of this Timeline.
237 : shard_identity: ShardIdentity,
238 :
239 : pub pg_version: u32,
240 :
241 : /// The tuple has two elements.
242 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
243 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
244 : ///
245 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
246 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
247 : ///
248 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
249 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
250 : /// `PersistentLayerDesc`'s.
251 : ///
252 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
253 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
254 : /// runtime, e.g., during page reconstruction.
255 : ///
256 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
257 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
258 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
259 :
260 : last_freeze_at: AtomicLsn,
261 : // Atomic would be more appropriate here.
262 : last_freeze_ts: RwLock<Instant>,
263 :
264 : pub(crate) standby_horizon: AtomicLsn,
265 :
266 : // WAL redo manager. `None` only for broken tenants.
267 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
268 :
269 : /// Remote storage client.
270 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
271 : pub remote_client: Arc<RemoteTimelineClient>,
272 :
273 : // What page versions do we hold in the repository? If we get a
274 : // request > last_record_lsn, we need to wait until we receive all
275 : // the WAL up to the request. The SeqWait provides functions for
276 : // that. TODO: If we get a request for an old LSN, such that the
277 : // versions have already been garbage collected away, we should
278 : // throw an error, but we don't track that currently.
279 : //
280 : // last_record_lsn.load().last points to the end of last processed WAL record.
281 : //
282 : // We also remember the starting point of the previous record in
283 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
284 : // first WAL record when the node is started up. But here, we just
285 : // keep track of it.
286 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
287 :
288 : // All WAL records have been processed and stored durably on files on
289 : // local disk, up to this LSN. On crash and restart, we need to re-process
290 : // the WAL starting from this point.
291 : //
292 : // Some later WAL records might have been processed and also flushed to disk
293 : // already, so don't be surprised to see some, but there's no guarantee on
294 : // them yet.
295 : disk_consistent_lsn: AtomicLsn,
296 :
297 : // Parent timeline that this timeline was branched from, and the LSN
298 : // of the branch point.
299 : ancestor_timeline: Option<Arc<Timeline>>,
300 : ancestor_lsn: Lsn,
301 :
302 : pub(super) metrics: TimelineMetrics,
303 :
304 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
305 : // in `crate::page_service` writes these metrics.
306 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
307 :
308 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
309 :
310 : /// Ensures layers aren't frozen by checkpointer between
311 : /// [`Timeline::get_layer_for_write`] and layer reads.
312 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
313 : /// Must always be acquired before the layer map/individual layer lock
314 : /// to avoid deadlock.
315 : ///
316 : /// The state is cleared upon freezing.
317 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
318 :
319 : /// Used to avoid multiple `flush_loop` tasks running
320 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
321 :
322 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
323 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
324 : /// The flush cycle counter is sent back on the layer_flush_done channel when
325 : /// the flush finishes. You can use that to wait for the flush to finish.
326 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
327 : /// read by whoever sends an update
328 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
329 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
330 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
331 :
332 : // Needed to ensure that we can't create a branch at a point that was already garbage collected
333 : pub latest_gc_cutoff_lsn: Rcu<Lsn>,
334 :
335 : // List of child timelines and their branch points. This is needed to avoid
336 : // garbage collecting data that is still needed by the child timelines.
337 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
338 :
339 : // It may change across major versions so for simplicity
340 : // keep it after running initdb for a timeline.
341 : // It is needed in checks when we want to error on some operations
342 : // when they are requested for pre-initdb lsn.
343 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
344 : // though let's keep them both for better error visibility.
345 : pub initdb_lsn: Lsn,
346 :
347 : /// When did we last calculate the partitioning? Make it pub to test cases.
348 : pub(super) partitioning: tokio::sync::Mutex<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
349 :
350 : /// Configuration: how often should the partitioning be recalculated.
351 : repartition_threshold: u64,
352 :
353 : last_image_layer_creation_check_at: AtomicLsn,
354 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
355 :
356 : /// Current logical size of the "datadir", at the last LSN.
357 : current_logical_size: LogicalSize,
358 :
359 : /// Information about the last processed message by the WAL receiver,
360 : /// or None if WAL receiver has not received anything for this timeline
361 : /// yet.
362 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
363 : pub walreceiver: Mutex<Option<WalReceiver>>,
364 :
365 : /// Relation size cache
366 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
367 :
368 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
369 :
370 : state: watch::Sender<TimelineState>,
371 :
372 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
373 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
374 : pub delete_progress: TimelineDeleteProgress,
375 :
376 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
377 :
378 : /// Load or creation time information about the disk_consistent_lsn and when the loading
379 : /// happened. Used for consumption metrics.
380 : pub(crate) loaded_at: (Lsn, SystemTime),
381 :
382 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
383 : pub(crate) gate: Gate,
384 :
385 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
386 : /// to the timeline should drop out when this token fires.
387 : pub(crate) cancel: CancellationToken,
388 :
389 : /// Make sure we only have one running compaction at a time in tests.
390 : ///
391 : /// Must only be taken in two places:
392 : /// - [`Timeline::compact`] (this file)
393 : /// - [`delete::delete_local_timeline_directory`]
394 : ///
395 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
396 : compaction_lock: tokio::sync::Mutex<()>,
397 :
398 : /// Make sure we only have one running gc at a time.
399 : ///
400 : /// Must only be taken in two places:
401 : /// - [`Timeline::gc`] (this file)
402 : /// - [`delete::delete_local_timeline_directory`]
403 : ///
404 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
405 : gc_lock: tokio::sync::Mutex<()>,
406 :
407 : /// Cloned from [`super::Tenant::timeline_get_throttle`] on construction.
408 : timeline_get_throttle:
409 : Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
410 :
411 : /// Size estimator for aux file v2
412 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
413 :
414 : /// Some test cases directly place keys into the timeline without actually modifying the directory
415 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
416 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
417 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
418 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
419 : #[cfg(test)]
420 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
421 :
422 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
423 :
424 : pub(crate) handles: handle::PerTimelineState<crate::page_service::TenantManagerTypes>,
425 :
426 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
427 : }
428 :
429 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
430 :
431 : pub struct WalReceiverInfo {
432 : pub wal_source_connconf: PgConnectionConfig,
433 : pub last_received_msg_lsn: Lsn,
434 : pub last_received_msg_ts: u128,
435 : }
436 :
437 : /// Information about how much history needs to be retained, needed by
438 : /// Garbage Collection.
439 : #[derive(Default)]
440 : pub(crate) struct GcInfo {
441 : /// Specific LSNs that are needed.
442 : ///
443 : /// Currently, this includes all points where child branches have
444 : /// been forked off from. In the future, could also include
445 : /// explicit user-defined snapshot points.
446 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
447 :
448 : /// The cutoff coordinates, which are combined by selecting the minimum.
449 : pub(crate) cutoffs: GcCutoffs,
450 :
451 : /// Leases granted to particular LSNs.
452 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
453 :
454 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
455 : pub(crate) within_ancestor_pitr: bool,
456 : }
457 :
458 : impl GcInfo {
459 226 : pub(crate) fn min_cutoff(&self) -> Lsn {
460 226 : self.cutoffs.select_min()
461 226 : }
462 :
463 228 : pub(super) fn insert_child(
464 228 : &mut self,
465 228 : child_id: TimelineId,
466 228 : child_lsn: Lsn,
467 228 : is_offloaded: MaybeOffloaded,
468 228 : ) {
469 228 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
470 228 : self.retain_lsns.sort_by_key(|i| i.0);
471 228 : }
472 :
473 2 : pub(super) fn remove_child(&mut self, child_id: TimelineId) {
474 2 : self.retain_lsns.retain(|i| i.1 != child_id);
475 2 : }
476 : }
477 :
478 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
479 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
480 : /// between time-based and space-based retention for observability and consumption metrics purposes.
481 : #[derive(Debug, Clone)]
482 : pub(crate) struct GcCutoffs {
483 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
484 : /// history we must keep to retain a specified number of bytes of WAL.
485 : pub(crate) space: Lsn,
486 :
487 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
488 : /// history we must keep to enable reading back at least the PITR interval duration.
489 : pub(crate) time: Lsn,
490 : }
491 :
492 : impl Default for GcCutoffs {
493 410 : fn default() -> Self {
494 410 : Self {
495 410 : space: Lsn::INVALID,
496 410 : time: Lsn::INVALID,
497 410 : }
498 410 : }
499 : }
500 :
501 : impl GcCutoffs {
502 252 : fn select_min(&self) -> Lsn {
503 252 : std::cmp::min(self.space, self.time)
504 252 : }
505 : }
506 :
507 : pub(crate) struct TimelineVisitOutcome {
508 : completed_keyspace: KeySpace,
509 : image_covered_keyspace: KeySpace,
510 : }
511 :
512 : /// An error happened in a get() operation.
513 2 : #[derive(thiserror::Error, Debug)]
514 : pub(crate) enum PageReconstructError {
515 : #[error(transparent)]
516 : Other(anyhow::Error),
517 :
518 : #[error("Ancestor LSN wait error: {0}")]
519 : AncestorLsnTimeout(WaitLsnError),
520 :
521 : #[error("timeline shutting down")]
522 : Cancelled,
523 :
524 : /// An error happened replaying WAL records
525 : #[error(transparent)]
526 : WalRedo(anyhow::Error),
527 :
528 : #[error("{0}")]
529 : MissingKey(MissingKeyError),
530 : }
531 :
532 : impl From<anyhow::Error> for PageReconstructError {
533 0 : fn from(value: anyhow::Error) -> Self {
534 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
535 0 : match value.downcast::<PageReconstructError>() {
536 0 : Ok(pre) => pre,
537 0 : Err(other) => PageReconstructError::Other(other),
538 : }
539 0 : }
540 : }
541 :
542 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
543 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
544 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
545 0 : }
546 : }
547 :
548 : impl From<layer_manager::Shutdown> for PageReconstructError {
549 0 : fn from(_: layer_manager::Shutdown) -> Self {
550 0 : PageReconstructError::Cancelled
551 0 : }
552 : }
553 :
554 : impl GetVectoredError {
555 : #[cfg(test)]
556 6 : pub(crate) fn is_missing_key_error(&self) -> bool {
557 6 : matches!(self, Self::MissingKey(_))
558 6 : }
559 : }
560 :
561 : impl From<layer_manager::Shutdown> for GetVectoredError {
562 0 : fn from(_: layer_manager::Shutdown) -> Self {
563 0 : GetVectoredError::Cancelled
564 0 : }
565 : }
566 :
567 : #[derive(thiserror::Error)]
568 : pub struct MissingKeyError {
569 : key: Key,
570 : shard: ShardNumber,
571 : cont_lsn: Lsn,
572 : request_lsn: Lsn,
573 : ancestor_lsn: Option<Lsn>,
574 : backtrace: Option<std::backtrace::Backtrace>,
575 : }
576 :
577 : impl std::fmt::Debug for MissingKeyError {
578 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
579 0 : write!(f, "{}", self)
580 0 : }
581 : }
582 :
583 : impl std::fmt::Display for MissingKeyError {
584 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
585 0 : write!(
586 0 : f,
587 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
588 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
589 0 : )?;
590 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
591 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
592 0 : }
593 :
594 0 : if let Some(ref backtrace) = self.backtrace {
595 0 : write!(f, "\n{}", backtrace)?;
596 0 : }
597 :
598 0 : Ok(())
599 0 : }
600 : }
601 :
602 : impl PageReconstructError {
603 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
604 0 : pub(crate) fn is_stopping(&self) -> bool {
605 : use PageReconstructError::*;
606 0 : match self {
607 0 : Cancelled => true,
608 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
609 : }
610 0 : }
611 : }
612 :
613 0 : #[derive(thiserror::Error, Debug)]
614 : pub(crate) enum CreateImageLayersError {
615 : #[error("timeline shutting down")]
616 : Cancelled,
617 :
618 : #[error("read failed")]
619 : GetVectoredError(#[source] GetVectoredError),
620 :
621 : #[error("reconstruction failed")]
622 : PageReconstructError(#[source] PageReconstructError),
623 :
624 : #[error(transparent)]
625 : Other(#[from] anyhow::Error),
626 : }
627 :
628 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
629 0 : fn from(_: layer_manager::Shutdown) -> Self {
630 0 : CreateImageLayersError::Cancelled
631 0 : }
632 : }
633 :
634 0 : #[derive(thiserror::Error, Debug, Clone)]
635 : pub(crate) enum FlushLayerError {
636 : /// Timeline cancellation token was cancelled
637 : #[error("timeline shutting down")]
638 : Cancelled,
639 :
640 : /// We tried to flush a layer while the Timeline is in an unexpected state
641 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
642 : NotRunning(FlushLoopState),
643 :
644 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
645 : // loop via a watch channel, where we can only borrow it.
646 : #[error("create image layers (shared)")]
647 : CreateImageLayersError(Arc<CreateImageLayersError>),
648 :
649 : #[error("other (shared)")]
650 : Other(#[from] Arc<anyhow::Error>),
651 : }
652 :
653 : impl FlushLayerError {
654 : // When crossing from generic anyhow errors to this error type, we explicitly check
655 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
656 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
657 0 : let cancelled = timeline.cancel.is_cancelled()
658 : // The upload queue might have been shut down before the official cancellation of the timeline.
659 0 : || err
660 0 : .downcast_ref::<NotInitialized>()
661 0 : .map(NotInitialized::is_stopping)
662 0 : .unwrap_or_default();
663 0 : if cancelled {
664 0 : Self::Cancelled
665 : } else {
666 0 : Self::Other(Arc::new(err))
667 : }
668 0 : }
669 : }
670 :
671 : impl From<layer_manager::Shutdown> for FlushLayerError {
672 0 : fn from(_: layer_manager::Shutdown) -> Self {
673 0 : FlushLayerError::Cancelled
674 0 : }
675 : }
676 :
677 0 : #[derive(thiserror::Error, Debug)]
678 : pub(crate) enum GetVectoredError {
679 : #[error("timeline shutting down")]
680 : Cancelled,
681 :
682 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
683 : Oversized(u64),
684 :
685 : #[error("requested at invalid LSN: {0}")]
686 : InvalidLsn(Lsn),
687 :
688 : #[error("requested key not found: {0}")]
689 : MissingKey(MissingKeyError),
690 :
691 : #[error("ancestry walk")]
692 : GetReadyAncestorError(#[source] GetReadyAncestorError),
693 :
694 : #[error(transparent)]
695 : Other(#[from] anyhow::Error),
696 : }
697 :
698 : impl From<GetReadyAncestorError> for GetVectoredError {
699 2 : fn from(value: GetReadyAncestorError) -> Self {
700 : use GetReadyAncestorError::*;
701 2 : match value {
702 0 : Cancelled => GetVectoredError::Cancelled,
703 : AncestorLsnTimeout(_) | BadState { .. } => {
704 2 : GetVectoredError::GetReadyAncestorError(value)
705 : }
706 : }
707 2 : }
708 : }
709 :
710 2 : #[derive(thiserror::Error, Debug)]
711 : pub(crate) enum GetReadyAncestorError {
712 : #[error("ancestor LSN wait error")]
713 : AncestorLsnTimeout(#[from] WaitLsnError),
714 :
715 : #[error("bad state on timeline {timeline_id}: {state:?}")]
716 : BadState {
717 : timeline_id: TimelineId,
718 : state: TimelineState,
719 : },
720 :
721 : #[error("cancelled")]
722 : Cancelled,
723 : }
724 :
725 : #[derive(Clone, Copy)]
726 : pub enum LogicalSizeCalculationCause {
727 : Initial,
728 : ConsumptionMetricsSyntheticSize,
729 : EvictionTaskImitation,
730 : TenantSizeHandler,
731 : }
732 :
733 : pub enum GetLogicalSizePriority {
734 : User,
735 : Background,
736 : }
737 :
738 0 : #[derive(enumset::EnumSetType)]
739 : pub(crate) enum CompactFlags {
740 : ForceRepartition,
741 : ForceImageLayerCreation,
742 : ForceL0Compaction,
743 : EnhancedGcBottomMostCompaction,
744 : DryRun,
745 : }
746 :
747 : impl std::fmt::Debug for Timeline {
748 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
749 0 : write!(f, "Timeline<{}>", self.timeline_id)
750 0 : }
751 : }
752 :
753 0 : #[derive(thiserror::Error, Debug)]
754 : pub(crate) enum WaitLsnError {
755 : // Called on a timeline which is shutting down
756 : #[error("Shutdown")]
757 : Shutdown,
758 :
759 : // Called on an timeline not in active state or shutting down
760 : #[error("Bad timeline state: {0:?}")]
761 : BadState(TimelineState),
762 :
763 : // Timeout expired while waiting for LSN to catch up with goal.
764 : #[error("{0}")]
765 : Timeout(String),
766 : }
767 :
768 : // The impls below achieve cancellation mapping for errors.
769 : // Perhaps there's a way of achieving this with less cruft.
770 :
771 : impl From<CreateImageLayersError> for CompactionError {
772 0 : fn from(e: CreateImageLayersError) -> Self {
773 0 : match e {
774 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
775 0 : CreateImageLayersError::Other(e) => {
776 0 : CompactionError::Other(e.context("create image layers"))
777 : }
778 0 : _ => CompactionError::Other(e.into()),
779 : }
780 0 : }
781 : }
782 :
783 : impl From<CreateImageLayersError> for FlushLayerError {
784 0 : fn from(e: CreateImageLayersError) -> Self {
785 0 : match e {
786 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
787 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
788 : }
789 0 : }
790 : }
791 :
792 : impl From<PageReconstructError> for CreateImageLayersError {
793 0 : fn from(e: PageReconstructError) -> Self {
794 0 : match e {
795 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
796 0 : _ => CreateImageLayersError::PageReconstructError(e),
797 : }
798 0 : }
799 : }
800 :
801 : impl From<GetVectoredError> for CreateImageLayersError {
802 0 : fn from(e: GetVectoredError) -> Self {
803 0 : match e {
804 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
805 0 : _ => CreateImageLayersError::GetVectoredError(e),
806 : }
807 0 : }
808 : }
809 :
810 : impl From<GetVectoredError> for PageReconstructError {
811 6 : fn from(e: GetVectoredError) -> Self {
812 6 : match e {
813 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
814 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
815 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
816 4 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
817 2 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
818 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
819 : }
820 6 : }
821 : }
822 :
823 : impl From<GetReadyAncestorError> for PageReconstructError {
824 2 : fn from(e: GetReadyAncestorError) -> Self {
825 : use GetReadyAncestorError::*;
826 2 : match e {
827 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
828 2 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
829 0 : Cancelled => PageReconstructError::Cancelled,
830 : }
831 2 : }
832 : }
833 :
834 : pub(crate) enum WaitLsnWaiter<'a> {
835 : Timeline(&'a Timeline),
836 : Tenant,
837 : PageService,
838 : }
839 :
840 : /// Argument to [`Timeline::shutdown`].
841 : #[derive(Debug, Clone, Copy)]
842 : pub(crate) enum ShutdownMode {
843 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
844 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
845 : ///
846 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
847 : /// the call to [`Timeline::shutdown`].
848 : FreezeAndFlush,
849 : /// Shut down immediately, without waiting for any open layers to flush.
850 : Hard,
851 : }
852 :
853 : struct ImageLayerCreationOutcome {
854 : image: Option<ResidentLayer>,
855 : next_start_key: Key,
856 : }
857 :
858 : /// Public interface functions
859 : impl Timeline {
860 : /// Get the LSN where this branch was created
861 2 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
862 2 : self.ancestor_lsn
863 2 : }
864 :
865 : /// Get the ancestor's timeline id
866 4 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
867 4 : self.ancestor_timeline
868 4 : .as_ref()
869 4 : .map(|ancestor| ancestor.timeline_id)
870 4 : }
871 :
872 : /// Get the ancestor timeline
873 0 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
874 0 : self.ancestor_timeline.as_ref()
875 0 : }
876 :
877 : /// Get the bytes written since the PITR cutoff on this branch, and
878 : /// whether this branch's ancestor_lsn is within its parent's PITR.
879 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
880 0 : let gc_info = self.gc_info.read().unwrap();
881 0 : let history = self
882 0 : .get_last_record_lsn()
883 0 : .checked_sub(gc_info.cutoffs.time)
884 0 : .unwrap_or(Lsn(0))
885 0 : .0;
886 0 : (history, gc_info.within_ancestor_pitr)
887 0 : }
888 :
889 : /// Lock and get timeline's GC cutoff
890 246 : pub(crate) fn get_latest_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
891 246 : self.latest_gc_cutoff_lsn.read()
892 246 : }
893 :
894 : /// Look up given page version.
895 : ///
896 : /// If a remote layer file is needed, it is downloaded as part of this
897 : /// call.
898 : ///
899 : /// This method enforces [`Self::timeline_get_throttle`] internally.
900 : ///
901 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
902 : /// abstraction above this needs to store suitable metadata to track what
903 : /// data exists with what keys, in separate metadata entries. If a
904 : /// non-existent key is requested, we may incorrectly return a value from
905 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
906 : /// non-existing key.
907 : ///
908 : /// # Cancel-Safety
909 : ///
910 : /// This method is cancellation-safe.
911 : #[inline(always)]
912 625019 : pub(crate) async fn get(
913 625019 : &self,
914 625019 : key: Key,
915 625019 : lsn: Lsn,
916 625019 : ctx: &RequestContext,
917 625019 : ) -> Result<Bytes, PageReconstructError> {
918 625019 : if !lsn.is_valid() {
919 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
920 625019 : }
921 625019 :
922 625019 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
923 625019 : // already checked the key against the shard_identity when looking up the Timeline from
924 625019 : // page_service.
925 625019 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
926 :
927 625019 : self.timeline_get_throttle.throttle(ctx, 1).await;
928 :
929 625019 : let keyspace = KeySpace {
930 625019 : ranges: vec![key..key.next()],
931 625019 : };
932 625019 :
933 625019 : // Initialise the reconstruct state for the key with the cache
934 625019 : // entry returned above.
935 625019 : let mut reconstruct_state = ValuesReconstructState::new();
936 :
937 625019 : let vectored_res = self
938 625019 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
939 182416 : .await;
940 :
941 625019 : let key_value = vectored_res?.pop_first();
942 625013 : match key_value {
943 625001 : Some((got_key, value)) => {
944 625001 : if got_key != key {
945 0 : error!(
946 0 : "Expected {}, but singular vectored get returned {}",
947 : key, got_key
948 : );
949 0 : Err(PageReconstructError::Other(anyhow!(
950 0 : "Singular vectored get returned wrong key"
951 0 : )))
952 : } else {
953 625001 : value
954 : }
955 : }
956 12 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
957 12 : key,
958 12 : shard: self.shard_identity.get_shard_number(&key),
959 12 : cont_lsn: Lsn(0),
960 12 : request_lsn: lsn,
961 12 : ancestor_lsn: None,
962 12 : backtrace: None,
963 12 : })),
964 : }
965 625019 : }
966 :
967 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
968 : pub(crate) const VEC_GET_LAYERS_VISITED_WARN_THRESH: f64 = 512.0;
969 :
970 : /// Look up multiple page versions at a given LSN
971 : ///
972 : /// This naive implementation will be replaced with a more efficient one
973 : /// which actually vectorizes the read path.
974 1108 : pub(crate) async fn get_vectored(
975 1108 : &self,
976 1108 : keyspace: KeySpace,
977 1108 : lsn: Lsn,
978 1108 : ctx: &RequestContext,
979 1108 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
980 1108 : if !lsn.is_valid() {
981 0 : return Err(GetVectoredError::InvalidLsn(lsn));
982 1108 : }
983 1108 :
984 1108 : let key_count = keyspace.total_raw_size().try_into().unwrap();
985 1108 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
986 0 : return Err(GetVectoredError::Oversized(key_count));
987 1108 : }
988 :
989 2216 : for range in &keyspace.ranges {
990 1108 : let mut key = range.start;
991 2416 : while key != range.end {
992 1308 : assert!(!self.shard_identity.is_key_disposable(&key));
993 1308 : key = key.next();
994 : }
995 : }
996 :
997 1108 : trace!(
998 0 : "get vectored request for {:?}@{} from task kind {:?}",
999 0 : keyspace,
1000 0 : lsn,
1001 0 : ctx.task_kind(),
1002 : );
1003 :
1004 1108 : let start = crate::metrics::GET_VECTORED_LATENCY
1005 1108 : .for_task_kind(ctx.task_kind())
1006 1108 : .map(|metric| (metric, Instant::now()));
1007 :
1008 : // start counting after throttle so that throttle time
1009 : // is always less than observation time
1010 1108 : let throttled = self
1011 1108 : .timeline_get_throttle
1012 1108 : .throttle(ctx, key_count as usize)
1013 0 : .await;
1014 :
1015 1108 : let res = self
1016 1108 : .get_vectored_impl(
1017 1108 : keyspace.clone(),
1018 1108 : lsn,
1019 1108 : &mut ValuesReconstructState::new(),
1020 1108 : ctx,
1021 1108 : )
1022 62 : .await;
1023 :
1024 1108 : if let Some((metric, start)) = start {
1025 0 : let elapsed = start.elapsed();
1026 0 : let ex_throttled = if let Some(throttled) = throttled {
1027 0 : elapsed.checked_sub(throttled)
1028 : } else {
1029 0 : Some(elapsed)
1030 : };
1031 :
1032 0 : if let Some(ex_throttled) = ex_throttled {
1033 0 : metric.observe(ex_throttled.as_secs_f64());
1034 0 : } else {
1035 0 : use utils::rate_limit::RateLimit;
1036 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1037 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
1038 0 : let mut rate_limit = LOGGED.lock().unwrap();
1039 0 : rate_limit.call(|| {
1040 0 : warn!("error deducting time spent throttled; this message is logged at a global rate limit");
1041 0 : });
1042 0 : }
1043 1108 : }
1044 :
1045 1108 : res
1046 1108 : }
1047 :
1048 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1049 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1050 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1051 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1052 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1053 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1054 : /// the scan operation will not cause OOM in the future.
