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 heatmap_layers_downloader;
8 : pub(crate) mod import_pgdata;
9 : mod init;
10 : pub mod layer_manager;
11 : pub(crate) mod logical_size;
12 : pub mod offload;
13 : pub mod span;
14 : pub mod uninit;
15 : mod walreceiver;
16 :
17 : use hashlink::LruCache;
18 : use std::array;
19 : use std::cmp::{max, min};
20 : use std::collections::btree_map::Entry;
21 : use std::collections::{BTreeMap, HashMap, HashSet};
22 : use std::ops::{ControlFlow, Deref, Range};
23 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering as AtomicOrdering};
24 : use std::sync::{Arc, Mutex, OnceLock, RwLock, Weak};
25 : use std::time::{Duration, Instant, SystemTime};
26 :
27 : use anyhow::{Context, Result, anyhow, bail, ensure};
28 : use arc_swap::{ArcSwap, ArcSwapOption};
29 : use bytes::Bytes;
30 : use camino::Utf8Path;
31 : use chrono::{DateTime, Utc};
32 : use compaction::{CompactionOutcome, GcCompactionCombinedSettings};
33 : use enumset::EnumSet;
34 : use fail::fail_point;
35 : use futures::stream::FuturesUnordered;
36 : use futures::{FutureExt, StreamExt};
37 : use handle::ShardTimelineId;
38 : use layer_manager::{
39 : LayerManagerLockHolder, LayerManagerReadGuard, LayerManagerWriteGuard, LockedLayerManager,
40 : Shutdown,
41 : };
42 :
43 : use once_cell::sync::Lazy;
44 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
45 : use pageserver_api::key::{
46 : KEY_SIZE, Key, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
47 : SPARSE_RANGE,
48 : };
49 : use pageserver_api::keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning};
50 : use pageserver_api::models::{
51 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
52 : DetachBehavior, DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest,
53 : EvictionPolicy, InMemoryLayerInfo, LayerMapInfo, LsnLease, PageTraceEvent, RelSizeMigration,
54 : TimelineState,
55 : };
56 : use pageserver_api::reltag::{BlockNumber, RelTag};
57 : use pageserver_api::shard::{ShardIdentity, ShardIndex, ShardNumber, TenantShardId};
58 : use postgres_connection::PgConnectionConfig;
59 : use postgres_ffi::v14::xlog_utils;
60 : use postgres_ffi::{PgMajorVersion, WAL_SEGMENT_SIZE, to_pg_timestamp};
61 : use rand::Rng;
62 : use remote_storage::DownloadError;
63 : use serde_with::serde_as;
64 : use storage_broker::BrokerClientChannel;
65 : use tokio::runtime::Handle;
66 : use tokio::sync::mpsc::Sender;
67 : use tokio::sync::{Notify, oneshot, watch};
68 : use tokio_util::sync::CancellationToken;
69 : use tracing::*;
70 : use utils::generation::Generation;
71 : use utils::guard_arc_swap::GuardArcSwap;
72 : use utils::id::TimelineId;
73 : use utils::logging::{MonitorSlowFutureCallback, monitor_slow_future};
74 : use utils::lsn::{AtomicLsn, Lsn, RecordLsn};
75 : use utils::postgres_client::PostgresClientProtocol;
76 : use utils::rate_limit::RateLimit;
77 : use utils::seqwait::SeqWait;
78 : use utils::simple_rcu::{Rcu, RcuReadGuard};
79 : use utils::sync::gate::{Gate, GateGuard};
80 : use utils::{completion, critical_timeline, fs_ext, pausable_failpoint};
81 : #[cfg(test)]
82 : use wal_decoder::models::value::Value;
83 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
84 :
85 : use self::delete::DeleteTimelineFlow;
86 : pub(super) use self::eviction_task::EvictionTaskTenantState;
87 : use self::eviction_task::EvictionTaskTimelineState;
88 : use self::logical_size::LogicalSize;
89 : use self::walreceiver::{WalReceiver, WalReceiverConf};
90 : use super::remote_timeline_client::RemoteTimelineClient;
91 : use super::remote_timeline_client::index::{GcCompactionState, IndexPart};
92 : use super::secondary::heatmap::HeatMapLayer;
93 : use super::storage_layer::{LayerFringe, LayerVisibilityHint, ReadableLayer};
94 : use super::tasks::log_compaction_error;
95 : use super::upload_queue::NotInitialized;
96 : use super::{
97 : AttachedTenantConf, GcError, HeatMapTimeline, MaybeOffloaded,
98 : debug_assert_current_span_has_tenant_and_timeline_id,
99 : };
100 : use crate::PERF_TRACE_TARGET;
101 : use crate::aux_file::AuxFileSizeEstimator;
102 : use crate::basebackup_cache::BasebackupCache;
103 : use crate::config::PageServerConf;
104 : use crate::context::{
105 : DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
106 : };
107 : use crate::disk_usage_eviction_task::{DiskUsageEvictionInfo, EvictionCandidate, finite_f32};
108 : use crate::feature_resolver::TenantFeatureResolver;
109 : use crate::keyspace::{KeyPartitioning, KeySpace};
110 : use crate::l0_flush::{self, L0FlushGlobalState};
111 : use crate::metrics::{
112 : DELTAS_PER_READ_GLOBAL, LAYERS_PER_READ_AMORTIZED_GLOBAL, LAYERS_PER_READ_BATCH_GLOBAL,
113 : LAYERS_PER_READ_GLOBAL, ScanLatencyOngoingRecording, TimelineMetrics,
114 : };
115 : use crate::page_service::TenantManagerTypes;
116 : use crate::pgdatadir_mapping::{
117 : CalculateLogicalSizeError, CollectKeySpaceError, DirectoryKind, LsnForTimestamp,
118 : MAX_AUX_FILE_V2_DELTAS, MetricsUpdate,
119 : };
120 : use crate::task_mgr::TaskKind;
121 : use crate::tenant::gc_result::GcResult;
122 : use crate::tenant::layer_map::LayerMap;
123 : use crate::tenant::metadata::TimelineMetadata;
124 : use crate::tenant::storage_layer::delta_layer::DeltaEntry;
125 : use crate::tenant::storage_layer::inmemory_layer::IndexEntry;
126 : use crate::tenant::storage_layer::{
127 : AsLayerDesc, BatchLayerWriter, DeltaLayerWriter, EvictionError, ImageLayerName,
128 : ImageLayerWriter, InMemoryLayer, IoConcurrency, Layer, LayerAccessStatsReset, LayerName,
129 : PersistentLayerDesc, PersistentLayerKey, ResidentLayer, ValueReconstructSituation,
130 : ValueReconstructState, ValuesReconstructState,
131 : };
132 : use crate::tenant::tasks::BackgroundLoopKind;
133 : use crate::tenant::timeline::logical_size::CurrentLogicalSize;
134 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
135 : use crate::walingest::WalLagCooldown;
136 : use crate::walredo::RedoAttemptType;
137 : use crate::{ZERO_PAGE, task_mgr, walredo};
138 :
139 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
140 : pub(crate) enum FlushLoopState {
141 : NotStarted,
142 : Running {
143 : #[cfg(test)]
144 : expect_initdb_optimization: bool,
145 : #[cfg(test)]
146 : initdb_optimization_count: usize,
147 : },
148 : Exited,
149 : }
150 :
151 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
152 : pub enum ImageLayerCreationMode {
153 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
154 : Try,
155 : /// Force creating the image layers if possible. For now, no image layers will be created
156 : /// for metadata keys. Used in compaction code path with force flag enabled.
157 : Force,
158 : /// Initial ingestion of the data, and no data should be dropped in this function. This
159 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
160 : /// code path.
161 : Initial,
162 : }
163 :
164 : #[derive(Clone, Debug, Default)]
165 : pub enum LastImageLayerCreationStatus {
166 : Incomplete {
167 : /// The last key of the partition (exclusive) that was processed in the last
168 : /// image layer creation attempt. We will continue from this key in the next
169 : /// attempt.
170 : last_key: Key,
171 : },
172 : Complete,
173 : #[default]
174 : Initial,
175 : }
176 :
177 : impl std::fmt::Display for ImageLayerCreationMode {
178 191 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
179 191 : write!(f, "{self:?}")
180 191 : }
181 : }
182 :
183 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
184 : /// Can be removed after all refactors are done.
185 23 : fn drop_layer_manager_rlock(rlock: LayerManagerReadGuard<'_>) {
186 23 : drop(rlock)
187 23 : }
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.
191 214 : fn drop_layer_manager_wlock(rlock: LayerManagerWriteGuard<'_>) {
192 214 : drop(rlock)
193 214 : }
194 :
195 : /// The outward-facing resources required to build a Timeline
196 : pub struct TimelineResources {
197 : pub remote_client: RemoteTimelineClient,
198 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
199 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
200 : pub l0_compaction_trigger: Arc<Notify>,
201 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
202 : pub basebackup_cache: Arc<BasebackupCache>,
203 : pub feature_resolver: Arc<TenantFeatureResolver>,
204 : }
205 :
206 : pub struct Timeline {
207 : pub(crate) conf: &'static PageServerConf,
208 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
209 :
210 : myself: Weak<Self>,
211 :
212 : pub(crate) tenant_shard_id: TenantShardId,
213 : pub timeline_id: TimelineId,
214 :
215 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
216 : /// Never changes for the lifetime of this [`Timeline`] object.
217 : ///
218 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
219 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
220 : pub(crate) generation: Generation,
221 :
222 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
223 : /// to shards, and is constant through the lifetime of this Timeline.
224 : shard_identity: ShardIdentity,
225 :
226 : pub pg_version: PgMajorVersion,
227 :
228 : /// The tuple has two elements.
229 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
230 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
231 : ///
232 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
233 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
234 : ///
235 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
236 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
237 : /// `PersistentLayerDesc`'s.
238 : ///
239 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
240 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
241 : /// runtime, e.g., during page reconstruction.
242 : ///
243 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
244 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
245 : pub(crate) layers: LockedLayerManager,
246 :
247 : last_freeze_at: AtomicLsn,
248 : // Atomic would be more appropriate here.
249 : last_freeze_ts: RwLock<Instant>,
250 :
251 : pub(crate) standby_horizon: AtomicLsn,
252 :
253 : // WAL redo manager. `None` only for broken tenants.
254 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
255 :
256 : /// Remote storage client.
257 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
258 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
259 :
260 : // What page versions do we hold in the repository? If we get a
261 : // request > last_record_lsn, we need to wait until we receive all
262 : // the WAL up to the request. The SeqWait provides functions for
263 : // that. TODO: If we get a request for an old LSN, such that the
264 : // versions have already been garbage collected away, we should
265 : // throw an error, but we don't track that currently.
266 : //
267 : // last_record_lsn.load().last points to the end of last processed WAL record.
268 : //
269 : // We also remember the starting point of the previous record in
270 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
271 : // first WAL record when the node is started up. But here, we just
272 : // keep track of it.
273 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
274 :
275 : // All WAL records have been processed and stored durably on files on
276 : // local disk, up to this LSN. On crash and restart, we need to re-process
277 : // the WAL starting from this point.
278 : //
279 : // Some later WAL records might have been processed and also flushed to disk
280 : // already, so don't be surprised to see some, but there's no guarantee on
281 : // them yet.
282 : disk_consistent_lsn: AtomicLsn,
283 :
284 : // Parent timeline that this timeline was branched from, and the LSN
285 : // of the branch point.
286 : ancestor_timeline: Option<Arc<Timeline>>,
287 : ancestor_lsn: Lsn,
288 :
289 : // The LSN of gc-compaction that was last applied to this timeline.
290 : gc_compaction_state: ArcSwap<Option<GcCompactionState>>,
291 :
292 : pub(crate) metrics: Arc<TimelineMetrics>,
293 :
294 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
295 : // in `crate::page_service` writes these metrics.
296 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
297 :
298 : directory_metrics_inited: [AtomicBool; DirectoryKind::KINDS_NUM],
299 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
300 :
301 : /// Ensures layers aren't frozen by checkpointer between
302 : /// [`Timeline::get_layer_for_write`] and layer reads.
303 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
304 : /// Must always be acquired before the layer map/individual layer lock
305 : /// to avoid deadlock.
306 : ///
307 : /// The state is cleared upon freezing.
308 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
309 :
310 : /// Used to avoid multiple `flush_loop` tasks running
311 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
312 :
313 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
314 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
315 : /// The flush cycle counter is sent back on the layer_flush_done channel when
316 : /// the flush finishes. You can use that to wait for the flush to finish.
317 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
318 : /// read by whoever sends an update
319 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
320 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
321 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
322 :
323 : // The LSN at which we have executed GC: whereas [`Self::gc_info`] records the LSN at which
324 : // we _intend_ to GC (i.e. the PITR cutoff), this LSN records where we actually last did it.
325 : // Because PITR interval is mutable, it's possible for this LSN to be earlier or later than
326 : // the planned GC cutoff.
327 : pub applied_gc_cutoff_lsn: Rcu<Lsn>,
328 :
329 : pub(crate) gc_compaction_layer_update_lock: tokio::sync::RwLock<()>,
330 :
331 : // List of child timelines and their branch points. This is needed to avoid
332 : // garbage collecting data that is still needed by the child timelines.
333 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
334 :
335 : pub(crate) last_image_layer_creation_status: ArcSwap<LastImageLayerCreationStatus>,
336 :
337 : // It may change across major versions so for simplicity
338 : // keep it after running initdb for a timeline.
339 : // It is needed in checks when we want to error on some operations
340 : // when they are requested for pre-initdb lsn.
341 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
342 : // though let's keep them both for better error visibility.
343 : pub initdb_lsn: Lsn,
344 :
345 : /// The repartitioning result. Allows a single writer and multiple readers.
346 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
347 :
348 : /// Configuration: how often should the partitioning be recalculated.
349 : repartition_threshold: u64,
350 :
351 : last_image_layer_creation_check_at: AtomicLsn,
352 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
353 :
354 : /// Current logical size of the "datadir", at the last LSN.
355 : current_logical_size: LogicalSize,
356 :
357 : /// Information about the last processed message by the WAL receiver,
358 : /// or None if WAL receiver has not received anything for this timeline
359 : /// yet.
360 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
361 : pub walreceiver: Mutex<Option<WalReceiver>>,
362 :
363 : /// Relation size cache
364 : pub(crate) rel_size_latest_cache: RwLock<HashMap<RelTag, (Lsn, BlockNumber)>>,
365 : pub(crate) rel_size_snapshot_cache: Mutex<LruCache<(Lsn, RelTag), BlockNumber>>,
366 :
367 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
368 :
369 : state: watch::Sender<TimelineState>,
370 :
371 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
372 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
373 : pub delete_progress: TimelineDeleteProgress,
374 :
375 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
376 :
377 : /// Load or creation time information about the disk_consistent_lsn and when the loading
378 : /// happened. Used for consumption metrics.
379 : pub(crate) loaded_at: (Lsn, SystemTime),
380 :
381 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
382 : pub(crate) gate: Gate,
383 :
384 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
385 : /// to the timeline should drop out when this token fires.
386 : pub(crate) cancel: CancellationToken,
387 :
388 : /// Make sure we only have one running compaction at a time in tests.
389 : ///
390 : /// Must only be taken in two places:
391 : /// - [`Timeline::compact`] (this file)
392 : /// - [`delete::delete_local_timeline_directory`]
393 : ///
394 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
395 : compaction_lock: tokio::sync::Mutex<()>,
396 :
397 : /// If true, the last compaction failed.
398 : compaction_failed: AtomicBool,
399 :
400 : /// Notifies the tenant compaction loop that there is pending L0 compaction work.
401 : l0_compaction_trigger: Arc<Notify>,
402 :
403 : /// Make sure we only have one running gc at a time.
404 : ///
405 : /// Must only be taken in two places:
406 : /// - [`Timeline::gc`] (this file)
407 : /// - [`delete::delete_local_timeline_directory`]
408 : ///
409 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
410 : gc_lock: tokio::sync::Mutex<()>,
411 :
412 : /// Cloned from [`super::TenantShard::pagestream_throttle`] on construction.
413 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
414 :
415 : /// Size estimator for aux file v2
416 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
417 :
418 : /// Some test cases directly place keys into the timeline without actually modifying the directory
419 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
420 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
421 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
422 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
423 : #[cfg(test)]
424 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
425 :
426 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
427 :
428 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
429 :
430 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
431 :
432 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
433 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
434 :
435 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
436 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
437 :
438 : pub(super) previous_heatmap: ArcSwapOption<PreviousHeatmap>,
439 :
440 : /// May host a background Tokio task which downloads all the layers from the current
441 : /// heatmap on demand.
442 : heatmap_layers_downloader: Mutex<Option<heatmap_layers_downloader::HeatmapLayersDownloader>>,
443 :
444 : pub(crate) rel_size_v2_status: ArcSwapOption<RelSizeMigration>,
445 :
446 : wait_lsn_log_slow: tokio::sync::Semaphore,
447 :
448 : /// A channel to send async requests to prepare a basebackup for the basebackup cache.
449 : basebackup_cache: Arc<BasebackupCache>,
450 :
451 : feature_resolver: Arc<TenantFeatureResolver>,
452 : }
453 :
454 : pub(crate) enum PreviousHeatmap {
455 : Active {
456 : heatmap: HeatMapTimeline,
457 : read_at: std::time::Instant,
458 : // End LSN covered by the heatmap if known
459 : end_lsn: Option<Lsn>,
460 : },
461 : Obsolete,
462 : }
463 :
464 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
465 :
466 : pub struct WalReceiverInfo {
467 : pub wal_source_connconf: PgConnectionConfig,
468 : pub last_received_msg_lsn: Lsn,
469 : pub last_received_msg_ts: u128,
470 : }
471 :
472 : /// Information about how much history needs to be retained, needed by
473 : /// Garbage Collection.
474 : #[derive(Default)]
475 : pub(crate) struct GcInfo {
476 : /// Specific LSNs that are needed.
477 : ///
478 : /// Currently, this includes all points where child branches have
479 : /// been forked off from. In the future, could also include
480 : /// explicit user-defined snapshot points.
481 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
482 :
483 : /// The cutoff coordinates, which are combined by selecting the minimum.
484 : pub(crate) cutoffs: GcCutoffs,
485 :
486 : /// Leases granted to particular LSNs.
487 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
488 :
489 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
490 : pub(crate) within_ancestor_pitr: bool,
491 : }
492 :
493 : impl GcInfo {
494 154 : pub(crate) fn min_cutoff(&self) -> Lsn {
495 154 : self.cutoffs.select_min()
496 154 : }
497 :
498 119 : pub(super) fn insert_child(
499 119 : &mut self,
500 119 : child_id: TimelineId,
501 119 : child_lsn: Lsn,
502 119 : is_offloaded: MaybeOffloaded,
503 119 : ) {
504 119 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
505 119 : self.retain_lsns.sort_by_key(|i| i.0);
506 119 : }
507 :
508 2 : pub(super) fn remove_child_maybe_offloaded(
509 2 : &mut self,
510 2 : child_id: TimelineId,
511 2 : maybe_offloaded: MaybeOffloaded,
512 2 : ) -> bool {
513 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
514 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
515 2 : let mut removed = false;
516 3 : self.retain_lsns.retain(|i| {
517 3 : if removed {
518 1 : return true;
519 2 : }
520 2 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
521 2 : removed |= remove;
522 2 : !remove
523 3 : });
524 2 : removed
525 2 : }
526 :
527 2 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
528 2 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
529 2 : }
530 :
531 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
532 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
533 0 : }
534 119 : pub(crate) fn lsn_covered_by_lease(&self, lsn: Lsn) -> bool {
535 119 : self.leases.contains_key(&lsn)
536 119 : }
537 : }
538 :
539 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
540 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
541 : /// between time-based and space-based retention for observability and consumption metrics purposes.
542 : #[derive(Clone, Debug, Default)]
543 : pub(crate) struct GcCutoffs {
544 : /// Calculated from the [`pageserver_api::models::TenantConfig::gc_horizon`], this LSN indicates how much
545 : /// history we must keep to retain a specified number of bytes of WAL.
546 : pub(crate) space: Lsn,
547 :
548 : /// Calculated from [`pageserver_api::models::TenantConfig::pitr_interval`], this LSN indicates
549 : /// how much history we must keep to enable reading back at least the PITR interval duration.
550 : ///
551 : /// None indicates that the PITR cutoff has not been computed. A PITR interval of 0 will yield
552 : /// Some(last_record_lsn).
553 : pub(crate) time: Option<Lsn>,
554 : }
555 :
556 : impl GcCutoffs {
557 154 : fn select_min(&self) -> Lsn {
558 : // NB: if we haven't computed the PITR cutoff yet, we can't GC anything.
559 154 : self.space.min(self.time.unwrap_or_default())
560 154 : }
561 : }
562 :
563 : pub(crate) struct TimelineVisitOutcome {
564 : completed_keyspace: KeySpace,
565 : image_covered_keyspace: KeySpace,
566 : }
567 :
568 : /// An error happened in a get() operation.
569 : #[derive(thiserror::Error, Debug)]
570 : pub(crate) enum PageReconstructError {
571 : #[error(transparent)]
572 : Other(anyhow::Error),
573 :
574 : #[error("Ancestor LSN wait error: {0}")]
575 : AncestorLsnTimeout(WaitLsnError),
576 :
577 : #[error("timeline shutting down")]
578 : Cancelled,
579 :
580 : /// An error happened replaying WAL records
581 : #[error(transparent)]
582 : WalRedo(anyhow::Error),
583 :
584 : #[error("{0}")]
585 : MissingKey(Box<MissingKeyError>),
586 : }
587 :
588 : impl PageReconstructError {
589 0 : pub(crate) fn is_cancel(&self) -> bool {
590 0 : match self {
591 0 : PageReconstructError::Other(_) => false,
592 0 : PageReconstructError::AncestorLsnTimeout(e) => e.is_cancel(),
593 0 : PageReconstructError::Cancelled => true,
594 0 : PageReconstructError::WalRedo(_) => false,
595 0 : PageReconstructError::MissingKey(_) => false,
596 : }
597 0 : }
598 : #[allow(dead_code)] // we use the is_cancel + into_anyhow pattern in quite a few places, this one will follow soon enough
599 0 : pub(crate) fn into_anyhow(self) -> anyhow::Error {
600 0 : match self {
601 0 : PageReconstructError::Other(e) => e,
602 0 : PageReconstructError::AncestorLsnTimeout(e) => e.into_anyhow(),
603 0 : PageReconstructError::Cancelled => anyhow::Error::new(self),
604 0 : PageReconstructError::WalRedo(e) => e,
605 0 : PageReconstructError::MissingKey(_) => anyhow::Error::new(self),
606 : }
607 0 : }
608 : }
609 :
610 : impl From<anyhow::Error> for PageReconstructError {
611 1 : fn from(value: anyhow::Error) -> Self {
612 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
613 1 : match value.downcast::<PageReconstructError>() {
614 0 : Ok(pre) => pre,
615 1 : Err(other) => PageReconstructError::Other(other),
616 : }
617 1 : }
618 : }
619 :
620 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
621 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
622 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
623 0 : }
624 : }
625 :
626 : impl From<layer_manager::Shutdown> for PageReconstructError {
627 0 : fn from(_: layer_manager::Shutdown) -> Self {
628 0 : PageReconstructError::Cancelled
629 0 : }
630 : }
631 :
632 : impl GetVectoredError {
633 : #[cfg(test)]
634 3 : pub(crate) fn is_missing_key_error(&self) -> bool {
635 3 : matches!(self, Self::MissingKey(_))
636 3 : }
637 : }
638 :
639 : impl From<layer_manager::Shutdown> for GetVectoredError {
640 0 : fn from(_: layer_manager::Shutdown) -> Self {
641 0 : GetVectoredError::Cancelled
642 0 : }
643 : }
644 :
645 : /// A layer identifier when used in the [`ReadPath`] structure. This enum is for observability purposes
646 : /// only and not used by the "real read path".
647 : pub enum ReadPathLayerId {
648 : PersistentLayer(PersistentLayerKey),
649 : InMemoryLayer(Range<Lsn>),
650 : }
651 :
652 : impl std::fmt::Display for ReadPathLayerId {
653 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
654 0 : match self {
655 0 : ReadPathLayerId::PersistentLayer(key) => write!(f, "{key}"),
656 0 : ReadPathLayerId::InMemoryLayer(range) => {
657 0 : write!(f, "in-mem {}..{}", range.start, range.end)
658 : }
659 : }
660 0 : }
661 : }
662 : pub struct ReadPath {
663 : keyspace: KeySpace,
664 : lsn: Lsn,
665 : path: Vec<(ReadPathLayerId, KeySpace, Range<Lsn>)>,
666 : }
667 :
668 : impl ReadPath {
669 312456 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
670 312456 : Self {
671 312456 : keyspace,
672 312456 : lsn,
673 312456 : path: Vec::new(),
674 312456 : }
675 312456 : }
676 :
677 446449 : pub fn record_layer_visit(
678 446449 : &mut self,
679 446449 : layer_to_read: &ReadableLayer,
680 446449 : keyspace_to_read: &KeySpace,
681 446449 : lsn_range: &Range<Lsn>,
682 446449 : ) {
683 446449 : let id = match layer_to_read {
684 139128 : ReadableLayer::PersistentLayer(layer) => {
685 139128 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
686 : }
687 307321 : ReadableLayer::InMemoryLayer(layer) => {
688 307321 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
689 : }
690 : };
691 446449 : self.path
692 446449 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
693 446449 : }
694 : }
695 :
696 : impl std::fmt::Display for ReadPath {
697 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
698 0 : writeln!(f, "Read path for {} at lsn {}:", self.keyspace, self.lsn)?;
699 0 : for (idx, (layer_id, keyspace, lsn_range)) in self.path.iter().enumerate() {
700 0 : writeln!(
701 0 : f,
702 0 : "{}: {} {}..{} {}",
703 : idx, layer_id, lsn_range.start, lsn_range.end, keyspace
704 0 : )?;
705 : }
706 0 : Ok(())
707 0 : }
708 : }
709 :
710 : #[derive(thiserror::Error)]
711 : pub struct MissingKeyError {
712 : keyspace: KeySpace,
713 : shard: ShardNumber,
714 : query: Option<VersionedKeySpaceQuery>,
715 : // This is largest request LSN from the get page request batch
716 : original_hwm_lsn: Lsn,
717 : ancestor_lsn: Option<Lsn>,
718 : /// Debug information about the read path if there's an error
719 : read_path: Option<ReadPath>,
720 : backtrace: Option<std::backtrace::Backtrace>,
721 : }
722 :
723 : impl MissingKeyError {
724 7 : fn enrich(&mut self, query: VersionedKeySpaceQuery) {
725 7 : self.query = Some(query);
726 7 : }
727 : }
728 :
729 : impl std::fmt::Debug for MissingKeyError {
730 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
731 0 : write!(f, "{self}")
732 0 : }
733 : }
734 :
735 : impl std::fmt::Display for MissingKeyError {
736 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
737 0 : write!(
738 0 : f,
739 0 : "could not find data for key {} (shard {:?}), original HWM LSN {}",
740 : self.keyspace, self.shard, self.original_hwm_lsn
741 0 : )?;
742 :
743 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
744 0 : write!(f, ", ancestor {ancestor_lsn}")?;
745 0 : }
746 :
747 0 : if let Some(ref query) = self.query {
748 0 : write!(f, ", query {query}")?;
749 0 : }
750 :
751 0 : if let Some(ref read_path) = self.read_path {
752 0 : write!(f, "\n{read_path}")?;
753 0 : }
754 :
755 0 : if let Some(ref backtrace) = self.backtrace {
756 0 : write!(f, "\n{backtrace}")?;
757 0 : }
758 :
759 0 : Ok(())
760 0 : }
761 : }
762 :
763 : #[derive(thiserror::Error, Debug)]
764 : pub(crate) enum CreateImageLayersError {
765 : #[error("timeline shutting down")]
766 : Cancelled,
767 :
768 : #[error("read failed")]
769 : GetVectoredError(#[source] GetVectoredError),
770 :
771 : #[error("reconstruction failed")]
772 : PageReconstructError(#[source] PageReconstructError),
773 :
774 : #[error(transparent)]
775 : Other(anyhow::Error),
776 : }
777 :
778 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
779 0 : fn from(_: layer_manager::Shutdown) -> Self {
780 0 : CreateImageLayersError::Cancelled
781 0 : }
782 : }
783 :
784 : #[derive(thiserror::Error, Debug, Clone)]
785 : pub(crate) enum FlushLayerError {
786 : /// Timeline cancellation token was cancelled
787 : #[error("timeline shutting down")]
788 : Cancelled,
789 :
790 : /// We tried to flush a layer while the Timeline is in an unexpected state
791 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
792 : NotRunning(FlushLoopState),
793 :
794 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
795 : // loop via a watch channel, where we can only borrow it.
796 : #[error("create image layers (shared)")]
797 : CreateImageLayersError(Arc<CreateImageLayersError>),
798 :
799 : #[error("other (shared)")]
800 : Other(#[from] Arc<anyhow::Error>),
801 : }
802 :
803 : impl FlushLayerError {
804 : // When crossing from generic anyhow errors to this error type, we explicitly check
805 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
806 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
807 0 : let cancelled = timeline.cancel.is_cancelled()
808 : // The upload queue might have been shut down before the official cancellation of the timeline.
809 0 : || err
810 0 : .downcast_ref::<NotInitialized>()
811 0 : .map(NotInitialized::is_stopping)
812 0 : .unwrap_or_default();
813 0 : if cancelled {
814 0 : Self::Cancelled
815 : } else {
816 0 : Self::Other(Arc::new(err))
817 : }
818 0 : }
819 : }
820 :
821 : impl From<layer_manager::Shutdown> for FlushLayerError {
822 0 : fn from(_: layer_manager::Shutdown) -> Self {
823 0 : FlushLayerError::Cancelled
824 0 : }
825 : }
826 :
827 : #[derive(thiserror::Error, Debug)]
828 : pub enum GetVectoredError {
829 : #[error("timeline shutting down")]
830 : Cancelled,
831 :
832 : #[error("requested too many keys: {0} > {1}")]
833 : Oversized(u64, u64),
834 :
835 : #[error("requested at invalid LSN: {0}")]
836 : InvalidLsn(Lsn),
837 :
838 : #[error("requested key not found: {0}")]
839 : MissingKey(Box<MissingKeyError>),
840 :
841 : #[error("ancestry walk")]
842 : GetReadyAncestorError(#[source] GetReadyAncestorError),
843 :
844 : #[error(transparent)]
845 : Other(#[from] anyhow::Error),
846 : }
847 :
848 : impl From<GetReadyAncestorError> for GetVectoredError {
849 1 : fn from(value: GetReadyAncestorError) -> Self {
850 : use GetReadyAncestorError::*;
851 1 : match value {
852 0 : Cancelled => GetVectoredError::Cancelled,
853 : AncestorLsnTimeout(_) | BadState { .. } => {
854 1 : GetVectoredError::GetReadyAncestorError(value)
855 : }
856 : }
857 1 : }
858 : }
859 :
860 : #[derive(thiserror::Error, Debug)]
861 : pub enum GetReadyAncestorError {
862 : #[error("ancestor LSN wait error")]
863 : AncestorLsnTimeout(#[from] WaitLsnError),
864 :
865 : #[error("bad state on timeline {timeline_id}: {state:?}")]
866 : BadState {
867 : timeline_id: TimelineId,
868 : state: TimelineState,
869 : },
870 :
871 : #[error("cancelled")]
872 : Cancelled,
873 : }
874 :
875 : #[derive(Clone, Copy)]
876 : pub enum LogicalSizeCalculationCause {
877 : Initial,
878 : ConsumptionMetricsSyntheticSize,
879 : EvictionTaskImitation,
880 : TenantSizeHandler,
881 : }
882 :
883 : pub enum GetLogicalSizePriority {
884 : User,
885 : Background,
886 : }
887 :
888 : #[derive(Debug, enumset::EnumSetType)]
889 : pub(crate) enum CompactFlags {
890 : ForceRepartition,
891 : ForceImageLayerCreation,
892 : ForceL0Compaction,
893 : OnlyL0Compaction,
894 : EnhancedGcBottomMostCompaction,
895 : DryRun,
896 : /// Makes image compaction yield if there's pending L0 compaction. This should always be used in
897 : /// the background compaction task, since we want to aggressively compact down L0 to bound
898 : /// read amplification.
899 : ///
900 : /// It only makes sense to use this when `compaction_l0_first` is enabled (such that we yield to
901 : /// an L0 compaction pass), and without `OnlyL0Compaction` (L0 compaction shouldn't yield for L0
902 : /// compaction).
903 : YieldForL0,
904 : }
905 :
906 : #[serde_with::serde_as]
907 0 : #[derive(Debug, Clone, serde::Deserialize)]
908 : pub(crate) struct CompactRequest {
909 : pub compact_key_range: Option<CompactKeyRange>,
910 : pub compact_lsn_range: Option<CompactLsnRange>,
911 : /// Whether the compaction job should be scheduled.
912 : #[serde(default)]
913 : pub scheduled: bool,
914 : /// Whether the compaction job should be split across key ranges.
915 : #[serde(default)]
916 : pub sub_compaction: bool,
917 : /// Max job size for each subcompaction job.
918 : pub sub_compaction_max_job_size_mb: Option<u64>,
919 : }
920 :
921 0 : #[derive(Debug, Clone, serde::Deserialize)]
922 : pub(crate) struct MarkInvisibleRequest {
923 : #[serde(default)]
924 : pub is_visible: Option<bool>,
925 : }
926 :
927 : #[derive(Debug, Clone, Default)]
928 : pub(crate) struct CompactOptions {
929 : pub flags: EnumSet<CompactFlags>,
930 : /// If set, the compaction will only compact the key range specified by this option.
931 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
932 : pub compact_key_range: Option<CompactKeyRange>,
933 : /// If set, the compaction will only compact the LSN within this value.
934 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
935 : pub compact_lsn_range: Option<CompactLsnRange>,
936 : /// Enable sub-compaction (split compaction job across key ranges).
937 : /// This option is only used by GC compaction.
938 : pub sub_compaction: bool,
939 : /// Set job size for the GC compaction.
940 : /// This option is only used by GC compaction.
941 : pub sub_compaction_max_job_size_mb: Option<u64>,
942 : }
943 :
944 : impl std::fmt::Debug for Timeline {
945 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
946 0 : write!(f, "Timeline<{}>", self.timeline_id)
947 0 : }
948 : }
949 :
950 : #[derive(thiserror::Error, Debug, Clone)]
951 : pub enum WaitLsnError {
952 : // Called on a timeline which is shutting down
953 : #[error("Shutdown")]
954 : Shutdown,
955 :
956 : // Called on an timeline not in active state or shutting down
957 : #[error("Bad timeline state: {0:?}")]
958 : BadState(TimelineState),
959 :
960 : // Timeout expired while waiting for LSN to catch up with goal.
961 : #[error("{0}")]
962 : Timeout(String),
963 : }
964 :
965 : impl WaitLsnError {
966 0 : pub(crate) fn is_cancel(&self) -> bool {
967 0 : match self {
968 0 : WaitLsnError::Shutdown => true,
969 0 : WaitLsnError::BadState(timeline_state) => match timeline_state {
970 0 : TimelineState::Loading => false,
971 0 : TimelineState::Active => false,
972 0 : TimelineState::Stopping => true,
973 0 : TimelineState::Broken { .. } => false,
974 : },
975 0 : WaitLsnError::Timeout(_) => false,
976 : }
977 0 : }
978 0 : pub(crate) fn into_anyhow(self) -> anyhow::Error {
979 0 : match self {
980 0 : WaitLsnError::Shutdown => anyhow::Error::new(self),
981 0 : WaitLsnError::BadState(_) => anyhow::Error::new(self),
982 0 : WaitLsnError::Timeout(_) => anyhow::Error::new(self),
983 : }
984 0 : }
985 : }
986 :
987 : impl From<WaitLsnError> for tonic::Status {
988 0 : fn from(err: WaitLsnError) -> Self {
989 : use tonic::Code;
990 0 : let code = if err.is_cancel() {
991 0 : Code::Unavailable
992 : } else {
993 0 : Code::Internal
994 : };
995 0 : tonic::Status::new(code, err.to_string())
996 0 : }
997 : }
998 :
999 : // The impls below achieve cancellation mapping for errors.
1000 : // Perhaps there's a way of achieving this with less cruft.
1001 :
1002 : impl From<CreateImageLayersError> for CompactionError {
1003 0 : fn from(e: CreateImageLayersError) -> Self {
1004 0 : match e {
1005 0 : CreateImageLayersError::Cancelled => CompactionError::new_cancelled(),
1006 0 : CreateImageLayersError::Other(e) => {
1007 0 : CompactionError::Other(e.context("create image layers"))
1008 : }
1009 0 : _ => CompactionError::Other(e.into()),
1010 : }
1011 0 : }
1012 : }
1013 :
1014 : impl From<CreateImageLayersError> for FlushLayerError {
1015 0 : fn from(e: CreateImageLayersError) -> Self {
1016 0 : match e {
1017 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
1018 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
1019 : }
1020 0 : }
1021 : }
1022 :
1023 : impl From<PageReconstructError> for CreateImageLayersError {
1024 0 : fn from(e: PageReconstructError) -> Self {
1025 0 : match e {
1026 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
1027 0 : _ => CreateImageLayersError::PageReconstructError(e),
1028 : }
1029 0 : }
1030 : }
1031 :
1032 : impl From<super::storage_layer::errors::PutError> for CreateImageLayersError {
1033 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
1034 0 : if e.is_cancel() {
1035 0 : CreateImageLayersError::Cancelled
1036 : } else {
1037 0 : CreateImageLayersError::Other(e.into_anyhow())
1038 : }
1039 0 : }
1040 : }
1041 :
1042 : impl From<GetVectoredError> for CreateImageLayersError {
1043 0 : fn from(e: GetVectoredError) -> Self {
1044 0 : match e {
1045 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
1046 0 : _ => CreateImageLayersError::GetVectoredError(e),
1047 : }
1048 0 : }
1049 : }
1050 :
1051 : impl From<GetVectoredError> for PageReconstructError {
1052 3 : fn from(e: GetVectoredError) -> Self {
1053 3 : match e {
1054 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
1055 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
1056 0 : err @ GetVectoredError::Oversized(_, _) => PageReconstructError::Other(err.into()),
1057 2 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
1058 1 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
1059 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
1060 : }
1061 3 : }
1062 : }
1063 :
1064 : impl From<GetReadyAncestorError> for PageReconstructError {
1065 1 : fn from(e: GetReadyAncestorError) -> Self {
1066 : use GetReadyAncestorError::*;
1067 1 : match e {
1068 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
1069 1 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
1070 0 : Cancelled => PageReconstructError::Cancelled,
1071 : }
1072 1 : }
1073 : }
1074 :
1075 : pub(crate) enum WaitLsnTimeout {
1076 : Custom(Duration),
1077 : // Use the [`PageServerConf::wait_lsn_timeout`] default
1078 : Default,
1079 : }
1080 :
1081 : pub(crate) enum WaitLsnWaiter<'a> {
1082 : Timeline(&'a Timeline),
1083 : Tenant,
1084 : PageService,
1085 : HttpEndpoint,
1086 : BaseBackupCache,
1087 : }
1088 :
1089 : /// Argument to [`Timeline::shutdown`].
1090 : #[derive(Debug, Clone, Copy)]
1091 : pub(crate) enum ShutdownMode {
1092 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk. This method can
1093 : /// take multiple seconds for a busy timeline.
1094 : ///
1095 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1096 : /// the call to [`Timeline::shutdown`].
1097 : FreezeAndFlush,
1098 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1099 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1100 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1101 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1102 : Reload,
1103 : /// Shut down immediately, without waiting for any open layers to flush.
1104 : Hard,
1105 : }
1106 :
1107 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1108 : enum ImageLayerCreationOutcome {
1109 : /// We generated an image layer
1110 : Generated {
1111 : unfinished_image_layer: ImageLayerWriter,
1112 : },
1113 : /// The key range is empty
1114 : Empty,
1115 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1116 : /// the image layer creation.