1055 12 : pub(crate) async fn scan(
1056 12 : &self,
1057 12 : keyspace: KeySpace,
1058 12 : lsn: Lsn,
1059 12 : ctx: &RequestContext,
1060 12 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1061 12 : if !lsn.is_valid() {
1062 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1063 12 : }
1064 12 :
1065 12 : trace!(
1066 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1067 0 : keyspace,
1068 0 : lsn,
1069 0 : ctx.task_kind()
1070 : );
1071 :
1072 : // We should generalize this into Keyspace::contains in the future.
1073 24 : for range in &keyspace.ranges {
1074 12 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1075 12 : || range.end.field1 > METADATA_KEY_END_PREFIX
1076 : {
1077 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1078 0 : "only metadata keyspace can be scanned"
1079 0 : )));
1080 12 : }
1081 : }
1082 :
1083 12 : let start = crate::metrics::SCAN_LATENCY
1084 12 : .for_task_kind(ctx.task_kind())
1085 12 : .map(ScanLatencyOngoingRecording::start_recording);
1086 :
1087 : // start counting after throttle so that throttle time
1088 : // is always less than observation time
1089 12 : let throttled = self
1090 12 : .timeline_get_throttle
1091 12 : // assume scan = 1 quota for now until we find a better way to process this
1092 12 : .throttle(ctx, 1)
1093 0 : .await;
1094 :
1095 12 : let vectored_res = self
1096 12 : .get_vectored_impl(
1097 12 : keyspace.clone(),
1098 12 : lsn,
1099 12 : &mut ValuesReconstructState::default(),
1100 12 : ctx,
1101 12 : )
1102 0 : .await;
1103 :
1104 12 : if let Some(recording) = start {
1105 0 : recording.observe(throttled);
1106 12 : }
1107 :
1108 12 : vectored_res
1109 12 : }
1110 :
1111 626435 : pub(super) async fn get_vectored_impl(
1112 626435 : &self,
1113 626435 : keyspace: KeySpace,
1114 626435 : lsn: Lsn,
1115 626435 : reconstruct_state: &mut ValuesReconstructState,
1116 626435 : ctx: &RequestContext,
1117 626435 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1118 626435 : let get_kind = if keyspace.total_raw_size() == 1 {
1119 626011 : GetKind::Singular
1120 : } else {
1121 424 : GetKind::Vectored
1122 : };
1123 :
1124 626435 : let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
1125 626435 : .for_get_kind(get_kind)
1126 626435 : .start_timer();
1127 626435 : self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1128 186430 : .await?;
1129 626419 : get_data_timer.stop_and_record();
1130 626419 :
1131 626419 : let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
1132 626419 : .for_get_kind(get_kind)
1133 626419 : .start_timer();
1134 626419 : let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
1135 626419 : let layers_visited = reconstruct_state.get_layers_visited();
1136 :
1137 666745 : for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
1138 666745 : match res {
1139 0 : Err(err) => {
1140 0 : results.insert(key, Err(err));
1141 0 : }
1142 666745 : Ok(state) => {
1143 666745 : let state = ValueReconstructState::from(state);
1144 :
1145 666745 : let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
1146 666745 : results.insert(key, reconstruct_res);
1147 : }
1148 : }
1149 : }
1150 626419 : reconstruct_timer.stop_and_record();
1151 626419 :
1152 626419 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1153 626419 : // when they're missing. Instead they are omitted from the resulting btree
1154 626419 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1155 626419 : // to avoid infinite results.
1156 626419 : if !results.is_empty() {
1157 626219 : let avg = layers_visited as f64 / results.len() as f64;
1158 626219 : if avg >= Self::VEC_GET_LAYERS_VISITED_WARN_THRESH {
1159 0 : use utils::rate_limit::RateLimit;
1160 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1161 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1162 0 : let mut rate_limit = LOGGED.lock().unwrap();
1163 0 : rate_limit.call(|| {
1164 0 : tracing::info!(
1165 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1166 0 : lsn = %lsn,
1167 0 : "Vectored read for {} visited {} layers on average per key and {} in total. {}/{} pages were returned",
1168 0 : keyspace, avg, layers_visited, results.len(), keyspace.total_raw_size());
1169 0 : });
1170 626219 : }
1171 :
1172 : // Note that this is an approximation. Tracking the exact number of layers visited
1173 : // per key requires virtually unbounded memory usage and is inefficient
1174 : // (i.e. segment tree tracking each range queried from a layer)
1175 626219 : crate::metrics::VEC_READ_NUM_LAYERS_VISITED.observe(avg);
1176 200 : }
1177 :
1178 626419 : Ok(results)
1179 626435 : }
1180 :
1181 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1182 274250 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1183 274250 : self.last_record_lsn.load().last
1184 274250 : }
1185 :
1186 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1187 0 : self.last_record_lsn.load().prev
1188 0 : }
1189 :
1190 : /// Atomically get both last and prev.
1191 226 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1192 226 : self.last_record_lsn.load()
1193 226 : }
1194 :
1195 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1196 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1197 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1198 0 : self.last_record_lsn.status_receiver()
1199 0 : }
1200 :
1201 406 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1202 406 : self.disk_consistent_lsn.load()
1203 406 : }
1204 :
1205 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1206 : /// not validated with control plane yet.
1207 : /// See [`Self::get_remote_consistent_lsn_visible`].
1208 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1209 0 : self.remote_client.remote_consistent_lsn_projected()
1210 0 : }
1211 :
1212 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1213 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1214 : /// generation validation in the deletion queue.
1215 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1216 0 : self.remote_client.remote_consistent_lsn_visible()
1217 0 : }
1218 :
1219 : /// The sum of the file size of all historic layers in the layer map.
1220 : /// This method makes no distinction between local and remote layers.
1221 : /// Hence, the result **does not represent local filesystem usage**.
1222 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1223 0 : let guard = self.layers.read().await;
1224 0 : guard.layer_size_sum()
1225 0 : }
1226 :
1227 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1228 0 : self.metrics.resident_physical_size_get()
1229 0 : }
1230 :
1231 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1232 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1233 0 : }
1234 :
1235 : ///
1236 : /// Wait until WAL has been received and processed up to this LSN.
1237 : ///
1238 : /// You should call this before any of the other get_* or list_* functions. Calling
1239 : /// those functions with an LSN that has been processed yet is an error.
1240 : ///
1241 225369 : pub(crate) async fn wait_lsn(
1242 225369 : &self,
1243 225369 : lsn: Lsn,
1244 225369 : who_is_waiting: WaitLsnWaiter<'_>,
1245 225369 : ctx: &RequestContext, /* Prepare for use by cancellation */
1246 225369 : ) -> Result<(), WaitLsnError> {
1247 225369 : let state = self.current_state();
1248 225369 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1249 0 : return Err(WaitLsnError::Shutdown);
1250 225369 : } else if !matches!(state, TimelineState::Active) {
1251 0 : return Err(WaitLsnError::BadState(state));
1252 225369 : }
1253 225369 :
1254 225369 : if cfg!(debug_assertions) {
1255 225369 : match ctx.task_kind() {
1256 : TaskKind::WalReceiverManager
1257 : | TaskKind::WalReceiverConnectionHandler
1258 : | TaskKind::WalReceiverConnectionPoller => {
1259 0 : let is_myself = match who_is_waiting {
1260 0 : WaitLsnWaiter::Timeline(waiter) => Weak::ptr_eq(&waiter.myself, &self.myself),
1261 0 : WaitLsnWaiter::Tenant | WaitLsnWaiter::PageService => unreachable!("tenant or page_service context are not expected to have task kind {:?}", ctx.task_kind()),
1262 : };
1263 0 : if is_myself {
1264 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1265 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1266 0 : panic!("this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock");
1267 0 : }
1268 0 : } else {
1269 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1270 0 : // our walreceiver task can make progress independent of theirs
1271 0 : }
1272 : }
1273 225369 : _ => {}
1274 : }
1275 0 : }
1276 :
1277 225369 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1278 225369 :
1279 225369 : match self
1280 225369 : .last_record_lsn
1281 225369 : .wait_for_timeout(lsn, self.conf.wait_lsn_timeout)
1282 0 : .await
1283 : {
1284 225369 : Ok(()) => Ok(()),
1285 0 : Err(e) => {
1286 : use utils::seqwait::SeqWaitError::*;
1287 0 : match e {
1288 0 : Shutdown => Err(WaitLsnError::Shutdown),
1289 : Timeout => {
1290 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1291 0 : drop(_timer);
1292 0 : let walreceiver_status = self.walreceiver_status();
1293 0 : Err(WaitLsnError::Timeout(format!(
1294 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1295 0 : lsn,
1296 0 : self.get_last_record_lsn(),
1297 0 : self.get_disk_consistent_lsn(),
1298 0 : walreceiver_status,
1299 0 : )))
1300 : }
1301 : }
1302 : }
1303 : }
1304 225369 : }
1305 :
1306 0 : pub(crate) fn walreceiver_status(&self) -> String {
1307 0 : match &*self.walreceiver.lock().unwrap() {
1308 0 : None => "stopping or stopped".to_string(),
1309 0 : Some(walreceiver) => match walreceiver.status() {
1310 0 : Some(status) => status.to_human_readable_string(),
1311 0 : None => "Not active".to_string(),
1312 : },
1313 : }
1314 0 : }
1315 :
1316 : /// Check that it is valid to request operations with that lsn.
1317 230 : pub(crate) fn check_lsn_is_in_scope(
1318 230 : &self,
1319 230 : lsn: Lsn,
1320 230 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1321 230 : ) -> anyhow::Result<()> {
1322 230 : ensure!(
1323 230 : lsn >= **latest_gc_cutoff_lsn,
1324 4 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1325 4 : lsn,
1326 4 : **latest_gc_cutoff_lsn,
1327 : );
1328 226 : Ok(())
1329 230 : }
1330 :
1331 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1332 10 : pub(crate) fn init_lsn_lease(
1333 10 : &self,
1334 10 : lsn: Lsn,
1335 10 : length: Duration,
1336 10 : ctx: &RequestContext,
1337 10 : ) -> anyhow::Result<LsnLease> {
1338 10 : self.make_lsn_lease(lsn, length, true, ctx)
1339 10 : }
1340 :
1341 : /// 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.
1342 4 : pub(crate) fn renew_lsn_lease(
1343 4 : &self,
1344 4 : lsn: Lsn,
1345 4 : length: Duration,
1346 4 : ctx: &RequestContext,
1347 4 : ) -> anyhow::Result<LsnLease> {
1348 4 : self.make_lsn_lease(lsn, length, false, ctx)
1349 4 : }
1350 :
1351 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1352 : ///
1353 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1354 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1355 : ///
1356 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1357 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1358 14 : fn make_lsn_lease(
1359 14 : &self,
1360 14 : lsn: Lsn,
1361 14 : length: Duration,
1362 14 : init: bool,
1363 14 : _ctx: &RequestContext,
1364 14 : ) -> anyhow::Result<LsnLease> {
1365 12 : let lease = {
1366 : // Normalize the requested LSN to be aligned, and move to the first record
1367 : // if it points to the beginning of the page (header).
1368 14 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1369 14 :
1370 14 : let mut gc_info = self.gc_info.write().unwrap();
1371 14 :
1372 14 : let valid_until = SystemTime::now() + length;
1373 14 :
1374 14 : let entry = gc_info.leases.entry(lsn);
1375 14 :
1376 14 : match entry {
1377 6 : Entry::Occupied(mut occupied) => {
1378 6 : let existing_lease = occupied.get_mut();
1379 6 : if valid_until > existing_lease.valid_until {
1380 2 : existing_lease.valid_until = valid_until;
1381 2 : let dt: DateTime<Utc> = valid_until.into();
1382 2 : info!("lease extended to {}", dt);
1383 : } else {
1384 4 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1385 4 : info!("existing lease covers greater length, valid until {}", dt);
1386 : }
1387 :
1388 6 : existing_lease.clone()
1389 : }
1390 8 : Entry::Vacant(vacant) => {
1391 : // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
1392 : // not blocked by the lsn lease deadline.
1393 8 : let validate = {
1394 8 : let conf = self.tenant_conf.load();
1395 8 : conf.location.attach_mode == AttachmentMode::Single
1396 8 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1397 : };
1398 :
1399 8 : if init || validate {
1400 8 : let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
1401 8 : if lsn < *latest_gc_cutoff_lsn {
1402 2 : bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
1403 6 : }
1404 0 : }
1405 :
1406 6 : let dt: DateTime<Utc> = valid_until.into();
1407 6 : info!("lease created, valid until {}", dt);
1408 6 : vacant.insert(LsnLease { valid_until }).clone()
1409 : }
1410 : }
1411 : };
1412 :
1413 12 : Ok(lease)
1414 14 : }
1415 :
1416 : /// Flush to disk all data that was written with the put_* functions
1417 1086 : #[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))]
1418 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1419 : self.freeze_and_flush0().await
1420 : }
1421 :
1422 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1423 : // polluting the span hierarchy.
1424 1086 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1425 1086 : let token = {
1426 : // Freeze the current open in-memory layer. It will be written to disk on next
1427 : // iteration.
1428 1086 : let mut g = self.write_lock.lock().await;
1429 :
1430 1086 : let to_lsn = self.get_last_record_lsn();
1431 1086 : self.freeze_inmem_layer_at(to_lsn, &mut g).await?
1432 : };
1433 1086 : self.wait_flush_completion(token).await
1434 1086 : }
1435 :
1436 : // Check if an open ephemeral layer should be closed: this provides
1437 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1438 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1439 : // ephemeral layer bytes has been breached.
1440 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1441 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1442 : // If the write lock is held, there is an active wal receiver: rolling open layers
1443 : // is their responsibility while they hold this lock.
1444 0 : return;
1445 : };
1446 :
1447 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1448 : // time, and this was missed.
1449 : // if write_guard.is_none() { return; }
1450 :
1451 0 : let Ok(layers_guard) = self.layers.try_read() else {
1452 : // Don't block if the layer lock is busy
1453 0 : return;
1454 : };
1455 :
1456 0 : let Ok(lm) = layers_guard.layer_map() else {
1457 0 : return;
1458 : };
1459 :
1460 0 : let Some(open_layer) = &lm.open_layer else {
1461 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1462 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1463 : // that didn't result in writes to this shard.
1464 :
1465 : // Must not hold the layers lock while waiting for a flush.
1466 0 : drop(layers_guard);
1467 0 :
1468 0 : let last_record_lsn = self.get_last_record_lsn();
1469 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1470 0 : if last_record_lsn > disk_consistent_lsn {
1471 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1472 : // we are a sharded tenant and have skipped some WAL
1473 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1474 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1475 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1476 : // without any data ingested (yet) doesn't write a remote index as soon as it
1477 : // sees its LSN advance: we only do this if we've been layer-less
1478 : // for some time.
1479 0 : tracing::debug!(
1480 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1481 : disk_consistent_lsn,
1482 : last_record_lsn
1483 : );
1484 :
1485 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1486 : // We know there is no open layer, so we can request freezing without actually
1487 : // freezing anything. This is true even if we have dropped the layers_guard, we
1488 : // still hold the write_guard.
1489 0 : let _ = async {
1490 0 : let token = self
1491 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1492 0 : .await?;
1493 0 : self.wait_flush_completion(token).await
1494 0 : }
1495 0 : .await;
1496 0 : }
1497 0 : }
1498 :
1499 0 : return;
1500 : };
1501 :
1502 0 : let Some(current_size) = open_layer.try_len() else {
1503 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1504 : // read lock to get size should always succeed.
1505 0 : tracing::warn!("Lock conflict while reading size of open layer");
1506 0 : return;
1507 : };
1508 :
1509 0 : let current_lsn = self.get_last_record_lsn();
1510 :
1511 0 : let checkpoint_distance_override = open_layer.tick().await;
1512 :
1513 0 : if let Some(size_override) = checkpoint_distance_override {
1514 0 : if current_size > size_override {
1515 : // This is not harmful, but it only happens in relatively rare cases where
1516 : // time-based checkpoints are not happening fast enough to keep the amount of
1517 : // ephemeral data within configured limits. It's a sign of stress on the system.
1518 0 : tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
1519 0 : }
1520 0 : }
1521 :
1522 0 : let checkpoint_distance =
1523 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1524 0 :
1525 0 : if self.should_roll(
1526 0 : current_size,
1527 0 : current_size,
1528 0 : checkpoint_distance,
1529 0 : self.get_last_record_lsn(),
1530 0 : self.last_freeze_at.load(),
1531 0 : open_layer.get_opened_at(),
1532 0 : ) {
1533 0 : match open_layer.info() {
1534 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1535 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1536 0 : // happens asynchronously in the background.
1537 0 : tracing::debug!(
1538 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1539 : );
1540 : }
1541 : InMemoryLayerInfo::Open { .. } => {
1542 : // Upgrade to a write lock and freeze the layer
1543 0 : drop(layers_guard);
1544 0 : let res = self
1545 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1546 0 : .await;
1547 :
1548 0 : if let Err(e) = res {
1549 0 : tracing::info!(
1550 0 : "failed to flush frozen layer after background freeze: {e:#}"
1551 : );
1552 0 : }
1553 : }
1554 : }
1555 0 : }
1556 0 : }
1557 :
1558 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1559 : ///
1560 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1561 : /// child timelines that are not offloaded yet.
1562 0 : pub(crate) fn can_offload(&self) -> bool {
1563 0 : if self.remote_client.is_archived() != Some(true) {
1564 0 : return false;
1565 0 : }
1566 0 :
1567 0 : true
1568 0 : }
1569 :
1570 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1571 : /// compaction tasks.
1572 364 : pub(crate) async fn compact(
1573 364 : self: &Arc<Self>,
1574 364 : cancel: &CancellationToken,
1575 364 : flags: EnumSet<CompactFlags>,
1576 364 : ctx: &RequestContext,
1577 364 : ) -> Result<bool, CompactionError> {
1578 364 : // most likely the cancellation token is from background task, but in tests it could be the
1579 364 : // request task as well.
1580 364 :
1581 364 : let prepare = async move {
1582 364 : let guard = self.compaction_lock.lock().await;
1583 :
1584 364 : let permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
1585 364 : BackgroundLoopKind::Compaction,
1586 364 : ctx,
1587 364 : )
1588 0 : .await;
1589 :
1590 364 : (guard, permit)
1591 364 : };
1592 :
1593 : // this wait probably never needs any "long time spent" logging, because we already nag if
1594 : // compaction task goes over it's period (20s) which is quite often in production.
1595 364 : let (_guard, _permit) = tokio::select! {
1596 364 : tuple = prepare => { tuple },
1597 364 : _ = self.cancel.cancelled() => return Ok(false),
1598 364 : _ = cancel.cancelled() => return Ok(false),
1599 : };
1600 :
1601 364 : let last_record_lsn = self.get_last_record_lsn();
1602 364 :
1603 364 : // Last record Lsn could be zero in case the timeline was just created
1604 364 : if !last_record_lsn.is_valid() {
1605 0 : warn!("Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}");
1606 0 : return Ok(false);
1607 364 : }
1608 364 :
1609 364 : match self.get_compaction_algorithm_settings().kind {
1610 : CompactionAlgorithm::Tiered => {
1611 0 : self.compact_tiered(cancel, ctx).await?;
1612 0 : Ok(false)
1613 : }
1614 37710 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, flags, ctx).await,
1615 : }
1616 364 : }
1617 :
1618 : /// Mutate the timeline with a [`TimelineWriter`].
1619 5133156 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1620 5133156 : TimelineWriter {
1621 5133156 : tl: self,
1622 5133156 : write_guard: self.write_lock.lock().await,
1623 : }
1624 5133156 : }
1625 :
1626 0 : pub(crate) fn activate(
1627 0 : self: &Arc<Self>,
1628 0 : parent: Arc<crate::tenant::Tenant>,
1629 0 : broker_client: BrokerClientChannel,
1630 0 : background_jobs_can_start: Option<&completion::Barrier>,
1631 0 : ctx: &RequestContext,
1632 0 : ) {
1633 0 : if self.tenant_shard_id.is_shard_zero() {
1634 0 : // Logical size is only maintained accurately on shard zero.
1635 0 : self.spawn_initial_logical_size_computation_task(ctx);
1636 0 : }
1637 0 : self.launch_wal_receiver(ctx, broker_client);
1638 0 : self.set_state(TimelineState::Active);
1639 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1640 0 : }
1641 :
1642 : /// After this function returns, there are no timeline-scoped tasks are left running.
1643 : ///
1644 : /// The preferred pattern for is:
1645 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1646 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1647 : /// go the extra mile and keep track of JoinHandles
1648 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1649 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1650 : ///
1651 : /// For legacy reasons, we still have multiple tasks spawned using
1652 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1653 : /// We refer to these as "timeline-scoped task_mgr tasks".
1654 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1655 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1656 : /// or [`task_mgr::shutdown_watcher`].
1657 : /// We want to gradually convert the code base away from these.
1658 : ///
1659 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1660 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1661 : /// ones that aren't mentioned here):
1662 : /// - [`TaskKind::TimelineDeletionWorker`]
1663 : /// - NB: also used for tenant deletion
1664 : /// - [`TaskKind::RemoteUploadTask`]`
1665 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1666 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1667 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1668 : /// - [`TaskKind::Eviction`]
1669 : /// - [`TaskKind::LayerFlushTask`]
1670 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1671 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1672 8 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1673 8 : debug_assert_current_span_has_tenant_and_timeline_id();
1674 :
1675 8 : let try_freeze_and_flush = match mode {
1676 6 : ShutdownMode::FreezeAndFlush => true,
1677 2 : ShutdownMode::Hard => false,
1678 : };
1679 :
1680 : // Regardless of whether we're going to try_freeze_and_flush
1681 : // or not, stop ingesting any more data. Walreceiver only provides
1682 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1683 : // So, only after the self.gate.close() below will we know for sure that
1684 : // no walreceiver tasks are left.
1685 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1686 : // data during the call to `self.freeze_and_flush()` below.
1687 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1688 : // which is what we'd need to properly model early shutdown of the walreceiver
1689 : // task sub-tree before the other Timeline task sub-trees.
1690 8 : let walreceiver = self.walreceiver.lock().unwrap().take();
1691 8 : tracing::debug!(
1692 0 : is_some = walreceiver.is_some(),
1693 0 : "Waiting for WalReceiverManager..."
1694 : );
1695 8 : if let Some(walreceiver) = walreceiver {
1696 0 : walreceiver.cancel();
1697 8 : }
1698 : // ... and inform any waiters for newer LSNs that there won't be any.
1699 8 : self.last_record_lsn.shutdown();
1700 8 :
1701 8 : if try_freeze_and_flush {
1702 6 : if let Some((open, frozen)) = self
1703 6 : .layers
1704 6 : .read()
1705 0 : .await
1706 6 : .layer_map()
1707 6 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1708 6 : .ok()
1709 6 : .filter(|(open, frozen)| *open || *frozen > 0)
1710 : {
1711 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
1712 6 : } else {
1713 6 : // this is double-shutdown, ignore it
1714 6 : }
1715 :
1716 : // we shut down walreceiver above, so, we won't add anything more
1717 : // to the InMemoryLayer; freeze it and wait for all frozen layers
1718 : // to reach the disk & upload queue, then shut the upload queue and
1719 : // wait for it to drain.
1720 6 : match self.freeze_and_flush().await {
1721 : Ok(_) => {
1722 : // drain the upload queue
1723 : // if we did not wait for completion here, it might be our shutdown process
1724 : // didn't wait for remote uploads to complete at all, as new tasks can forever
1725 : // be spawned.
1726 : //
1727 : // what is problematic is the shutting down of RemoteTimelineClient, because
1728 : // obviously it does not make sense to stop while we wait for it, but what
1729 : // about corner cases like s3 suddenly hanging up?
1730 6 : self.remote_client.shutdown().await;
1731 : }
1732 : Err(FlushLayerError::Cancelled) => {
1733 : // this is likely the second shutdown, ignore silently.