1117 : Skip,
1118 : }
1119 :
1120 : enum RepartitionError {
1121 : Other(anyhow::Error),
1122 : CollectKeyspace(CollectKeySpaceError),
1123 : }
1124 :
1125 : impl RepartitionError {
1126 0 : fn is_cancel(&self) -> bool {
1127 0 : match self {
1128 0 : RepartitionError::Other(_) => false,
1129 0 : RepartitionError::CollectKeyspace(e) => e.is_cancel(),
1130 : }
1131 0 : }
1132 0 : fn into_anyhow(self) -> anyhow::Error {
1133 0 : match self {
1134 0 : RepartitionError::Other(e) => e,
1135 0 : RepartitionError::CollectKeyspace(e) => e.into_anyhow(),
1136 : }
1137 0 : }
1138 : }
1139 :
1140 : /// Public interface functions
1141 : impl Timeline {
1142 : /// Get the LSN where this branch was created
1143 80 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1144 80 : self.ancestor_lsn
1145 80 : }
1146 :
1147 : /// Get the ancestor's timeline id
1148 416 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1149 416 : self.ancestor_timeline
1150 416 : .as_ref()
1151 416 : .map(|ancestor| ancestor.timeline_id)
1152 416 : }
1153 :
1154 : /// Get the ancestor timeline
1155 1 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1156 1 : self.ancestor_timeline.as_ref()
1157 1 : }
1158 :
1159 : /// Get the bytes written since the PITR cutoff on this branch, and
1160 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1161 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1162 : // TODO: for backwards compatibility, we return the full history back to 0 when the PITR
1163 : // cutoff has not yet been initialized. This should return None instead, but this is exposed
1164 : // in external HTTP APIs and callers may not handle a null value.
1165 0 : let gc_info = self.gc_info.read().unwrap();
1166 0 : let history = self
1167 0 : .get_last_record_lsn()
1168 0 : .checked_sub(gc_info.cutoffs.time.unwrap_or_default())
1169 0 : .unwrap_or_default()
1170 0 : .0;
1171 0 : (history, gc_info.within_ancestor_pitr)
1172 0 : }
1173 :
1174 : /// Read timeline's GC cutoff: this is the LSN at which GC has started to happen
1175 427630 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1176 427630 : self.applied_gc_cutoff_lsn.read()
1177 427630 : }
1178 :
1179 : /// Read timeline's planned GC cutoff: this is the logical end of history that users are allowed
1180 : /// to read (based on configured PITR), even if physically we have more history. Returns None
1181 : /// if the PITR cutoff has not yet been initialized.
1182 0 : pub(crate) fn get_gc_cutoff_lsn(&self) -> Option<Lsn> {
1183 0 : self.gc_info.read().unwrap().cutoffs.time
1184 0 : }
1185 :
1186 : /// Look up given page version.
1187 : ///
1188 : /// If a remote layer file is needed, it is downloaded as part of this
1189 : /// call.
1190 : ///
1191 : /// This method enforces [`Self::pagestream_throttle`] internally.
1192 : ///
1193 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1194 : /// abstraction above this needs to store suitable metadata to track what
1195 : /// data exists with what keys, in separate metadata entries. If a
1196 : /// non-existent key is requested, we may incorrectly return a value from
1197 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1198 : /// non-existing key.
1199 : ///
1200 : /// # Cancel-Safety
1201 : ///
1202 : /// This method is cancellation-safe.
1203 : #[inline(always)]
1204 301318 : pub(crate) async fn get(
1205 301318 : &self,
1206 301318 : key: Key,
1207 301318 : lsn: Lsn,
1208 301318 : ctx: &RequestContext,
1209 301318 : ) -> Result<Bytes, PageReconstructError> {
1210 301318 : if !lsn.is_valid() {
1211 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1212 301318 : }
1213 :
1214 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1215 : // already checked the key against the shard_identity when looking up the Timeline from
1216 : // page_service.
1217 301318 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1218 :
1219 301318 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1220 :
1221 301318 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
1222 :
1223 301318 : let vectored_res = self
1224 301318 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
1225 301318 : .await;
1226 :
1227 301318 : let key_value = vectored_res?.pop_first();
1228 301315 : match key_value {
1229 301309 : Some((got_key, value)) => {
1230 301309 : if got_key != key {
1231 0 : error!(
1232 0 : "Expected {}, but singular vectored get returned {}",
1233 : key, got_key
1234 : );
1235 0 : Err(PageReconstructError::Other(anyhow!(
1236 0 : "Singular vectored get returned wrong key"
1237 0 : )))
1238 : } else {
1239 301309 : value
1240 : }
1241 : }
1242 6 : None => Err(PageReconstructError::MissingKey(Box::new(
1243 6 : MissingKeyError {
1244 6 : keyspace: KeySpace::single(key..key.next()),
1245 6 : shard: self.shard_identity.get_shard_number(&key),
1246 6 : original_hwm_lsn: lsn,
1247 6 : ancestor_lsn: None,
1248 6 : backtrace: None,
1249 6 : read_path: None,
1250 6 : query: None,
1251 6 : },
1252 6 : ))),
1253 : }
1254 301318 : }
1255 :
1256 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1257 :
1258 : /// Look up multiple page versions at a given LSN
1259 : ///
1260 : /// This naive implementation will be replaced with a more efficient one
1261 : /// which actually vectorizes the read path.
1262 10894 : pub(crate) async fn get_vectored(
1263 10894 : &self,
1264 10894 : query: VersionedKeySpaceQuery,
1265 10894 : io_concurrency: super::storage_layer::IoConcurrency,
1266 10894 : ctx: &RequestContext,
1267 10894 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1268 10894 : let total_keyspace = query.total_keyspace();
1269 :
1270 10894 : let key_count = total_keyspace.total_raw_size();
1271 10894 : if key_count > self.conf.max_get_vectored_keys.get() {
1272 0 : return Err(GetVectoredError::Oversized(
1273 0 : key_count as u64,
1274 0 : self.conf.max_get_vectored_keys.get() as u64,
1275 0 : ));
1276 10894 : }
1277 :
1278 34207 : for range in &total_keyspace.ranges {
1279 23313 : let mut key = range.start;
1280 65342 : while key != range.end {
1281 42029 : assert!(!self.shard_identity.is_key_disposable(&key));
1282 42029 : key = key.next();
1283 : }
1284 : }
1285 :
1286 10894 : trace!(
1287 0 : "get vectored query {} from task kind {:?}",
1288 : query,
1289 0 : ctx.task_kind(),
1290 : );
1291 :
1292 10894 : let start = crate::metrics::GET_VECTORED_LATENCY
1293 10894 : .for_task_kind(ctx.task_kind())
1294 10894 : .map(|metric| (metric, Instant::now()));
1295 :
1296 10894 : let res = self
1297 10894 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1298 10894 : .await;
1299 :
1300 10894 : if let Some((metric, start)) = start {
1301 0 : let elapsed = start.elapsed();
1302 0 : metric.observe(elapsed.as_secs_f64());
1303 10894 : }
1304 :
1305 10894 : res
1306 10894 : }
1307 :
1308 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1309 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1310 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1311 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1312 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1313 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1314 : /// the scan operation will not cause OOM in the future.
1315 8 : pub(crate) async fn scan(
1316 8 : &self,
1317 8 : keyspace: KeySpace,
1318 8 : lsn: Lsn,
1319 8 : ctx: &RequestContext,
1320 8 : io_concurrency: super::storage_layer::IoConcurrency,
1321 8 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1322 8 : if !lsn.is_valid() {
1323 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1324 8 : }
1325 :
1326 8 : trace!(
1327 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1328 : keyspace,
1329 : lsn,
1330 0 : ctx.task_kind()
1331 : );
1332 :
1333 : // We should generalize this into Keyspace::contains in the future.
1334 16 : for range in &keyspace.ranges {
1335 8 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1336 8 : || range.end.field1 > METADATA_KEY_END_PREFIX
1337 : {
1338 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1339 0 : "only metadata keyspace can be scanned"
1340 0 : )));
1341 8 : }
1342 : }
1343 :
1344 8 : let start = crate::metrics::SCAN_LATENCY
1345 8 : .for_task_kind(ctx.task_kind())
1346 8 : .map(ScanLatencyOngoingRecording::start_recording);
1347 :
1348 8 : let query = VersionedKeySpaceQuery::uniform(keyspace, lsn);
1349 :
1350 8 : let vectored_res = self
1351 8 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1352 8 : .await;
1353 :
1354 8 : if let Some(recording) = start {
1355 0 : recording.observe();
1356 8 : }
1357 :
1358 8 : vectored_res
1359 8 : }
1360 :
1361 312456 : pub(super) async fn get_vectored_impl(
1362 312456 : &self,
1363 312456 : query: VersionedKeySpaceQuery,
1364 312456 : reconstruct_state: &mut ValuesReconstructState,
1365 312456 : ctx: &RequestContext,
1366 312456 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1367 312456 : if query.is_empty() {
1368 0 : return Ok(BTreeMap::default());
1369 312456 : }
1370 :
1371 312456 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1372 312456 : Some(ReadPath::new(
1373 312456 : query.total_keyspace(),
1374 312456 : query.high_watermark_lsn()?,
1375 : ))
1376 : } else {
1377 0 : None
1378 : };
1379 :
1380 312456 : reconstruct_state.read_path = read_path;
1381 :
1382 312456 : let redo_attempt_type = if ctx.task_kind() == TaskKind::Compaction {
1383 0 : RedoAttemptType::LegacyCompaction
1384 : } else {
1385 312456 : RedoAttemptType::ReadPage
1386 : };
1387 :
1388 312456 : let traversal_res: Result<(), _> = {
1389 312456 : let ctx = RequestContextBuilder::from(ctx)
1390 312456 : .perf_span(|crnt_perf_span| {
1391 0 : info_span!(
1392 : target: PERF_TRACE_TARGET,
1393 0 : parent: crnt_perf_span,
1394 : "PLAN_IO",
1395 : )
1396 0 : })
1397 312456 : .attached_child();
1398 :
1399 312456 : self.get_vectored_reconstruct_data(query.clone(), reconstruct_state, &ctx)
1400 312456 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1401 312456 : .await
1402 : };
1403 :
1404 312456 : if let Err(err) = traversal_res {
1405 : // Wait for all the spawned IOs to complete.
1406 : // See comments on `spawn_io` inside `storage_layer` for more details.
1407 8 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1408 8 : .into_values()
1409 8 : .map(|state| state.collect_pending_ios())
1410 8 : .collect::<FuturesUnordered<_>>();
1411 8 : while collect_futs.next().await.is_some() {}
1412 :
1413 : // Enrich the missing key error with the original query.
1414 8 : if let GetVectoredError::MissingKey(mut missing_err) = err {
1415 7 : missing_err.enrich(query.clone());
1416 7 : return Err(GetVectoredError::MissingKey(missing_err));
1417 1 : }
1418 :
1419 1 : return Err(err);
1420 312448 : };
1421 :
1422 312448 : let layers_visited = reconstruct_state.get_layers_visited();
1423 :
1424 312448 : let ctx = RequestContextBuilder::from(ctx)
1425 312448 : .perf_span(|crnt_perf_span| {
1426 0 : info_span!(
1427 : target: PERF_TRACE_TARGET,
1428 0 : parent: crnt_perf_span,
1429 : "RECONSTRUCT",
1430 : )
1431 0 : })
1432 312448 : .attached_child();
1433 :
1434 312448 : let futs = FuturesUnordered::new();
1435 363577 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1436 363577 : let req_lsn_for_key = query.map_key_to_lsn(&key);
1437 :
1438 363577 : futs.push({
1439 363577 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1440 363577 : let ctx = RequestContextBuilder::from(&ctx)
1441 363577 : .perf_span(|crnt_perf_span| {
1442 0 : info_span!(
1443 : target: PERF_TRACE_TARGET,
1444 0 : parent: crnt_perf_span,
1445 : "RECONSTRUCT_KEY",
1446 : key = %key,
1447 : )
1448 0 : })
1449 363577 : .attached_child();
1450 :
1451 363577 : async move {
1452 363577 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1453 :
1454 363577 : let res = state
1455 363577 : .collect_pending_ios()
1456 363577 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1457 0 : info_span!(
1458 : target: PERF_TRACE_TARGET,
1459 0 : parent: crnt_perf_span,
1460 : "WAIT_FOR_IO_COMPLETIONS",
1461 : )
1462 0 : })
1463 363577 : .await;
1464 :
1465 363577 : let converted = match res {
1466 363577 : Ok(ok) => ok,
1467 0 : Err(err) => {
1468 0 : return (key, Err(err));
1469 : }
1470 : };
1471 363577 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1472 :
1473 : // The walredo module expects the records to be descending in terms of Lsn.
1474 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1475 363577 : debug_assert!(
1476 363577 : converted
1477 363577 : .records
1478 1403254 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1479 0 : "{converted:?}"
1480 : );
1481 :
1482 363577 : let walredo_deltas = converted.num_deltas();
1483 363577 : let walredo_res = walredo_self
1484 363577 : .reconstruct_value(key, req_lsn_for_key, converted, redo_attempt_type)
1485 363577 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1486 0 : info_span!(
1487 : target: PERF_TRACE_TARGET,
1488 0 : parent: crnt_perf_span,
1489 : "WALREDO",
1490 : deltas = %walredo_deltas,
1491 : )
1492 0 : })
1493 363577 : .await;
1494 :
1495 363577 : (key, walredo_res)
1496 363577 : }
1497 : });
1498 : }
1499 :
1500 312448 : let results = futs
1501 312448 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1502 312448 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1503 312448 : .await;
1504 :
1505 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1506 : // when they're missing. Instead they are omitted from the resulting btree
1507 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1508 : // to avoid infinite results.
1509 312448 : if !results.is_empty() {
1510 312269 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1511 0 : let total_keyspace = query.total_keyspace();
1512 0 : let max_request_lsn = query.high_watermark_lsn().expect("Validated previously");
1513 :
1514 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1515 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1516 0 : LOG_PACER.lock().unwrap().call(|| {
1517 0 : let num_keys = total_keyspace.total_raw_size();
1518 0 : let num_pages = results.len();
1519 0 : tracing::info!(
1520 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1521 : lsn = %max_request_lsn,
1522 0 : "Vectored read for {total_keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1523 : );
1524 0 : });
1525 312269 : }
1526 :
1527 : // Records the number of layers visited in a few different ways:
1528 : //
1529 : // * LAYERS_PER_READ: all layers count towards every read in the batch, because each
1530 : // layer directly affects its observed latency.
1531 : //
1532 : // * LAYERS_PER_READ_BATCH: all layers count towards each batch, to get the per-batch
1533 : // layer visits and access cost.
1534 : //
1535 : // * LAYERS_PER_READ_AMORTIZED: the average layer count per read, to get the amortized
1536 : // read amplification after batching.
1537 312269 : let layers_visited = layers_visited as f64;
1538 312269 : let avg_layers_visited = layers_visited / results.len() as f64;
1539 312269 : LAYERS_PER_READ_BATCH_GLOBAL.observe(layers_visited);
1540 675846 : for _ in &results {
1541 363577 : self.metrics.layers_per_read.observe(layers_visited);
1542 363577 : LAYERS_PER_READ_GLOBAL.observe(layers_visited);
1543 363577 : LAYERS_PER_READ_AMORTIZED_GLOBAL.observe(avg_layers_visited);
1544 363577 : }
1545 179 : }
1546 :
1547 312448 : Ok(results)
1548 312456 : }
1549 :
1550 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1551 138202 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1552 138202 : self.last_record_lsn.load().last
1553 138202 : }
1554 :
1555 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1556 0 : self.last_record_lsn.load().prev
1557 0 : }
1558 :
1559 : /// Atomically get both last and prev.
1560 117 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1561 117 : self.last_record_lsn.load()
1562 117 : }
1563 :
1564 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1565 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1566 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1567 0 : self.last_record_lsn.status_receiver()
1568 0 : }
1569 :
1570 610 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1571 610 : self.disk_consistent_lsn.load()
1572 610 : }
1573 :
1574 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1575 : /// not validated with control plane yet.
1576 : /// See [`Self::get_remote_consistent_lsn_visible`].
1577 2 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1578 2 : self.remote_client.remote_consistent_lsn_projected()
1579 2 : }
1580 :
1581 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1582 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1583 : /// generation validation in the deletion queue.
1584 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1585 0 : self.remote_client.remote_consistent_lsn_visible()
1586 0 : }
1587 :
1588 : /// The sum of the file size of all historic layers in the layer map.
1589 : /// This method makes no distinction between local and remote layers.
1590 : /// Hence, the result **does not represent local filesystem usage**.
1591 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1592 0 : let guard = self
1593 0 : .layers
1594 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
1595 0 : .await;
1596 0 : guard.layer_size_sum()
1597 0 : }
1598 :
1599 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1600 0 : self.metrics.resident_physical_size_get()
1601 0 : }
1602 :
1603 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1604 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1605 0 : }
1606 :
1607 : ///
1608 : /// Wait until WAL has been received and processed up to this LSN.
1609 : ///
1610 : /// You should call this before any of the other get_* or list_* functions. Calling
1611 : /// those functions with an LSN that has been processed yet is an error.
1612 : ///
1613 114636 : pub(crate) async fn wait_lsn(
1614 114636 : &self,
1615 114636 : lsn: Lsn,
1616 114636 : who_is_waiting: WaitLsnWaiter<'_>,
1617 114636 : timeout: WaitLsnTimeout,
1618 114636 : ctx: &RequestContext, /* Prepare for use by cancellation */
1619 114636 : ) -> Result<(), WaitLsnError> {
1620 114636 : let state = self.current_state();
1621 114636 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1622 0 : return Err(WaitLsnError::Shutdown);
1623 114636 : } else if !matches!(state, TimelineState::Active) {
1624 0 : return Err(WaitLsnError::BadState(state));
1625 114636 : }
1626 :
1627 114636 : if cfg!(debug_assertions) {
1628 114636 : match ctx.task_kind() {
1629 : TaskKind::WalReceiverManager
1630 : | TaskKind::WalReceiverConnectionHandler
1631 : | TaskKind::WalReceiverConnectionPoller => {
1632 0 : let is_myself = match who_is_waiting {
1633 0 : WaitLsnWaiter::Timeline(waiter) => {
1634 0 : Weak::ptr_eq(&waiter.myself, &self.myself)
1635 : }
1636 : WaitLsnWaiter::Tenant
1637 : | WaitLsnWaiter::PageService
1638 : | WaitLsnWaiter::HttpEndpoint
1639 0 : | WaitLsnWaiter::BaseBackupCache => unreachable!(
1640 : "tenant or page_service context are not expected to have task kind {:?}",
1641 0 : ctx.task_kind()
1642 : ),
1643 : };
1644 0 : if is_myself {
1645 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1646 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1647 0 : panic!(
1648 0 : "this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock"
1649 : );
1650 0 : }
1651 0 : } else {
1652 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1653 0 : // our walreceiver task can make progress independent of theirs
1654 0 : }
1655 : }
1656 114636 : _ => {}
1657 : }
1658 0 : }
1659 :
1660 114636 : let timeout = match timeout {
1661 0 : WaitLsnTimeout::Custom(t) => t,
1662 114636 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1663 : };
1664 :
1665 114636 : let timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1666 114636 : let start_finish_counterpair_guard = self.metrics.wait_lsn_start_finish_counterpair.guard();
1667 :
1668 114636 : let wait_for_timeout = self.last_record_lsn.wait_for_timeout(lsn, timeout);
1669 114636 : let wait_for_timeout = std::pin::pin!(wait_for_timeout);
1670 : // Use threshold of 1 because even 1 second of wait for ingest is very much abnormal.
1671 114636 : let log_slow_threshold = Duration::from_secs(1);
1672 : // Use period of 10 to avoid flooding logs during an outage that affects all timelines.
1673 114636 : let log_slow_period = Duration::from_secs(10);
1674 114636 : let mut logging_permit = None;
1675 114636 : let wait_for_timeout = monitor_slow_future(
1676 114636 : log_slow_threshold,
1677 114636 : log_slow_period,
1678 114636 : wait_for_timeout,
1679 : |MonitorSlowFutureCallback {
1680 : ready,
1681 : is_slow,
1682 : elapsed_total,
1683 : elapsed_since_last_callback,
1684 114636 : }| {
1685 114636 : self.metrics
1686 114636 : .wait_lsn_in_progress_micros
1687 114636 : .inc_by(u64::try_from(elapsed_since_last_callback.as_micros()).unwrap());
1688 114636 : if !is_slow {
1689 114636 : return;
1690 0 : }
1691 : // It's slow, see if we should log it.
1692 : // (We limit the logging to one per invocation per timeline to avoid excessive
1693 : // logging during an extended broker / networking outage that affects all timelines.)
1694 0 : if logging_permit.is_none() {
1695 0 : logging_permit = self.wait_lsn_log_slow.try_acquire().ok();
1696 0 : }
1697 0 : if logging_permit.is_none() {
1698 0 : return;
1699 0 : }
1700 : // We log it.