1734 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
1735 0 : debug_assert!(self.cancel.is_cancelled());
1736 : }
1737 0 : Err(e) => {
1738 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
1739 0 : // we have some extra WAL replay to do next time the timeline starts.
1740 0 : warn!("failed to freeze and flush: {e:#}");
1741 : }
1742 : }
1743 2 : }
1744 :
1745 : // Signal any subscribers to our cancellation token to drop out
1746 8 : tracing::debug!("Cancelling CancellationToken");
1747 8 : self.cancel.cancel();
1748 8 :
1749 8 : // Ensure Prevent new page service requests from starting.
1750 8 : self.handles.shutdown();
1751 8 :
1752 8 : // Transition the remote_client into a state where it's only useful for timeline deletion.
1753 8 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
1754 8 : self.remote_client.stop();
1755 8 :
1756 8 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
1757 8 : // to shut down the upload queue tasks.
1758 8 : // TODO: fix that, task management should be encapsulated inside remote_client.
1759 8 : task_mgr::shutdown_tasks(
1760 8 : Some(TaskKind::RemoteUploadTask),
1761 8 : Some(self.tenant_shard_id),
1762 8 : Some(self.timeline_id),
1763 8 : )
1764 0 : .await;
1765 :
1766 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
1767 8 : tracing::debug!("Waiting for tasks...");
1768 8 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
1769 :
1770 : {
1771 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
1772 : // open.
1773 8 : let mut write_guard = self.write_lock.lock().await;
1774 8 : self.layers.write().await.shutdown(&mut write_guard);
1775 8 : }
1776 8 :
1777 8 : // Finally wait until any gate-holders are complete.
1778 8 : //
1779 8 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
1780 8 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
1781 8 : self.gate.close().await;
1782 :
1783 8 : self.metrics.shutdown();
1784 8 : }
1785 :
1786 410 : pub(crate) fn set_state(&self, new_state: TimelineState) {
1787 410 : match (self.current_state(), new_state) {
1788 410 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
1789 2 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
1790 : }
1791 0 : (st, TimelineState::Loading) => {
1792 0 : error!("ignoring transition from {st:?} into Loading state");
1793 : }
1794 0 : (TimelineState::Broken { .. }, new_state) => {
1795 0 : error!("Ignoring state update {new_state:?} for broken timeline");
1796 : }
1797 : (TimelineState::Stopping, TimelineState::Active) => {
1798 0 : error!("Not activating a Stopping timeline");
1799 : }
1800 408 : (_, new_state) => {
1801 408 : self.state.send_replace(new_state);
1802 408 : }
1803 : }
1804 410 : }
1805 :
1806 2 : pub(crate) fn set_broken(&self, reason: String) {
1807 2 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
1808 2 : let broken_state = TimelineState::Broken {
1809 2 : reason,
1810 2 : backtrace: backtrace_str,
1811 2 : };
1812 2 : self.set_state(broken_state);
1813 2 :
1814 2 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
1815 2 : // later when this tenant is detach or the process shuts down), firing the cancellation token
1816 2 : // here avoids the need for other tasks to watch for the Broken state explicitly.
1817 2 : self.cancel.cancel();
1818 2 : }
1819 :
1820 226369 : pub(crate) fn current_state(&self) -> TimelineState {
1821 226369 : self.state.borrow().clone()
1822 226369 : }
1823 :
1824 6 : pub(crate) fn is_broken(&self) -> bool {
1825 6 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
1826 6 : }
1827 :
1828 222 : pub(crate) fn is_active(&self) -> bool {
1829 222 : self.current_state() == TimelineState::Active
1830 222 : }
1831 :
1832 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
1833 0 : self.remote_client.is_archived()
1834 0 : }
1835 :
1836 368 : pub(crate) fn is_stopping(&self) -> bool {
1837 368 : self.current_state() == TimelineState::Stopping
1838 368 : }
1839 :
1840 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
1841 0 : self.state.subscribe()
1842 0 : }
1843 :
1844 225371 : pub(crate) async fn wait_to_become_active(
1845 225371 : &self,
1846 225371 : _ctx: &RequestContext, // Prepare for use by cancellation
1847 225371 : ) -> Result<(), TimelineState> {
1848 225371 : let mut receiver = self.state.subscribe();
1849 : loop {
1850 225371 : let current_state = receiver.borrow().clone();
1851 225371 : match current_state {
1852 : TimelineState::Loading => {
1853 0 : receiver
1854 0 : .changed()
1855 0 : .await
1856 0 : .expect("holding a reference to self");
1857 : }
1858 : TimelineState::Active { .. } => {
1859 225369 : return Ok(());
1860 : }
1861 : TimelineState::Broken { .. } | TimelineState::Stopping => {
1862 : // There's no chance the timeline can transition back into ::Active
1863 2 : return Err(current_state);
1864 : }
1865 : }
1866 : }
1867 225371 : }
1868 :
1869 0 : pub(crate) async fn layer_map_info(
1870 0 : &self,
1871 0 : reset: LayerAccessStatsReset,
1872 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
1873 0 : let guard = self.layers.read().await;
1874 0 : let layer_map = guard.layer_map()?;
1875 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
1876 0 : if let Some(open_layer) = &layer_map.open_layer {
1877 0 : in_memory_layers.push(open_layer.info());
1878 0 : }
1879 0 : for frozen_layer in &layer_map.frozen_layers {
1880 0 : in_memory_layers.push(frozen_layer.info());
1881 0 : }
1882 :
1883 0 : let historic_layers = layer_map
1884 0 : .iter_historic_layers()
1885 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
1886 0 : .collect();
1887 0 :
1888 0 : Ok(LayerMapInfo {
1889 0 : in_memory_layers,
1890 0 : historic_layers,
1891 0 : })
1892 0 : }
1893 :
1894 0 : #[instrument(skip_all, fields(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id))]
1895 : pub(crate) async fn download_layer(
1896 : &self,
1897 : layer_file_name: &LayerName,
1898 : ) -> anyhow::Result<Option<bool>> {
1899 : let Some(layer) = self.find_layer(layer_file_name).await? else {
1900 : return Ok(None);
1901 : };
1902 :
1903 : layer.download().await?;
1904 :
1905 : Ok(Some(true))
1906 : }
1907 :
1908 : /// Evict just one layer.
1909 : ///
1910 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
1911 0 : pub(crate) async fn evict_layer(
1912 0 : &self,
1913 0 : layer_file_name: &LayerName,
1914 0 : ) -> anyhow::Result<Option<bool>> {
1915 0 : let _gate = self
1916 0 : .gate
1917 0 : .enter()
1918 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
1919 :
1920 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
1921 0 : return Ok(None);
1922 : };
1923 :
1924 : // curl has this by default
1925 0 : let timeout = std::time::Duration::from_secs(120);
1926 0 :
1927 0 : match local_layer.evict_and_wait(timeout).await {
1928 0 : Ok(()) => Ok(Some(true)),
1929 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
1930 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
1931 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
1932 : }
1933 0 : }
1934 :
1935 4803010 : fn should_roll(
1936 4803010 : &self,
1937 4803010 : layer_size: u64,
1938 4803010 : projected_layer_size: u64,
1939 4803010 : checkpoint_distance: u64,
1940 4803010 : projected_lsn: Lsn,
1941 4803010 : last_freeze_at: Lsn,
1942 4803010 : opened_at: Instant,
1943 4803010 : ) -> bool {
1944 4803010 : let distance = projected_lsn.widening_sub(last_freeze_at);
1945 4803010 :
1946 4803010 : // Rolling the open layer can be triggered by:
1947 4803010 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
1948 4803010 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
1949 4803010 : // account for how writes are distributed across shards: we expect each node to consume
1950 4803010 : // 1/count of the LSN on average.
1951 4803010 : // 2. The size of the currently open layer.
1952 4803010 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
1953 4803010 : // up and suspend activity.
1954 4803010 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
1955 0 : info!(
1956 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
1957 : projected_lsn, layer_size, distance
1958 : );
1959 :
1960 0 : true
1961 4803010 : } else if projected_layer_size >= checkpoint_distance {
1962 : // NB: this check is relied upon by:
1963 80 : let _ = IndexEntry::validate_checkpoint_distance;
1964 80 : info!(
1965 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
1966 : projected_lsn, layer_size, projected_layer_size
1967 : );
1968 :
1969 80 : true
1970 4802930 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
1971 0 : info!(
1972 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
1973 0 : projected_lsn,
1974 0 : layer_size,
1975 0 : opened_at.elapsed()
1976 : );
1977 :
1978 0 : true
1979 : } else {
1980 4802930 : false
1981 : }
1982 4803010 : }
1983 : }
1984 :
1985 : /// Number of times we will compute partition within a checkpoint distance.
1986 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
1987 :
1988 : // Private functions
1989 : impl Timeline {
1990 12 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
1991 12 : let tenant_conf = self.tenant_conf.load();
1992 12 : tenant_conf
1993 12 : .tenant_conf
1994 12 : .lsn_lease_length
1995 12 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
1996 12 : }
1997 :
1998 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
1999 0 : let tenant_conf = self.tenant_conf.load();
2000 0 : tenant_conf
2001 0 : .tenant_conf
2002 0 : .lsn_lease_length_for_ts
2003 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2004 0 : }
2005 :
2006 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2007 0 : let tenant_conf = self.tenant_conf.load();
2008 0 : tenant_conf
2009 0 : .tenant_conf
2010 0 : .lazy_slru_download
2011 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2012 0 : }
2013 :
2014 4804596 : fn get_checkpoint_distance(&self) -> u64 {
2015 4804596 : let tenant_conf = self.tenant_conf.load();
2016 4804596 : tenant_conf
2017 4804596 : .tenant_conf
2018 4804596 : .checkpoint_distance
2019 4804596 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2020 4804596 : }
2021 :
2022 4802930 : fn get_checkpoint_timeout(&self) -> Duration {
2023 4802930 : let tenant_conf = self.tenant_conf.load();
2024 4802930 : tenant_conf
2025 4802930 : .tenant_conf
2026 4802930 : .checkpoint_timeout
2027 4802930 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2028 4802930 : }
2029 :
2030 584 : fn get_compaction_target_size(&self) -> u64 {
2031 584 : let tenant_conf = self.tenant_conf.load();
2032 584 : tenant_conf
2033 584 : .tenant_conf
2034 584 : .compaction_target_size
2035 584 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2036 584 : }
2037 :
2038 392 : fn get_compaction_threshold(&self) -> usize {
2039 392 : let tenant_conf = self.tenant_conf.load();
2040 392 : tenant_conf
2041 392 : .tenant_conf
2042 392 : .compaction_threshold
2043 392 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2044 392 : }
2045 :
2046 14 : fn get_image_creation_threshold(&self) -> usize {
2047 14 : let tenant_conf = self.tenant_conf.load();
2048 14 : tenant_conf
2049 14 : .tenant_conf
2050 14 : .image_creation_threshold
2051 14 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2052 14 : }
2053 :
2054 364 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2055 364 : let tenant_conf = &self.tenant_conf.load();
2056 364 : tenant_conf
2057 364 : .tenant_conf
2058 364 : .compaction_algorithm
2059 364 : .as_ref()
2060 364 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2061 364 : .clone()
2062 364 : }
2063 :
2064 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2065 0 : let tenant_conf = self.tenant_conf.load();
2066 0 : tenant_conf
2067 0 : .tenant_conf
2068 0 : .eviction_policy
2069 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2070 0 : }
2071 :
2072 410 : fn get_evictions_low_residence_duration_metric_threshold(
2073 410 : tenant_conf: &TenantConfOpt,
2074 410 : default_tenant_conf: &TenantConf,
2075 410 : ) -> Duration {
2076 410 : tenant_conf
2077 410 : .evictions_low_residence_duration_metric_threshold
2078 410 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2079 410 : }
2080 :
2081 704 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2082 704 : let tenant_conf = self.tenant_conf.load();
2083 704 : tenant_conf
2084 704 : .tenant_conf
2085 704 : .image_layer_creation_check_threshold
2086 704 : .unwrap_or(
2087 704 : self.conf
2088 704 : .default_tenant_conf
2089 704 : .image_layer_creation_check_threshold,
2090 704 : )
2091 704 : }
2092 :
2093 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2094 0 : // NB: Most tenant conf options are read by background loops, so,
2095 0 : // changes will automatically be picked up.
2096 0 :
2097 0 : // The threshold is embedded in the metric. So, we need to update it.
2098 0 : {
2099 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2100 0 : new_conf,
2101 0 : &self.conf.default_tenant_conf,
2102 0 : );
2103 0 :
2104 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2105 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2106 0 :
2107 0 : let timeline_id_str = self.timeline_id.to_string();
2108 0 : self.metrics
2109 0 : .evictions_with_low_residence_duration
2110 0 : .write()
2111 0 : .unwrap()
2112 0 : .change_threshold(
2113 0 : &tenant_id_str,
2114 0 : &shard_id_str,
2115 0 : &timeline_id_str,
2116 0 : new_threshold,
2117 0 : );
2118 0 : }
2119 0 : }
2120 :
2121 : /// Open a Timeline handle.
2122 : ///
2123 : /// Loads the metadata for the timeline into memory, but not the layer map.
2124 : #[allow(clippy::too_many_arguments)]
2125 410 : pub(super) fn new(
2126 410 : conf: &'static PageServerConf,
2127 410 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2128 410 : metadata: &TimelineMetadata,
2129 410 : ancestor: Option<Arc<Timeline>>,
2130 410 : timeline_id: TimelineId,
2131 410 : tenant_shard_id: TenantShardId,
2132 410 : generation: Generation,
2133 410 : shard_identity: ShardIdentity,
2134 410 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2135 410 : resources: TimelineResources,
2136 410 : pg_version: u32,
2137 410 : state: TimelineState,
2138 410 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2139 410 : cancel: CancellationToken,
2140 410 : ) -> Arc<Self> {
2141 410 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2142 410 : let (state, _) = watch::channel(state);
2143 410 :
2144 410 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2145 410 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2146 410 :
2147 410 : let evictions_low_residence_duration_metric_threshold = {
2148 410 : let loaded_tenant_conf = tenant_conf.load();
2149 410 : Self::get_evictions_low_residence_duration_metric_threshold(
2150 410 : &loaded_tenant_conf.tenant_conf,
2151 410 : &conf.default_tenant_conf,
2152 410 : )
2153 : };
2154 :
2155 410 : if let Some(ancestor) = &ancestor {
2156 228 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2157 228 : // If we construct an explicit timeline object, it's obviously not offloaded
2158 228 : let is_offloaded = MaybeOffloaded::No;
2159 228 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2160 228 : }
2161 :
2162 410 : Arc::new_cyclic(|myself| {
2163 410 : let metrics = TimelineMetrics::new(
2164 410 : &tenant_shard_id,
2165 410 : &timeline_id,
2166 410 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2167 410 : "mtime",
2168 410 : evictions_low_residence_duration_metric_threshold,
2169 410 : ),
2170 410 : );
2171 410 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2172 :
2173 410 : let mut result = Timeline {
2174 410 : conf,
2175 410 : tenant_conf,
2176 410 : myself: myself.clone(),
2177 410 : timeline_id,
2178 410 : tenant_shard_id,
2179 410 : generation,
2180 410 : shard_identity,
2181 410 : pg_version,
2182 410 : layers: Default::default(),
2183 410 :
2184 410 : walredo_mgr,
2185 410 : walreceiver: Mutex::new(None),
2186 410 :
2187 410 : remote_client: Arc::new(resources.remote_client),
2188 410 :
2189 410 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2190 410 : last_record_lsn: SeqWait::new(RecordLsn {
2191 410 : last: disk_consistent_lsn,
2192 410 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2193 410 : }),
2194 410 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2195 410 :
2196 410 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2197 410 : last_freeze_ts: RwLock::new(Instant::now()),
2198 410 :
2199 410 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2200 410 :
2201 410 : ancestor_timeline: ancestor,
2202 410 : ancestor_lsn: metadata.ancestor_lsn(),
2203 410 :
2204 410 : metrics,
2205 410 :
2206 410 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2207 410 : &tenant_shard_id,
2208 410 : &timeline_id,
2209 410 : ),
2210 410 :
2211 2870 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2212 410 :
2213 410 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2214 410 :
2215 410 : layer_flush_start_tx,
2216 410 : layer_flush_done_tx,
2217 410 :
2218 410 : write_lock: tokio::sync::Mutex::new(None),
2219 410 :
2220 410 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2221 410 :
2222 410 : latest_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2223 410 : initdb_lsn: metadata.initdb_lsn(),
2224 410 :
2225 410 : current_logical_size: if disk_consistent_lsn.is_valid() {
2226 : // we're creating timeline data with some layer files existing locally,
2227 : // need to recalculate timeline's logical size based on data in the layers.
2228 232 : LogicalSize::deferred_initial(disk_consistent_lsn)
2229 : } else {
2230 : // we're creating timeline data without any layers existing locally,
2231 : // initial logical size is 0.
2232 178 : LogicalSize::empty_initial()
2233 : },
2234 410 : partitioning: tokio::sync::Mutex::new((
2235 410 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2236 410 : Lsn(0),
2237 410 : )),
2238 410 : repartition_threshold: 0,
2239 410 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2240 410 : last_image_layer_creation_check_instant: Mutex::new(None),
2241 410 :
2242 410 : last_received_wal: Mutex::new(None),
2243 410 : rel_size_cache: RwLock::new(RelSizeCache {
2244 410 : complete_as_of: disk_consistent_lsn,
2245 410 : map: HashMap::new(),
2246 410 : }),
2247 410 :
2248 410 : download_all_remote_layers_task_info: RwLock::new(None),
2249 410 :
2250 410 : state,
2251 410 :
2252 410 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2253 410 : EvictionTaskTimelineState::default(),
2254 410 : ),
2255 410 : delete_progress: TimelineDeleteProgress::default(),
2256 410 :
2257 410 : cancel,
2258 410 : gate: Gate::default(),
2259 410 :
2260 410 : compaction_lock: tokio::sync::Mutex::default(),
2261 410 : gc_lock: tokio::sync::Mutex::default(),
2262 410 :
2263 410 : standby_horizon: AtomicLsn::new(0),
2264 410 :
2265 410 : timeline_get_throttle: resources.timeline_get_throttle,
2266 410 :
2267 410 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2268 410 :
2269 410 : #[cfg(test)]
2270 410 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2271 410 :
2272 410 : l0_flush_global_state: resources.l0_flush_global_state,
2273 410 :
2274 410 : handles: Default::default(),
2275 410 :
2276 410 : attach_wal_lag_cooldown,
2277 410 : };
2278 410 :
2279 410 : result.repartition_threshold =
2280 410 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2281 410 :
2282 410 : result
2283 410 : .metrics
2284 410 : .last_record_gauge
2285 410 : .set(disk_consistent_lsn.0 as i64);
2286 410 : result
2287 410 : })
2288 410 : }
2289 :
2290 574 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2291 574 : let Ok(guard) = self.gate.enter() else {
2292 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2293 0 : return;
2294 : };
2295 574 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2296 574 : match *flush_loop_state {
2297 404 : FlushLoopState::NotStarted => (),
2298 : FlushLoopState::Running { .. } => {
2299 170 : info!(
2300 0 : "skipping attempt to start flush_loop twice {}/{}",
2301 0 : self.tenant_shard_id, self.timeline_id
2302 : );
2303 170 : return;
2304 : }
2305 : FlushLoopState::Exited => {
2306 0 : warn!(
2307 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2308 0 : self.tenant_shard_id, self.timeline_id
2309 : );
2310 0 : return;
2311 : }
2312 : }
2313 :
2314 404 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2315 404 : let self_clone = Arc::clone(self);
2316 404 :
2317 404 : debug!("spawning flush loop");
2318 404 : *flush_loop_state = FlushLoopState::Running {
2319 404 : #[cfg(test)]
2320 404 : expect_initdb_optimization: false,
2321 404 : #[cfg(test)]
2322 404 : initdb_optimization_count: 0,
2323 404 : };
2324 404 : task_mgr::spawn(
2325 404 : task_mgr::BACKGROUND_RUNTIME.handle(),
2326 404 : task_mgr::TaskKind::LayerFlushTask,
2327 404 : self.tenant_shard_id,
2328 404 : Some(self.timeline_id),
2329 404 : "layer flush task",
2330 404 : async move {
2331 404 : let _guard = guard;
2332 404 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error);
2333 17929 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2334 8 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2335 8 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2336 8 : *flush_loop_state = FlushLoopState::Exited;
2337 8 : Ok(())
2338 8 : }
2339 404 : .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))
2340 : );
2341 574 : }
2342 :
2343 : /// Creates and starts the wal receiver.
2344 : ///
2345 : /// This function is expected to be called at most once per Timeline's lifecycle
2346 : /// when the timeline is activated.
2347 0 : fn launch_wal_receiver(
2348 0 : self: &Arc<Self>,
2349 0 : ctx: &RequestContext,
2350 0 : broker_client: BrokerClientChannel,
2351 0 : ) {
2352 0 : info!(
2353 0 : "launching WAL receiver for timeline {} of tenant {}",
2354 0 : self.timeline_id, self.tenant_shard_id
2355 : );
2356 :
2357 0 : let tenant_conf = self.tenant_conf.load();
2358 0 : let wal_connect_timeout = tenant_conf
2359 0 : .tenant_conf
2360 0 : .walreceiver_connect_timeout
2361 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2362 0 : let lagging_wal_timeout = tenant_conf
2363 0 : .tenant_conf
2364 0 : .lagging_wal_timeout
2365 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2366 0 : let max_lsn_wal_lag = tenant_conf
2367 0 : .tenant_conf
2368 0 : .max_lsn_wal_lag
2369 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2370 0 :
2371 0 : let mut guard = self.walreceiver.lock().unwrap();
2372 0 : assert!(
2373 0 : guard.is_none(),
2374 0 : "multiple launches / re-launches of WAL receiver are not supported"
2375 : );
2376 0 : *guard = Some(WalReceiver::start(
2377 0 : Arc::clone(self),
2378 0 : WalReceiverConf {
2379 0 : wal_connect_timeout,
2380 0 : lagging_wal_timeout,
2381 0 : max_lsn_wal_lag,
2382 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2383 0 : availability_zone: self.conf.availability_zone.clone(),
2384 0 : ingest_batch_size: self.conf.ingest_batch_size,
2385 0 : },
2386 0 : broker_client,
2387 0 : ctx,
2388 0 : ));
2389 0 : }
2390 :
2391 : /// Initialize with an empty layer map. Used when creating a new timeline.
2392 404 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2393 404 : let mut layers = self.layers.try_write().expect(
2394 404 : "in the context where we call this function, no other task has access to the object",
2395 404 : );
2396 404 : layers
2397 404 : .open_mut()
2398 404 : .expect("in this context the LayerManager must still be open")
2399 404 : .initialize_empty(Lsn(start_lsn.0));
2400 404 : }
2401 :
2402 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2403 : /// files.
2404 6 : pub(super) async fn load_layer_map(
2405 6 : &self,
2406 6 : disk_consistent_lsn: Lsn,
2407 6 : index_part: IndexPart,
2408 6 : ) -> anyhow::Result<()> {
2409 : use init::{Decision::*, Discovered, DismissedLayer};
2410 : use LayerName::*;
2411 :
2412 6 : let mut guard = self.layers.write().await;
2413 :
2414 6 : let timer = self.metrics.load_layer_map_histo.start_timer();
2415 6 :
2416 6 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2417 6 : // structs representing all files on disk
2418 6 : let timeline_path = self
2419 6 : .conf
2420 6 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2421 6 : let conf = self.conf;
2422 6 : let span = tracing::Span::current();
2423 6 :
2424 6 : // Copy to move into the task we're about to spawn
2425 6 : let this = self.myself.upgrade().expect("&self method holds the arc");
2426 :
2427 6 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
2428 6 : move || {
2429 6 : let _g = span.entered();
2430 6 : let discovered = init::scan_timeline_dir(&timeline_path)?;
2431 6 : let mut discovered_layers = Vec::with_capacity(discovered.len());
2432 6 : let mut unrecognized_files = Vec::new();
2433 6 :
2434 6 : let mut path = timeline_path;
2435 :
2436 22 : for discovered in discovered {
2437 16 : let (name, kind) = match discovered {
2438 16 : Discovered::Layer(layer_file_name, local_metadata) => {
2439 16 : discovered_layers.push((layer_file_name, local_metadata));
2440 16 : continue;
2441 : }
2442 0 : Discovered::IgnoredBackup(path) => {
2443 0 : std::fs::remove_file(path)
2444 0 : .or_else(fs_ext::ignore_not_found)
2445 0 : .fatal_err("Removing .old file");
2446 0 : continue;
2447 : }
2448 0 : Discovered::Unknown(file_name) => {
2449 0 : // we will later error if there are any
2450 0 : unrecognized_files.push(file_name);
2451 0 : continue;
2452 : }
2453 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
2454 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
2455 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
2456 : };
2457 0 : path.push(Utf8Path::new(&name));
2458 0 : init::cleanup(&path, kind)?;
2459 0 : path.pop();
2460 : }
2461 :
2462 6 : if !unrecognized_files.is_empty() {
2463 : // assume that if there are any there are many many.