1701 0 : if ready {
1702 0 : info!(
1703 0 : "slow wait_lsn completed after {:.3}s",
1704 0 : elapsed_total.as_secs_f64()
1705 : );
1706 : } else {
1707 0 : info!(
1708 0 : "slow wait_lsn still running for {:.3}s",
1709 0 : elapsed_total.as_secs_f64()
1710 : );
1711 : }
1712 114636 : },
1713 : );
1714 114636 : let res = wait_for_timeout.await;
1715 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1716 114636 : drop(logging_permit);
1717 114636 : drop(start_finish_counterpair_guard);
1718 114636 : drop(timer);
1719 114636 : match res {
1720 114636 : Ok(()) => Ok(()),
1721 0 : Err(e) => {
1722 : use utils::seqwait::SeqWaitError::*;
1723 0 : match e {
1724 0 : Shutdown => Err(WaitLsnError::Shutdown),
1725 : Timeout => {
1726 0 : let walreceiver_status = self.walreceiver_status();
1727 0 : Err(WaitLsnError::Timeout(format!(
1728 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1729 0 : lsn,
1730 0 : self.get_last_record_lsn(),
1731 0 : self.get_disk_consistent_lsn(),
1732 0 : walreceiver_status,
1733 0 : )))
1734 : }
1735 : }
1736 : }
1737 : }
1738 114636 : }
1739 :
1740 0 : pub(crate) fn walreceiver_status(&self) -> String {
1741 0 : match &*self.walreceiver.lock().unwrap() {
1742 0 : None => "stopping or stopped".to_string(),
1743 0 : Some(walreceiver) => match walreceiver.status() {
1744 0 : Some(status) => status.to_human_readable_string(),
1745 0 : None => "Not active".to_string(),
1746 : },
1747 : }
1748 0 : }
1749 :
1750 : /// Check that it is valid to request operations with that lsn.
1751 119 : pub(crate) fn check_lsn_is_in_scope(
1752 119 : &self,
1753 119 : lsn: Lsn,
1754 119 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1755 119 : ) -> anyhow::Result<()> {
1756 119 : ensure!(
1757 119 : lsn >= **latest_gc_cutoff_lsn,
1758 2 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1759 : lsn,
1760 2 : **latest_gc_cutoff_lsn,
1761 : );
1762 117 : Ok(())
1763 119 : }
1764 :
1765 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1766 5 : pub(crate) fn init_lsn_lease(
1767 5 : &self,
1768 5 : lsn: Lsn,
1769 5 : length: Duration,
1770 5 : ctx: &RequestContext,
1771 5 : ) -> anyhow::Result<LsnLease> {
1772 5 : self.make_lsn_lease(lsn, length, true, ctx)
1773 5 : }
1774 :
1775 : /// 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.
1776 2 : pub(crate) fn renew_lsn_lease(
1777 2 : &self,
1778 2 : lsn: Lsn,
1779 2 : length: Duration,
1780 2 : ctx: &RequestContext,
1781 2 : ) -> anyhow::Result<LsnLease> {
1782 2 : self.make_lsn_lease(lsn, length, false, ctx)
1783 2 : }
1784 :
1785 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1786 : ///
1787 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1788 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1789 : ///
1790 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1791 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1792 7 : fn make_lsn_lease(
1793 7 : &self,
1794 7 : lsn: Lsn,
1795 7 : length: Duration,
1796 7 : init: bool,
1797 7 : _ctx: &RequestContext,
1798 7 : ) -> anyhow::Result<LsnLease> {
1799 6 : let lease = {
1800 : // Normalize the requested LSN to be aligned, and move to the first record
1801 : // if it points to the beginning of the page (header).
1802 7 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1803 :
1804 7 : let mut gc_info = self.gc_info.write().unwrap();
1805 7 : let planned_cutoff = gc_info.min_cutoff();
1806 :
1807 7 : let valid_until = SystemTime::now() + length;
1808 :
1809 7 : let entry = gc_info.leases.entry(lsn);
1810 :
1811 7 : match entry {
1812 3 : Entry::Occupied(mut occupied) => {
1813 3 : let existing_lease = occupied.get_mut();
1814 3 : if valid_until > existing_lease.valid_until {
1815 1 : existing_lease.valid_until = valid_until;
1816 1 : let dt: DateTime<Utc> = valid_until.into();
1817 1 : info!("lease extended to {}", dt);
1818 : } else {
1819 2 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1820 2 : info!("existing lease covers greater length, valid until {}", dt);
1821 : }
1822 :
1823 3 : existing_lease.clone()
1824 : }
1825 4 : Entry::Vacant(vacant) => {
1826 : // Never allow a lease to be requested for an LSN below the applied GC cutoff. The data could have been deleted.
1827 4 : let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
1828 4 : if lsn < *latest_gc_cutoff_lsn {
1829 1 : bail!(
1830 1 : "tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}",
1831 : lsn,
1832 1 : *latest_gc_cutoff_lsn
1833 : );
1834 3 : }
1835 :
1836 : // We allow create lease for those below the planned gc cutoff if we are still within the grace period
1837 : // of GC blocking.
1838 3 : let validate = {
1839 3 : let conf = self.tenant_conf.load();
1840 3 : !conf.is_gc_blocked_by_lsn_lease_deadline()
1841 : };
1842 :
1843 : // Do not allow initial lease creation to be below the planned gc cutoff. The client (compute_ctl) determines
1844 : // whether it is a initial lease creation or a renewal.
1845 3 : if (init || validate) && lsn < planned_cutoff {
1846 0 : bail!(
1847 0 : "tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}",
1848 : lsn,
1849 : planned_cutoff
1850 : );
1851 3 : }
1852 :
1853 3 : let dt: DateTime<Utc> = valid_until.into();
1854 3 : info!("lease created, valid until {}", dt);
1855 3 : vacant.insert(LsnLease { valid_until }).clone()
1856 : }
1857 : }
1858 : };
1859 :
1860 6 : Ok(lease)
1861 7 : }
1862 :
1863 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1864 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1865 : #[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))]
1866 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1867 : self.freeze0().await
1868 : }
1869 :
1870 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1871 : #[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))]
1872 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1873 : self.freeze_and_flush0().await
1874 : }
1875 :
1876 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1877 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1878 571 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1879 571 : let mut g = self.write_lock.lock().await;
1880 571 : let to_lsn = self.get_last_record_lsn();
1881 571 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1882 571 : }
1883 :
1884 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1885 : // polluting the span hierarchy.
1886 571 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1887 571 : let token = self.freeze0().await?;
1888 571 : self.wait_flush_completion(token).await
1889 571 : }
1890 :
1891 : // Check if an open ephemeral layer should be closed: this provides
1892 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1893 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1894 : // ephemeral layer bytes has been breached.
1895 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1896 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1897 : // If the write lock is held, there is an active wal receiver: rolling open layers
1898 : // is their responsibility while they hold this lock.
1899 0 : return;
1900 : };
1901 :
1902 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1903 : // time, and this was missed.
1904 : // if write_guard.is_none() { return; }
1905 :
1906 0 : let Ok(layers_guard) = self.layers.try_read(LayerManagerLockHolder::TryFreezeLayer) else {
1907 : // Don't block if the layer lock is busy
1908 0 : return;
1909 : };
1910 :
1911 0 : let Ok(lm) = layers_guard.layer_map() else {
1912 0 : return;
1913 : };
1914 :
1915 0 : let Some(open_layer) = &lm.open_layer else {
1916 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1917 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1918 : // that didn't result in writes to this shard.
1919 :
1920 : // Must not hold the layers lock while waiting for a flush.
1921 0 : drop(layers_guard);
1922 :
1923 0 : let last_record_lsn = self.get_last_record_lsn();
1924 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1925 0 : if last_record_lsn > disk_consistent_lsn {
1926 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1927 : // we are a sharded tenant and have skipped some WAL
1928 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1929 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1930 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1931 : // without any data ingested (yet) doesn't write a remote index as soon as it
1932 : // sees its LSN advance: we only do this if we've been layer-less
1933 : // for some time.
1934 0 : tracing::debug!(
1935 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1936 : disk_consistent_lsn,
1937 : last_record_lsn
1938 : );
1939 :
1940 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1941 : // We know there is no open layer, so we can request freezing without actually
1942 : // freezing anything. This is true even if we have dropped the layers_guard, we
1943 : // still hold the write_guard.
1944 0 : let _ = async {
1945 0 : let token = self
1946 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1947 0 : .await?;
1948 0 : self.wait_flush_completion(token).await
1949 0 : }
1950 0 : .await;
1951 0 : }
1952 0 : }
1953 :
1954 0 : return;
1955 : };
1956 :
1957 0 : let current_size = open_layer.len();
1958 :
1959 0 : let current_lsn = self.get_last_record_lsn();
1960 :
1961 0 : let checkpoint_distance_override = open_layer.tick();
1962 :
1963 0 : if let Some(size_override) = checkpoint_distance_override {
1964 0 : if current_size > size_override {
1965 : // This is not harmful, but it only happens in relatively rare cases where
1966 : // time-based checkpoints are not happening fast enough to keep the amount of
1967 : // ephemeral data within configured limits. It's a sign of stress on the system.
1968 0 : tracing::info!(
1969 0 : "Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure"
1970 : );
1971 0 : }
1972 0 : }
1973 :
1974 0 : let checkpoint_distance =
1975 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1976 :
1977 0 : if self.should_roll(
1978 0 : current_size,
1979 0 : current_size,
1980 0 : checkpoint_distance,
1981 0 : self.get_last_record_lsn(),
1982 0 : self.last_freeze_at.load(),
1983 0 : open_layer.get_opened_at(),
1984 : ) {
1985 0 : match open_layer.info() {
1986 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1987 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1988 : // happens asynchronously in the background.
1989 0 : tracing::debug!(
1990 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1991 : );
1992 : }
1993 : InMemoryLayerInfo::Open { .. } => {
1994 : // Upgrade to a write lock and freeze the layer
1995 0 : drop(layers_guard);
1996 0 : let res = self
1997 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1998 0 : .await;
1999 :
2000 0 : if let Err(e) = res {
2001 0 : tracing::info!(
2002 0 : "failed to flush frozen layer after background freeze: {e:#}"
2003 : );
2004 0 : }
2005 : }
2006 : }
2007 0 : }
2008 0 : }
2009 :
2010 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
2011 : ///
2012 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
2013 : /// child timelines that are not offloaded yet.
2014 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
2015 0 : if self.remote_client.is_archived() != Some(true) {
2016 0 : return (false, "the timeline is not archived");
2017 0 : }
2018 0 : if !self.remote_client.no_pending_work() {
2019 : // if the remote client is still processing some work, we can't offload
2020 0 : return (false, "the upload queue is not drained yet");
2021 0 : }
2022 :
2023 0 : (true, "ok")
2024 0 : }
2025 :
2026 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
2027 : /// compaction tasks.
2028 192 : pub(crate) async fn compact(
2029 192 : self: &Arc<Self>,
2030 192 : cancel: &CancellationToken,
2031 192 : flags: EnumSet<CompactFlags>,
2032 192 : ctx: &RequestContext,
2033 192 : ) -> Result<CompactionOutcome, CompactionError> {
2034 192 : let res = self
2035 192 : .compact_with_options(
2036 192 : cancel,
2037 192 : CompactOptions {
2038 192 : flags,
2039 192 : compact_key_range: None,
2040 192 : compact_lsn_range: None,
2041 192 : sub_compaction: false,
2042 192 : sub_compaction_max_job_size_mb: None,
2043 192 : },
2044 192 : ctx,
2045 192 : )
2046 192 : .await;
2047 192 : if let Err(err) = &res {
2048 0 : log_compaction_error(err, None, cancel.is_cancelled(), false);
2049 192 : }
2050 192 : res
2051 192 : }
2052 :
2053 : /// Outermost timeline compaction operation; downloads needed layers.
2054 : ///
2055 : /// NB: the cancellation token is usually from a background task, but can also come from a
2056 : /// request task.
2057 192 : pub(crate) async fn compact_with_options(
2058 192 : self: &Arc<Self>,
2059 192 : cancel: &CancellationToken,
2060 192 : options: CompactOptions,
2061 192 : ctx: &RequestContext,
2062 192 : ) -> Result<CompactionOutcome, CompactionError> {
2063 : // Acquire the compaction lock and task semaphore.
2064 : //
2065 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
2066 : // out by other background tasks (including image compaction). We request this via
2067 : // `BackgroundLoopKind::L0Compaction`.
2068 : //
2069 : // Yield for pending L0 compaction while waiting for the semaphore.
2070 192 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
2071 192 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
2072 0 : true => BackgroundLoopKind::L0Compaction,
2073 192 : false => BackgroundLoopKind::Compaction,
2074 : };
2075 192 : let yield_for_l0 = options.flags.contains(CompactFlags::YieldForL0);
2076 192 : if yield_for_l0 {
2077 : // If this is an L0 pass, it doesn't make sense to yield for L0.
2078 0 : debug_assert!(!is_l0_only, "YieldForL0 during L0 pass");
2079 : // If `compaction_l0_first` is disabled, there's no point yielding.
2080 0 : debug_assert!(self.get_compaction_l0_first(), "YieldForL0 without L0 pass");
2081 192 : }
2082 :
2083 192 : let acquire = async move {
2084 192 : let guard = self.compaction_lock.lock().await;
2085 192 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
2086 192 : (guard, permit)
2087 192 : };
2088 :
2089 192 : let (_guard, _permit) = tokio::select! {
2090 192 : (guard, permit) = acquire => (guard, permit),
2091 192 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
2092 0 : return Ok(CompactionOutcome::YieldForL0);
2093 : }
2094 192 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2095 192 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2096 : };
2097 :
2098 192 : let last_record_lsn = self.get_last_record_lsn();
2099 :
2100 : // Last record Lsn could be zero in case the timeline was just created
2101 192 : if !last_record_lsn.is_valid() {
2102 0 : warn!(
2103 0 : "Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}"
2104 : );
2105 0 : return Ok(CompactionOutcome::Skipped);
2106 192 : }
2107 :
2108 192 : let result = match self.get_compaction_algorithm_settings().kind {
2109 : CompactionAlgorithm::Tiered => {
2110 0 : self.compact_tiered(cancel, ctx).await?;
2111 0 : Ok(CompactionOutcome::Done)
2112 : }
2113 192 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
2114 : };
2115 :
2116 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
2117 0 : match &result {
2118 192 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
2119 0 : Err(e) if e.is_cancel() => {}
2120 0 : Err(_) => self.compaction_failed.store(true, AtomicOrdering::Relaxed),
2121 : };
2122 :
2123 192 : result
2124 192 : }
2125 :
2126 : /// Mutate the timeline with a [`TimelineWriter`].
2127 2566609 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
2128 : TimelineWriter {
2129 2566609 : tl: self,
2130 2566609 : write_guard: self.write_lock.lock().await,
2131 : }
2132 2566609 : }
2133 :
2134 0 : pub(crate) fn activate(
2135 0 : self: &Arc<Self>,
2136 0 : parent: Arc<crate::tenant::TenantShard>,
2137 0 : broker_client: BrokerClientChannel,
2138 0 : background_jobs_can_start: Option<&completion::Barrier>,
2139 0 : ctx: &RequestContext,
2140 0 : ) {
2141 0 : if self.tenant_shard_id.is_shard_zero() {
2142 0 : // Logical size is only maintained accurately on shard zero.
2143 0 : self.spawn_initial_logical_size_computation_task(ctx);
2144 0 : }
2145 0 : self.launch_wal_receiver(ctx, broker_client);
2146 0 : self.set_state(TimelineState::Active);
2147 0 : self.launch_eviction_task(parent, background_jobs_can_start);
2148 0 : }
2149 :
2150 : /// After this function returns, there are no timeline-scoped tasks are left running.
2151 : ///
2152 : /// The preferred pattern for is:
2153 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
2154 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
2155 : /// go the extra mile and keep track of JoinHandles
2156 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
2157 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
2158 : ///
2159 : /// For legacy reasons, we still have multiple tasks spawned using
2160 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
2161 : /// We refer to these as "timeline-scoped task_mgr tasks".
2162 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
2163 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
2164 : /// or [`task_mgr::shutdown_watcher`].
2165 : /// We want to gradually convert the code base away from these.
2166 : ///
2167 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
2168 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
2169 : /// ones that aren't mentioned here):
2170 : /// - [`TaskKind::TimelineDeletionWorker`]
2171 : /// - NB: also used for tenant deletion
2172 : /// - [`TaskKind::RemoteUploadTask`]`
2173 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
2174 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
2175 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
2176 : /// - [`TaskKind::Eviction`]
2177 : /// - [`TaskKind::LayerFlushTask`]
2178 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
2179 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
2180 5 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
2181 5 : debug_assert_current_span_has_tenant_and_timeline_id();
2182 :
2183 : // Regardless of whether we're going to try_freeze_and_flush
2184 : // cancel walreceiver to stop ingesting more data asap.
2185 : //
2186 : // Note that we're accepting a race condition here where we may
2187 : // do the final flush below, before walreceiver observes the
2188 : // cancellation and exits.
2189 : // This means we may open a new InMemoryLayer after the final flush below.
2190 : // Flush loop is also still running for a short while, so, in theory, it
2191 : // could also make its way into the upload queue.
2192 : //
2193 : // If we wait for the shutdown of the walreceiver before moving on to the
2194 : // flush, then that would be avoided. But we don't do it because the
2195 : // walreceiver entertains reads internally, which means that it possibly
2196 : // depends on the download of layers. Layer download is only sensitive to
2197 : // the cancellation of the entire timeline, so cancelling the walreceiver
2198 : // will have no effect on the individual get requests.
2199 : // This would cause problems when there is a lot of ongoing downloads or
2200 : // there is S3 unavailabilities, i.e. detach, deletion, etc would hang,
2201 : // and we can't deallocate resources of the timeline, etc.
2202 5 : let walreceiver = self.walreceiver.lock().unwrap().take();
2203 5 : tracing::debug!(
2204 0 : is_some = walreceiver.is_some(),
2205 0 : "Waiting for WalReceiverManager..."
2206 : );
2207 5 : if let Some(walreceiver) = walreceiver {
2208 0 : walreceiver.cancel().await;
2209 5 : }
2210 : // ... and inform any waiters for newer LSNs that there won't be any.
2211 5 : self.last_record_lsn.shutdown();
2212 :
2213 5 : if let ShutdownMode::FreezeAndFlush = mode {
2214 3 : let do_flush = if let Some((open, frozen)) = self
2215 3 : .layers
2216 3 : .read(LayerManagerLockHolder::Shutdown)
2217 3 : .await
2218 3 : .layer_map()
2219 3 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
2220 3 : .ok()
2221 3 : .filter(|(open, frozen)| *open || *frozen > 0)
2222 : {
2223 0 : if self.remote_client.is_archived() == Some(true) {
2224 : // No point flushing on shutdown for an archived timeline: it is not important
2225 : // to have it nice and fresh after our restart, and trying to flush here might
2226 : // race with trying to offload it (which also stops the flush loop)
2227 0 : false
2228 : } else {
2229 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
2230 0 : true
2231 : }
2232 : } else {
2233 : // this is double-shutdown, it'll be a no-op
2234 3 : true
2235 : };
2236 :
2237 : // we shut down walreceiver above, so, we won't add anything more
2238 : // to the InMemoryLayer; freeze it and wait for all frozen layers
2239 : // to reach the disk & upload queue, then shut the upload queue and
2240 : // wait for it to drain.
2241 3 : if do_flush {
2242 3 : match self.freeze_and_flush().await {
2243 : Ok(_) => {
2244 : // drain the upload queue
2245 : // if we did not wait for completion here, it might be our shutdown process
2246 : // didn't wait for remote uploads to complete at all, as new tasks can forever
2247 : // be spawned.
2248 : //
2249 : // what is problematic is the shutting down of RemoteTimelineClient, because
2250 : // obviously it does not make sense to stop while we wait for it, but what
2251 : // about corner cases like s3 suddenly hanging up?
2252 3 : self.remote_client.shutdown().await;
2253 : }
2254 : Err(FlushLayerError::Cancelled) => {
2255 : // this is likely the second shutdown, ignore silently.
2256 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2257 0 : debug_assert!(self.cancel.is_cancelled());
2258 : }
2259 0 : Err(e) => {
2260 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
2261 : // we have some extra WAL replay to do next time the timeline starts.
2262 0 : warn!("failed to freeze and flush: {e:#}");
2263 : }
2264 : }
2265 :
2266 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2267 : // we also do a final check here to ensure that the queue is empty.
2268 3 : if !self.remote_client.no_pending_work() {
2269 0 : warn!(
2270 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2271 : );
2272 3 : }
2273 0 : }
2274 2 : }
2275 :
2276 5 : if let ShutdownMode::Reload = mode {
2277 : // drain the upload queue
2278 1 : self.remote_client.shutdown().await;
2279 1 : if !self.remote_client.no_pending_work() {
2280 0 : warn!(
2281 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2282 : );
2283 1 : }
2284 4 : }
2285 :
2286 : // Signal any subscribers to our cancellation token to drop out
2287 5 : tracing::debug!("Cancelling CancellationToken");
2288 5 : self.cancel.cancel();
2289 :
2290 : // If we have a background task downloading heatmap layers stop it.
2291 : // The background downloads are sensitive to timeline cancellation (done above),
2292 : // so the drain will be immediate.
2293 5 : self.stop_and_drain_heatmap_layers_download().await;
2294 :
2295 : // Ensure Prevent new page service requests from starting.
2296 5 : self.handles.shutdown();
2297 :
2298 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2299 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2300 5 : self.remote_client.stop();
2301 :
2302 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2303 : // to shut down the upload queue tasks.
2304 : // TODO: fix that, task management should be encapsulated inside remote_client.
2305 5 : task_mgr::shutdown_tasks(
2306 5 : Some(TaskKind::RemoteUploadTask),
2307 5 : Some(self.tenant_shard_id),
2308 5 : Some(self.timeline_id),
2309 5 : )
2310 5 : .await;
2311 :
2312 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2313 5 : tracing::debug!("Waiting for tasks...");
2314 5 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2315 :
2316 : {
2317 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2318 : // open.
2319 5 : let mut write_guard = self.write_lock.lock().await;
2320 5 : self.layers
2321 5 : .write(LayerManagerLockHolder::Shutdown)
2322 5 : .await
2323 5 : .shutdown(&mut write_guard);
2324 : }
2325 :
2326 : // Finally wait until any gate-holders are complete.
2327 : //
2328 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2329 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2330 5 : self.gate.close().await;
2331 :
2332 5 : self.metrics.shutdown();
2333 5 : }
2334 :
2335 236 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2336 236 : match (self.current_state(), new_state) {
2337 236 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2338 1 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2339 : }
2340 0 : (st, TimelineState::Loading) => {
2341 0 : error!("ignoring transition from {st:?} into Loading state");
2342 : }
2343 0 : (TimelineState::Broken { .. }, new_state) => {
2344 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2345 : }
2346 : (TimelineState::Stopping, TimelineState::Active) => {
2347 0 : error!("Not activating a Stopping timeline");
2348 : }
2349 235 : (_, new_state) => {
2350 235 : self.state.send_replace(new_state);
2351 235 : }
2352 : }
2353 236 : }
2354 :
2355 1 : pub(crate) fn set_broken(&self, reason: String) {
2356 1 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2357 1 : let broken_state = TimelineState::Broken {
2358 1 : reason,
2359 1 : backtrace: backtrace_str,
2360 1 : };
2361 1 : self.set_state(broken_state);
2362 :
2363 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2364 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2365 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2366 1 : self.cancel.cancel();
2367 1 : }
2368 :
2369 115567 : pub(crate) fn current_state(&self) -> TimelineState {
2370 115567 : self.state.borrow().clone()
2371 115567 : }
2372 :
2373 3 : pub(crate) fn is_broken(&self) -> bool {
2374 3 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2375 3 : }
2376 :
2377 126 : pub(crate) fn is_active(&self) -> bool {
2378 126 : self.current_state() == TimelineState::Active
2379 126 : }
2380 :
2381 8 : pub(crate) fn is_archived(&self) -> Option<bool> {
2382 8 : self.remote_client.is_archived()
2383 8 : }
2384 :
2385 8 : pub(crate) fn is_invisible(&self) -> Option<bool> {
2386 8 : self.remote_client.is_invisible()
2387 8 : }
2388 :
2389 569 : pub(crate) fn is_stopping(&self) -> bool {
2390 569 : self.current_state() == TimelineState::Stopping
2391 569 : }
2392 :
2393 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2394 0 : self.state.subscribe()
2395 0 : }
2396 :
2397 114637 : pub(crate) async fn wait_to_become_active(
2398 114637 : &self,
2399 114637 : _ctx: &RequestContext, // Prepare for use by cancellation
2400 114637 : ) -> Result<(), TimelineState> {
2401 114637 : let mut receiver = self.state.subscribe();
2402 : loop {
2403 114637 : let current_state = receiver.borrow().clone();
2404 114637 : match current_state {
2405 : TimelineState::Loading => {
2406 0 : receiver
2407 0 : .changed()
2408 0 : .await
2409 0 : .expect("holding a reference to self");
2410 : }
2411 : TimelineState::Active => {
2412 114636 : return Ok(());
2413 : }
2414 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2415 : // There's no chance the timeline can transition back into ::Active
2416 1 : return Err(current_state);
2417 : }
2418 : }
2419 : }
2420 114637 : }
2421 :
2422 0 : pub(crate) async fn layer_map_info(
2423 0 : &self,
2424 0 : reset: LayerAccessStatsReset,
2425 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2426 0 : let guard = self
2427 0 : .layers
2428 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
2429 0 : .await;
2430 0 : let layer_map = guard.layer_map()?;
2431 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2432 0 : if let Some(open_layer) = &layer_map.open_layer {
2433 0 : in_memory_layers.push(open_layer.info());
2434 0 : }
2435 0 : for frozen_layer in &layer_map.frozen_layers {
2436 0 : in_memory_layers.push(frozen_layer.info());
2437 0 : }
2438 :
2439 0 : let historic_layers = layer_map
2440 0 : .iter_historic_layers()
2441 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2442 0 : .collect();
2443 :
2444 0 : Ok(LayerMapInfo {
2445 0 : in_memory_layers,
2446 0 : historic_layers,
2447 0 : })
2448 0 : }
2449 :
2450 : #[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))]
2451 : pub(crate) async fn download_layer(
2452 : &self,
2453 : layer_file_name: &LayerName,
2454 : ctx: &RequestContext,
2455 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2456 : let Some(layer) = self
2457 : .find_layer(layer_file_name)
2458 : .await
2459 0 : .map_err(|e| match e {
2460 : layer_manager::Shutdown => {
2461 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2462 : }
2463 0 : })?
2464 : else {
2465 : return Ok(None);
2466 : };
2467 :
2468 : layer.download(ctx).await?;
2469 :
2470 : Ok(Some(true))
2471 : }
2472 :
2473 : /// Evict just one layer.
2474 : ///
2475 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2476 0 : pub(crate) async fn evict_layer(
2477 0 : &self,
2478 0 : layer_file_name: &LayerName,
2479 0 : ) -> anyhow::Result<Option<bool>> {
2480 0 : let _gate = self
2481 0 : .gate
2482 0 : .enter()
2483 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2484 :
2485 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2486 0 : return Ok(None);
2487 : };
2488 :
2489 : // curl has this by default
2490 0 : let timeout = std::time::Duration::from_secs(120);
2491 :
2492 0 : match local_layer.evict_and_wait(timeout).await {
2493 0 : Ok(()) => Ok(Some(true)),
2494 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2495 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2496 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2497 : }
2498 0 : }
2499 :
2500 2401506 : fn should_roll(
2501 2401506 : &self,
2502 2401506 : layer_size: u64,
2503 2401506 : projected_layer_size: u64,
2504 2401506 : checkpoint_distance: u64,
2505 2401506 : projected_lsn: Lsn,
2506 2401506 : last_freeze_at: Lsn,
2507 2401506 : opened_at: Instant,
2508 2401506 : ) -> bool {
2509 2401506 : let distance = projected_lsn.widening_sub(last_freeze_at);
2510 :
2511 : // Rolling the open layer can be triggered by:
2512 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2513 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2514 : // account for how writes are distributed across shards: we expect each node to consume
2515 : // 1/count of the LSN on average.
2516 : // 2. The size of the currently open layer.
2517 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2518 : // up and suspend activity.
2519 2401506 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2520 0 : info!(
2521 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2522 : projected_lsn, layer_size, distance
2523 : );
2524 :
2525 0 : true
2526 2401506 : } else if projected_layer_size >= checkpoint_distance {
2527 : // NB: this check is relied upon by:
2528 40 : let _ = IndexEntry::validate_checkpoint_distance;
2529 40 : info!(
2530 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2531 : projected_lsn, layer_size, projected_layer_size
2532 : );
2533 :
2534 40 : true
2535 2401466 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2536 0 : info!(
2537 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2538 : projected_lsn,
2539 : layer_size,
2540 0 : opened_at.elapsed()
2541 : );
2542 :
2543 0 : true
2544 : } else {
2545 2401466 : false
2546 : }
2547 2401506 : }
2548 :
2549 1 : pub(crate) fn is_basebackup_cache_enabled(&self) -> bool {
2550 1 : let tenant_conf = self.tenant_conf.load();
2551 1 : tenant_conf
2552 1 : .tenant_conf
2553 1 : .basebackup_cache_enabled
2554 1 : .unwrap_or(self.conf.default_tenant_conf.basebackup_cache_enabled)
2555 1 : }
2556 :
2557 : /// Try to get a basebackup from the on-disk cache.
2558 0 : pub(crate) async fn get_cached_basebackup(&self, lsn: Lsn) -> Option<tokio::fs::File> {
2559 0 : self.basebackup_cache
2560 0 : .get(self.tenant_shard_id.tenant_id, self.timeline_id, lsn)
2561 0 : .await
2562 0 : }
2563 :
2564 : /// Convenience method to attempt fetching a basebackup for the timeline if enabled and safe for
2565 : /// the given request parameters.
2566 : ///
2567 : /// TODO: consider moving this onto GrpcPageServiceHandler once the libpq handler is gone.
2568 0 : pub async fn get_cached_basebackup_if_enabled(
2569 0 : &self,
2570 0 : lsn: Option<Lsn>,
2571 0 : prev_lsn: Option<Lsn>,
2572 0 : full: bool,
2573 0 : replica: bool,
2574 0 : gzip: bool,
2575 0 : ) -> Option<tokio::fs::File> {
2576 0 : if !self.is_basebackup_cache_enabled() || !self.basebackup_cache.is_enabled() {
2577 0 : return None;
2578 0 : }
2579 : // We have to know which LSN to fetch the basebackup for.
2580 0 : let lsn = lsn?;
2581 : // We only cache gzipped, non-full basebackups for primary computes with automatic prev_lsn.
2582 0 : if prev_lsn.is_some() || full || replica || !gzip {
2583 0 : return None;
2584 0 : }
2585 0 : self.get_cached_basebackup(lsn).await
2586 0 : }
2587 :
2588 : /// Prepare basebackup for the given LSN and store it in the basebackup cache.
2589 : /// The method is asynchronous and returns immediately.
2590 : /// The actual basebackup preparation is performed in the background
2591 : /// by the basebackup cache on a best-effort basis.
2592 1 : pub(crate) fn prepare_basebackup(&self, lsn: Lsn) {
2593 1 : if !self.is_basebackup_cache_enabled() {
2594 1 : return;
2595 0 : }
2596 0 : if !self.tenant_shard_id.is_shard_zero() {
2597 : // In theory we should never get here, but just in case check it.
2598 : // Preparing basebackup doesn't make sense for shards other than shard zero.
2599 0 : return;
2600 0 : }
2601 0 : if !self.is_active() {
2602 : // May happen during initial timeline creation.
2603 : // Such timeline is not in the global timeline map yet,
2604 : // so basebackup cache will not be able to find it.
2605 : // TODO(diko): We can prepare such timelines in finish_creation().
2606 0 : return;
2607 0 : }
2608 :
2609 0 : self.basebackup_cache
2610 0 : .send_prepare(self.tenant_shard_id, self.timeline_id, lsn);
2611 1 : }
2612 : }
2613 :
2614 : /// Number of times we will compute partition within a checkpoint distance.
2615 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2616 :
2617 : // Private functions
2618 : impl Timeline {
2619 6 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2620 6 : let tenant_conf = self.tenant_conf.load();
2621 6 : tenant_conf
2622 6 : .tenant_conf
2623 6 : .lsn_lease_length
2624 6 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2625 6 : }
2626 :
2627 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2628 0 : let tenant_conf = self.tenant_conf.load();
2629 0 : tenant_conf
2630 0 : .tenant_conf
2631 0 : .lsn_lease_length_for_ts
2632 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2633 0 : }
2634 :
2635 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2636 0 : let tenant_conf = self.tenant_conf.load();
2637 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2638 0 : }
2639 :
2640 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2641 0 : let tenant_conf = self.tenant_conf.load();
2642 0 : tenant_conf
2643 0 : .tenant_conf
2644 0 : .lazy_slru_download
2645 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2646 0 : }
2647 :
2648 : /// Checks if a get page request should get perf tracing
2649 : ///
2650 : /// The configuration priority is: tenant config override, default tenant config,
2651 : /// pageserver config.
2652 0 : pub(crate) fn is_get_page_request_sampled(&self) -> bool {
2653 0 : let tenant_conf = self.tenant_conf.load();
2654 0 : let ratio = tenant_conf
2655 0 : .tenant_conf
2656 0 : .sampling_ratio
2657 0 : .flatten()
2658 0 : .or(self.conf.default_tenant_conf.sampling_ratio)
2659 0 : .or(self.conf.tracing.as_ref().map(|t| t.sampling_ratio));
2660 :
2661 0 : match ratio {
2662 0 : Some(r) => {
2663 0 : if r.numerator == 0 {
2664 0 : false
2665 : } else {
2666 0 : rand::thread_rng().gen_range(0..r.denominator) < r.numerator
2667 : }
2668 : }
2669 0 : None => false,
2670 : }
2671 0 : }
2672 :
2673 2402187 : fn get_checkpoint_distance(&self) -> u64 {
2674 2402187 : let tenant_conf = self.tenant_conf.load();
2675 2402187 : tenant_conf
2676 2402187 : .tenant_conf
2677 2402187 : .checkpoint_distance
2678 2402187 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2679 2402187 : }
2680 :
2681 2401522 : fn get_checkpoint_timeout(&self) -> Duration {
2682 2401522 : let tenant_conf = self.tenant_conf.load();
2683 2401522 : tenant_conf
2684 2401522 : .tenant_conf
2685 2401522 : .checkpoint_timeout
2686 2401522 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2687 2401522 : }
2688 :
2689 1 : pub(crate) fn get_pitr_interval(&self) -> Duration {
2690 1 : let tenant_conf = &self.tenant_conf.load().tenant_conf;
2691 1 : tenant_conf
2692 1 : .pitr_interval
2693 1 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2694 1 : }
2695 :
2696 1273 : fn get_compaction_period(&self) -> Duration {
2697 1273 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2698 1273 : tenant_conf
2699 1273 : .compaction_period
2700 1273 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2701 1273 : }
2702 :
2703 352 : fn get_compaction_target_size(&self) -> u64 {
2704 352 : let tenant_conf = self.tenant_conf.load();
2705 352 : tenant_conf
2706 352 : .tenant_conf
2707 352 : .compaction_target_size
2708 352 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2709 352 : }
2710 :
2711 820 : fn get_compaction_threshold(&self) -> usize {
2712 820 : let tenant_conf = self.tenant_conf.load();
2713 820 : tenant_conf
2714 820 : .tenant_conf
2715 820 : .compaction_threshold
2716 820 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2717 820 : }
2718 :
2719 : /// Returns `true` if the rel_size_v2 config is enabled. NOTE: the write path and read path
2720 : /// should look at `get_rel_size_v2_status()` to get the actual status of the timeline. It is
2721 : /// possible that the index part persists the state while the config doesn't get persisted.
2722 973 : pub(crate) fn get_rel_size_v2_enabled(&self) -> bool {
2723 973 : let tenant_conf = self.tenant_conf.load();
2724 973 : tenant_conf
2725 973 : .tenant_conf
2726 973 : .rel_size_v2_enabled
2727 973 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
2728 973 : }
2729 :
2730 1099 : pub(crate) fn get_rel_size_v2_status(&self) -> RelSizeMigration {
2731 1099 : self.rel_size_v2_status
2732 1099 : .load()
2733 1099 : .as_ref()
2734 1099 : .map(|s| s.as_ref().clone())
2735 1099 : .unwrap_or(RelSizeMigration::Legacy)
2736 1099 : }
2737 :
2738 23 : fn get_compaction_upper_limit(&self) -> usize {
2739 23 : let tenant_conf = self.tenant_conf.load();
2740 23 : tenant_conf
2741 23 : .tenant_conf
2742 23 : .compaction_upper_limit
2743 23 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2744 23 : }
2745 :
2746 0 : pub fn get_compaction_l0_first(&self) -> bool {
2747 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2748 0 : tenant_conf
2749 0 : .compaction_l0_first
2750 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2751 0 : }
2752 :
2753 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2754 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2755 0 : tenant_conf
2756 0 : .compaction_l0_semaphore
2757 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2758 0 : }
2759 :
2760 636 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2761 : // By default, delay L0 flushes at 3x the compaction threshold. The compaction threshold
2762 : // defaults to 10, and L0 compaction is generally able to keep L0 counts below 30.
2763 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 3;
2764 :
2765 : // If compaction is disabled, don't delay.
2766 636 : if self.get_compaction_period() == Duration::ZERO {
2767 632 : return None;
2768 4 : }
2769 :
2770 4 : let compaction_threshold = self.get_compaction_threshold();
2771 4 : let tenant_conf = self.tenant_conf.load();
2772 4 : let l0_flush_delay_threshold = tenant_conf
2773 4 : .tenant_conf
2774 4 : .l0_flush_delay_threshold
2775 4 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2776 4 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2777 :
2778 : // 0 disables backpressure.
2779 4 : if l0_flush_delay_threshold == 0 {
2780 0 : return None;
2781 4 : }
2782 :
2783 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2784 : // backpressure flushes below this.
2785 : // TODO: the tenant config should have validation to prevent this instead.
2786 4 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2787 4 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2788 636 : }
2789 :
2790 637 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2791 : // Disable L0 stalls by default. Stalling can cause unavailability if L0 compaction isn't
2792 : // responsive, and it can e.g. block on other compaction via the compaction semaphore or
2793 : // sibling timelines. We need more confidence before enabling this.
2794 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2795 :
2796 : // If compaction is disabled, don't stall.
2797 637 : if self.get_compaction_period() == Duration::ZERO {
2798 632 : return None;
2799 5 : }
2800 :
2801 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2802 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2803 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2804 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2805 5 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2806 0 : return None;
2807 5 : }
2808 :
2809 5 : let compaction_threshold = self.get_compaction_threshold();
2810 5 : let tenant_conf = self.tenant_conf.load();
2811 5 : let l0_flush_stall_threshold = tenant_conf
2812 5 : .tenant_conf
2813 5 : .l0_flush_stall_threshold
2814 5 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2815 :
2816 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2817 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2818 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2819 5 : if cfg!(feature = "testing")
2820 5 : && compaction_threshold == 1
2821 0 : && l0_flush_stall_threshold.is_none()
2822 : {
2823 0 : return None;
2824 5 : }
2825 :
2826 5 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2827 5 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2828 :
2829 : // 0 disables backpressure.
2830 5 : if l0_flush_stall_threshold == 0 {
2831 5 : return None;
2832 0 : }
2833 :
2834 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2835 : // backpressure flushes below this.
2836 : // TODO: the tenant config should have validation to prevent this instead.