2464 0 : let n = unrecognized_files.len();
2465 0 : let first = &unrecognized_files[..n.min(10)];
2466 0 : anyhow::bail!(
2467 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
2468 0 : );
2469 6 : }
2470 6 :
2471 6 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
2472 6 :
2473 6 : let mut loaded_layers = Vec::new();
2474 6 : let mut needs_cleanup = Vec::new();
2475 6 : let mut total_physical_size = 0;
2476 :
2477 22 : for (name, decision) in decided {
2478 16 : let decision = match decision {
2479 16 : Ok(decision) => decision,
2480 0 : Err(DismissedLayer::Future { local }) => {
2481 0 : if let Some(local) = local {
2482 0 : init::cleanup_future_layer(
2483 0 : &local.local_path,
2484 0 : &name,
2485 0 : disk_consistent_lsn,
2486 0 : )?;
2487 0 : }
2488 0 : needs_cleanup.push(name);
2489 0 : continue;
2490 : }
2491 0 : Err(DismissedLayer::LocalOnly(local)) => {
2492 0 : init::cleanup_local_only_file(&name, &local)?;
2493 : // this file never existed remotely, we will have to do rework
2494 0 : continue;
2495 : }
2496 0 : Err(DismissedLayer::BadMetadata(local)) => {
2497 0 : init::cleanup_local_file_for_remote(&local)?;
2498 : // this file never existed remotely, we will have to do rework
2499 0 : continue;
2500 : }
2501 : };
2502 :
2503 16 : match &name {
2504 12 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
2505 4 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
2506 : }
2507 :
2508 16 : tracing::debug!(layer=%name, ?decision, "applied");
2509 :
2510 16 : let layer = match decision {
2511 16 : Resident { local, remote } => {
2512 16 : total_physical_size += local.file_size;
2513 16 : Layer::for_resident(conf, &this, local.local_path, name, remote)
2514 16 : .drop_eviction_guard()
2515 : }
2516 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
2517 : };
2518 :
2519 16 : loaded_layers.push(layer);
2520 : }
2521 6 : Ok((loaded_layers, needs_cleanup, total_physical_size))
2522 6 : }
2523 6 : })
2524 6 : .await
2525 6 : .map_err(anyhow::Error::new)
2526 6 : .and_then(|x| x)?;
2527 :
2528 6 : let num_layers = loaded_layers.len();
2529 6 :
2530 6 : guard
2531 6 : .open_mut()
2532 6 : .expect("layermanager must be open during init")
2533 6 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
2534 6 :
2535 6 : self.remote_client
2536 6 : .schedule_layer_file_deletion(&needs_cleanup)?;
2537 6 : self.remote_client
2538 6 : .schedule_index_upload_for_file_changes()?;
2539 : // This barrier orders above DELETEs before any later operations.
2540 : // This is critical because code executing after the barrier might
2541 : // create again objects with the same key that we just scheduled for deletion.
2542 : // For example, if we just scheduled deletion of an image layer "from the future",
2543 : // later compaction might run again and re-create the same image layer.
2544 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
2545 : // "same" here means same key range and LSN.
2546 : //
2547 : // Without a barrier between above DELETEs and the re-creation's PUTs,
2548 : // the upload queue may execute the PUT first, then the DELETE.
2549 : // In our example, we will end up with an IndexPart referencing a non-existent object.
2550 : //
2551 : // 1. a future image layer is created and uploaded
2552 : // 2. ps restart
2553 : // 3. the future layer from (1) is deleted during load layer map
2554 : // 4. image layer is re-created and uploaded
2555 : // 5. deletion queue would like to delete (1) but actually deletes (4)
2556 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
2557 : //
2558 : // See https://github.com/neondatabase/neon/issues/5878
2559 : //
2560 : // NB: generation numbers naturally protect against this because they disambiguate
2561 : // (1) and (4)
2562 6 : self.remote_client.schedule_barrier()?;
2563 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
2564 : // on retry.
2565 :
2566 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
2567 6 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
2568 6 : self.update_layer_visibility().await?;
2569 :
2570 6 : info!(
2571 0 : "loaded layer map with {} layers at {}, total physical size: {}",
2572 : num_layers, disk_consistent_lsn, total_physical_size
2573 : );
2574 :
2575 6 : timer.stop_and_record();
2576 6 : Ok(())
2577 6 : }
2578 :
2579 : /// Retrieve current logical size of the timeline.
2580 : ///
2581 : /// The size could be lagging behind the actual number, in case
2582 : /// the initial size calculation has not been run (gets triggered on the first size access).
2583 : ///
2584 : /// return size and boolean flag that shows if the size is exact
2585 0 : pub(crate) fn get_current_logical_size(
2586 0 : self: &Arc<Self>,
2587 0 : priority: GetLogicalSizePriority,
2588 0 : ctx: &RequestContext,
2589 0 : ) -> logical_size::CurrentLogicalSize {
2590 0 : if !self.tenant_shard_id.is_shard_zero() {
2591 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
2592 : // when HTTP API is serving a GET for timeline zero, return zero
2593 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
2594 0 : }
2595 0 :
2596 0 : let current_size = self.current_logical_size.current_size();
2597 0 : debug!("Current size: {current_size:?}");
2598 :
2599 0 : match (current_size.accuracy(), priority) {
2600 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
2601 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
2602 0 : // background task will eventually deliver an exact value, we're in no rush
2603 0 : }
2604 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
2605 : // background task is not ready, but user is asking for it now;
2606 : // => make the background task skip the line
2607 : // (The alternative would be to calculate the size here, but,
2608 : // it can actually take a long time if the user has a lot of rels.
2609 : // And we'll inevitable need it again; So, let the background task do the work.)
2610 0 : match self
2611 0 : .current_logical_size
2612 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
2613 0 : .get()
2614 : {
2615 0 : Some(cancel) => cancel.cancel(),
2616 : None => {
2617 0 : let state = self.current_state();
2618 0 : if matches!(
2619 0 : state,
2620 : TimelineState::Broken { .. } | TimelineState::Stopping
2621 0 : ) {
2622 0 :
2623 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
2624 0 : // Don't make noise.
2625 0 : } else {
2626 0 : warn!("unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work");
2627 0 : debug_assert!(false);
2628 : }
2629 : }
2630 : };
2631 : }
2632 : }
2633 :
2634 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
2635 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
2636 0 : let first = self
2637 0 : .current_logical_size
2638 0 : .did_return_approximate_to_walreceiver
2639 0 : .compare_exchange(
2640 0 : false,
2641 0 : true,
2642 0 : AtomicOrdering::Relaxed,
2643 0 : AtomicOrdering::Relaxed,
2644 0 : )
2645 0 : .is_ok();
2646 0 : if first {
2647 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
2648 0 : }
2649 0 : }
2650 0 : }
2651 :
2652 0 : current_size
2653 0 : }
2654 :
2655 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
2656 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
2657 : // nothing to do for freshly created timelines;
2658 0 : assert_eq!(
2659 0 : self.current_logical_size.current_size().accuracy(),
2660 0 : logical_size::Accuracy::Exact,
2661 0 : );
2662 0 : self.current_logical_size.initialized.add_permits(1);
2663 0 : return;
2664 : };
2665 :
2666 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
2667 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
2668 0 : self.current_logical_size
2669 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
2670 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
2671 0 :
2672 0 : let self_clone = Arc::clone(self);
2673 0 : let background_ctx = ctx.detached_child(
2674 0 : TaskKind::InitialLogicalSizeCalculation,
2675 0 : DownloadBehavior::Download,
2676 0 : );
2677 0 : task_mgr::spawn(
2678 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2679 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
2680 0 : self.tenant_shard_id,
2681 0 : Some(self.timeline_id),
2682 0 : "initial size calculation",
2683 : // NB: don't log errors here, task_mgr will do that.
2684 0 : async move {
2685 0 : let cancel = task_mgr::shutdown_token();
2686 0 : self_clone
2687 0 : .initial_logical_size_calculation_task(
2688 0 : initial_part_end,
2689 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
2690 0 : cancel,
2691 0 : background_ctx,
2692 0 : )
2693 0 : .await;
2694 0 : Ok(())
2695 0 : }
2696 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)),
2697 : );
2698 0 : }
2699 :
2700 0 : async fn initial_logical_size_calculation_task(
2701 0 : self: Arc<Self>,
2702 0 : initial_part_end: Lsn,
2703 0 : skip_concurrency_limiter: CancellationToken,
2704 0 : cancel: CancellationToken,
2705 0 : background_ctx: RequestContext,
2706 0 : ) {
2707 0 : scopeguard::defer! {
2708 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
2709 0 : self.current_logical_size.initialized.add_permits(1);
2710 0 : }
2711 0 :
2712 0 : let try_once = |attempt: usize| {
2713 0 : let background_ctx = &background_ctx;
2714 0 : let self_ref = &self;
2715 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
2716 0 : async move {
2717 0 : let cancel = task_mgr::shutdown_token();
2718 0 : let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
2719 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
2720 0 : background_ctx,
2721 0 : );
2722 :
2723 : use crate::metrics::initial_logical_size::StartCircumstances;
2724 0 : let (_maybe_permit, circumstances) = tokio::select! {
2725 0 : permit = wait_for_permit => {
2726 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
2727 : }
2728 0 : _ = self_ref.cancel.cancelled() => {
2729 0 : return Err(CalculateLogicalSizeError::Cancelled);
2730 : }
2731 0 : _ = cancel.cancelled() => {
2732 0 : return Err(CalculateLogicalSizeError::Cancelled);
2733 : },
2734 0 : () = skip_concurrency_limiter.cancelled() => {
2735 : // Some action that is part of a end user interaction requested logical size
2736 : // => break out of the rate limit
2737 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
2738 : // but then again what happens if they cancel; also, we should just be using
2739 : // one runtime across the entire process, so, let's leave this for now.
2740 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
2741 : }
2742 : };
2743 :
2744 0 : let metrics_guard = if attempt == 1 {
2745 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
2746 : } else {
2747 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
2748 : };
2749 :
2750 0 : let calculated_size = self_ref
2751 0 : .logical_size_calculation_task(
2752 0 : initial_part_end,
2753 0 : LogicalSizeCalculationCause::Initial,
2754 0 : background_ctx,
2755 0 : )
2756 0 : .await?;
2757 :
2758 0 : self_ref
2759 0 : .trigger_aux_file_size_computation(initial_part_end, background_ctx)
2760 0 : .await?;
2761 :
2762 : // TODO: add aux file size to logical size
2763 :
2764 0 : Ok((calculated_size, metrics_guard))
2765 0 : }
2766 0 : };
2767 :
2768 0 : let retrying = async {
2769 0 : let mut attempt = 0;
2770 : loop {
2771 0 : attempt += 1;
2772 0 :
2773 0 : match try_once(attempt).await {
2774 0 : Ok(res) => return ControlFlow::Continue(res),
2775 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
2776 : Err(
2777 0 : e @ (CalculateLogicalSizeError::Decode(_)
2778 0 : | CalculateLogicalSizeError::PageRead(_)),
2779 0 : ) => {
2780 0 : warn!(attempt, "initial size calculation failed: {e:?}");
2781 : // exponential back-off doesn't make sense at these long intervals;
2782 : // use fixed retry interval with generous jitter instead
2783 0 : let sleep_duration = Duration::from_secs(
2784 0 : u64::try_from(
2785 0 : // 1hour base
2786 0 : (60_i64 * 60_i64)
2787 0 : // 10min jitter
2788 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
2789 0 : )
2790 0 : .expect("10min < 1hour"),
2791 0 : );
2792 0 : tokio::time::sleep(sleep_duration).await;
2793 : }
2794 : }
2795 : }
2796 0 : };
2797 :
2798 0 : let (calculated_size, metrics_guard) = tokio::select! {
2799 0 : res = retrying => {
2800 0 : match res {
2801 0 : ControlFlow::Continue(calculated_size) => calculated_size,
2802 0 : ControlFlow::Break(()) => return,
2803 : }
2804 : }
2805 0 : _ = cancel.cancelled() => {
2806 0 : return;
2807 : }
2808 : };
2809 :
2810 : // we cannot query current_logical_size.current_size() to know the current
2811 : // *negative* value, only truncated to u64.
2812 0 : let added = self
2813 0 : .current_logical_size
2814 0 : .size_added_after_initial
2815 0 : .load(AtomicOrdering::Relaxed);
2816 0 :
2817 0 : let sum = calculated_size.saturating_add_signed(added);
2818 0 :
2819 0 : // set the gauge value before it can be set in `update_current_logical_size`.
2820 0 : self.metrics.current_logical_size_gauge.set(sum);
2821 0 :
2822 0 : self.current_logical_size
2823 0 : .initial_logical_size
2824 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
2825 0 : .ok()
2826 0 : .expect("only this task sets it");
2827 0 : }
2828 :
2829 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
2830 0 : self: &Arc<Self>,
2831 0 : lsn: Lsn,
2832 0 : cause: LogicalSizeCalculationCause,
2833 0 : ctx: RequestContext,
2834 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
2835 0 : let (sender, receiver) = oneshot::channel();
2836 0 : let self_clone = Arc::clone(self);
2837 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
2838 0 : // we should stop the size calculation work and return an error.
2839 0 : // That would require restructuring this function's API to
2840 0 : // return the result directly, instead of a Receiver for the result.
2841 0 : let ctx = ctx.detached_child(
2842 0 : TaskKind::OndemandLogicalSizeCalculation,
2843 0 : DownloadBehavior::Download,
2844 0 : );
2845 0 : task_mgr::spawn(
2846 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
2847 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
2848 0 : self.tenant_shard_id,
2849 0 : Some(self.timeline_id),
2850 0 : "ondemand logical size calculation",
2851 0 : async move {
2852 0 : let res = self_clone
2853 0 : .logical_size_calculation_task(lsn, cause, &ctx)
2854 0 : .await;
2855 0 : let _ = sender.send(res).ok();
2856 0 : Ok(()) // Receiver is responsible for handling errors
2857 0 : }
2858 0 : .in_current_span(),
2859 0 : );
2860 0 : receiver
2861 0 : }
2862 :
2863 : /// # Cancel-Safety
2864 : ///
2865 : /// This method is cancellation-safe.
2866 0 : #[instrument(skip_all)]
2867 : async fn logical_size_calculation_task(
2868 : self: &Arc<Self>,
2869 : lsn: Lsn,
2870 : cause: LogicalSizeCalculationCause,
2871 : ctx: &RequestContext,
2872 : ) -> Result<u64, CalculateLogicalSizeError> {
2873 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
2874 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
2875 : // accurate relation sizes, and they do not emit consumption metrics.
2876 : debug_assert!(self.tenant_shard_id.is_shard_zero());
2877 :
2878 : let guard = self
2879 : .gate
2880 : .enter()
2881 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
2882 :
2883 : let self_calculation = Arc::clone(self);
2884 :
2885 0 : let mut calculation = pin!(async {
2886 0 : let ctx = ctx.attached_child();
2887 0 : self_calculation
2888 0 : .calculate_logical_size(lsn, cause, &guard, &ctx)
2889 0 : .await
2890 0 : });
2891 :
2892 : tokio::select! {
2893 : res = &mut calculation => { res }
2894 : _ = self.cancel.cancelled() => {
2895 : debug!("cancelling logical size calculation for timeline shutdown");
2896 : calculation.await
2897 : }
2898 : }
2899 : }
2900 :
2901 : /// Calculate the logical size of the database at the latest LSN.
2902 : ///
2903 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
2904 : /// especially if we need to download remote layers.
2905 : ///
2906 : /// # Cancel-Safety
2907 : ///
2908 : /// This method is cancellation-safe.
2909 0 : async fn calculate_logical_size(
2910 0 : &self,
2911 0 : up_to_lsn: Lsn,
2912 0 : cause: LogicalSizeCalculationCause,
2913 0 : _guard: &GateGuard,
2914 0 : ctx: &RequestContext,
2915 0 : ) -> Result<u64, CalculateLogicalSizeError> {
2916 0 : info!(
2917 0 : "Calculating logical size for timeline {} at {}",
2918 : self.timeline_id, up_to_lsn
2919 : );
2920 :
2921 0 : pausable_failpoint!("timeline-calculate-logical-size-pause");
2922 :
2923 : // See if we've already done the work for initial size calculation.
2924 : // This is a short-cut for timelines that are mostly unused.
2925 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
2926 0 : return Ok(size);
2927 0 : }
2928 0 : let storage_time_metrics = match cause {
2929 : LogicalSizeCalculationCause::Initial
2930 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
2931 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
2932 : LogicalSizeCalculationCause::EvictionTaskImitation => {
2933 0 : &self.metrics.imitate_logical_size_histo
2934 : }
2935 : };
2936 0 : let timer = storage_time_metrics.start_timer();
2937 0 : let logical_size = self
2938 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
2939 0 : .await?;
2940 0 : debug!("calculated logical size: {logical_size}");
2941 0 : timer.stop_and_record();
2942 0 : Ok(logical_size)
2943 0 : }
2944 :
2945 : /// Update current logical size, adding `delta' to the old value.
2946 270570 : fn update_current_logical_size(&self, delta: i64) {
2947 270570 : let logical_size = &self.current_logical_size;
2948 270570 : logical_size.increment_size(delta);
2949 270570 :
2950 270570 : // Also set the value in the prometheus gauge. Note that
2951 270570 : // there is a race condition here: if this is is called by two
2952 270570 : // threads concurrently, the prometheus gauge might be set to
2953 270570 : // one value while current_logical_size is set to the
2954 270570 : // other.
2955 270570 : match logical_size.current_size() {
2956 270570 : CurrentLogicalSize::Exact(ref new_current_size) => self
2957 270570 : .metrics
2958 270570 : .current_logical_size_gauge
2959 270570 : .set(new_current_size.into()),
2960 0 : CurrentLogicalSize::Approximate(_) => {
2961 0 : // don't update the gauge yet, this allows us not to update the gauge back and
2962 0 : // forth between the initial size calculation task.
2963 0 : }
2964 : }
2965 270570 : }
2966 :
2967 2804 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: u64) {
2968 2804 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
2969 2804 : let aux_metric =
2970 2804 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
2971 2804 :
2972 2804 : let sum_of_entries = self
2973 2804 : .directory_metrics
2974 2804 : .iter()
2975 19628 : .map(|v| v.load(AtomicOrdering::Relaxed))
2976 2804 : .sum();
2977 : // Set a high general threshold and a lower threshold for the auxiliary files,
2978 : // as we can have large numbers of relations in the db directory.
2979 : const SUM_THRESHOLD: u64 = 5000;
2980 : const AUX_THRESHOLD: u64 = 1000;
2981 2804 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
2982 0 : self.metrics
2983 0 : .directory_entries_count_gauge
2984 0 : .set(sum_of_entries);
2985 2804 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
2986 0 : metric.set(sum_of_entries);
2987 2804 : }
2988 2804 : }
2989 :
2990 0 : async fn find_layer(
2991 0 : &self,
2992 0 : layer_name: &LayerName,
2993 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
2994 0 : let guard = self.layers.read().await;
2995 0 : let layer = guard
2996 0 : .layer_map()?
2997 0 : .iter_historic_layers()
2998 0 : .find(|l| &l.layer_name() == layer_name)
2999 0 : .map(|found| guard.get_from_desc(&found));
3000 0 : Ok(layer)
3001 0 : }
3002 :
3003 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3004 : /// indicating which layers are currently on-disk on the primary.
3005 : ///
3006 : /// None is returned if the Timeline is in a state where uploading a heatmap
3007 : /// doesn't make sense, such as shutting down or initializing. The caller
3008 : /// should treat this as a cue to simply skip doing any heatmap uploading
3009 : /// for this timeline.
3010 2 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3011 2 : if !self.is_active() {
3012 0 : return None;
3013 2 : }
3014 :
3015 2 : let guard = self.layers.read().await;
3016 :
3017 10 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3018 10 : match layer.visibility() {
3019 : LayerVisibilityHint::Visible => {
3020 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3021 8 : let last_activity_ts = layer.latest_activity();
3022 8 : Some((layer.layer_desc(), layer.metadata(), last_activity_ts))
3023 : }
3024 : LayerVisibilityHint::Covered => {
3025 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3026 2 : None
3027 : }
3028 : }
3029 10 : });
3030 2 :
3031 2 : let mut layers = resident.collect::<Vec<_>>();
3032 2 :
3033 2 : // Sort layers in order of which to download first. For a large set of layers to download, we
3034 2 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3035 2 : // or hours later:
3036 2 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3037 2 : // only exist for a few minutes before being compacted into L1s.
3038 2 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3039 2 : // the layer is likely to be covered by an image layer during compaction.
3040 20 : layers.sort_by_key(|(desc, _meta, _atime)| {
3041 20 : std::cmp::Reverse((
3042 20 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3043 20 : desc.lsn_range.end,
3044 20 : ))
3045 20 : });
3046 2 :
3047 2 : let layers = layers
3048 2 : .into_iter()
3049 8 : .map(|(desc, meta, atime)| HeatMapLayer::new(desc.layer_name(), meta, atime))
3050 2 : .collect();
3051 2 :
3052 2 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3053 2 : }
3054 :
3055 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3056 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3057 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3058 0 : // branchpoint in the value in IndexPart::lineage
3059 0 : self.ancestor_lsn == lsn
3060 0 : || (self.ancestor_lsn == Lsn::INVALID
3061 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3062 0 : }
3063 : }
3064 :
3065 : impl Timeline {
3066 : #[allow(clippy::doc_lazy_continuation)]
3067 : /// Get the data needed to reconstruct all keys in the provided keyspace
3068 : ///
3069 : /// The algorithm is as follows:
3070 : /// 1. While some keys are still not done and there's a timeline to visit:
3071 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3072 : /// 2.1: Build the fringe for the current keyspace
3073 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3074 : /// intersects
3075 : /// 2.3. Pop the timeline from the fringe
3076 : /// 2.4. If the fringe is empty, go back to 1
3077 626435 : async fn get_vectored_reconstruct_data(
3078 626435 : &self,
3079 626435 : mut keyspace: KeySpace,
3080 626435 : request_lsn: Lsn,
3081 626435 : reconstruct_state: &mut ValuesReconstructState,
3082 626435 : ctx: &RequestContext,
3083 626435 : ) -> Result<(), GetVectoredError> {
3084 626435 : let mut timeline_owned: Arc<Timeline>;
3085 626435 : let mut timeline = self;
3086 626435 :
3087 626435 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3088 :
3089 626433 : let missing_keyspace = loop {
3090 851804 : if self.cancel.is_cancelled() {
3091 0 : return Err(GetVectoredError::Cancelled);
3092 851804 : }
3093 :
3094 : let TimelineVisitOutcome {
3095 851804 : completed_keyspace: completed,
3096 851804 : image_covered_keyspace,
3097 851804 : } = Self::get_vectored_reconstruct_data_timeline(
3098 851804 : timeline,
3099 851804 : keyspace.clone(),
3100 851804 : cont_lsn,
3101 851804 : reconstruct_state,
3102 851804 : &self.cancel,
3103 851804 : ctx,
3104 851804 : )
3105 186430 : .await?;
3106 :
3107 851804 : keyspace.remove_overlapping_with(&completed);
3108 851804 :
3109 851804 : // Do not descend into the ancestor timeline for aux files.
3110 851804 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3111 851804 : // stalling compaction.
3112 851804 : keyspace.remove_overlapping_with(&KeySpace {
3113 851804 : ranges: vec![NON_INHERITED_RANGE, NON_INHERITED_SPARSE_RANGE],
3114 851804 : });
3115 851804 :
3116 851804 : // Keyspace is fully retrieved
3117 851804 : if keyspace.is_empty() {
3118 626419 : break None;
3119 225385 : }
3120 :
3121 225385 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3122 : // Not fully retrieved but no ancestor timeline.
3123 14 : break Some(keyspace);
3124 : };
3125 :
3126 : // Now we see if there are keys covered by the image layer but does not exist in the
3127 : // image layer, which means that the key does not exist.
3128 :
3129 : // The block below will stop the vectored search if any of the keys encountered an image layer
3130 : // which did not contain a snapshot for said key. Since we have already removed all completed
3131 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3132 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3133 : // and stop the search as a result of that.
3134 225371 : let removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3135 225371 : if !removed.is_empty() {
3136 0 : break Some(removed);
3137 225371 : }
3138 225371 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3139 225371 : // keyspace.