2837 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2838 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2839 637 : }
2840 :
2841 57 : fn get_image_creation_threshold(&self) -> usize {
2842 57 : let tenant_conf = self.tenant_conf.load();
2843 57 : tenant_conf
2844 57 : .tenant_conf
2845 57 : .image_creation_threshold
2846 57 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2847 57 : }
2848 :
2849 : // HADRON
2850 193 : fn get_image_layer_force_creation_period(&self) -> Option<Duration> {
2851 193 : let tenant_conf = self.tenant_conf.load();
2852 193 : tenant_conf
2853 193 : .tenant_conf
2854 193 : .image_layer_force_creation_period
2855 193 : .or(self
2856 193 : .conf
2857 193 : .default_tenant_conf
2858 193 : .image_layer_force_creation_period)
2859 193 : }
2860 :
2861 192 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2862 192 : let tenant_conf = &self.tenant_conf.load();
2863 192 : tenant_conf
2864 192 : .tenant_conf
2865 192 : .compaction_algorithm
2866 192 : .as_ref()
2867 192 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2868 192 : .clone()
2869 192 : }
2870 :
2871 192 : pub fn get_compaction_shard_ancestor(&self) -> bool {
2872 192 : let tenant_conf = self.tenant_conf.load();
2873 192 : tenant_conf
2874 192 : .tenant_conf
2875 192 : .compaction_shard_ancestor
2876 192 : .unwrap_or(self.conf.default_tenant_conf.compaction_shard_ancestor)
2877 192 : }
2878 :
2879 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2880 0 : let tenant_conf = self.tenant_conf.load();
2881 0 : tenant_conf
2882 0 : .tenant_conf
2883 0 : .eviction_policy
2884 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2885 0 : }
2886 :
2887 235 : fn get_evictions_low_residence_duration_metric_threshold(
2888 235 : tenant_conf: &pageserver_api::models::TenantConfig,
2889 235 : default_tenant_conf: &pageserver_api::config::TenantConfigToml,
2890 235 : ) -> Duration {
2891 235 : tenant_conf
2892 235 : .evictions_low_residence_duration_metric_threshold
2893 235 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2894 235 : }
2895 :
2896 191 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2897 191 : let tenant_conf = self.tenant_conf.load();
2898 191 : tenant_conf
2899 191 : .tenant_conf
2900 191 : .image_layer_creation_check_threshold
2901 191 : .unwrap_or(
2902 191 : self.conf
2903 191 : .default_tenant_conf
2904 191 : .image_layer_creation_check_threshold,
2905 : )
2906 191 : }
2907 :
2908 27 : fn get_gc_compaction_settings(&self) -> GcCompactionCombinedSettings {
2909 27 : let tenant_conf = &self.tenant_conf.load();
2910 27 : let gc_compaction_enabled = tenant_conf
2911 27 : .tenant_conf
2912 27 : .gc_compaction_enabled
2913 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_enabled);
2914 27 : let gc_compaction_verification = tenant_conf
2915 27 : .tenant_conf
2916 27 : .gc_compaction_verification
2917 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_verification);
2918 27 : let gc_compaction_initial_threshold_kb = tenant_conf
2919 27 : .tenant_conf
2920 27 : .gc_compaction_initial_threshold_kb
2921 27 : .unwrap_or(
2922 27 : self.conf
2923 27 : .default_tenant_conf
2924 27 : .gc_compaction_initial_threshold_kb,
2925 : );
2926 27 : let gc_compaction_ratio_percent = tenant_conf
2927 27 : .tenant_conf
2928 27 : .gc_compaction_ratio_percent
2929 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_ratio_percent);
2930 27 : GcCompactionCombinedSettings {
2931 27 : gc_compaction_enabled,
2932 27 : gc_compaction_verification,
2933 27 : gc_compaction_initial_threshold_kb,
2934 27 : gc_compaction_ratio_percent,
2935 27 : }
2936 27 : }
2937 :
2938 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
2939 0 : let tenant_conf = self.tenant_conf.load();
2940 0 : tenant_conf
2941 0 : .tenant_conf
2942 0 : .image_creation_preempt_threshold
2943 0 : .unwrap_or(
2944 0 : self.conf
2945 0 : .default_tenant_conf
2946 0 : .image_creation_preempt_threshold,
2947 : )
2948 0 : }
2949 :
2950 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2951 : // NB: Most tenant conf options are read by background loops, so,
2952 : // changes will automatically be picked up.
2953 :
2954 : // The threshold is embedded in the metric. So, we need to update it.
2955 : {
2956 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2957 0 : &new_conf.tenant_conf,
2958 0 : &self.conf.default_tenant_conf,
2959 : );
2960 :
2961 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2962 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2963 :
2964 0 : let timeline_id_str = self.timeline_id.to_string();
2965 :
2966 0 : self.remote_client.update_config(&new_conf.location);
2967 :
2968 0 : let mut rel_size_cache = self.rel_size_snapshot_cache.lock().unwrap();
2969 0 : if let Some(new_capacity) = new_conf.tenant_conf.relsize_snapshot_cache_capacity {
2970 0 : if new_capacity != rel_size_cache.capacity() {
2971 0 : rel_size_cache.set_capacity(new_capacity);
2972 0 : }
2973 0 : }
2974 :
2975 0 : self.metrics
2976 0 : .evictions_with_low_residence_duration
2977 0 : .write()
2978 0 : .unwrap()
2979 0 : .change_threshold(
2980 0 : &tenant_id_str,
2981 0 : &shard_id_str,
2982 0 : &timeline_id_str,
2983 0 : new_threshold,
2984 0 : );
2985 : }
2986 0 : }
2987 :
2988 : /// Open a Timeline handle.
2989 : ///
2990 : /// Loads the metadata for the timeline into memory, but not the layer map.
2991 : #[allow(clippy::too_many_arguments)]
2992 235 : pub(super) fn new(
2993 235 : conf: &'static PageServerConf,
2994 235 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2995 235 : metadata: &TimelineMetadata,
2996 235 : previous_heatmap: Option<PreviousHeatmap>,
2997 235 : ancestor: Option<Arc<Timeline>>,
2998 235 : timeline_id: TimelineId,
2999 235 : tenant_shard_id: TenantShardId,
3000 235 : generation: Generation,
3001 235 : shard_identity: ShardIdentity,
3002 235 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
3003 235 : resources: TimelineResources,
3004 235 : pg_version: PgMajorVersion,
3005 235 : state: TimelineState,
3006 235 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
3007 235 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
3008 235 : gc_compaction_state: Option<GcCompactionState>,
3009 235 : rel_size_v2_status: Option<RelSizeMigration>,
3010 235 : cancel: CancellationToken,
3011 235 : ) -> Arc<Self> {
3012 235 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
3013 235 : let (state, _) = watch::channel(state);
3014 :
3015 235 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
3016 235 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
3017 :
3018 235 : let evictions_low_residence_duration_metric_threshold = {
3019 235 : let loaded_tenant_conf = tenant_conf.load();
3020 235 : Self::get_evictions_low_residence_duration_metric_threshold(
3021 235 : &loaded_tenant_conf.tenant_conf,
3022 235 : &conf.default_tenant_conf,
3023 : )
3024 : };
3025 :
3026 235 : if let Some(ancestor) = &ancestor {
3027 118 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
3028 118 : // If we construct an explicit timeline object, it's obviously not offloaded
3029 118 : let is_offloaded = MaybeOffloaded::No;
3030 118 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
3031 118 : }
3032 :
3033 235 : let relsize_snapshot_cache_capacity = {
3034 235 : let loaded_tenant_conf = tenant_conf.load();
3035 235 : loaded_tenant_conf
3036 235 : .tenant_conf
3037 235 : .relsize_snapshot_cache_capacity
3038 235 : .unwrap_or(conf.default_tenant_conf.relsize_snapshot_cache_capacity)
3039 : };
3040 :
3041 235 : Arc::new_cyclic(|myself| {
3042 235 : let metrics = Arc::new(TimelineMetrics::new(
3043 235 : &tenant_shard_id,
3044 235 : &timeline_id,
3045 235 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
3046 : "mtime",
3047 235 : evictions_low_residence_duration_metric_threshold,
3048 : ),
3049 : ));
3050 235 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
3051 :
3052 235 : let mut result = Timeline {
3053 235 : conf,
3054 235 : tenant_conf,
3055 235 : myself: myself.clone(),
3056 235 : timeline_id,
3057 235 : tenant_shard_id,
3058 235 : generation,
3059 235 : shard_identity,
3060 235 : pg_version,
3061 235 : layers: Default::default(),
3062 235 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
3063 :
3064 235 : walredo_mgr,
3065 235 : walreceiver: Mutex::new(None),
3066 :
3067 235 : remote_client: Arc::new(resources.remote_client),
3068 :
3069 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
3070 235 : last_record_lsn: SeqWait::new(RecordLsn {
3071 235 : last: disk_consistent_lsn,
3072 235 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
3073 235 : }),
3074 235 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
3075 :
3076 235 : gc_compaction_state: ArcSwap::new(Arc::new(gc_compaction_state)),
3077 :
3078 235 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
3079 235 : last_freeze_ts: RwLock::new(Instant::now()),
3080 :
3081 235 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
3082 :
3083 235 : ancestor_timeline: ancestor,
3084 235 : ancestor_lsn: metadata.ancestor_lsn(),
3085 :
3086 235 : metrics,
3087 :
3088 235 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
3089 235 : &tenant_shard_id,
3090 235 : &timeline_id,
3091 235 : resources.pagestream_throttle_metrics,
3092 : ),
3093 :
3094 1880 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
3095 1880 : directory_metrics_inited: array::from_fn(|_| AtomicBool::new(false)),
3096 :
3097 235 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
3098 :
3099 235 : layer_flush_start_tx,
3100 235 : layer_flush_done_tx,
3101 :
3102 235 : write_lock: tokio::sync::Mutex::new(None),
3103 :
3104 235 : gc_info: std::sync::RwLock::new(GcInfo::default()),
3105 :
3106 235 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
3107 235 : LastImageLayerCreationStatus::default(),
3108 : )),
3109 :
3110 235 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
3111 235 : initdb_lsn: metadata.initdb_lsn(),
3112 :
3113 235 : current_logical_size: if disk_consistent_lsn.is_valid() {
3114 : // we're creating timeline data with some layer files existing locally,
3115 : // need to recalculate timeline's logical size based on data in the layers.
3116 120 : LogicalSize::deferred_initial(disk_consistent_lsn)
3117 : } else {
3118 : // we're creating timeline data without any layers existing locally,
3119 : // initial logical size is 0.
3120 115 : LogicalSize::empty_initial()
3121 : },
3122 :
3123 235 : partitioning: GuardArcSwap::new((
3124 235 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
3125 235 : Lsn(0),
3126 235 : )),
3127 : repartition_threshold: 0,
3128 235 : last_image_layer_creation_check_at: AtomicLsn::new(0),
3129 235 : last_image_layer_creation_check_instant: Mutex::new(None),
3130 235 : last_received_wal: Mutex::new(None),
3131 235 : rel_size_latest_cache: RwLock::new(HashMap::new()),
3132 235 : rel_size_snapshot_cache: Mutex::new(LruCache::new(relsize_snapshot_cache_capacity)),
3133 :
3134 235 : download_all_remote_layers_task_info: RwLock::new(None),
3135 :
3136 235 : state,
3137 :
3138 235 : eviction_task_timeline_state: tokio::sync::Mutex::new(
3139 235 : EvictionTaskTimelineState::default(),
3140 : ),
3141 235 : delete_progress: TimelineDeleteProgress::default(),
3142 :
3143 235 : cancel,
3144 235 : gate: Gate::default(),
3145 :
3146 235 : compaction_lock: tokio::sync::Mutex::default(),
3147 235 : compaction_failed: AtomicBool::default(),
3148 235 : l0_compaction_trigger: resources.l0_compaction_trigger,
3149 235 : gc_lock: tokio::sync::Mutex::default(),
3150 :
3151 235 : standby_horizon: AtomicLsn::new(0),
3152 :
3153 235 : pagestream_throttle: resources.pagestream_throttle,
3154 :
3155 235 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
3156 :
3157 : #[cfg(test)]
3158 235 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
3159 :
3160 235 : l0_flush_global_state: resources.l0_flush_global_state,
3161 :
3162 235 : handles: Default::default(),
3163 :
3164 235 : attach_wal_lag_cooldown,
3165 :
3166 235 : create_idempotency,
3167 :
3168 235 : page_trace: Default::default(),
3169 :
3170 235 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
3171 :
3172 235 : heatmap_layers_downloader: Mutex::new(None),
3173 :
3174 235 : rel_size_v2_status: ArcSwapOption::from_pointee(rel_size_v2_status),
3175 :
3176 235 : wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
3177 :
3178 235 : basebackup_cache: resources.basebackup_cache,
3179 :
3180 235 : feature_resolver: resources.feature_resolver.clone(),
3181 : };
3182 :
3183 235 : result.repartition_threshold =
3184 235 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
3185 :
3186 235 : result
3187 235 : .metrics
3188 235 : .last_record_lsn_gauge
3189 235 : .set(disk_consistent_lsn.0 as i64);
3190 235 : result
3191 235 : })
3192 235 : }
3193 :
3194 343 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
3195 343 : let Ok(guard) = self.gate.enter() else {
3196 0 : info!("cannot start flush loop when the timeline gate has already been closed");
3197 0 : return;
3198 : };
3199 343 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
3200 343 : match *flush_loop_state {
3201 232 : FlushLoopState::NotStarted => (),
3202 : FlushLoopState::Running { .. } => {
3203 111 : info!(
3204 0 : "skipping attempt to start flush_loop twice {}/{}",
3205 0 : self.tenant_shard_id, self.timeline_id
3206 : );
3207 111 : return;
3208 : }
3209 : FlushLoopState::Exited => {
3210 0 : info!(
3211 0 : "ignoring attempt to restart exited flush_loop {}/{}",
3212 0 : self.tenant_shard_id, self.timeline_id
3213 : );
3214 0 : return;
3215 : }
3216 : }
3217 :
3218 232 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
3219 232 : let self_clone = Arc::clone(self);
3220 :
3221 232 : debug!("spawning flush loop");
3222 232 : *flush_loop_state = FlushLoopState::Running {
3223 232 : #[cfg(test)]
3224 232 : expect_initdb_optimization: false,
3225 232 : #[cfg(test)]
3226 232 : initdb_optimization_count: 0,
3227 232 : };
3228 232 : task_mgr::spawn(
3229 232 : task_mgr::BACKGROUND_RUNTIME.handle(),
3230 232 : task_mgr::TaskKind::LayerFlushTask,
3231 232 : self.tenant_shard_id,
3232 232 : Some(self.timeline_id),
3233 232 : "layer flush task",
3234 232 : async move {
3235 232 : let _guard = guard;
3236 232 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error).with_scope_timeline(&self_clone);
3237 232 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
3238 5 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
3239 5 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
3240 5 : *flush_loop_state = FlushLoopState::Exited;
3241 5 : Ok(())
3242 5 : }
3243 232 : .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))
3244 : );
3245 343 : }
3246 :
3247 0 : pub(crate) fn update_gc_compaction_state(
3248 0 : &self,
3249 0 : gc_compaction_state: GcCompactionState,
3250 0 : ) -> anyhow::Result<()> {
3251 0 : self.gc_compaction_state
3252 0 : .store(Arc::new(Some(gc_compaction_state.clone())));
3253 0 : self.remote_client
3254 0 : .schedule_index_upload_for_gc_compaction_state_update(gc_compaction_state)
3255 0 : }
3256 :
3257 0 : pub(crate) fn update_rel_size_v2_status(
3258 0 : &self,
3259 0 : rel_size_v2_status: RelSizeMigration,
3260 0 : ) -> anyhow::Result<()> {
3261 0 : self.rel_size_v2_status
3262 0 : .store(Some(Arc::new(rel_size_v2_status.clone())));
3263 0 : self.remote_client
3264 0 : .schedule_index_upload_for_rel_size_v2_status_update(rel_size_v2_status)
3265 0 : }
3266 :
3267 0 : pub(crate) fn get_gc_compaction_state(&self) -> Option<GcCompactionState> {
3268 0 : self.gc_compaction_state.load_full().as_ref().clone()
3269 0 : }
3270 :
3271 : /// Creates and starts the wal receiver.
3272 : ///
3273 : /// This function is expected to be called at most once per Timeline's lifecycle
3274 : /// when the timeline is activated.
3275 0 : fn launch_wal_receiver(
3276 0 : self: &Arc<Self>,
3277 0 : ctx: &RequestContext,
3278 0 : broker_client: BrokerClientChannel,
3279 0 : ) {
3280 0 : info!(
3281 0 : "launching WAL receiver for timeline {} of tenant {}",
3282 0 : self.timeline_id, self.tenant_shard_id
3283 : );
3284 :
3285 0 : let tenant_conf = self.tenant_conf.load();
3286 0 : let wal_connect_timeout = tenant_conf
3287 0 : .tenant_conf
3288 0 : .walreceiver_connect_timeout
3289 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
3290 0 : let lagging_wal_timeout = tenant_conf
3291 0 : .tenant_conf
3292 0 : .lagging_wal_timeout
3293 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
3294 0 : let max_lsn_wal_lag = tenant_conf
3295 0 : .tenant_conf
3296 0 : .max_lsn_wal_lag
3297 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
3298 :
3299 0 : let mut guard = self.walreceiver.lock().unwrap();
3300 0 : assert!(
3301 0 : guard.is_none(),
3302 0 : "multiple launches / re-launches of WAL receiver are not supported"
3303 : );
3304 :
3305 0 : let protocol = PostgresClientProtocol::Interpreted {
3306 0 : format: utils::postgres_client::InterpretedFormat::Protobuf,
3307 0 : compression: Some(utils::postgres_client::Compression::Zstd { level: 1 }),
3308 0 : };
3309 :
3310 0 : *guard = Some(WalReceiver::start(
3311 0 : Arc::clone(self),
3312 0 : WalReceiverConf {
3313 0 : protocol,
3314 0 : wal_connect_timeout,
3315 0 : lagging_wal_timeout,
3316 0 : max_lsn_wal_lag,
3317 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
3318 0 : availability_zone: self.conf.availability_zone.clone(),
3319 0 : ingest_batch_size: self.conf.ingest_batch_size,
3320 0 : validate_wal_contiguity: self.conf.validate_wal_contiguity,
3321 0 : },
3322 0 : broker_client,
3323 0 : ctx,
3324 0 : ));
3325 0 : }
3326 :
3327 : /// Initialize with an empty layer map. Used when creating a new timeline.
3328 232 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
3329 232 : let mut layers = self.layers.try_write(LayerManagerLockHolder::Init).expect(
3330 232 : "in the context where we call this function, no other task has access to the object",
3331 : );
3332 232 : layers
3333 232 : .open_mut()
3334 232 : .expect("in this context the LayerManager must still be open")
3335 232 : .initialize_empty(Lsn(start_lsn.0));
3336 232 : }
3337 :
3338 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
3339 : /// files.
3340 3 : pub(super) async fn load_layer_map(
3341 3 : &self,
3342 3 : disk_consistent_lsn: Lsn,
3343 3 : index_part: IndexPart,
3344 3 : ) -> anyhow::Result<()> {
3345 : use LayerName::*;
3346 : use init::Decision::*;
3347 : use init::{Discovered, DismissedLayer};
3348 :
3349 3 : let mut guard = self
3350 3 : .layers
3351 3 : .write(LayerManagerLockHolder::LoadLayerMap)
3352 3 : .await;
3353 :
3354 3 : let timer = self.metrics.load_layer_map_histo.start_timer();
3355 :
3356 : // Scan timeline directory and create ImageLayerName and DeltaFilename
3357 : // structs representing all files on disk
3358 3 : let timeline_path = self
3359 3 : .conf
3360 3 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
3361 3 : let conf = self.conf;
3362 3 : let span = tracing::Span::current();
3363 :
3364 : // Copy to move into the task we're about to spawn
3365 3 : let this = self.myself.upgrade().expect("&self method holds the arc");
3366 :
3367 3 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
3368 3 : move || {
3369 3 : let _g = span.entered();
3370 3 : let discovered = init::scan_timeline_dir(&timeline_path)?;
3371 3 : let mut discovered_layers = Vec::with_capacity(discovered.len());
3372 3 : let mut unrecognized_files = Vec::new();
3373 :
3374 3 : let mut path = timeline_path;
3375 :
3376 11 : for discovered in discovered {
3377 8 : let (name, kind) = match discovered {
3378 8 : Discovered::Layer(layer_file_name, local_metadata) => {
3379 8 : discovered_layers.push((layer_file_name, local_metadata));
3380 8 : continue;
3381 : }
3382 0 : Discovered::IgnoredBackup(path) => {
3383 0 : std::fs::remove_file(path)
3384 0 : .or_else(fs_ext::ignore_not_found)
3385 0 : .fatal_err("Removing .old file");
3386 0 : continue;
3387 : }
3388 0 : Discovered::Unknown(file_name) => {
3389 : // we will later error if there are any
3390 0 : unrecognized_files.push(file_name);
3391 0 : continue;
3392 : }
3393 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
3394 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
3395 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
3396 : };
3397 0 : path.push(Utf8Path::new(&name));
3398 0 : init::cleanup(&path, kind)?;
3399 0 : path.pop();
3400 : }
3401 :
3402 3 : if !unrecognized_files.is_empty() {
3403 : // assume that if there are any there are many many.