3140 225371 :
3141 225371 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3142 225371 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3143 225371 : timeline_owned = timeline
3144 225371 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3145 2 : .await?;
3146 225369 : timeline = &*timeline_owned;
3147 : };
3148 :
3149 626433 : if let Some(missing_keyspace) = missing_keyspace {
3150 14 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3151 14 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3152 14 : shard: self
3153 14 : .shard_identity
3154 14 : .get_shard_number(&missing_keyspace.start().unwrap()),
3155 14 : cont_lsn,
3156 14 : request_lsn,
3157 14 : ancestor_lsn: Some(timeline.ancestor_lsn),
3158 14 : backtrace: None,
3159 14 : }));
3160 626419 : }
3161 626419 :
3162 626419 : Ok(())
3163 626435 : }
3164 :
3165 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3166 : ///
3167 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3168 : /// the current keyspace. The current keyspace of the search at any given timeline
3169 : /// is the original keyspace minus all the keys that have been completed minus
3170 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3171 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3172 : ///
3173 : /// This is basically a depth-first search visitor implementation where a vertex
3174 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3175 : ///
3176 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3177 : /// and get all the required reconstruct data from the layer in one go.
3178 : ///
3179 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3180 : /// decides how to deal with these two keyspaces.
3181 851804 : async fn get_vectored_reconstruct_data_timeline(
3182 851804 : timeline: &Timeline,
3183 851804 : keyspace: KeySpace,
3184 851804 : mut cont_lsn: Lsn,
3185 851804 : reconstruct_state: &mut ValuesReconstructState,
3186 851804 : cancel: &CancellationToken,
3187 851804 : ctx: &RequestContext,
3188 851804 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3189 851804 : let mut unmapped_keyspace = keyspace.clone();
3190 851804 : let mut fringe = LayerFringe::new();
3191 851804 :
3192 851804 : let mut completed_keyspace = KeySpace::default();
3193 851804 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3194 :
3195 : loop {
3196 1669957 : if cancel.is_cancelled() {
3197 0 : return Err(GetVectoredError::Cancelled);
3198 1669957 : }
3199 1669957 :
3200 1669957 : let (keys_done_last_step, keys_with_image_coverage) =
3201 1669957 : reconstruct_state.consume_done_keys();
3202 1669957 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3203 1669957 : completed_keyspace.merge(&keys_done_last_step);
3204 1669957 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3205 8032 : unmapped_keyspace
3206 8032 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3207 8032 : image_covered_keyspace.add_range(keys_with_image_coverage);
3208 1661925 : }
3209 :
3210 : // Do not descent any further if the last layer we visited
3211 : // completed all keys in the keyspace it inspected. This is not
3212 : // required for correctness, but avoids visiting extra layers
3213 : // which turns out to be a perf bottleneck in some cases.
3214 1669957 : if !unmapped_keyspace.is_empty() {
3215 1043722 : let guard = timeline.layers.read().await;
3216 1043722 : let layers = guard.layer_map()?;
3217 :
3218 1043722 : let in_memory_layer = layers.find_in_memory_layer(|l| {
3219 911610 : let start_lsn = l.get_lsn_range().start;
3220 911610 : cont_lsn > start_lsn
3221 1043722 : });
3222 1043722 :
3223 1043722 : match in_memory_layer {
3224 606177 : Some(l) => {
3225 606177 : let lsn_range = l.get_lsn_range().start..cont_lsn;
3226 606177 : fringe.update(
3227 606177 : ReadableLayer::InMemoryLayer(l),
3228 606177 : unmapped_keyspace.clone(),
3229 606177 : lsn_range,
3230 606177 : );
3231 606177 : }
3232 : None => {
3233 509161 : for range in unmapped_keyspace.ranges.iter() {
3234 509161 : let results = layers.range_search(range.clone(), cont_lsn);
3235 509161 :
3236 509161 : results
3237 509161 : .found
3238 509161 : .into_iter()
3239 509161 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3240 275568 : (
3241 275568 : ReadableLayer::PersistentLayer(guard.get_from_desc(&layer)),
3242 275568 : keyspace_accum.to_keyspace(),
3243 275568 : lsn_floor..cont_lsn,
3244 275568 : )
3245 509161 : })
3246 509161 : .for_each(|(layer, keyspace, lsn_range)| {
3247 275568 : fringe.update(layer, keyspace, lsn_range)
3248 509161 : });
3249 509161 : }
3250 : }
3251 : }
3252 :
3253 : // It's safe to drop the layer map lock after planning the next round of reads.
3254 : // The fringe keeps readable handles for the layers which are safe to read even
3255 : // if layers were compacted or flushed.
3256 : //
3257 : // The more interesting consideration is: "Why is the read algorithm still correct
3258 : // if the layer map changes while it is operating?". Doing a vectored read on a
3259 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3260 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3261 : // range at *a particular point in time*. It is fine for the answer to be different
3262 : // at two different time points.
3263 1043722 : drop(guard);
3264 626235 : }
3265 :
3266 1669957 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3267 818153 : let next_cont_lsn = lsn_range.start;
3268 818153 : layer_to_read
3269 818153 : .get_values_reconstruct_data(
3270 818153 : keyspace_to_read.clone(),
3271 818153 : lsn_range,
3272 818153 : reconstruct_state,
3273 818153 : ctx,
3274 818153 : )
3275 177375 : .await?;
3276 :
3277 818153 : unmapped_keyspace = keyspace_to_read;
3278 818153 : cont_lsn = next_cont_lsn;
3279 818153 :
3280 818153 : reconstruct_state.on_layer_visited(&layer_to_read);
3281 : } else {
3282 851804 : break;
3283 851804 : }
3284 851804 : }
3285 851804 :
3286 851804 : Ok(TimelineVisitOutcome {
3287 851804 : completed_keyspace,
3288 851804 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
3289 851804 : })
3290 851804 : }
3291 :
3292 225371 : async fn get_ready_ancestor_timeline(
3293 225371 : &self,
3294 225371 : ancestor: &Arc<Timeline>,
3295 225371 : ctx: &RequestContext,
3296 225371 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
3297 225371 : // It's possible that the ancestor timeline isn't active yet, or
3298 225371 : // is active but hasn't yet caught up to the branch point. Wait
3299 225371 : // for it.
3300 225371 : //
3301 225371 : // This cannot happen while the pageserver is running normally,
3302 225371 : // because you cannot create a branch from a point that isn't
3303 225371 : // present in the pageserver yet. However, we don't wait for the
3304 225371 : // branch point to be uploaded to cloud storage before creating
3305 225371 : // a branch. I.e., the branch LSN need not be remote consistent
3306 225371 : // for the branching operation to succeed.
3307 225371 : //
3308 225371 : // Hence, if we try to load a tenant in such a state where
3309 225371 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
3310 225371 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
3311 225371 : // then we will need to wait for the ancestor timeline to
3312 225371 : // re-stream WAL up to branch_lsn before we access it.
3313 225371 : //
3314 225371 : // How can a tenant get in such a state?
3315 225371 : // - ungraceful pageserver process exit
3316 225371 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
3317 225371 : //
3318 225371 : // NB: this could be avoided by requiring
3319 225371 : // branch_lsn >= remote_consistent_lsn
3320 225371 : // during branch creation.
3321 225371 : match ancestor.wait_to_become_active(ctx).await {
3322 225369 : Ok(()) => {}
3323 : Err(TimelineState::Stopping) => {
3324 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
3325 0 : return Err(GetReadyAncestorError::Cancelled);
3326 : }
3327 2 : Err(state) => {
3328 2 : return Err(GetReadyAncestorError::BadState {
3329 2 : timeline_id: ancestor.timeline_id,
3330 2 : state,
3331 2 : });
3332 : }
3333 : }
3334 225369 : ancestor
3335 225369 : .wait_lsn(self.ancestor_lsn, WaitLsnWaiter::Timeline(self), ctx)
3336 0 : .await
3337 225369 : .map_err(|e| match e {
3338 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
3339 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
3340 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
3341 0 : timeline_id: ancestor.timeline_id,
3342 0 : state,
3343 0 : },
3344 225369 : })?;
3345 :
3346 225369 : Ok(ancestor.clone())
3347 225371 : }
3348 :
3349 5452 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
3350 5452 : &self.shard_identity
3351 5452 : }
3352 :
3353 : #[inline(always)]
3354 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
3355 0 : ShardTimelineId {
3356 0 : shard_index: ShardIndex {
3357 0 : shard_number: self.shard_identity.number,
3358 0 : shard_count: self.shard_identity.count,
3359 0 : },
3360 0 : timeline_id: self.timeline_id,
3361 0 : }
3362 0 : }
3363 :
3364 : /// Returns a non-frozen open in-memory layer for ingestion.
3365 : ///
3366 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
3367 : /// this function without holding the mutex.
3368 1262 : async fn get_layer_for_write(
3369 1262 : &self,
3370 1262 : lsn: Lsn,
3371 1262 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3372 1262 : ctx: &RequestContext,
3373 1262 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
3374 1262 : let mut guard = self.layers.write().await;
3375 1262 : let gate_guard = self.gate.enter().context("enter gate for inmem layer")?;
3376 :
3377 1262 : let last_record_lsn = self.get_last_record_lsn();
3378 1262 : ensure!(
3379 1262 : lsn > last_record_lsn,
3380 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
3381 : lsn,
3382 : last_record_lsn,
3383 : );
3384 :
3385 1262 : let layer = guard
3386 1262 : .open_mut()?
3387 1262 : .get_layer_for_write(
3388 1262 : lsn,
3389 1262 : self.conf,
3390 1262 : self.timeline_id,
3391 1262 : self.tenant_shard_id,
3392 1262 : gate_guard,
3393 1262 : ctx,
3394 1262 : )
3395 713 : .await?;
3396 1262 : Ok(layer)
3397 1262 : }
3398 :
3399 5279058 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
3400 5279058 : assert!(new_lsn.is_aligned());
3401 :
3402 5279058 : self.metrics.last_record_gauge.set(new_lsn.0 as i64);
3403 5279058 : self.last_record_lsn.advance(new_lsn);
3404 5279058 : }
3405 :
3406 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
3407 : ///
3408 : /// Unconditional flush loop notification is given because in sharded cases we will want to
3409 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
3410 1166 : async fn freeze_inmem_layer_at(
3411 1166 : &self,
3412 1166 : at: Lsn,
3413 1166 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
3414 1166 : ) -> Result<u64, FlushLayerError> {
3415 1166 : let frozen = {
3416 1166 : let mut guard = self.layers.write().await;
3417 1166 : guard
3418 1166 : .open_mut()?
3419 1166 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock)
3420 3 : .await
3421 : };
3422 :
3423 1166 : if frozen {
3424 1138 : let now = Instant::now();
3425 1138 : *(self.last_freeze_ts.write().unwrap()) = now;
3426 1138 : }
3427 :
3428 : // Increment the flush cycle counter and wake up the flush task.
3429 : // Remember the new value, so that when we listen for the flush
3430 : // to finish, we know when the flush that we initiated has
3431 : // finished, instead of some other flush that was started earlier.
3432 1166 : let mut my_flush_request = 0;
3433 1166 :
3434 1166 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
3435 1166 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
3436 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
3437 1166 : }
3438 1166 :
3439 1166 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
3440 1166 : my_flush_request = *counter + 1;
3441 1166 : *counter = my_flush_request;
3442 1166 : *lsn = std::cmp::max(at, *lsn);
3443 1166 : });
3444 1166 :
3445 1166 : assert_ne!(my_flush_request, 0);
3446 :
3447 1166 : Ok(my_flush_request)
3448 1166 : }
3449 :
3450 : /// Layer flusher task's main loop.
3451 404 : async fn flush_loop(
3452 404 : self: &Arc<Self>,
3453 404 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
3454 404 : ctx: &RequestContext,
3455 404 : ) {
3456 404 : info!("started flush loop");
3457 : loop {
3458 1532 : tokio::select! {
3459 1532 : _ = self.cancel.cancelled() => {
3460 8 : info!("shutting down layer flush task due to Timeline::cancel");
3461 8 : break;
3462 : },
3463 1532 : _ = layer_flush_start_rx.changed() => {}
3464 1128 : }
3465 1128 : trace!("waking up");
3466 1128 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
3467 1128 :
3468 1128 : // The highest LSN to which we flushed in the loop over frozen layers
3469 1128 : let mut flushed_to_lsn = Lsn(0);
3470 :
3471 1128 : let result = loop {
3472 2266 : if self.cancel.is_cancelled() {
3473 0 : info!("dropping out of flush loop for timeline shutdown");
3474 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
3475 : // anyone waiting on that will respect self.cancel as well: they will stop
3476 : // waiting at the same time we as drop out of this loop.
3477 0 : return;
3478 2266 : }
3479 2266 :
3480 2266 : let timer = self.metrics.flush_time_histo.start_timer();
3481 :
3482 2266 : let layer_to_flush = {
3483 2266 : let guard = self.layers.read().await;
3484 2266 : let Ok(lm) = guard.layer_map() else {
3485 0 : info!("dropping out of flush loop for timeline shutdown");
3486 0 : return;
3487 : };
3488 2266 : lm.frozen_layers.front().cloned()
3489 : // drop 'layers' lock to allow concurrent reads and writes
3490 : };
3491 2266 : let Some(layer_to_flush) = layer_to_flush else {
3492 1128 : break Ok(());
3493 : };
3494 16973 : match self.flush_frozen_layer(layer_to_flush, ctx).await {
3495 1138 : Ok(this_layer_to_lsn) => {
3496 1138 : flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
3497 1138 : }
3498 : Err(FlushLayerError::Cancelled) => {
3499 0 : info!("dropping out of flush loop for timeline shutdown");
3500 0 : return;
3501 : }
3502 0 : err @ Err(
3503 0 : FlushLayerError::NotRunning(_)
3504 0 : | FlushLayerError::Other(_)
3505 0 : | FlushLayerError::CreateImageLayersError(_),
3506 0 : ) => {
3507 0 : error!("could not flush frozen layer: {err:?}");
3508 0 : break err.map(|_| ());
3509 : }
3510 : }
3511 1138 : timer.stop_and_record();
3512 : };
3513 :
3514 : // Unsharded tenants should never advance their LSN beyond the end of the
3515 : // highest layer they write: such gaps between layer data and the frozen LSN
3516 : // are only legal on sharded tenants.
3517 1128 : debug_assert!(
3518 1128 : self.shard_identity.count.count() > 1
3519 1128 : || flushed_to_lsn >= frozen_to_lsn
3520 67 : || !flushed_to_lsn.is_valid()
3521 : );
3522 :
3523 1128 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
3524 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
3525 : // to us via layer_flush_start_rx, then advance it here.
3526 : //
3527 : // This path is only taken for tenants with multiple shards: single sharded tenants should
3528 : // never encounter a gap in the wal.
3529 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
3530 0 : tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
3531 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
3532 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
3533 0 : tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
3534 0 : }
3535 0 : }
3536 1128 : }
3537 :
3538 : // Notify any listeners that we're done
3539 1128 : let _ = self
3540 1128 : .layer_flush_done_tx
3541 1128 : .send_replace((flush_counter, result));
3542 : }
3543 8 : }
3544 :
3545 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
3546 1086 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
3547 1086 : let mut rx = self.layer_flush_done_tx.subscribe();
3548 : loop {
3549 : {
3550 2170 : let (last_result_counter, last_result) = &*rx.borrow();
3551 2170 : if *last_result_counter >= request {
3552 1086 : if let Err(err) = last_result {
3553 : // We already logged the original error in
3554 : // flush_loop. We cannot propagate it to the caller
3555 : // here, because it might not be Cloneable
3556 0 : return Err(err.clone());
3557 : } else {
3558 1086 : return Ok(());
3559 : }
3560 1084 : }
3561 1084 : }
3562 1084 : trace!("waiting for flush to complete");
3563 1084 : tokio::select! {
3564 1084 : rx_e = rx.changed() => {
3565 1084 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
3566 : },
3567 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
3568 : // the notification from [`flush_loop`] that it completed.
3569 1084 : _ = self.cancel.cancelled() => {
3570 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
3571 0 : return Ok(())
3572 : }
3573 : };
3574 1084 : trace!("done")
3575 : }
3576 1086 : }
3577 :
3578 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
3579 : ///
3580 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
3581 1138 : #[instrument(skip_all, fields(layer=%frozen_layer))]
3582 : async fn flush_frozen_layer(
3583 : self: &Arc<Self>,
3584 : frozen_layer: Arc<InMemoryLayer>,
3585 : ctx: &RequestContext,
3586 : ) -> Result<Lsn, FlushLayerError> {
3587 : debug_assert_current_span_has_tenant_and_timeline_id();
3588 :
3589 : // As a special case, when we have just imported an image into the repository,
3590 : // instead of writing out a L0 delta layer, we directly write out image layer
3591 : // files instead. This is possible as long as *all* the data imported into the
3592 : // repository have the same LSN.
3593 : let lsn_range = frozen_layer.get_lsn_range();
3594 :
3595 : // Whether to directly create image layers for this flush, or flush them as delta layers
3596 : let create_image_layer =
3597 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
3598 :
3599 : #[cfg(test)]
3600 : {
3601 : match &mut *self.flush_loop_state.lock().unwrap() {
3602 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
3603 : panic!("flush loop not running")
3604 : }
3605 : FlushLoopState::Running {
3606 : expect_initdb_optimization,
3607 : initdb_optimization_count,
3608 : ..
3609 : } => {
3610 : if create_image_layer {
3611 : *initdb_optimization_count += 1;
3612 : } else {
3613 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
3614 : }
3615 : }
3616 : }
3617 : }
3618 :
3619 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
3620 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
3621 : // require downloading anything during initial import.
3622 : let ((rel_partition, metadata_partition), _lsn) = self
3623 : .repartition(
3624 : self.initdb_lsn,
3625 : self.get_compaction_target_size(),
3626 : EnumSet::empty(),
3627 : ctx,
3628 : )
3629 : .await
3630 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
3631 :
3632 : if self.cancel.is_cancelled() {
3633 : return Err(FlushLayerError::Cancelled);
3634 : }
3635 :
3636 : let mut layers_to_upload = Vec::new();
3637 : layers_to_upload.extend(
3638 : self.create_image_layers(
3639 : &rel_partition,
3640 : self.initdb_lsn,
3641 : ImageLayerCreationMode::Initial,
3642 : ctx,
3643 : )
3644 : .await?,
3645 : );
3646 : if !metadata_partition.parts.is_empty() {
3647 : assert_eq!(
3648 : metadata_partition.parts.len(),
3649 : 1,
3650 : "currently sparse keyspace should only contain a single metadata keyspace"
3651 : );
3652 : layers_to_upload.extend(
3653 : self.create_image_layers(
3654 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
3655 : // every single key within the keyspace, and therefore, it's safe to force converting it
3656 : // into a dense keyspace before calling this function.
3657 : &metadata_partition.into_dense(),
3658 : self.initdb_lsn,
3659 : ImageLayerCreationMode::Initial,
3660 : ctx,
3661 : )
3662 : .await?,
3663 : );
3664 : }
3665 :
3666 : (layers_to_upload, None)
3667 : } else {
3668 : // Normal case, write out a L0 delta layer file.
3669 : // `create_delta_layer` will not modify the layer map.
3670 : // We will remove frozen layer and add delta layer in one atomic operation later.
3671 : let Some(layer) = self
3672 : .create_delta_layer(&frozen_layer, None, ctx)
3673 : .await
3674 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
3675 : else {
3676 : panic!("delta layer cannot be empty if no filter is applied");
3677 : };
3678 : (
3679 : // FIXME: even though we have a single image and single delta layer assumption
3680 : // we push them to vec
3681 : vec![layer.clone()],
3682 : Some(layer),
3683 : )
3684 : };
3685 :
3686 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
3687 :
3688 : if self.cancel.is_cancelled() {
3689 : return Err(FlushLayerError::Cancelled);
3690 : }
3691 :
3692 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
3693 :
3694 : // The new on-disk layers are now in the layer map. We can remove the
3695 : // in-memory layer from the map now. The flushed layer is stored in
3696 : // the mapping in `create_delta_layer`.
3697 : {
3698 : let mut guard = self.layers.write().await;
3699 :
3700 : guard.open_mut()?.finish_flush_l0_layer(
3701 : delta_layer_to_add.as_ref(),
3702 : &frozen_layer,
3703 : &self.metrics,
3704 : );
3705 :
3706 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
3707 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
3708 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
3709 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
3710 : }
3711 : // release lock on 'layers'
3712 : };
3713 :
3714 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
3715 : // This makes us refuse ingest until the new layers have been persisted to the remote.
3716 : self.remote_client
3717 : .wait_completion()
3718 : .await
3719 0 : .map_err(|e| match e {
3720 : WaitCompletionError::UploadQueueShutDownOrStopped
3721 : | WaitCompletionError::NotInitialized(
3722 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
3723 0 : ) => FlushLayerError::Cancelled,
3724 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
3725 0 : FlushLayerError::Other(anyhow!(e).into())
3726 : }
3727 0 : })?;
3728 :
3729 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
3730 : // a compaction can delete the file and then it won't be available for uploads any more.
3731 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
3732 : // race situation.
3733 : // See https://github.com/neondatabase/neon/issues/4526
3734 : pausable_failpoint!("flush-frozen-pausable");
3735 :
3736 : // This failpoint is used by another test case `test_pageserver_recovery`.
3737 : fail_point!("flush-frozen-exit");
3738 :
3739 : Ok(Lsn(lsn_range.end.0 - 1))
3740 : }
3741 :
3742 : /// Return true if the value changed
3743 : ///
3744 : /// This function must only be used from the layer flush task.
3745 1138 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
3746 1138 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
3747 1138 : assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
3748 1138 : new_value != old_value
3749 1138 : }
3750 :
3751 : /// Update metadata file
3752 1140 : fn schedule_uploads(
3753 1140 : &self,
3754 1140 : disk_consistent_lsn: Lsn,
3755 1140 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
3756 1140 : ) -> anyhow::Result<()> {
3757 1140 : // We can only save a valid 'prev_record_lsn' value on disk if we
3758 1140 : // flushed *all* in-memory changes to disk. We only track
3759 1140 : // 'prev_record_lsn' in memory for the latest processed record, so we
3760 1140 : // don't remember what the correct value that corresponds to some old
3761 1140 : // LSN is. But if we flush everything, then the value corresponding
3762 1140 : // current 'last_record_lsn' is correct and we can store it on disk.
3763 1140 : let RecordLsn {
3764 1140 : last: last_record_lsn,
3765 1140 : prev: prev_record_lsn,
3766 1140 : } = self.last_record_lsn.load();
3767 1140 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
3768 1060 : Some(prev_record_lsn)
3769 : } else {
3770 80 : None
3771 : };
3772 :
3773 1140 : let update = crate::tenant::metadata::MetadataUpdate::new(
3774 1140 : disk_consistent_lsn,
3775 1140 : ondisk_prev_record_lsn,
3776 1140 : *self.latest_gc_cutoff_lsn.read(),
3777 1140 : );
3778 1140 :
3779 1140 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
3780 0 : "{}",
3781 0 : x.unwrap()
3782 1140 : ));
3783 :
3784 2290 : for layer in layers_to_upload {
3785 1150 : self.remote_client.schedule_layer_file_upload(layer)?;
3786 : }
3787 1140 : self.remote_client
3788 1140 : .schedule_index_upload_for_metadata_update(&update)?;
3789 :
3790 1140 : Ok(())
3791 1140 : }
3792 :
3793 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
3794 0 : self.remote_client
3795 0 : .preserve_initdb_archive(
3796 0 : &self.tenant_shard_id.tenant_id,
3797 0 : &self.timeline_id,
3798 0 : &self.cancel,
3799 0 : )
3800 0 : .await
3801 0 : }
3802 :
3803 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
3804 : // in layer map immediately. The caller is responsible to put it into the layer map.
3805 968 : async fn create_delta_layer(
3806 968 : self: &Arc<Self>,
3807 968 : frozen_layer: &Arc<InMemoryLayer>,
3808 968 : key_range: Option<Range<Key>>,
3809 968 : ctx: &RequestContext,
3810 968 : ) -> anyhow::Result<Option<ResidentLayer>> {
3811 968 : let self_clone = Arc::clone(self);
3812 968 : let frozen_layer = Arc::clone(frozen_layer);
3813 968 : let ctx = ctx.attached_child();
3814 968 : let work = async move {
3815 968 : let Some((desc, path)) = frozen_layer
3816 968 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
3817 10247 : .await?
3818 : else {
3819 0 : return Ok(None);
3820 : };
3821 968 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
3822 :
3823 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
3824 : // We just need to fsync the directory in which these inodes are linked,
3825 : // which we know to be the timeline directory.
3826 : //
3827 : // We use fatal_err() below because the after write_to_disk returns with success,
3828 : // the in-memory state of the filesystem already has the layer file in its final place,
3829 : // and subsequent pageserver code could think it's durable while it really isn't.