3404 0 : let n = unrecognized_files.len();
3405 0 : let first = &unrecognized_files[..n.min(10)];
3406 0 : anyhow::bail!(
3407 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
3408 : );
3409 3 : }
3410 :
3411 3 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
3412 :
3413 3 : let mut loaded_layers = Vec::new();
3414 3 : let mut needs_cleanup = Vec::new();
3415 3 : let mut total_physical_size = 0;
3416 :
3417 11 : for (name, decision) in decided {
3418 8 : let decision = match decision {
3419 8 : Ok(decision) => decision,
3420 0 : Err(DismissedLayer::Future { local }) => {
3421 0 : if let Some(local) = local {
3422 0 : init::cleanup_future_layer(
3423 0 : &local.local_path,
3424 0 : &name,
3425 0 : disk_consistent_lsn,
3426 0 : )?;
3427 0 : }
3428 0 : needs_cleanup.push(name);
3429 0 : continue;
3430 : }
3431 0 : Err(DismissedLayer::LocalOnly(local)) => {
3432 0 : init::cleanup_local_only_file(&name, &local)?;
3433 : // this file never existed remotely, we will have to do rework
3434 0 : continue;
3435 : }
3436 0 : Err(DismissedLayer::BadMetadata(local)) => {
3437 0 : init::cleanup_local_file_for_remote(&local)?;
3438 : // this file never existed remotely, we will have to do rework
3439 0 : continue;
3440 : }
3441 : };
3442 :
3443 8 : match &name {
3444 6 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
3445 2 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
3446 : }
3447 :
3448 8 : tracing::debug!(layer=%name, ?decision, "applied");
3449 :
3450 8 : let layer = match decision {
3451 8 : Resident { local, remote } => {
3452 8 : total_physical_size += local.file_size;
3453 8 : Layer::for_resident(conf, &this, local.local_path, name, remote)
3454 8 : .drop_eviction_guard()
3455 : }
3456 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
3457 : };
3458 :
3459 8 : loaded_layers.push(layer);
3460 : }
3461 3 : Ok((loaded_layers, needs_cleanup, total_physical_size))
3462 3 : }
3463 : })
3464 3 : .await
3465 3 : .map_err(anyhow::Error::new)
3466 3 : .and_then(|x| x)?;
3467 :
3468 3 : let num_layers = loaded_layers.len();
3469 :
3470 3 : guard
3471 3 : .open_mut()
3472 3 : .expect("layermanager must be open during init")
3473 3 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3474 :
3475 3 : self.remote_client
3476 3 : .schedule_layer_file_deletion(&needs_cleanup)?;
3477 3 : self.remote_client
3478 3 : .schedule_index_upload_for_file_changes()?;
3479 : // This barrier orders above DELETEs before any later operations.
3480 : // This is critical because code executing after the barrier might
3481 : // create again objects with the same key that we just scheduled for deletion.
3482 : // For example, if we just scheduled deletion of an image layer "from the future",
3483 : // later compaction might run again and re-create the same image layer.
3484 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3485 : // "same" here means same key range and LSN.
3486 : //
3487 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3488 : // the upload queue may execute the PUT first, then the DELETE.
3489 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3490 : //
3491 : // 1. a future image layer is created and uploaded
3492 : // 2. ps restart
3493 : // 3. the future layer from (1) is deleted during load layer map
3494 : // 4. image layer is re-created and uploaded
3495 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3496 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3497 : //
3498 : // See https://github.com/neondatabase/neon/issues/5878
3499 : //
3500 : // NB: generation numbers naturally protect against this because they disambiguate
3501 : // (1) and (4)
3502 : // TODO: this is basically a no-op now, should we remove it?
3503 3 : self.remote_client.schedule_barrier()?;
3504 : // TenantShard::create_timeline will wait for these uploads to happen before returning, or
3505 : // on retry.
3506 :
3507 3 : info!(
3508 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3509 : num_layers, disk_consistent_lsn, total_physical_size
3510 : );
3511 :
3512 3 : timer.stop_and_record();
3513 3 : Ok(())
3514 3 : }
3515 :
3516 : /// Retrieve current logical size of the timeline.
3517 : ///
3518 : /// The size could be lagging behind the actual number, in case
3519 : /// the initial size calculation has not been run (gets triggered on the first size access).
3520 : ///
3521 : /// return size and boolean flag that shows if the size is exact
3522 0 : pub(crate) fn get_current_logical_size(
3523 0 : self: &Arc<Self>,
3524 0 : priority: GetLogicalSizePriority,
3525 0 : ctx: &RequestContext,
3526 0 : ) -> logical_size::CurrentLogicalSize {
3527 0 : if !self.tenant_shard_id.is_shard_zero() {
3528 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3529 : // when HTTP API is serving a GET for timeline zero, return zero
3530 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3531 0 : }
3532 :
3533 0 : let current_size = self.current_logical_size.current_size();
3534 0 : debug!("Current size: {current_size:?}");
3535 :
3536 0 : match (current_size.accuracy(), priority) {
3537 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3538 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3539 0 : // background task will eventually deliver an exact value, we're in no rush
3540 0 : }
3541 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3542 : // background task is not ready, but user is asking for it now;
3543 : // => make the background task skip the line
3544 : // (The alternative would be to calculate the size here, but,
3545 : // it can actually take a long time if the user has a lot of rels.
3546 : // And we'll inevitable need it again; So, let the background task do the work.)
3547 0 : match self
3548 0 : .current_logical_size
3549 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3550 0 : .get()
3551 : {
3552 0 : Some(cancel) => cancel.cancel(),
3553 : None => {
3554 0 : match self.current_state() {
3555 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3556 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3557 0 : // Don't make noise.
3558 0 : }
3559 : TimelineState::Loading => {
3560 : // Import does not return an activated timeline.
3561 0 : info!(
3562 0 : "discarding priority boost for logical size calculation because timeline is not yet active"
3563 : );
3564 : }
3565 : TimelineState::Active => {
3566 : // activation should be setting the once cell
3567 0 : warn!(
3568 0 : "unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work"
3569 : );
3570 0 : debug_assert!(false);
3571 : }
3572 : }
3573 : }
3574 : }
3575 : }
3576 : }
3577 :
3578 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3579 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3580 0 : let first = self
3581 0 : .current_logical_size
3582 0 : .did_return_approximate_to_walreceiver
3583 0 : .compare_exchange(
3584 0 : false,
3585 0 : true,
3586 0 : AtomicOrdering::Relaxed,
3587 0 : AtomicOrdering::Relaxed,
3588 0 : )
3589 0 : .is_ok();
3590 0 : if first {
3591 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3592 0 : }
3593 0 : }
3594 0 : }
3595 :
3596 0 : current_size
3597 0 : }
3598 :
3599 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3600 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3601 : // nothing to do for freshly created timelines;
3602 0 : assert_eq!(
3603 0 : self.current_logical_size.current_size().accuracy(),
3604 : logical_size::Accuracy::Exact,
3605 : );
3606 0 : self.current_logical_size.initialized.add_permits(1);
3607 0 : return;
3608 : };
3609 :
3610 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3611 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3612 0 : self.current_logical_size
3613 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3614 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3615 :
3616 0 : let self_clone = Arc::clone(self);
3617 0 : let background_ctx = ctx.detached_child(
3618 0 : TaskKind::InitialLogicalSizeCalculation,
3619 0 : DownloadBehavior::Download,
3620 : );
3621 0 : task_mgr::spawn(
3622 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3623 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3624 0 : self.tenant_shard_id,
3625 0 : Some(self.timeline_id),
3626 0 : "initial size calculation",
3627 : // NB: don't log errors here, task_mgr will do that.
3628 0 : async move {
3629 0 : self_clone
3630 0 : .initial_logical_size_calculation_task(
3631 0 : initial_part_end,
3632 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3633 0 : background_ctx,
3634 0 : )
3635 0 : .await;
3636 0 : Ok(())
3637 0 : }
3638 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)),
3639 : );
3640 0 : }
3641 :
3642 : /// # Cancellation
3643 : ///
3644 : /// This method is sensitive to `Timeline::cancel`.
3645 : ///
3646 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3647 : ///
3648 : /// # Cancel-Safety
3649 : ///
3650 : /// It does Timeline IO, hence this should be polled to completion because
3651 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3652 : /// to reason about.
3653 0 : async fn initial_logical_size_calculation_task(
3654 0 : self: Arc<Self>,
3655 0 : initial_part_end: Lsn,
3656 0 : skip_concurrency_limiter: CancellationToken,
3657 0 : background_ctx: RequestContext,
3658 0 : ) {
3659 0 : scopeguard::defer! {
3660 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3661 : self.current_logical_size.initialized.add_permits(1);
3662 : }
3663 :
3664 0 : let try_once = |attempt: usize| {
3665 0 : let background_ctx = &background_ctx;
3666 0 : let self_ref = &self;
3667 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3668 0 : async move {
3669 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3670 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3671 0 : background_ctx,
3672 : );
3673 :
3674 : use crate::metrics::initial_logical_size::StartCircumstances;
3675 0 : let (_maybe_permit, circumstances) = tokio::select! {
3676 0 : permit = wait_for_permit => {
3677 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3678 : }
3679 0 : _ = self_ref.cancel.cancelled() => {
3680 0 : return Err(CalculateLogicalSizeError::Cancelled);
3681 : }
3682 0 : () = skip_concurrency_limiter.cancelled() => {
3683 : // Some action that is part of a end user interaction requested logical size
3684 : // => break out of the rate limit
3685 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3686 : // but then again what happens if they cancel; also, we should just be using
3687 : // one runtime across the entire process, so, let's leave this for now.
3688 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3689 : }
3690 : };
3691 :
3692 0 : let metrics_guard = if attempt == 1 {
3693 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3694 : } else {
3695 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3696 : };
3697 :
3698 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3699 0 : self_ref.conf.get_vectored_concurrent_io,
3700 0 : self_ref
3701 0 : .gate
3702 0 : .enter()
3703 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3704 : );
3705 :
3706 0 : let calculated_size = self_ref
3707 0 : .logical_size_calculation_task(
3708 0 : initial_part_end,
3709 0 : LogicalSizeCalculationCause::Initial,
3710 0 : background_ctx,
3711 0 : )
3712 0 : .await?;
3713 :
3714 0 : self_ref
3715 0 : .trigger_aux_file_size_computation(
3716 0 : initial_part_end,
3717 0 : background_ctx,
3718 0 : io_concurrency,
3719 0 : )
3720 0 : .await?;
3721 :
3722 : // TODO: add aux file size to logical size
3723 :
3724 0 : Ok((calculated_size, metrics_guard))
3725 0 : }
3726 0 : };
3727 :
3728 0 : let retrying = async {
3729 0 : let mut attempt = 0;
3730 : loop {
3731 0 : attempt += 1;
3732 :
3733 0 : match try_once(attempt).await {
3734 0 : Ok(res) => return ControlFlow::Continue(res),
3735 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3736 : Err(
3737 0 : e @ (CalculateLogicalSizeError::Decode(_)
3738 : | CalculateLogicalSizeError::PageRead(_)),
3739 : ) => {
3740 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3741 : // exponential back-off doesn't make sense at these long intervals;
3742 : // use fixed retry interval with generous jitter instead
3743 0 : let sleep_duration = Duration::from_secs(
3744 0 : u64::try_from(
3745 : // 1hour base
3746 0 : (60_i64 * 60_i64)
3747 0 : // 10min jitter
3748 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3749 : )
3750 0 : .expect("10min < 1hour"),
3751 : );
3752 0 : tokio::select! {
3753 0 : _ = tokio::time::sleep(sleep_duration) => {}
3754 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3755 : }
3756 : }
3757 : }
3758 : }
3759 0 : };
3760 :
3761 0 : let (calculated_size, metrics_guard) = match retrying.await {
3762 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3763 0 : ControlFlow::Break(()) => return,
3764 : };
3765 :
3766 : // we cannot query current_logical_size.current_size() to know the current
3767 : // *negative* value, only truncated to u64.
3768 0 : let added = self
3769 0 : .current_logical_size
3770 0 : .size_added_after_initial
3771 0 : .load(AtomicOrdering::Relaxed);
3772 :
3773 0 : let sum = calculated_size.saturating_add_signed(added);
3774 :
3775 : // set the gauge value before it can be set in `update_current_logical_size`.
3776 0 : self.metrics.current_logical_size_gauge.set(sum);
3777 :
3778 0 : self.current_logical_size
3779 0 : .initial_logical_size
3780 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3781 0 : .ok()
3782 0 : .expect("only this task sets it");
3783 0 : }
3784 :
3785 7 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3786 7 : self: &Arc<Self>,
3787 7 : lsn: Lsn,
3788 7 : cause: LogicalSizeCalculationCause,
3789 7 : ctx: RequestContext,
3790 7 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3791 7 : let (sender, receiver) = oneshot::channel();
3792 7 : let self_clone = Arc::clone(self);
3793 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3794 : // we should stop the size calculation work and return an error.
3795 : // That would require restructuring this function's API to
3796 : // return the result directly, instead of a Receiver for the result.
3797 7 : let ctx = ctx.detached_child(
3798 7 : TaskKind::OndemandLogicalSizeCalculation,
3799 7 : DownloadBehavior::Download,
3800 : );
3801 7 : task_mgr::spawn(
3802 7 : task_mgr::BACKGROUND_RUNTIME.handle(),
3803 7 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3804 7 : self.tenant_shard_id,
3805 7 : Some(self.timeline_id),
3806 7 : "ondemand logical size calculation",
3807 7 : async move {
3808 7 : let res = self_clone
3809 7 : .logical_size_calculation_task(lsn, cause, &ctx)
3810 7 : .await;
3811 7 : let _ = sender.send(res).ok();
3812 7 : Ok(()) // Receiver is responsible for handling errors
3813 7 : }
3814 7 : .in_current_span(),
3815 : );
3816 7 : receiver
3817 7 : }
3818 :
3819 : #[instrument(skip_all)]
3820 : async fn logical_size_calculation_task(
3821 : self: &Arc<Self>,
3822 : lsn: Lsn,
3823 : cause: LogicalSizeCalculationCause,
3824 : ctx: &RequestContext,
3825 : ) -> Result<u64, CalculateLogicalSizeError> {
3826 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3827 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3828 : // accurate relation sizes, and they do not emit consumption metrics.
3829 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3830 :
3831 : let guard = self
3832 : .gate
3833 : .enter()
3834 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3835 :
3836 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3837 : }
3838 :
3839 : /// Calculate the logical size of the database at the latest LSN.
3840 : ///
3841 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3842 : /// especially if we need to download remote layers.
3843 7 : async fn calculate_logical_size(
3844 7 : &self,
3845 7 : up_to_lsn: Lsn,
3846 7 : cause: LogicalSizeCalculationCause,
3847 7 : _guard: &GateGuard,
3848 7 : ctx: &RequestContext,
3849 7 : ) -> Result<u64, CalculateLogicalSizeError> {
3850 7 : info!(
3851 0 : "Calculating logical size for timeline {} at {}",
3852 : self.timeline_id, up_to_lsn
3853 : );
3854 :
3855 7 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
3856 : {
3857 0 : return Err(CalculateLogicalSizeError::Cancelled);
3858 7 : }
3859 :
3860 : // See if we've already done the work for initial size calculation.
3861 : // This is a short-cut for timelines that are mostly unused.
3862 7 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3863 0 : return Ok(size);
3864 7 : }
3865 7 : let storage_time_metrics = match cause {
3866 : LogicalSizeCalculationCause::Initial
3867 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3868 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3869 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3870 7 : &self.metrics.imitate_logical_size_histo
3871 : }
3872 : };
3873 7 : let timer = storage_time_metrics.start_timer();
3874 7 : let logical_size = self
3875 7 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3876 7 : .await?;
3877 7 : debug!("calculated logical size: {logical_size}");
3878 7 : timer.stop_and_record();
3879 7 : Ok(logical_size)
3880 7 : }
3881 :
3882 : /// Update current logical size, adding `delta' to the old value.
3883 135285 : fn update_current_logical_size(&self, delta: i64) {
3884 135285 : let logical_size = &self.current_logical_size;
3885 135285 : logical_size.increment_size(delta);
3886 :
3887 : // Also set the value in the prometheus gauge. Note that
3888 : // there is a race condition here: if this is is called by two
3889 : // threads concurrently, the prometheus gauge might be set to
3890 : // one value while current_logical_size is set to the
3891 : // other.
3892 135285 : match logical_size.current_size() {
3893 135285 : CurrentLogicalSize::Exact(ref new_current_size) => self
3894 135285 : .metrics
3895 135285 : .current_logical_size_gauge
3896 135285 : .set(new_current_size.into()),
3897 0 : CurrentLogicalSize::Approximate(_) => {
3898 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3899 0 : // forth between the initial size calculation task.
3900 0 : }
3901 : }
3902 135285 : }
3903 :
3904 1527 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: MetricsUpdate) {
3905 : // TODO: this directory metrics is not correct -- we could have multiple reldirs in the system
3906 : // for each of the database, but we only store one value, and therefore each pgdirmodification
3907 : // would overwrite the previous value if they modify different databases.
3908 :
3909 1527 : match count {
3910 566 : MetricsUpdate::Set(count) => {
3911 566 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3912 566 : self.directory_metrics_inited[kind.offset()].store(true, AtomicOrdering::Relaxed);
3913 566 : }
3914 960 : MetricsUpdate::Add(count) => {
3915 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3916 : // it's fine.
3917 960 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3918 960 : // The metrics has been initialized with `MetricsUpdate::Set` before, so we can add/sub
3919 960 : // the value reliably.
3920 960 : self.directory_metrics[kind.offset()].fetch_add(count, AtomicOrdering::Relaxed);
3921 960 : }
3922 : // Otherwise, ignore this update
3923 : }
3924 1 : MetricsUpdate::Sub(count) => {
3925 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3926 : // it's fine.
3927 1 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3928 1 : // The metrics has been initialized with `MetricsUpdate::Set` before.
3929 1 : // The operation could overflow so we need to normalize the value.
3930 1 : let prev_val =
3931 1 : self.directory_metrics[kind.offset()].load(AtomicOrdering::Relaxed);
3932 1 : let res = prev_val.saturating_sub(count);
3933 1 : self.directory_metrics[kind.offset()].store(res, AtomicOrdering::Relaxed);
3934 1 : }
3935 : // Otherwise, ignore this update
3936 : }
3937 : };
3938 :
3939 : // TODO: remove this, there's no place in the code that updates this aux metrics.
3940 1527 : let aux_metric =
3941 1527 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3942 :
3943 1527 : let sum_of_entries = self
3944 1527 : .directory_metrics
3945 1527 : .iter()
3946 12216 : .map(|v| v.load(AtomicOrdering::Relaxed))
3947 1527 : .sum();
3948 : // Set a high general threshold and a lower threshold for the auxiliary files,
3949 : // as we can have large numbers of relations in the db directory.
3950 : const SUM_THRESHOLD: u64 = 5000;
3951 : const AUX_THRESHOLD: u64 = 1000;
3952 1527 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3953 0 : self.metrics
3954 0 : .directory_entries_count_gauge
3955 0 : .set(sum_of_entries);
3956 1527 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3957 0 : metric.set(sum_of_entries);
3958 1527 : }
3959 1527 : }
3960 :
3961 0 : async fn find_layer(
3962 0 : &self,
3963 0 : layer_name: &LayerName,
3964 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3965 0 : let guard = self
3966 0 : .layers
3967 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
3968 0 : .await;
3969 0 : let layer = guard
3970 0 : .layer_map()?
3971 0 : .iter_historic_layers()
3972 0 : .find(|l| &l.layer_name() == layer_name)
3973 0 : .map(|found| guard.get_from_desc(&found));
3974 0 : Ok(layer)
3975 0 : }
3976 :
3977 0 : pub(super) fn should_keep_previous_heatmap(&self, new_heatmap_end_lsn: Lsn) -> bool {
3978 0 : let crnt = self.previous_heatmap.load();
3979 0 : match crnt.as_deref() {
3980 0 : Some(PreviousHeatmap::Active { end_lsn, .. }) => match end_lsn {
3981 0 : Some(crnt_end_lsn) => *crnt_end_lsn > new_heatmap_end_lsn,
3982 0 : None => true,
3983 : },
3984 0 : Some(PreviousHeatmap::Obsolete) => false,
3985 0 : None => false,
3986 : }
3987 0 : }
3988 :
3989 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3990 : /// indicating which layers are currently on-disk on the primary.
3991 : ///
3992 : /// None is returned if the Timeline is in a state where uploading a heatmap
3993 : /// doesn't make sense, such as shutting down or initializing. The caller
3994 : /// should treat this as a cue to simply skip doing any heatmap uploading
3995 : /// for this timeline.
3996 8 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3997 8 : if !self.is_active() {
3998 0 : return None;
3999 8 : }
4000 :
4001 8 : let guard = self
4002 8 : .layers
4003 8 : .read(LayerManagerLockHolder::GenerateHeatmap)
4004 8 : .await;
4005 :
4006 : // Firstly, if there's any heatmap left over from when this location
4007 : // was a secondary, take that into account. Keep layers that are:
4008 : // * present in the layer map
4009 : // * visible
4010 : // * non-resident
4011 : // * not evicted since we read the heatmap
4012 : //
4013 : // Without this, a new cold, attached location would clobber the previous
4014 : // heatamp.