3830 968 : let timeline_dir = VirtualFile::open(
3831 968 : &self_clone
3832 968 : .conf
3833 968 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
3834 968 : &ctx,
3835 968 : )
3836 490 : .await
3837 968 : .fatal_err("VirtualFile::open for timeline dir fsync");
3838 968 : timeline_dir
3839 968 : .sync_all()
3840 484 : .await
3841 968 : .fatal_err("VirtualFile::sync_all timeline dir");
3842 968 : anyhow::Ok(Some(new_delta))
3843 968 : };
3844 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
3845 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
3846 : use crate::virtual_file::io_engine::IoEngine;
3847 968 : match crate::virtual_file::io_engine::get() {
3848 0 : IoEngine::NotSet => panic!("io engine not set"),
3849 : IoEngine::StdFs => {
3850 484 : let span = tracing::info_span!("blocking");
3851 484 : tokio::task::spawn_blocking({
3852 484 : move || Handle::current().block_on(work.instrument(span))
3853 484 : })
3854 484 : .await
3855 484 : .context("spawn_blocking")
3856 484 : .and_then(|x| x)
3857 : }
3858 : #[cfg(target_os = "linux")]
3859 11216 : IoEngine::TokioEpollUring => work.await,
3860 : }
3861 968 : }
3862 :
3863 534 : async fn repartition(
3864 534 : &self,
3865 534 : lsn: Lsn,
3866 534 : partition_size: u64,
3867 534 : flags: EnumSet<CompactFlags>,
3868 534 : ctx: &RequestContext,
3869 534 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
3870 534 : let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
3871 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
3872 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
3873 : // and hence before the compaction task starts.
3874 0 : return Err(CompactionError::Other(anyhow!(
3875 0 : "repartition() called concurrently, this should not happen"
3876 0 : )));
3877 : };
3878 534 : let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
3879 534 : if lsn < *partition_lsn {
3880 0 : return Err(CompactionError::Other(anyhow!(
3881 0 : "repartition() called with LSN going backwards, this should not happen"
3882 0 : )));
3883 534 : }
3884 534 :
3885 534 : let distance = lsn.0 - partition_lsn.0;
3886 534 : if *partition_lsn != Lsn(0)
3887 262 : && distance <= self.repartition_threshold
3888 262 : && !flags.contains(CompactFlags::ForceRepartition)
3889 : {
3890 248 : debug!(
3891 : distance,
3892 : threshold = self.repartition_threshold,
3893 0 : "no repartitioning needed"
3894 : );
3895 248 : return Ok((
3896 248 : (dense_partition.clone(), sparse_partition.clone()),
3897 248 : *partition_lsn,
3898 248 : ));
3899 286 : }
3900 :
3901 15799 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
3902 286 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
3903 286 : let sparse_partitioning = SparseKeyPartitioning {
3904 286 : parts: vec![sparse_ks],
3905 286 : }; // no partitioning for metadata keys for now
3906 286 : *partitioning_guard = ((dense_partitioning, sparse_partitioning), lsn);
3907 286 :
3908 286 : Ok((partitioning_guard.0.clone(), partitioning_guard.1))
3909 534 : }
3910 :
3911 : // Is it time to create a new image layer for the given partition?
3912 14 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
3913 14 : let threshold = self.get_image_creation_threshold();
3914 :
3915 14 : let guard = self.layers.read().await;
3916 14 : let Ok(layers) = guard.layer_map() else {
3917 0 : return false;
3918 : };
3919 :
3920 14 : let mut max_deltas = 0;
3921 28 : for part_range in &partition.ranges {
3922 14 : let image_coverage = layers.image_coverage(part_range, lsn);
3923 28 : for (img_range, last_img) in image_coverage {
3924 14 : let img_lsn = if let Some(last_img) = last_img {
3925 0 : last_img.get_lsn_range().end
3926 : } else {
3927 14 : Lsn(0)
3928 : };
3929 : // Let's consider an example:
3930 : //
3931 : // delta layer with LSN range 71-81
3932 : // delta layer with LSN range 81-91
3933 : // delta layer with LSN range 91-101
3934 : // image layer at LSN 100
3935 : //
3936 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
3937 : // there's no need to create a new one. We check this case explicitly, to avoid passing
3938 : // a bogus range to count_deltas below, with start > end. It's even possible that there
3939 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
3940 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
3941 14 : if img_lsn < lsn {
3942 14 : let num_deltas =
3943 14 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
3944 14 :
3945 14 : max_deltas = max_deltas.max(num_deltas);
3946 14 : if num_deltas >= threshold {
3947 0 : debug!(
3948 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
3949 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
3950 : );
3951 0 : return true;
3952 14 : }
3953 0 : }
3954 : }
3955 : }
3956 :
3957 14 : debug!(
3958 : max_deltas,
3959 0 : "none of the partitioned ranges had >= {threshold} deltas"
3960 : );
3961 14 : false
3962 14 : }
3963 :
3964 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
3965 : /// so that at most one image layer will be produced from this function.
3966 196 : async fn create_image_layer_for_rel_blocks(
3967 196 : self: &Arc<Self>,
3968 196 : partition: &KeySpace,
3969 196 : mut image_layer_writer: ImageLayerWriter,
3970 196 : lsn: Lsn,
3971 196 : ctx: &RequestContext,
3972 196 : img_range: Range<Key>,
3973 196 : start: Key,
3974 196 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
3975 196 : let mut wrote_keys = false;
3976 196 :
3977 196 : let mut key_request_accum = KeySpaceAccum::new();
3978 1302 : for range in &partition.ranges {
3979 1106 : let mut key = range.start;
3980 2394 : while key < range.end {
3981 : // Decide whether to retain this key: usually we do, but sharded tenants may
3982 : // need to drop keys that don't belong to them. If we retain the key, add it
3983 : // to `key_request_accum` for later issuing a vectored get
3984 1288 : if self.shard_identity.is_key_disposable(&key) {
3985 0 : debug!(
3986 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
3987 0 : key,
3988 0 : self.shard_identity.get_shard_number(&key)
3989 : );
3990 1288 : } else {
3991 1288 : key_request_accum.add_key(key);
3992 1288 : }
3993 :
3994 1288 : let last_key_in_range = key.next() == range.end;
3995 1288 : key = key.next();
3996 1288 :
3997 1288 : // Maybe flush `key_rest_accum`
3998 1288 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
3999 1288 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4000 : {
4001 1106 : let results = self
4002 1106 : .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
4003 49 : .await?;
4004 :
4005 1106 : if self.cancel.is_cancelled() {
4006 0 : return Err(CreateImageLayersError::Cancelled);
4007 1106 : }
4008 :
4009 2394 : for (img_key, img) in results {
4010 1288 : let img = match img {
4011 1288 : Ok(img) => img,
4012 0 : Err(err) => {
4013 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4014 0 : // page without losing any actual user data. That seems better
4015 0 : // than failing repeatedly and getting stuck.
4016 0 : //
4017 0 : // We had a bug at one point, where we truncated the FSM and VM
4018 0 : // in the pageserver, but the Postgres didn't know about that
4019 0 : // and continued to generate incremental WAL records for pages
4020 0 : // that didn't exist in the pageserver. Trying to replay those
4021 0 : // WAL records failed to find the previous image of the page.
4022 0 : // This special case allows us to recover from that situation.
4023 0 : // See https://github.com/neondatabase/neon/issues/2601.
4024 0 : //
4025 0 : // Unfortunately we cannot do this for the main fork, or for
4026 0 : // any metadata keys, keys, as that would lead to actual data
4027 0 : // loss.
4028 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4029 0 : warn!("could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}");
4030 0 : ZERO_PAGE.clone()
4031 : } else {
4032 0 : return Err(CreateImageLayersError::from(err));
4033 : }
4034 : }
4035 : };
4036 :
4037 : // Write all the keys we just read into our new image layer.
4038 1404 : image_layer_writer.put_image(img_key, img, ctx).await?;
4039 1288 : wrote_keys = true;
4040 : }
4041 182 : }
4042 : }
4043 : }
4044 :
4045 196 : if wrote_keys {
4046 : // Normal path: we have written some data into the new image layer for this
4047 : // partition, so flush it to disk.
4048 396 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4049 196 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4050 196 : info!("created image layer for rel {}", image_layer.local_path());
4051 196 : Ok(ImageLayerCreationOutcome {
4052 196 : image: Some(image_layer),
4053 196 : next_start_key: img_range.end,
4054 196 : })
4055 : } else {
4056 : // Special case: the image layer may be empty if this is a sharded tenant and the
4057 : // partition does not cover any keys owned by this shard. In this case, to ensure
4058 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4059 : // layer we write will cover the key range that we just scanned.
4060 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4061 0 : Ok(ImageLayerCreationOutcome {
4062 0 : image: None,
4063 0 : next_start_key: start,
4064 0 : })
4065 : }
4066 196 : }
4067 :
4068 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4069 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4070 : /// would not be too large to fit in a single image layer.
4071 : #[allow(clippy::too_many_arguments)]
4072 186 : async fn create_image_layer_for_metadata_keys(
4073 186 : self: &Arc<Self>,
4074 186 : partition: &KeySpace,
4075 186 : mut image_layer_writer: ImageLayerWriter,
4076 186 : lsn: Lsn,
4077 186 : ctx: &RequestContext,
4078 186 : img_range: Range<Key>,
4079 186 : mode: ImageLayerCreationMode,
4080 186 : start: Key,
4081 186 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4082 186 : // Metadata keys image layer creation.
4083 186 : let mut reconstruct_state = ValuesReconstructState::default();
4084 186 : let data = self
4085 186 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4086 1636 : .await?;
4087 186 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4088 186 : let mut new_data = BTreeMap::new();
4089 186 : let mut total_kb_retrieved = 0;
4090 186 : let mut total_keys_retrieved = 0;
4091 10198 : for (k, v) in data {
4092 10012 : let v = v?;
4093 10012 : total_kb_retrieved += KEY_SIZE + v.len();
4094 10012 : total_keys_retrieved += 1;
4095 10012 : new_data.insert(k, v);
4096 : }
4097 186 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4098 186 : };
4099 186 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4100 186 :
4101 186 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4102 186 : debug!(
4103 : trigger_generation,
4104 : delta_files_accessed,
4105 : total_kb_retrieved,
4106 : total_keys_retrieved,
4107 0 : "generate metadata images"
4108 : );
4109 :
4110 186 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4111 2 : return Ok(ImageLayerCreationOutcome {
4112 2 : image: None,
4113 2 : next_start_key: img_range.end,
4114 2 : });
4115 184 : }
4116 184 : if self.cancel.is_cancelled() {
4117 0 : return Err(CreateImageLayersError::Cancelled);
4118 184 : }
4119 184 : let mut wrote_any_image = false;
4120 10196 : for (k, v) in data {
4121 10012 : if v.is_empty() {
4122 : // the key has been deleted, it does not need an image
4123 : // in metadata keyspace, an empty image == tombstone
4124 8 : continue;
4125 10004 : }
4126 10004 : wrote_any_image = true;
4127 10004 :
4128 10004 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4129 10004 :
4130 10004 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4131 10004 : // on the normal data path either.
4132 10159 : image_layer_writer.put_image(k, v, ctx).await?;
4133 : }
4134 :
4135 184 : if wrote_any_image {
4136 : // Normal path: we have written some data into the new image layer for this
4137 : // partition, so flush it to disk.
4138 24 : let (desc, path) = image_layer_writer.finish(ctx).await?;
4139 12 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
4140 12 : info!(
4141 0 : "created image layer for metadata {}",
4142 0 : image_layer.local_path()
4143 : );
4144 12 : Ok(ImageLayerCreationOutcome {
4145 12 : image: Some(image_layer),
4146 12 : next_start_key: img_range.end,
4147 12 : })
4148 : } else {
4149 : // Special case: the image layer may be empty if this is a sharded tenant and the
4150 : // partition does not cover any keys owned by this shard. In this case, to ensure
4151 : // we don't leave gaps between image layers, leave `start` where it is, so that the next
4152 : // layer we write will cover the key range that we just scanned.
4153 172 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4154 172 : Ok(ImageLayerCreationOutcome {
4155 172 : image: None,
4156 172 : next_start_key: start,
4157 172 : })
4158 : }
4159 186 : }
4160 :
4161 : /// Predicate function which indicates whether we should check if new image layers
4162 : /// are required. Since checking if new image layers are required is expensive in
4163 : /// terms of CPU, we only do it in the following cases:
4164 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4165 : /// 2. If enough time has passed since the last check:
4166 : /// 1. For large tenants, we wish to perform the check more often since they
4167 : /// suffer from the lack of image layers
4168 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4169 704 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4170 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
4171 :
4172 704 : let last_checks_at = self.last_image_layer_creation_check_at.load();
4173 704 : let distance = lsn
4174 704 : .checked_sub(last_checks_at)
4175 704 : .expect("Attempt to compact with LSN going backwards");
4176 704 : let min_distance =
4177 704 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
4178 704 :
4179 704 : let distance_based_decision = distance.0 >= min_distance;
4180 704 :
4181 704 : let mut time_based_decision = false;
4182 704 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
4183 704 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
4184 602 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
4185 : {
4186 0 : self.get_checkpoint_timeout()
4187 : } else {
4188 602 : Duration::from_secs(3600 * 48)
4189 : };
4190 :
4191 602 : time_based_decision = match *last_check_instant {
4192 432 : Some(last_check) => {
4193 432 : let elapsed = last_check.elapsed();
4194 432 : elapsed >= check_required_after
4195 : }
4196 170 : None => true,
4197 : };
4198 102 : }
4199 :
4200 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
4201 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
4202 : // check.
4203 704 : let decision = distance_based_decision || time_based_decision;
4204 :
4205 704 : if decision {
4206 172 : self.last_image_layer_creation_check_at.store(lsn);
4207 172 : *last_check_instant = Some(Instant::now());
4208 532 : }
4209 :
4210 704 : decision
4211 704 : }
4212 :
4213 704 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
4214 : async fn create_image_layers(
4215 : self: &Arc<Timeline>,
4216 : partitioning: &KeyPartitioning,
4217 : lsn: Lsn,
4218 : mode: ImageLayerCreationMode,
4219 : ctx: &RequestContext,
4220 : ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
4221 : let timer = self.metrics.create_images_time_histo.start_timer();
4222 : let mut image_layers = Vec::new();
4223 :
4224 : // We need to avoid holes between generated image layers.
4225 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
4226 : // image layer with hole between them. In this case such layer can not be utilized by GC.
4227 : //
4228 : // How such hole between partitions can appear?
4229 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
4230 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
4231 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
4232 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
4233 : let mut start = Key::MIN;
4234 :
4235 : let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
4236 :
4237 : for partition in partitioning.parts.iter() {
4238 : if self.cancel.is_cancelled() {
4239 : return Err(CreateImageLayersError::Cancelled);
4240 : }
4241 :
4242 : let img_range = start..partition.ranges.last().unwrap().end;
4243 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
4244 : if compact_metadata {
4245 : for range in &partition.ranges {
4246 : assert!(
4247 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
4248 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
4249 : "metadata keys must be partitioned separately"
4250 : );
4251 : }
4252 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
4253 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
4254 : // might mess up with evictions.
4255 : start = img_range.end;
4256 : continue;
4257 : }
4258 : // For initial and force modes, we always generate image layers for metadata keys.
4259 : } else if let ImageLayerCreationMode::Try = mode {
4260 : // check_for_image_layers = false -> skip
4261 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
4262 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
4263 : start = img_range.end;
4264 : continue;
4265 : }
4266 : }
4267 : if let ImageLayerCreationMode::Force = mode {
4268 : // When forced to create image layers, we might try and create them where they already
4269 : // exist. This mode is only used in tests/debug.
4270 : let layers = self.layers.read().await;
4271 : if layers.contains_key(&PersistentLayerKey {
4272 : key_range: img_range.clone(),
4273 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
4274 : is_delta: false,
4275 : }) {
4276 : tracing::info!(
4277 : "Skipping image layer at {lsn} {}..{}, already exists",
4278 : img_range.start,
4279 : img_range.end
4280 : );
4281 : start = img_range.end;
4282 : continue;
4283 : }
4284 : }
4285 :
4286 : let image_layer_writer = ImageLayerWriter::new(
4287 : self.conf,
4288 : self.timeline_id,
4289 : self.tenant_shard_id,
4290 : &img_range,
4291 : lsn,
4292 : ctx,
4293 : )
4294 : .await?;
4295 :
4296 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
4297 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
4298 0 : "failpoint image-layer-writer-fail-before-finish"
4299 0 : )))
4300 0 : });
4301 :
4302 : if !compact_metadata {
4303 : let ImageLayerCreationOutcome {
4304 : image,
4305 : next_start_key,
4306 : } = self
4307 : .create_image_layer_for_rel_blocks(
4308 : partition,
4309 : image_layer_writer,
4310 : lsn,
4311 : ctx,
4312 : img_range,
4313 : start,
4314 : )
4315 : .await?;
4316 :
4317 : start = next_start_key;
4318 : image_layers.extend(image);
4319 : } else {
4320 : let ImageLayerCreationOutcome {
4321 : image,
4322 : next_start_key,
4323 : } = self
4324 : .create_image_layer_for_metadata_keys(
4325 : partition,
4326 : image_layer_writer,
4327 : lsn,
4328 : ctx,
4329 : img_range,
4330 : mode,
4331 : start,
4332 : )
4333 : .await?;
4334 : start = next_start_key;
4335 : image_layers.extend(image);
4336 : }
4337 : }
4338 :
4339 : let mut guard = self.layers.write().await;
4340 :
4341 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
4342 : // now they are being scheduled outside of write lock; current way is inconsistent with
4343 : // compaction lock order.
4344 : guard
4345 : .open_mut()?
4346 : .track_new_image_layers(&image_layers, &self.metrics);
4347 : drop_wlock(guard);
4348 : timer.stop_and_record();
4349 :
4350 : // Creating image layers may have caused some previously visible layers to be covered
4351 : if !image_layers.is_empty() {
4352 : self.update_layer_visibility().await?;
4353 : }
4354 :
4355 : Ok(image_layers)
4356 : }
4357 :
4358 : /// Wait until the background initial logical size calculation is complete, or
4359 : /// this Timeline is shut down. Calling this function will cause the initial
4360 : /// logical size calculation to skip waiting for the background jobs barrier.
4361 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
4362 0 : if !self.shard_identity.is_shard_zero() {
4363 : // We don't populate logical size on shard >0: skip waiting for it.
4364 0 : return;
4365 0 : }
4366 0 :
4367 0 : if self.remote_client.is_deleting() {
4368 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
4369 0 : return;
4370 0 : }
4371 0 :
4372 0 : if self.current_logical_size.current_size().is_exact() {
4373 : // root timelines are initialized with exact count, but never start the background
4374 : // calculation
4375 0 : return;
4376 0 : }
4377 :
4378 0 : if let Some(await_bg_cancel) = self
4379 0 : .current_logical_size
4380 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
4381 0 : .get()
4382 0 : {
4383 0 : await_bg_cancel.cancel();
4384 0 : } else {
4385 : // We should not wait if we were not able to explicitly instruct
4386 : // the logical size cancellation to skip the concurrency limit semaphore.
4387 : // TODO: this is an unexpected case. We should restructure so that it
4388 : // can't happen.
4389 0 : tracing::warn!(
4390 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
4391 : );
4392 0 : debug_assert!(false);
4393 : }
4394 :
4395 0 : tokio::select!(
4396 0 : _ = self.current_logical_size.initialized.acquire() => {},
4397 0 : _ = self.cancel.cancelled() => {}
4398 : )
4399 0 : }
4400 :
4401 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
4402 : /// Timelines layers up to the ancestor_lsn.
4403 : ///
4404 : /// Requires a timeline that:
4405 : /// - has an ancestor to detach from
4406 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
4407 : /// a technical requirement
4408 : ///
4409 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
4410 : /// polled again until completion.
4411 : ///
4412 : /// During the operation all timelines sharing the data with this timeline will be reparented
4413 : /// from our ancestor to be branches of this timeline.
4414 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
4415 0 : self: &Arc<Timeline>,
4416 0 : tenant: &crate::tenant::Tenant,
4417 0 : options: detach_ancestor::Options,
4418 0 : ctx: &RequestContext,
4419 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
4420 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
4421 0 : }
4422 :
4423 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
4424 : /// reparents any reparentable children of previous ancestor.
4425 : ///
4426 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
4427 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
4428 : /// successfully, tenant must be reloaded.
4429 : ///
4430 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
4431 : /// resetting the tenant.
4432 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
4433 0 : self: &Arc<Timeline>,
4434 0 : tenant: &crate::tenant::Tenant,
4435 0 : prepared: detach_ancestor::PreparedTimelineDetach,
4436 0 : ctx: &RequestContext,
4437 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
4438 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
4439 0 : }
4440 :
4441 : /// Final step which unblocks the GC.
4442 : ///
4443 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
4444 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
4445 0 : self: &Arc<Timeline>,
4446 0 : tenant: &crate::tenant::Tenant,
4447 0 : attempt: detach_ancestor::Attempt,
4448 0 : ctx: &RequestContext,
4449 0 : ) -> Result<(), detach_ancestor::Error> {
4450 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
4451 0 : }
4452 : }
4453 :
4454 : impl Drop for Timeline {
4455 8 : fn drop(&mut self) {
4456 8 : if let Some(ancestor) = &self.ancestor_timeline {
4457 : // This lock should never be poisoned, but in case it is we do a .map() instead of
4458 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
4459 2 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
4460 2 : gc_info.remove_child(self.timeline_id)
4461 0 : }
4462 6 : }
4463 8 : }
4464 : }
4465 :
4466 : /// Top-level failure to compact.
4467 0 : #[derive(Debug, thiserror::Error)]
4468 : pub(crate) enum CompactionError {
4469 : #[error("The timeline or pageserver is shutting down")]
4470 : ShuttingDown,
4471 : /// Compaction cannot be done right now; page reconstruction and so on.
4472 : #[error(transparent)]
4473 : Other(anyhow::Error),
4474 : }
4475 :
4476 : impl CompactionError {
4477 0 : pub fn is_cancelled(&self) -> bool {
4478 0 : matches!(self, CompactionError::ShuttingDown)
4479 0 : }
4480 : }
4481 :
4482 : impl From<CollectKeySpaceError> for CompactionError {
4483 0 : fn from(err: CollectKeySpaceError) -> Self {
4484 0 : match err {
4485 : CollectKeySpaceError::Cancelled
4486 : | CollectKeySpaceError::PageRead(PageReconstructError::Cancelled) => {
4487 0 : CompactionError::ShuttingDown
4488 : }
4489 0 : e => CompactionError::Other(e.into()),
4490 : }
4491 0 : }
4492 : }
4493 :
4494 : impl From<super::upload_queue::NotInitialized> for CompactionError {
4495 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
4496 0 : match value {
4497 : super::upload_queue::NotInitialized::Uninitialized => {
4498 0 : CompactionError::Other(anyhow::anyhow!(value))
4499 : }
4500 : super::upload_queue::NotInitialized::ShuttingDown
4501 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
4502 : }
4503 0 : }
4504 : }
4505 :
4506 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
4507 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
4508 0 : match e {
4509 : super::storage_layer::layer::DownloadError::TimelineShutdown
4510 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
4511 0 : CompactionError::ShuttingDown
4512 : }
4513 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
4514 : | super::storage_layer::layer::DownloadError::DownloadRequired
4515 : | super::storage_layer::layer::DownloadError::NotFile(_)
4516 : | super::storage_layer::layer::DownloadError::DownloadFailed
4517 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
4518 0 : CompactionError::Other(anyhow::anyhow!(e))
4519 : }
4520 : #[cfg(test)]
4521 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
4522 0 : CompactionError::Other(anyhow::anyhow!(e))
4523 : }
4524 : }
4525 0 : }
4526 : }
4527 :
4528 : impl From<layer_manager::Shutdown> for CompactionError {
4529 0 : fn from(_: layer_manager::Shutdown) -> Self {
4530 0 : CompactionError::ShuttingDown
4531 0 : }
4532 : }
4533 :
4534 : #[serde_as]
4535 196 : #[derive(serde::Serialize)]
4536 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
4537 :
4538 : #[derive(Default)]
4539 : enum DurationRecorder {
4540 : #[default]
4541 : NotStarted,
4542 : Recorded(RecordedDuration, tokio::time::Instant),
4543 : }
4544 :
4545 : impl DurationRecorder {
4546 504 : fn till_now(&self) -> DurationRecorder {
4547 504 : match self {
4548 : DurationRecorder::NotStarted => {
4549 0 : panic!("must only call on recorded measurements")
4550 : }
4551 504 : DurationRecorder::Recorded(_, ended) => {
4552 504 : let now = tokio::time::Instant::now();
4553 504 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
4554 504 : }
4555 504 : }
4556 504 : }
4557 196 : fn into_recorded(self) -> Option<RecordedDuration> {
4558 196 : match self {
4559 0 : DurationRecorder::NotStarted => None,
4560 196 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
4561 : }
4562 196 : }
4563 : }
4564 :
4565 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
4566 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
4567 : /// the layer descriptor requires the user to provide the ranges, which should cover all
4568 : /// keys specified in the `data` field.
4569 : #[cfg(test)]
4570 : #[derive(Clone)]
4571 : pub struct DeltaLayerTestDesc {
4572 : pub lsn_range: Range<Lsn>,
4573 : pub key_range: Range<Key>,
4574 : pub data: Vec<(Key, Lsn, Value)>,
4575 : }
4576 :
4577 : #[cfg(test)]
4578 : impl DeltaLayerTestDesc {
4579 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
4580 2 : Self {
4581 2 : lsn_range,
4582 2 : key_range,
4583 2 : data,
4584 2 : }
4585 2 : }
4586 :
4587 58 : pub fn new_with_inferred_key_range(
4588 58 : lsn_range: Range<Lsn>,
4589 58 : data: Vec<(Key, Lsn, Value)>,
4590 58 : ) -> Self {
4591 144 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
4592 144 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
4593 58 : Self {
4594 58 : key_range: (*key_min)..(key_max.next()),
4595 58 : lsn_range,
4596 58 : data,
4597 58 : }
4598 58 : }
4599 :
4600 10 : pub(crate) fn layer_name(&self) -> LayerName {
4601 10 : LayerName::Delta(super::storage_layer::DeltaLayerName {
4602 10 : key_range: self.key_range.clone(),
4603 10 : lsn_range: self.lsn_range.clone(),
4604 10 : })
4605 10 : }
4606 : }
4607 :
4608 : impl Timeline {
4609 28 : async fn finish_compact_batch(
4610 28 : self: &Arc<Self>,
4611 28 : new_deltas: &[ResidentLayer],
4612 28 : new_images: &[ResidentLayer],
4613 28 : layers_to_remove: &[Layer],
4614 28 : ) -> Result<(), CompactionError> {
4615 28 : let mut guard = tokio::select! {
4616 28 : guard = self.layers.write() => guard,
4617 28 : _ = self.cancel.cancelled() => {
4618 0 : return Err(CompactionError::ShuttingDown);
4619 : }
4620 : };
4621 :
4622 28 : let mut duplicated_layers = HashSet::new();
4623 28 :
4624 28 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
4625 :
4626 336 : for l in new_deltas {
4627 308 : if guard.contains(l.as_ref()) {
4628 : // expected in tests
4629 0 : tracing::error!(layer=%l, "duplicated L1 layer");
4630 :
4631 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
4632 : // `guard` on self.layers. as of writing this, there are no error returns except
4633 : // for compact_level0_phase1 creating an L0, which does not happen in practice
4634 : // because we have not implemented L0 => L0 compaction.
4635 0 : duplicated_layers.insert(l.layer_desc().key());
4636 308 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
4637 0 : return Err(CompactionError::Other(anyhow::anyhow!("compaction generates a L0 layer file as output, which will cause infinite compaction.")));
4638 308 : } else {
4639 308 : insert_layers.push(l.clone());
4640 308 : }
4641 : }
4642 :
4643 : // only remove those inputs which were not outputs
4644 28 : let remove_layers: Vec<Layer> = layers_to_remove
4645 28 : .iter()
4646 402 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
4647 28 : .cloned()
4648 28 : .collect();
4649 28 :
4650 28 : if !new_images.is_empty() {
4651 0 : guard
4652 0 : .open_mut()?
4653 0 : .track_new_image_layers(new_images, &self.metrics);
4654 28 : }
4655 :
4656 28 : guard
4657 28 : .open_mut()?
4658 28 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
4659 28 :
4660 28 : self.remote_client
4661 28 : .schedule_compaction_update(&remove_layers, new_deltas)?;
4662 :
4663 28 : drop_wlock(guard);
4664 28 :
4665 28 : Ok(())
4666 28 : }
4667 :
4668 0 : async fn rewrite_layers(
4669 0 : self: &Arc<Self>,
4670 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
4671 0 : mut drop_layers: Vec<Layer>,
4672 0 : ) -> Result<(), CompactionError> {
4673 0 : let mut guard = self.layers.write().await;
4674 :
4675 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
4676 : // to avoid double-removing, and avoid rewriting something that was removed.
4677 0 : replace_layers.retain(|(l, _)| guard.contains(l));
4678 0 : drop_layers.retain(|l| guard.contains(l));
4679 0 :
4680 0 : guard
4681 0 : .open_mut()?
4682 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
4683 0 :
4684 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
4685 0 :
4686 0 : self.remote_client
4687 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
4688 :
4689 0 : Ok(())
4690 0 : }
4691 :
4692 : /// Schedules the uploads of the given image layers
4693 364 : fn upload_new_image_layers(
4694 364 : self: &Arc<Self>,
4695 364 : new_images: impl IntoIterator<Item = ResidentLayer>,
4696 364 : ) -> Result<(), super::upload_queue::NotInitialized> {
4697 390 : for layer in new_images {
4698 26 : self.remote_client.schedule_layer_file_upload(layer)?;
4699 : }
4700 : // should any new image layer been created, not uploading index_part will
4701 : // result in a mismatch between remote_physical_size and layermap calculated
4702 : // size, which will fail some tests, but should not be an issue otherwise.
4703 364 : self.remote_client
4704 364 : .schedule_index_upload_for_file_changes()?;
4705 364 : Ok(())
4706 364 : }
4707 :
4708 : /// Find the Lsns above which layer files need to be retained on
4709 : /// garbage collection.
4710 : ///
4711 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
4712 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
4713 : /// the space-based retention.
4714 : ///
4715 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
4716 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
4717 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
4718 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
4719 : /// in the response as the GC cutoff point for the timeline.
4720 4 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
4721 : pub(super) async fn find_gc_cutoffs(
4722 : &self,
4723 : space_cutoff: Lsn,
4724 : pitr: Duration,
4725 : cancel: &CancellationToken,
4726 : ctx: &RequestContext,
4727 : ) -> Result<GcCutoffs, PageReconstructError> {
4728 : let _timer = self
4729 : .metrics
4730 : .find_gc_cutoffs_histo
4731 : .start_timer()
4732 : .record_on_drop();
4733 :
4734 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
4735 :
4736 : if cfg!(test) {
4737 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
4738 : if pitr == Duration::ZERO {
4739 : return Ok(GcCutoffs {
4740 : time: self.get_last_record_lsn(),
4741 : space: space_cutoff,
4742 : });
4743 : }
4744 : }
4745 :
4746 : // Calculate a time-based limit on how much to retain:
4747 : // - if PITR interval is set, then this is our cutoff.
4748 : // - if PITR interval is not set, then we do a lookup
4749 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
4750 : let time_cutoff = {
4751 : let now = SystemTime::now();
4752 : let time_range = if pitr == Duration::ZERO {
4753 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
4754 : } else {
4755 : pitr
4756 : };
4757 :
4758 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
4759 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
4760 : let timestamp = to_pg_timestamp(time_cutoff);
4761 :
4762 : match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
4763 : LsnForTimestamp::Present(lsn) => Some(lsn),
4764 : LsnForTimestamp::Future(lsn) => {
4765 : // The timestamp is in the future. That sounds impossible,
4766 : // but what it really means is that there hasn't been
4767 : // any commits since the cutoff timestamp.
4768 : //
4769 : // In this case we should use the LSN of the most recent commit,
4770 : // which is implicitly the last LSN in the log.
4771 : debug!("future({})", lsn);
4772 : Some(self.get_last_record_lsn())
4773 : }
4774 : LsnForTimestamp::Past(lsn) => {
4775 : debug!("past({})", lsn);
4776 : None
4777 : }
4778 : LsnForTimestamp::NoData(lsn) => {
4779 : debug!("nodata({})", lsn);
4780 : None
4781 : }
4782 : }
4783 : };
4784 :
4785 : Ok(match (pitr, time_cutoff) {
4786 : (Duration::ZERO, Some(time_cutoff)) => {
4787 : // PITR is not set. Retain the size-based limit, or the default time retention,
4788 : // whichever requires less data.
4789 : GcCutoffs {
4790 : time: self.get_last_record_lsn(),
4791 : space: std::cmp::max(time_cutoff, space_cutoff),
4792 : }
4793 : }
4794 : (Duration::ZERO, None) => {
4795 : // PITR is not set, and time lookup failed
4796 : GcCutoffs {
4797 : time: self.get_last_record_lsn(),
4798 : space: space_cutoff,
4799 : }
4800 : }
4801 : (_, None) => {
4802 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
4803 : // cannot advance beyond what was already GC'd, and respect space-based retention
4804 : GcCutoffs {
4805 : time: *self.get_latest_gc_cutoff_lsn(),
4806 : space: space_cutoff,
4807 : }
4808 : }
4809 : (_, Some(time_cutoff)) => {
4810 : // PITR interval is set and we looked up timestamp successfully. Ignore
4811 : // size based retention and make time cutoff authoritative
4812 : GcCutoffs {
4813 : time: time_cutoff,
4814 : space: time_cutoff,
4815 : }
4816 : }
4817 : })
4818 : }
4819 :
4820 : /// Garbage collect layer files on a timeline that are no longer needed.
4821 : ///
4822 : /// Currently, we don't make any attempt at removing unneeded page versions
4823 : /// within a layer file. We can only remove the whole file if it's fully
4824 : /// obsolete.
4825 4 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
4826 : // this is most likely the background tasks, but it might be the spawned task from
4827 : // immediate_gc
4828 4 : let _g = tokio::select! {
4829 4 : guard = self.gc_lock.lock() => guard,
4830 4 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
4831 : };
4832 4 : let timer = self.metrics.garbage_collect_histo.start_timer();
4833 4 :
4834 4 : fail_point!("before-timeline-gc");
4835 4 :
4836 4 : // Is the timeline being deleted?
4837 4 : if self.is_stopping() {
4838 0 : return Err(GcError::TimelineCancelled);
4839 4 : }
4840 4 :
4841 4 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
4842 4 : let gc_info = self.gc_info.read().unwrap();
4843 4 :
4844 4 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
4845 4 : let time_cutoff = gc_info.cutoffs.time;
4846 4 : let retain_lsns = gc_info
4847 4 : .retain_lsns
4848 4 : .iter()
4849 4 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
4850 4 : .collect();
4851 4 :
4852 4 : // Gets the maximum LSN that holds the valid lease.
4853 4 : //
4854 4 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
4855 4 : // Here, we do not check for stale leases again.
4856 4 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
4857 4 :
4858 4 : (
4859 4 : space_cutoff,
4860 4 : time_cutoff,
4861 4 : retain_lsns,
4862 4 : max_lsn_with_valid_lease,
4863 4 : )
4864 4 : };
4865 4 :
4866 4 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
4867 4 : let standby_horizon = self.standby_horizon.load();
4868 4 : // Hold GC for the standby, but as a safety guard do it only within some
4869 4 : // reasonable lag.
4870 4 : if standby_horizon != Lsn::INVALID {
4871 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
4872 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
4873 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
4874 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
4875 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
4876 : } else {
4877 0 : warn!(
4878 0 : "standby is lagging for more than {}MB, not holding gc for it",
4879 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
4880 : )
4881 : }
4882 0 : }
4883 4 : }
4884 :
4885 : // Reset standby horizon to ignore it if it is not updated till next GC.
4886 : // It is an easy way to unset it when standby disappears without adding
4887 : // more conf options.
4888 4 : self.standby_horizon.store(Lsn::INVALID);
4889 4 : self.metrics
4890 4 : .standby_horizon_gauge
4891 4 : .set(Lsn::INVALID.0 as i64);
4892 :
4893 4 : let res = self
4894 4 : .gc_timeline(
4895 4 : space_cutoff,
4896 4 : time_cutoff,
4897 4 : retain_lsns,
4898 4 : max_lsn_with_valid_lease,
4899 4 : new_gc_cutoff,
4900 4 : )
4901 4 : .instrument(
4902 4 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
4903 : )
4904 0 : .await?;
4905 :
4906 : // only record successes
4907 4 : timer.stop_and_record();
4908 4 :
4909 4 : Ok(res)
4910 4 : }
4911 :
4912 4 : async fn gc_timeline(
4913 4 : &self,
4914 4 : space_cutoff: Lsn,
4915 4 : time_cutoff: Lsn,
4916 4 : retain_lsns: Vec<Lsn>,
4917 4 : max_lsn_with_valid_lease: Option<Lsn>,
4918 4 : new_gc_cutoff: Lsn,
4919 4 : ) -> Result<GcResult, GcError> {
4920 4 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
4921 4 :
4922 4 : let now = SystemTime::now();
4923 4 : let mut result: GcResult = GcResult::default();
4924 4 :
4925 4 : // Nothing to GC. Return early.
4926 4 : let latest_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
4927 4 : if latest_gc_cutoff >= new_gc_cutoff {
4928 0 : info!(
4929 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
4930 : );
4931 0 : return Ok(result);
4932 4 : }
4933 :
4934 : // We need to ensure that no one tries to read page versions or create
4935 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
4936 : // for details. This will block until the old value is no longer in use.
4937 : //
4938 : // The GC cutoff should only ever move forwards.
4939 4 : let waitlist = {
4940 4 : let write_guard = self.latest_gc_cutoff_lsn.lock_for_write();
4941 4 : if *write_guard > new_gc_cutoff {
4942 0 : return Err(GcError::BadLsn {
4943 0 : why: format!(
4944 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
4945 0 : *write_guard, new_gc_cutoff
4946 0 : ),
4947 0 : });
4948 4 : }
4949 4 :
4950 4 : write_guard.store_and_unlock(new_gc_cutoff)
4951 4 : };
4952 4 : waitlist.wait().await;
4953 :
4954 4 : info!("GC starting");
4955 :
4956 4 : debug!("retain_lsns: {:?}", retain_lsns);
4957 :
4958 4 : let mut layers_to_remove = Vec::new();
4959 :
4960 : // Scan all layers in the timeline (remote or on-disk).
4961 : //
4962 : // Garbage collect the layer if all conditions are satisfied:
4963 : // 1. it is older than cutoff LSN;
4964 : // 2. it is older than PITR interval;
4965 : // 3. it doesn't need to be retained for 'retain_lsns';
4966 : // 4. it does not need to be kept for LSNs holding valid leases.
4967 : // 5. newer on-disk image layers cover the layer's whole key range
4968 : //
4969 : // TODO holding a write lock is too agressive and avoidable
4970 4 : let mut guard = self.layers.write().await;
4971 4 : let layers = guard.layer_map()?;
4972 24 : 'outer: for l in layers.iter_historic_layers() {
4973 24 : result.layers_total += 1;
4974 24 :
4975 24 : // 1. Is it newer than GC horizon cutoff point?
4976 24 : if l.get_lsn_range().end > space_cutoff {
4977 2 : debug!(
4978 0 : "keeping {} because it's newer than space_cutoff {}",
4979 0 : l.layer_name(),
4980 : space_cutoff,
4981 : );
4982 2 : result.layers_needed_by_cutoff += 1;
4983 2 : continue 'outer;
4984 22 : }
4985 22 :
4986 22 : // 2. It is newer than PiTR cutoff point?
4987 22 : if l.get_lsn_range().end > time_cutoff {
4988 0 : debug!(
4989 0 : "keeping {} because it's newer than time_cutoff {}",
4990 0 : l.layer_name(),
4991 : time_cutoff,
4992 : );
4993 0 : result.layers_needed_by_pitr += 1;
4994 0 : continue 'outer;
4995 22 : }
4996 :
4997 : // 3. Is it needed by a child branch?
4998 : // NOTE With that we would keep data that
4999 : // might be referenced by child branches forever.
5000 : // We can track this in child timeline GC and delete parent layers when
5001 : // they are no longer needed. This might be complicated with long inheritance chains.
5002 : //
5003 : // TODO Vec is not a great choice for `retain_lsns`
5004 22 : for retain_lsn in &retain_lsns {
5005 : // start_lsn is inclusive
5006 0 : if &l.get_lsn_range().start <= retain_lsn {
5007 0 : debug!(
5008 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
5009 0 : l.layer_name(),
5010 0 : retain_lsn,
5011 0 : l.is_incremental(),
5012 : );
5013 0 : result.layers_needed_by_branches += 1;
5014 0 : continue 'outer;
5015 0 : }
5016 : }
5017 :
5018 : // 4. Is there a valid lease that requires us to keep this layer?
5019 22 : if let Some(lsn) = &max_lsn_with_valid_lease {
5020 : // keep if layer start <= any of the lease
5021 18 : if &l.get_lsn_range().start <= lsn {
5022 14 : debug!(
5023 0 : "keeping {} because there is a valid lease preventing GC at {}",
5024 0 : l.layer_name(),
5025 : lsn,
5026 : );
5027 14 : result.layers_needed_by_leases += 1;
5028 14 : continue 'outer;
5029 4 : }
5030 4 : }
5031 :
5032 : // 5. Is there a later on-disk layer for this relation?
5033 : //
5034 : // The end-LSN is exclusive, while disk_consistent_lsn is
5035 : // inclusive. For example, if disk_consistent_lsn is 100, it is
5036 : // OK for a delta layer to have end LSN 101, but if the end LSN
5037 : // is 102, then it might not have been fully flushed to disk
5038 : // before crash.
5039 : //
5040 : // For example, imagine that the following layers exist:
5041 : //
5042 : // 1000 - image (A)
5043 : // 1000-2000 - delta (B)
5044 : // 2000 - image (C)
5045 : // 2000-3000 - delta (D)
5046 : // 3000 - image (E)
5047 : //
5048 : // If GC horizon is at 2500, we can remove layers A and B, but
5049 : // we cannot remove C, even though it's older than 2500, because
5050 : // the delta layer 2000-3000 depends on it.
5051 8 : if !layers
5052 8 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
5053 : {
5054 6 : debug!("keeping {} because it is the latest layer", l.layer_name());
5055 6 : result.layers_not_updated += 1;
5056 6 : continue 'outer;
5057 2 : }
5058 2 :
5059 2 : // We didn't find any reason to keep this file, so remove it.
5060 2 : debug!(
5061 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
5062 0 : l.layer_name(),
5063 0 : l.is_incremental(),
5064 : );
5065 2 : layers_to_remove.push(l);
5066 : }
5067 :
5068 4 : if !layers_to_remove.is_empty() {
5069 : // Persist the new GC cutoff value before we actually remove anything.
5070 : // This unconditionally schedules also an index_part.json update, even though, we will
5071 : // be doing one a bit later with the unlinked gc'd layers.
5072 2 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
5073 2 : self.schedule_uploads(disk_consistent_lsn, None)
5074 2 : .map_err(|e| {
5075 0 : if self.cancel.is_cancelled() {
5076 0 : GcError::TimelineCancelled
5077 : } else {
5078 0 : GcError::Remote(e)
5079 : }
5080 2 : })?;
5081 :
5082 2 : let gc_layers = layers_to_remove
5083 2 : .iter()
5084 2 : .map(|x| guard.get_from_desc(x))
5085 2 : .collect::<Vec<Layer>>();
5086 2 :
5087 2 : result.layers_removed = gc_layers.len() as u64;
5088 2 :
5089 2 : self.remote_client.schedule_gc_update(&gc_layers)?;
5090 :
5091 2 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
5092 2 :
5093 2 : #[cfg(feature = "testing")]
5094 2 : {
5095 2 : result.doomed_layers = gc_layers;
5096 2 : }
5097 2 : }
5098 :
5099 4 : info!(
5100 0 : "GC completed removing {} layers, cutoff {}",
5101 : result.layers_removed, new_gc_cutoff
5102 : );
5103 :
5104 4 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
5105 4 : Ok(result)
5106 4 : }
5107 :
5108 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
5109 667163 : async fn reconstruct_value(
5110 667163 : &self,
5111 667163 : key: Key,
5112 667163 : request_lsn: Lsn,
5113 667163 : mut data: ValueReconstructState,
5114 667163 : ) -> Result<Bytes, PageReconstructError> {
5115 667163 : // Perform WAL redo if needed
5116 667163 : data.records.reverse();
5117 667163 :
5118 667163 : // If we have a page image, and no WAL, we're all set
5119 667163 : if data.records.is_empty() {
5120 666821 : if let Some((img_lsn, img)) = &data.img {
5121 666821 : trace!(
5122 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
5123 : key,
5124 : img_lsn,
5125 : request_lsn,
5126 : );
5127 666821 : Ok(img.clone())
5128 : } else {
5129 0 : Err(PageReconstructError::from(anyhow!(
5130 0 : "base image for {key} at {request_lsn} not found"
5131 0 : )))
5132 : }
5133 : } else {
5134 : // We need to do WAL redo.
5135 : //
5136 : // If we don't have a base image, then the oldest WAL record better initialize
5137 : // the page
5138 342 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
5139 0 : Err(PageReconstructError::from(anyhow!(
5140 0 : "Base image for {} at {} not found, but got {} WAL records",
5141 0 : key,
5142 0 : request_lsn,
5143 0 : data.records.len()
5144 0 : )))
5145 : } else {
5146 342 : if data.img.is_some() {
5147 338 : trace!(
5148 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
5149 0 : data.records.len(),
5150 : key,
5151 : request_lsn
5152 : );
5153 : } else {
5154 4 : trace!("found {} WAL records that will init the page for {} at {}, performing WAL redo", data.records.len(), key, request_lsn);
5155 : };
5156 342 : let res = self
5157 342 : .walredo_mgr
5158 342 : .as_ref()
5159 342 : .context("timeline has no walredo manager")
5160 342 : .map_err(PageReconstructError::WalRedo)?
5161 342 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
5162 0 : .await;
5163 342 : let img = match res {
5164 342 : Ok(img) => img,
5165 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
5166 0 : Err(walredo::Error::Other(e)) => {
5167 0 : return Err(PageReconstructError::WalRedo(
5168 0 : e.context("reconstruct a page image"),
5169 0 : ))
5170 : }
5171 : };
5172 342 : Ok(img)
5173 : }
5174 : }
5175 667163 : }
5176 :
5177 0 : pub(crate) async fn spawn_download_all_remote_layers(
5178 0 : self: Arc<Self>,
5179 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5180 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
5181 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5182 :
5183 : // this is not really needed anymore; it has tests which really check the return value from
5184 : // http api. it would be better not to maintain this anymore.
5185 :
5186 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
5187 0 : if let Some(st) = &*status_guard {
5188 0 : match &st.state {
5189 : DownloadRemoteLayersTaskState::Running => {
5190 0 : return Err(st.clone());
5191 : }
5192 : DownloadRemoteLayersTaskState::ShutDown
5193 0 : | DownloadRemoteLayersTaskState::Completed => {
5194 0 : *status_guard = None;
5195 0 : }
5196 : }
5197 0 : }
5198 :
5199 0 : let self_clone = Arc::clone(&self);
5200 0 : let task_id = task_mgr::spawn(
5201 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
5202 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
5203 0 : self.tenant_shard_id,
5204 0 : Some(self.timeline_id),
5205 0 : "download all remote layers task",
5206 0 : async move {
5207 0 : self_clone.download_all_remote_layers(request).await;
5208 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
5209 0 : match &mut *status_guard {
5210 : None => {
5211 0 : warn!("tasks status is supposed to be Some(), since we are running");
5212 : }
5213 0 : Some(st) => {
5214 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
5215 0 : if st.task_id != exp_task_id {
5216 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
5217 0 : } else {
5218 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5219 0 : }
5220 : }
5221 : };
5222 0 : Ok(())
5223 0 : }
5224 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))
5225 : );
5226 :
5227 0 : let initial_info = DownloadRemoteLayersTaskInfo {
5228 0 : task_id: format!("{task_id}"),
5229 0 : state: DownloadRemoteLayersTaskState::Running,
5230 0 : total_layer_count: 0,
5231 0 : successful_download_count: 0,
5232 0 : failed_download_count: 0,
5233 0 : };
5234 0 : *status_guard = Some(initial_info.clone());
5235 0 :
5236 0 : Ok(initial_info)
5237 0 : }
5238 :
5239 0 : async fn download_all_remote_layers(
5240 0 : self: &Arc<Self>,
5241 0 : request: DownloadRemoteLayersTaskSpawnRequest,
5242 0 : ) {
5243 : use pageserver_api::models::DownloadRemoteLayersTaskState;
5244 :
5245 0 : let remaining = {
5246 0 : let guard = self.layers.read().await;
5247 0 : let Ok(lm) = guard.layer_map() else {
5248 : // technically here we could look into iterating accessible layers, but downloading
5249 : // all layers of a shutdown timeline makes no sense regardless.
5250 0 : tracing::info!("attempted to download all layers of shutdown timeline");
5251 0 : return;
5252 : };
5253 0 : lm.iter_historic_layers()
5254 0 : .map(|desc| guard.get_from_desc(&desc))
5255 0 : .collect::<Vec<_>>()
5256 0 : };
5257 0 : let total_layer_count = remaining.len();
5258 :
5259 : macro_rules! lock_status {
5260 : ($st:ident) => {
5261 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
5262 : let st = st
5263 : .as_mut()
5264 : .expect("this function is only called after the task has been spawned");
5265 : assert_eq!(
5266 : st.task_id,
5267 : format!(
5268 : "{}",
5269 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
5270 : )
5271 : );
5272 : let $st = st;
5273 : };
5274 : }
5275 :
5276 : {
5277 0 : lock_status!(st);
5278 0 : st.total_layer_count = total_layer_count as u64;
5279 0 : }
5280 0 :
5281 0 : let mut remaining = remaining.into_iter();
5282 0 : let mut have_remaining = true;
5283 0 : let mut js = tokio::task::JoinSet::new();
5284 0 :
5285 0 : let cancel = task_mgr::shutdown_token();
5286 0 :
5287 0 : let limit = request.max_concurrent_downloads;
5288 :
5289 : loop {
5290 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
5291 0 : let Some(next) = remaining.next() else {
5292 0 : have_remaining = false;
5293 0 : break;
5294 : };
5295 :
5296 0 : let span = tracing::info_span!("download", layer = %next);
5297 :
5298 0 : js.spawn(
5299 0 : async move {
5300 0 : let res = next.download().await;
5301 0 : (next, res)
5302 0 : }
5303 0 : .instrument(span),
5304 0 : );
5305 0 : }
5306 :
5307 0 : while let Some(res) = js.join_next().await {
5308 0 : match res {
5309 : Ok((_, Ok(_))) => {
5310 0 : lock_status!(st);
5311 0 : st.successful_download_count += 1;
5312 : }
5313 0 : Ok((layer, Err(e))) => {
5314 0 : tracing::error!(%layer, "download failed: {e:#}");
5315 0 : lock_status!(st);
5316 0 : st.failed_download_count += 1;
5317 : }
5318 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
5319 0 : Err(je) if je.is_panic() => {
5320 0 : lock_status!(st);
5321 0 : st.failed_download_count += 1;
5322 : }
5323 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
5324 : }
5325 : }
5326 :
5327 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
5328 0 : break;
5329 0 : }
5330 : }
5331 :
5332 : {
5333 0 : lock_status!(st);
5334 0 : st.state = DownloadRemoteLayersTaskState::Completed;
5335 : }
5336 0 : }
5337 :
5338 0 : pub(crate) fn get_download_all_remote_layers_task_info(
5339 0 : &self,
5340 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
5341 0 : self.download_all_remote_layers_task_info
5342 0 : .read()
5343 0 : .unwrap()
5344 0 : .clone()
5345 0 : }
5346 : }
5347 :
5348 : impl Timeline {
5349 : /// Returns non-remote layers for eviction.
5350 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
5351 0 : let guard = self.layers.read().await;
5352 0 : let mut max_layer_size: Option<u64> = None;
5353 0 :
5354 0 : let resident_layers = guard
5355 0 : .likely_resident_layers()
5356 0 : .map(|layer| {
5357 0 : let file_size = layer.layer_desc().file_size;
5358 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
5359 0 :
5360 0 : let last_activity_ts = layer.latest_activity();
5361 0 :
5362 0 : EvictionCandidate {
5363 0 : layer: layer.to_owned().into(),
5364 0 : last_activity_ts,
5365 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
5366 0 : visibility: layer.visibility(),
5367 0 : }
5368 0 : })
5369 0 : .collect();
5370 0 :
5371 0 : DiskUsageEvictionInfo {
5372 0 : max_layer_size,
5373 0 : resident_layers,
5374 0 : }
5375 0 : }
5376 :
5377 1690 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
5378 1690 : ShardIndex {
5379 1690 : shard_number: self.tenant_shard_id.shard_number,
5380 1690 : shard_count: self.tenant_shard_id.shard_count,
5381 1690 : }
5382 1690 : }
5383 :
5384 : /// Persistently blocks gc for `Manual` reason.
5385 : ///
5386 : /// Returns true if no such block existed before, false otherwise.
5387 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
5388 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5389 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5390 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
5391 0 : }
5392 :
5393 : /// Persistently unblocks gc for `Manual` reason.
5394 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
5395 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
5396 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
5397 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
5398 0 : }
5399 :
5400 : #[cfg(test)]
5401 36 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
5402 36 : self.last_record_lsn.advance(new_lsn);
5403 36 : }
5404 :
5405 : #[cfg(test)]
5406 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
5407 2 : self.disk_consistent_lsn.store(new_value);
5408 2 : }
5409 :
5410 : /// Force create an image layer and place it into the layer map.
5411 : ///
5412 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5413 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5414 : /// placed into the layer map in one run AND be validated.
5415 : #[cfg(test)]
5416 50 : pub(super) async fn force_create_image_layer(
5417 50 : self: &Arc<Timeline>,
5418 50 : lsn: Lsn,
5419 50 : mut images: Vec<(Key, Bytes)>,
5420 50 : check_start_lsn: Option<Lsn>,
5421 50 : ctx: &RequestContext,
5422 50 : ) -> anyhow::Result<()> {
5423 50 : let last_record_lsn = self.get_last_record_lsn();
5424 50 : assert!(
5425 50 : lsn <= last_record_lsn,
5426 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
5427 : );
5428 50 : if let Some(check_start_lsn) = check_start_lsn {
5429 50 : assert!(lsn >= check_start_lsn);
5430 0 : }
5431 108 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
5432 50 : let min_key = *images.first().map(|(k, _)| k).unwrap();
5433 50 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
5434 50 : let mut image_layer_writer = ImageLayerWriter::new(
5435 50 : self.conf,
5436 50 : self.timeline_id,
5437 50 : self.tenant_shard_id,
5438 50 : &(min_key..end_key),
5439 50 : lsn,
5440 50 : ctx,
5441 50 : )
5442 25 : .await?;
5443 208 : for (key, img) in images {
5444 158 : image_layer_writer.put_image(key, img, ctx).await?;
5445 : }
5446 100 : let (desc, path) = image_layer_writer.finish(ctx).await?;
5447 50 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5448 50 : info!("force created image layer {}", image_layer.local_path());
5449 : {
5450 50 : let mut guard = self.layers.write().await;
5451 50 : guard.open_mut().unwrap().force_insert_layer(image_layer);
5452 50 : }
5453 50 :
5454 50 : Ok(())
5455 50 : }
5456 :
5457 : /// Force create a delta layer and place it into the layer map.
5458 : ///
5459 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
5460 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
5461 : /// placed into the layer map in one run AND be validated.
5462 : #[cfg(test)]
5463 60 : pub(super) async fn force_create_delta_layer(
5464 60 : self: &Arc<Timeline>,
5465 60 : mut deltas: DeltaLayerTestDesc,
5466 60 : check_start_lsn: Option<Lsn>,
5467 60 : ctx: &RequestContext,
5468 60 : ) -> anyhow::Result<()> {
5469 60 : let last_record_lsn = self.get_last_record_lsn();
5470 60 : deltas
5471 60 : .data
5472 86 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
5473 60 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
5474 60 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
5475 206 : for (_, lsn, _) in &deltas.data {
5476 146 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
5477 : }
5478 60 : assert!(
5479 60 : deltas.lsn_range.end <= last_record_lsn,
5480 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
5481 : deltas.lsn_range.end,
5482 : last_record_lsn
5483 : );
5484 60 : if let Some(check_start_lsn) = check_start_lsn {
5485 60 : assert!(deltas.lsn_range.start >= check_start_lsn);
5486 0 : }
5487 60 : let mut delta_layer_writer = DeltaLayerWriter::new(
5488 60 : self.conf,
5489 60 : self.timeline_id,
5490 60 : self.tenant_shard_id,
5491 60 : deltas.key_range.start,
5492 60 : deltas.lsn_range,
5493 60 : ctx,
5494 60 : )
5495 30 : .await?;
5496 206 : for (key, lsn, val) in deltas.data {
5497 146 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
5498 : }
5499 150 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
5500 60 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
5501 60 : info!("force created delta layer {}", delta_layer.local_path());
5502 : {
5503 60 : let mut guard = self.layers.write().await;
5504 60 : guard.open_mut().unwrap().force_insert_layer(delta_layer);
5505 60 : }
5506 60 :
5507 60 : Ok(())
5508 60 : }
5509 :
5510 : /// Return all keys at the LSN in the image layers
5511 : #[cfg(test)]
5512 6 : pub(crate) async fn inspect_image_layers(
5513 6 : self: &Arc<Timeline>,
5514 6 : lsn: Lsn,
5515 6 : ctx: &RequestContext,
5516 6 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
5517 6 : let mut all_data = Vec::new();
5518 6 : let guard = self.layers.read().await;
5519 34 : for layer in guard.layer_map()?.iter_historic_layers() {
5520 34 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
5521 8 : let layer = guard.get_from_desc(&layer);
5522 8 : let mut reconstruct_data = ValuesReconstructState::default();
5523 8 : layer
5524 8 : .get_values_reconstruct_data(
5525 8 : KeySpace::single(Key::MIN..Key::MAX),
5526 8 : lsn..Lsn(lsn.0 + 1),
5527 8 : &mut reconstruct_data,
5528 8 : ctx,
5529 8 : )
5530 13 : .await?;
5531 74 : for (k, v) in reconstruct_data.keys {
5532 66 : all_data.push((k, v?.img.unwrap().1));
5533 : }
5534 26 : }
5535 : }
5536 6 : all_data.sort();
5537 6 : Ok(all_data)
5538 6 : }
5539 :
5540 : /// Get all historic layer descriptors in the layer map
5541 : #[cfg(test)]
5542 2 : pub(crate) async fn inspect_historic_layers(
5543 2 : self: &Arc<Timeline>,
5544 2 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
5545 2 : let mut layers = Vec::new();
5546 2 : let guard = self.layers.read().await;
5547 6 : for layer in guard.layer_map()?.iter_historic_layers() {
5548 6 : layers.push(layer.key());
5549 6 : }
5550 2 : Ok(layers)
5551 2 : }
5552 :
5553 : #[cfg(test)]
5554 10 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
5555 10 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
5556 10 : keyspace.merge(&ks);
5557 10 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
5558 10 : }
5559 : }
5560 :
5561 : /// Tracking writes ingestion does to a particular in-memory layer.
5562 : ///
5563 : /// Cleared upon freezing a layer.
5564 : pub(crate) struct TimelineWriterState {
5565 : open_layer: Arc<InMemoryLayer>,
5566 : current_size: u64,
5567 : // Previous Lsn which passed through
5568 : prev_lsn: Option<Lsn>,
5569 : // Largest Lsn which passed through the current writer
5570 : max_lsn: Option<Lsn>,
5571 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
5572 : cached_last_freeze_at: Lsn,
5573 : }
5574 :
5575 : impl TimelineWriterState {
5576 1262 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
5577 1262 : Self {
5578 1262 : open_layer,
5579 1262 : current_size,
5580 1262 : prev_lsn: None,
5581 1262 : max_lsn: None,
5582 1262 : cached_last_freeze_at: last_freeze_at,
5583 1262 : }
5584 1262 : }
5585 : }
5586 :
5587 : /// Various functions to mutate the timeline.
5588 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
5589 : // This is probably considered a bad practice in Rust and should be fixed eventually,
5590 : // but will cause large code changes.
5591 : pub(crate) struct TimelineWriter<'a> {
5592 : tl: &'a Timeline,
5593 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
5594 : }
5595 :
5596 : impl Deref for TimelineWriter<'_> {
5597 : type Target = Timeline;
5598 :
5599 4807156 : fn deref(&self) -> &Self::Target {
5600 4807156 : self.tl
5601 4807156 : }
5602 : }
5603 :
5604 : #[derive(PartialEq)]
5605 : enum OpenLayerAction {
5606 : Roll,
5607 : Open,
5608 : None,
5609 : }
5610 :
5611 : impl<'a> TimelineWriter<'a> {
5612 4804198 : async fn handle_open_layer_action(
5613 4804198 : &mut self,
5614 4804198 : at: Lsn,
5615 4804198 : action: OpenLayerAction,
5616 4804198 : ctx: &RequestContext,
5617 4804198 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
5618 4804198 : match action {
5619 : OpenLayerAction::Roll => {
5620 80 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
5621 80 : self.roll_layer(freeze_at).await?;
5622 80 : self.open_layer(at, ctx).await?;
5623 : }
5624 1182 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
5625 : OpenLayerAction::None => {
5626 4802936 : assert!(self.write_guard.is_some());
5627 : }
5628 : }
5629 :
5630 4804198 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
5631 4804198 : }
5632 :
5633 1262 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
5634 1262 : let layer = self
5635 1262 : .tl
5636 1262 : .get_layer_for_write(at, &self.write_guard, ctx)
5637 713 : .await?;
5638 1262 : let initial_size = layer.size().await?;
5639 :
5640 1262 : let last_freeze_at = self.last_freeze_at.load();
5641 1262 : self.write_guard.replace(TimelineWriterState::new(
5642 1262 : layer,
5643 1262 : initial_size,
5644 1262 : last_freeze_at,
5645 1262 : ));
5646 1262 :
5647 1262 : Ok(())
5648 1262 : }
5649 :
5650 80 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
5651 80 : let current_size = self.write_guard.as_ref().unwrap().current_size;
5652 80 :
5653 80 : // self.write_guard will be taken by the freezing
5654 80 : self.tl
5655 80 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
5656 3 : .await?;
5657 :
5658 80 : assert!(self.write_guard.is_none());
5659 :
5660 80 : if current_size >= self.get_checkpoint_distance() * 2 {
5661 0 : warn!("Flushed oversized open layer with size {}", current_size)
5662 80 : }
5663 :
5664 80 : Ok(())
5665 80 : }
5666 :
5667 4804198 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
5668 4804198 : let state = &*self.write_guard;
5669 4804198 : let Some(state) = &state else {
5670 1182 : return OpenLayerAction::Open;
5671 : };
5672 :
5673 : #[cfg(feature = "testing")]
5674 4803016 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
5675 : // this check and assertion are not really needed because
5676 : // LayerManager::try_freeze_in_memory_layer will always clear out the
5677 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
5678 : // is no TimelineWriterState.
5679 0 : assert!(
5680 0 : state.open_layer.end_lsn.get().is_some(),
5681 0 : "our open_layer must be outdated"
5682 : );
5683 :
5684 : // this would be a memory leak waiting to happen because the in-memory layer always has
5685 : // an index
5686 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
5687 4803016 : }
5688 4803016 :
5689 4803016 : if state.prev_lsn == Some(lsn) {
5690 : // Rolling mid LSN is not supported by [downstream code].
5691 : // Hence, only roll at LSN boundaries.
5692 : //
5693 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
5694 6 : return OpenLayerAction::None;
5695 4803010 : }
5696 4803010 :
5697 4803010 : if state.current_size == 0 {
5698 : // Don't roll empty layers
5699 0 : return OpenLayerAction::None;
5700 4803010 : }
5701 4803010 :
5702 4803010 : if self.tl.should_roll(
5703 4803010 : state.current_size,
5704 4803010 : state.current_size + new_value_size,
5705 4803010 : self.get_checkpoint_distance(),
5706 4803010 : lsn,
5707 4803010 : state.cached_last_freeze_at,
5708 4803010 : state.open_layer.get_opened_at(),
5709 4803010 : ) {
5710 80 : OpenLayerAction::Roll
5711 : } else {
5712 4802930 : OpenLayerAction::None
5713 : }
5714 4804198 : }
5715 :
5716 : /// Put a batch of keys at the specified Lsns.
5717 4804196 : pub(crate) async fn put_batch(
5718 4804196 : &mut self,
5719 4804196 : batch: Vec<(CompactKey, Lsn, usize, Value)>,
5720 4804196 : ctx: &RequestContext,
5721 4804196 : ) -> anyhow::Result<()> {
5722 4804196 : if batch.is_empty() {
5723 0 : return Ok(());
5724 4804196 : }
5725 :
5726 4804196 : let serialized_batch = inmemory_layer::SerializedBatch::from_values(batch)?;
5727 4804196 : let batch_max_lsn = serialized_batch.max_lsn;
5728 4804196 : let buf_size: u64 = serialized_batch.raw.len() as u64;
5729 4804196 :
5730 4804196 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
5731 4804196 : let layer = self
5732 4804196 : .handle_open_layer_action(batch_max_lsn, action, ctx)
5733 716 : .await?;
5734 :
5735 4804196 : let res = layer.put_batch(serialized_batch, ctx).await;
5736 :
5737 4804196 : if res.is_ok() {
5738 4804196 : // Update the current size only when the entire write was ok.
5739 4804196 : // In case of failures, we may have had partial writes which
5740 4804196 : // render the size tracking out of sync. That's ok because
5741 4804196 : // the checkpoint distance should be significantly smaller
5742 4804196 : // than the S3 single shot upload limit of 5GiB.
5743 4804196 : let state = self.write_guard.as_mut().unwrap();
5744 4804196 :
5745 4804196 : state.current_size += buf_size;
5746 4804196 : state.prev_lsn = Some(batch_max_lsn);
5747 4804196 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
5748 4804196 : }
5749 :
5750 4804196 : res
5751 4804196 : }
5752 :
5753 : #[cfg(test)]
5754 : /// Test helper, for tests that would like to poke individual values without composing a batch
5755 4390154 : pub(crate) async fn put(
5756 4390154 : &mut self,
5757 4390154 : key: Key,
5758 4390154 : lsn: Lsn,
5759 4390154 : value: &Value,
5760 4390154 : ctx: &RequestContext,
5761 4390154 : ) -> anyhow::Result<()> {
5762 : use utils::bin_ser::BeSer;
5763 4390154 : if !key.is_valid_key_on_write_path() {
5764 0 : bail!(
5765 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
5766 0 : key
5767 0 : );
5768 4390154 : }
5769 4390154 : let val_ser_size = value.serialized_size().unwrap() as usize;
5770 4390154 : self.put_batch(
5771 4390154 : vec![(key.to_compact(), lsn, val_ser_size, value.clone())],
5772 4390154 : ctx,
5773 4390154 : )
5774 4836 : .await
5775 4390154 : }
5776 :
5777 2 : pub(crate) async fn delete_batch(
5778 2 : &mut self,
5779 2 : batch: &[(Range<Key>, Lsn)],
5780 2 : ctx: &RequestContext,
5781 2 : ) -> anyhow::Result<()> {
5782 2 : if let Some((_, lsn)) = batch.first() {
5783 2 : let action = self.get_open_layer_action(*lsn, 0);
5784 2 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
5785 2 : layer.put_tombstones(batch).await?;
5786 0 : }
5787 :
5788 2 : Ok(())
5789 2 : }
5790 :
5791 : /// Track the end of the latest digested WAL record.
5792 : /// Remember the (end of) last valid WAL record remembered in the timeline.
5793 : ///
5794 : /// Call this after you have finished writing all the WAL up to 'lsn'.
5795 : ///
5796 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
5797 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
5798 : /// the latest and can be read.
5799 5279058 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
5800 5279058 : self.tl.finish_write(new_lsn);
5801 5279058 : }
5802 :
5803 270570 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
5804 270570 : self.tl.update_current_logical_size(delta)
5805 270570 : }
5806 : }
5807 :
5808 : // We need TimelineWriter to be send in upcoming conversion of
5809 : // Timeline::layers to tokio::sync::RwLock.
5810 : #[test]
5811 2 : fn is_send() {
5812 2 : fn _assert_send<T: Send>() {}
5813 2 : _assert_send::<TimelineWriter<'_>>();
5814 2 : }
5815 :
5816 : #[cfg(test)]
5817 : mod tests {
5818 : use pageserver_api::key::Key;
5819 : use utils::{id::TimelineId, lsn::Lsn};
5820 :
5821 : use crate::{
5822 : repository::Value,
5823 : tenant::{
5824 : harness::{test_img, TenantHarness},
5825 : layer_map::LayerMap,
5826 : storage_layer::{Layer, LayerName},
5827 : timeline::{DeltaLayerTestDesc, EvictionError},
5828 : Timeline,
5829 : },
5830 : };
5831 :
5832 : #[tokio::test]
5833 2 : async fn test_heatmap_generation() {
5834 2 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
5835 2 :
5836 2 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
5837 2 : Lsn(0x10)..Lsn(0x20),
5838 2 : vec![(
5839 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
5840 2 : Lsn(0x11),
5841 2 : Value::Image(test_img("foo")),
5842 2 : )],
5843 2 : );
5844 2 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
5845 2 : Lsn(0x10)..Lsn(0x20),
5846 2 : vec![(
5847 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
5848 2 : Lsn(0x11),
5849 2 : Value::Image(test_img("foo")),
5850 2 : )],
5851 2 : );
5852 2 : let l0_delta = DeltaLayerTestDesc::new(
5853 2 : Lsn(0x20)..Lsn(0x30),
5854 2 : Key::from_hex("000000000000000000000000000000000000").unwrap()
5855 2 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
5856 2 : vec![(
5857 2 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
5858 2 : Lsn(0x25),
5859 2 : Value::Image(test_img("foo")),
5860 2 : )],
5861 2 : );
5862 2 : let delta_layers = vec![
5863 2 : covered_delta.clone(),
5864 2 : visible_delta.clone(),
5865 2 : l0_delta.clone(),
5866 2 : ];
5867 2 :
5868 2 : let image_layer = (
5869 2 : Lsn(0x40),
5870 2 : vec![(
5871 2 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
5872 2 : test_img("bar"),
5873 2 : )],
5874 2 : );
5875 2 : let image_layers = vec![image_layer];
5876 2 :
5877 10 : let (tenant, ctx) = harness.load().await;
5878 2 : let timeline = tenant
5879 2 : .create_test_timeline_with_layers(
5880 2 : TimelineId::generate(),
5881 2 : Lsn(0x10),
5882 2 : 14,
5883 2 : &ctx,
5884 2 : delta_layers,
5885 2 : image_layers,
5886 2 : Lsn(0x100),
5887 2 : )
5888 29 : .await
5889 2 : .unwrap();
5890 2 :
5891 2 : // Layer visibility is an input to heatmap generation, so refresh it first
5892 2 : timeline.update_layer_visibility().await.unwrap();
5893 2 :
5894 2 : let heatmap = timeline
5895 2 : .generate_heatmap()
5896 2 : .await
5897 2 : .expect("Infallible while timeline is not shut down");
5898 2 :
5899 2 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
5900 2 :
5901 2 : // L0 should come last
5902 2 : assert_eq!(heatmap.layers.last().unwrap().name, l0_delta.layer_name());
5903 2 :
5904 2 : let mut last_lsn = Lsn::MAX;
5905 10 : for layer in heatmap.layers {
5906 2 : // Covered layer should be omitted
5907 8 : assert!(layer.name != covered_delta.layer_name());
5908 2 :
5909 8 : let layer_lsn = match &layer.name {
5910 4 : LayerName::Delta(d) => d.lsn_range.end,
5911 4 : LayerName::Image(i) => i.lsn,
5912 2 : };
5913 2 :
5914 2 : // Apart from L0s, newest Layers should come first
5915 8 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
5916 6 : assert!(layer_lsn <= last_lsn);
5917 6 : last_lsn = layer_lsn;
5918 2 : }
5919 2 : }
5920 2 : }
5921 :
5922 : #[tokio::test]
5923 2 : async fn two_layer_eviction_attempts_at_the_same_time() {
5924 2 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
5925 2 : .await
5926 2 : .unwrap();
5927 2 :
5928 9 : let (tenant, ctx) = harness.load().await;
5929 2 : let timeline = tenant
5930 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
5931 4 : .await
5932 2 : .unwrap();
5933 2 :
5934 2 : let layer = find_some_layer(&timeline).await;
5935 2 : let layer = layer
5936 2 : .keep_resident()
5937 2 : .await
5938 2 : .expect("no download => no downloading errors")
5939 2 : .drop_eviction_guard();
5940 2 :
5941 2 : let forever = std::time::Duration::from_secs(120);
5942 2 :
5943 2 : let first = layer.evict_and_wait(forever);
5944 2 : let second = layer.evict_and_wait(forever);
5945 2 :
5946 2 : let (first, second) = tokio::join!(first, second);
5947 2 :
5948 2 : let res = layer.keep_resident().await;
5949 2 : assert!(res.is_none(), "{res:?}");
5950 2 :
5951 2 : match (first, second) {
5952 2 : (Ok(()), Ok(())) => {
5953 2 : // because there are no more timeline locks being taken on eviction path, we can
5954 2 : // witness all three outcomes here.
5955 2 : }
5956 2 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
5957 0 : // if one completes before the other, this is fine just as well.
5958 0 : }
5959 2 : other => unreachable!("unexpected {:?}", other),
5960 2 : }
5961 2 : }
5962 :
5963 2 : async fn find_some_layer(timeline: &Timeline) -> Layer {
5964 2 : let layers = timeline.layers.read().await;
5965 2 : let desc = layers
5966 2 : .layer_map()
5967 2 : .unwrap()
5968 2 : .iter_historic_layers()
5969 2 : .next()
5970 2 : .expect("must find one layer to evict");
5971 2 :
5972 2 : layers.get_from_desc(&desc)
5973 2 : }
5974 : }
|