4015 8 : let previous_heatmap = self.previous_heatmap.load();
4016 8 : let visible_non_resident = match previous_heatmap.as_deref() {
4017 : Some(PreviousHeatmap::Active {
4018 6 : heatmap, read_at, ..
4019 23 : }) => Some(heatmap.all_layers().filter_map(|hl| {
4020 23 : let desc: PersistentLayerDesc = hl.name.clone().into();
4021 23 : let layer = guard.try_get_from_key(&desc.key())?;
4022 :
4023 23 : if layer.visibility() == LayerVisibilityHint::Covered {
4024 0 : return None;
4025 23 : }
4026 :
4027 23 : if layer.is_likely_resident() {
4028 10 : return None;
4029 13 : }
4030 :
4031 13 : if layer.last_evicted_at().happened_after(*read_at) {
4032 3 : return None;
4033 10 : }
4034 :
4035 10 : Some((desc, hl.metadata.clone(), hl.access_time, hl.cold))
4036 23 : })),
4037 0 : Some(PreviousHeatmap::Obsolete) => None,
4038 2 : None => None,
4039 : };
4040 :
4041 : // Secondly, all currently visible, resident layers are included.
4042 18 : let resident = guard.likely_resident_layers().filter_map(|layer| {
4043 18 : match layer.visibility() {
4044 : LayerVisibilityHint::Visible => {
4045 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
4046 17 : let last_activity_ts = layer.latest_activity();
4047 17 : Some((
4048 17 : layer.layer_desc().clone(),
4049 17 : layer.metadata(),
4050 17 : last_activity_ts,
4051 17 : false, // these layers are not cold
4052 17 : ))
4053 : }
4054 : LayerVisibilityHint::Covered => {
4055 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
4056 1 : None
4057 : }
4058 : }
4059 18 : });
4060 :
4061 8 : let mut layers = match visible_non_resident {
4062 6 : Some(non_resident) => {
4063 6 : let mut non_resident = non_resident.peekable();
4064 6 : if non_resident.peek().is_none() {
4065 2 : tracing::info!(timeline_id=%self.timeline_id, "Previous heatmap now obsolete");
4066 2 : self.previous_heatmap
4067 2 : .store(Some(PreviousHeatmap::Obsolete.into()));
4068 4 : }
4069 :
4070 6 : non_resident.chain(resident).collect::<Vec<_>>()
4071 : }
4072 2 : None => resident.collect::<Vec<_>>(),
4073 : };
4074 :
4075 : // Sort layers in order of which to download first. For a large set of layers to download, we
4076 : // want to prioritize those layers which are most likely to still be in the resident many minutes
4077 : // or hours later:
4078 : // - Cold layers go last for convenience when a human inspects the heatmap.
4079 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
4080 : // only exist for a few minutes before being compacted into L1s.
4081 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
4082 : // the layer is likely to be covered by an image layer during compaction.
4083 62 : layers.sort_by_key(|(desc, _meta, _atime, cold)| {
4084 62 : std::cmp::Reverse((
4085 62 : *cold,
4086 62 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
4087 62 : desc.lsn_range.end,
4088 62 : ))
4089 62 : });
4090 :
4091 8 : let layers = layers
4092 8 : .into_iter()
4093 27 : .map(|(desc, meta, atime, cold)| {
4094 27 : HeatMapLayer::new(desc.layer_name(), meta, atime, cold)
4095 27 : })
4096 8 : .collect();
4097 :
4098 8 : Some(HeatMapTimeline::new(self.timeline_id, layers))
4099 8 : }
4100 :
4101 0 : pub(super) async fn generate_unarchival_heatmap(&self, end_lsn: Lsn) -> PreviousHeatmap {
4102 0 : let guard = self
4103 0 : .layers
4104 0 : .read(LayerManagerLockHolder::GenerateHeatmap)
4105 0 : .await;
4106 :
4107 0 : let now = SystemTime::now();
4108 0 : let mut heatmap_layers = Vec::default();
4109 0 : for vl in guard.visible_layers() {
4110 0 : if vl.layer_desc().get_lsn_range().start >= end_lsn {
4111 0 : continue;
4112 0 : }
4113 :
4114 0 : let hl = HeatMapLayer {
4115 0 : name: vl.layer_desc().layer_name(),
4116 0 : metadata: vl.metadata(),
4117 0 : access_time: now,
4118 0 : cold: true,
4119 0 : };
4120 0 : heatmap_layers.push(hl);
4121 : }
4122 :
4123 0 : tracing::info!(
4124 0 : "Generating unarchival heatmap with {} layers",
4125 0 : heatmap_layers.len()
4126 : );
4127 :
4128 0 : let heatmap = HeatMapTimeline::new(self.timeline_id, heatmap_layers);
4129 0 : PreviousHeatmap::Active {
4130 0 : heatmap,
4131 0 : read_at: Instant::now(),
4132 0 : end_lsn: Some(end_lsn),
4133 0 : }
4134 0 : }
4135 :
4136 : /// Returns true if the given lsn is or was an ancestor branchpoint.
4137 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
4138 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
4139 : // branchpoint in the value in IndexPart::lineage
4140 0 : self.ancestor_lsn == lsn
4141 0 : || (self.ancestor_lsn == Lsn::INVALID
4142 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
4143 0 : }
4144 : }
4145 :
4146 : #[derive(Clone)]
4147 : /// Type representing a query in the ([`Lsn`], [`Key`]) space.
4148 : /// In other words, a set of segments in a 2D space.
4149 : ///
4150 : /// This representation has the advatange of avoiding hash map
4151 : /// allocations for uniform queries.
4152 : pub(crate) enum VersionedKeySpaceQuery {
4153 : /// Variant for queries at a single [`Lsn`]
4154 : Uniform { keyspace: KeySpace, lsn: Lsn },
4155 : /// Variant for queries at multiple [`Lsn`]s
4156 : Scattered {
4157 : keyspaces_at_lsn: Vec<(Lsn, KeySpace)>,
4158 : },
4159 : }
4160 :
4161 : impl VersionedKeySpaceQuery {
4162 302264 : pub(crate) fn uniform(keyspace: KeySpace, lsn: Lsn) -> Self {
4163 302264 : Self::Uniform { keyspace, lsn }
4164 302264 : }
4165 :
4166 10192 : pub(crate) fn scattered(keyspaces_at_lsn: Vec<(Lsn, KeySpace)>) -> Self {
4167 10192 : Self::Scattered { keyspaces_at_lsn }
4168 10192 : }
4169 :
4170 : /// Returns the most recent (largest) LSN included in the query.
4171 : /// If any of the LSNs included in the query are invalid, returns
4172 : /// an error instead.
4173 624912 : fn high_watermark_lsn(&self) -> Result<Lsn, GetVectoredError> {
4174 624912 : match self {
4175 604528 : Self::Uniform { lsn, .. } => {
4176 604528 : if !lsn.is_valid() {
4177 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4178 604528 : }
4179 :
4180 604528 : Ok(*lsn)
4181 : }
4182 20384 : Self::Scattered { keyspaces_at_lsn } => {
4183 20384 : let mut max_lsn = None;
4184 42218 : for (lsn, _keyspace) in keyspaces_at_lsn.iter() {
4185 42218 : if !lsn.is_valid() {
4186 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4187 42218 : }
4188 42218 : max_lsn = std::cmp::max(max_lsn, Some(lsn));
4189 : }
4190 :
4191 20384 : if let Some(computed) = max_lsn {
4192 20384 : Ok(*computed)
4193 : } else {
4194 0 : Err(GetVectoredError::Other(anyhow!("empty input")))
4195 : }
4196 : }
4197 : }
4198 624912 : }
4199 :
4200 : /// Returns the total keyspace being queried: the result of projecting
4201 : /// everything in the key dimensions onto the key axis.
4202 323533 : fn total_keyspace(&self) -> KeySpace {
4203 323533 : match self {
4204 303149 : Self::Uniform { keyspace, .. } => keyspace.clone(),
4205 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4206 20384 : .iter()
4207 20384 : .map(|(_lsn, keyspace)| keyspace)
4208 42218 : .fold(KeySpace::default(), |mut acc, v| {
4209 42218 : acc.merge(v);
4210 42218 : acc
4211 42218 : }),
4212 : }
4213 323533 : }
4214 :
4215 : /// Returns LSN for a specific key.
4216 : ///
4217 : /// Invariant: requested key must be part of [`Self::total_keyspace`]
4218 395577 : pub(super) fn map_key_to_lsn(&self, key: &Key) -> Lsn {
4219 395577 : match self {
4220 322385 : Self::Uniform { lsn, .. } => *lsn,
4221 73192 : Self::Scattered { keyspaces_at_lsn } => {
4222 73192 : keyspaces_at_lsn
4223 73192 : .iter()
4224 420374 : .find(|(_lsn, keyspace)| keyspace.contains(key))
4225 73192 : .expect("Returned key was requested")
4226 : .0
4227 : }
4228 : }
4229 395577 : }
4230 :
4231 : /// Remove any parts of the query (segments) which overlap with the provided
4232 : /// key space (also segments).
4233 968821 : fn remove_overlapping_with(&mut self, to_remove: &KeySpace) -> KeySpace {
4234 968821 : match self {
4235 948437 : Self::Uniform { keyspace, .. } => keyspace.remove_overlapping_with(to_remove),
4236 20384 : Self::Scattered { keyspaces_at_lsn } => {
4237 20384 : let mut removed_accum = KeySpaceRandomAccum::new();
4238 42218 : keyspaces_at_lsn.iter_mut().for_each(|(_lsn, keyspace)| {
4239 42218 : let removed = keyspace.remove_overlapping_with(to_remove);
4240 42218 : removed_accum.add_keyspace(removed);
4241 42218 : });
4242 :
4243 20384 : removed_accum.to_keyspace()
4244 : }
4245 : }
4246 968821 : }
4247 :
4248 739548 : fn is_empty(&self) -> bool {
4249 739548 : match self {
4250 719164 : Self::Uniform { keyspace, .. } => keyspace.is_empty(),
4251 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4252 20384 : .iter()
4253 31301 : .all(|(_lsn, keyspace)| keyspace.is_empty()),
4254 : }
4255 739548 : }
4256 :
4257 : /// "Lower" the query on the LSN dimension
4258 114637 : fn lower(&mut self, to: Lsn) {
4259 114637 : match self {
4260 114637 : Self::Uniform { lsn, .. } => {
4261 114637 : // If the originally requested LSN is smaller than the starting
4262 114637 : // LSN of the ancestor we are descending into, we need to respect that.
4263 114637 : // Hence the min.
4264 114637 : *lsn = std::cmp::min(*lsn, to);
4265 114637 : }
4266 0 : Self::Scattered { keyspaces_at_lsn } => {
4267 0 : keyspaces_at_lsn.iter_mut().for_each(|(lsn, _keyspace)| {
4268 0 : *lsn = std::cmp::min(*lsn, to);
4269 0 : });
4270 : }
4271 : }
4272 114637 : }
4273 : }
4274 :
4275 : impl std::fmt::Display for VersionedKeySpaceQuery {
4276 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
4277 0 : write!(f, "[")?;
4278 :
4279 0 : match self {
4280 0 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4281 0 : write!(f, "{keyspace} @ {lsn}")?;
4282 : }
4283 0 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4284 0 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4285 0 : write!(f, "{keyspace} @ {lsn},")?;
4286 : }
4287 : }
4288 : }
4289 :
4290 0 : write!(f, "]")
4291 0 : }
4292 : }
4293 :
4294 : impl Timeline {
4295 : #[allow(clippy::doc_lazy_continuation)]
4296 : /// Get the data needed to reconstruct all keys in the provided keyspace
4297 : ///
4298 : /// The algorithm is as follows:
4299 : /// 1. While some keys are still not done and there's a timeline to visit:
4300 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
4301 : /// 2.1: Build the fringe for the current keyspace
4302 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
4303 : /// intersects
4304 : /// 2.3. Pop the timeline from the fringe
4305 : /// 2.4. If the fringe is empty, go back to 1
4306 312456 : async fn get_vectored_reconstruct_data(
4307 312456 : &self,
4308 312456 : mut query: VersionedKeySpaceQuery,
4309 312456 : reconstruct_state: &mut ValuesReconstructState,
4310 312456 : ctx: &RequestContext,
4311 312456 : ) -> Result<(), GetVectoredError> {
4312 312456 : let original_hwm_lsn = query.high_watermark_lsn().unwrap();
4313 :
4314 : let mut timeline_owned: Arc<Timeline>;
4315 312456 : let mut timeline = self;
4316 :
4317 312455 : let missing_keyspace = loop {
4318 427092 : if self.cancel.is_cancelled() {
4319 0 : return Err(GetVectoredError::Cancelled);
4320 427092 : }
4321 :
4322 : let TimelineVisitOutcome {
4323 427092 : completed_keyspace: completed,
4324 427092 : image_covered_keyspace,
4325 : } = {
4326 427092 : let ctx = RequestContextBuilder::from(ctx)
4327 427092 : .perf_span(|crnt_perf_span| {
4328 0 : info_span!(
4329 : target: PERF_TRACE_TARGET,
4330 0 : parent: crnt_perf_span,
4331 : "PLAN_IO_TIMELINE",
4332 : timeline = %timeline.timeline_id,
4333 0 : high_watermark_lsn = %query.high_watermark_lsn().unwrap(),
4334 : )
4335 0 : })
4336 427092 : .attached_child();
4337 :
4338 427092 : Self::get_vectored_reconstruct_data_timeline(
4339 427092 : timeline,
4340 427092 : &query,
4341 427092 : reconstruct_state,
4342 427092 : &self.cancel,
4343 427092 : &ctx,
4344 : )
4345 427092 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4346 427092 : .await?
4347 : };
4348 :
4349 427092 : query.remove_overlapping_with(&completed);
4350 :
4351 : // Do not descend into the ancestor timeline for aux files.
4352 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
4353 : // stalling compaction.
4354 427092 : query.remove_overlapping_with(&KeySpace {
4355 427092 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
4356 427092 : });
4357 :
4358 : // Keyspace is fully retrieved
4359 427092 : if query.is_empty() {
4360 312272 : break None;
4361 114820 : }
4362 :
4363 114820 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
4364 : // Not fully retrieved but no ancestor timeline.
4365 183 : break Some(query.total_keyspace());
4366 : };
4367 :
4368 : // Now we see if there are keys covered by the image layer but does not exist in the
4369 : // image layer, which means that the key does not exist.
4370 :
4371 : // The block below will stop the vectored search if any of the keys encountered an image layer
4372 : // which did not contain a snapshot for said key. Since we have already removed all completed
4373 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
4374 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
4375 : // and stop the search as a result of that.
4376 114637 : let mut removed = query.remove_overlapping_with(&image_covered_keyspace);
4377 : // Do not fire missing key error and end early for sparse keys. Note that we hava already removed
4378 : // non-inherited keyspaces before, so we can safely do a full `SPARSE_RANGE` remove instead of
4379 : // figuring out what is the inherited key range and do a fine-grained pruning.
4380 114637 : removed.remove_overlapping_with(&KeySpace {
4381 114637 : ranges: vec![SPARSE_RANGE],
4382 114637 : });
4383 114637 : if !removed.is_empty() {
4384 0 : break Some(removed);
4385 114637 : }
4386 :
4387 : // Each key range in the original query is at some point in the LSN space.
4388 : // When descending into the ancestor, lower all ranges in the LSN space
4389 : // such that new changes on the parent timeline are not visible.
4390 114637 : query.lower(timeline.ancestor_lsn);
4391 :
4392 114637 : let ctx = RequestContextBuilder::from(ctx)
4393 114637 : .perf_span(|crnt_perf_span| {
4394 0 : info_span!(
4395 : target: PERF_TRACE_TARGET,
4396 0 : parent: crnt_perf_span,
4397 : "GET_ANCESTOR",
4398 : timeline = %timeline.timeline_id,
4399 0 : ancestor = %ancestor_timeline.timeline_id,
4400 : ancestor_lsn = %timeline.ancestor_lsn
4401 : )
4402 0 : })
4403 114637 : .attached_child();
4404 :
4405 114637 : timeline_owned = timeline
4406 114637 : .get_ready_ancestor_timeline(ancestor_timeline, &ctx)
4407 114637 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4408 114637 : .await?;
4409 114636 : timeline = &*timeline_owned;
4410 : };
4411 :
4412 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
4413 312455 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
4414 183 : let mut missing_keyspace = missing_keyspace;
4415 183 : missing_keyspace.remove_overlapping_with(&KeySpace {
4416 183 : ranges: vec![SPARSE_RANGE],
4417 183 : });
4418 183 : if missing_keyspace.is_empty() {
4419 176 : None
4420 : } else {
4421 7 : Some(missing_keyspace)
4422 : }
4423 : } else {
4424 312272 : None
4425 : };
4426 :
4427 312455 : if let Some(missing_keyspace) = missing_keyspace {
4428 7 : return Err(GetVectoredError::MissingKey(Box::new(MissingKeyError {
4429 7 : keyspace: missing_keyspace, /* better if we can store the full keyspace */
4430 7 : shard: self.shard_identity.number,
4431 7 : original_hwm_lsn,
4432 7 : ancestor_lsn: Some(timeline.ancestor_lsn),
4433 7 : backtrace: None,
4434 7 : read_path: std::mem::take(&mut reconstruct_state.read_path),
4435 7 : query: None,
4436 7 : })));
4437 312448 : }
4438 :
4439 312448 : Ok(())
4440 312456 : }
4441 :
4442 427092 : async fn get_vectored_init_fringe(
4443 427092 : &self,
4444 427092 : query: &VersionedKeySpaceQuery,
4445 427092 : ) -> Result<LayerFringe, GetVectoredError> {
4446 427092 : let mut fringe = LayerFringe::new();
4447 427092 : let guard = self.layers.read(LayerManagerLockHolder::GetPage).await;
4448 :
4449 427092 : match query {
4450 416900 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4451 : // LSNs requested by the compute or determined by the pageserver
4452 : // are inclusive. Queries to the layer map use exclusive LSNs.
4453 : // Hence, bump the value before the query - same in the other
4454 : // match arm.
4455 416900 : let cont_lsn = Lsn(lsn.0 + 1);
4456 416900 : guard.update_search_fringe(keyspace, cont_lsn, &mut fringe)?;
4457 : }
4458 10192 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4459 21109 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4460 21109 : let cont_lsn_for_keyspace = Lsn(lsn.0 + 1);
4461 21109 : guard.update_search_fringe(keyspace, cont_lsn_for_keyspace, &mut fringe)?;
4462 : }
4463 : }
4464 : }
4465 :
4466 427092 : Ok(fringe)
4467 427092 : }
4468 :
4469 : /// Collect the reconstruct data for a keyspace from the specified timeline.
4470 : ///
4471 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
4472 : /// the current keyspace. The current keyspace of the search at any given timeline
4473 : /// is the original keyspace minus all the keys that have been completed minus
4474 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
4475 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
4476 : ///
4477 : /// This is basically a depth-first search visitor implementation where a vertex
4478 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
4479 : ///
4480 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
4481 : /// and get all the required reconstruct data from the layer in one go.
4482 : ///
4483 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
4484 : /// decides how to deal with these two keyspaces.
4485 427092 : async fn get_vectored_reconstruct_data_timeline(
4486 427092 : timeline: &Timeline,
4487 427092 : query: &VersionedKeySpaceQuery,
4488 427092 : reconstruct_state: &mut ValuesReconstructState,
4489 427092 : cancel: &CancellationToken,
4490 427092 : ctx: &RequestContext,
4491 427092 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
4492 : // Prevent GC from progressing while visiting the current timeline.
4493 : // If we are GC-ing because a new image layer was added while traversing
4494 : // the timeline, then it will remove layers that are required for fulfilling
4495 : // the current get request (read-path cannot "look back" and notice the new
4496 : // image layer).
4497 427092 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
4498 :
4499 : // See `compaction::compact_with_gc` for why we need this.
4500 427092 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
4501 :
4502 : // Initialize the fringe
4503 427092 : let mut fringe = timeline.get_vectored_init_fringe(query).await?;
4504 :
4505 427092 : let mut completed_keyspace = KeySpace::default();
4506 427092 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
4507 :
4508 873541 : while let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
4509 446449 : if cancel.is_cancelled() {
4510 0 : return Err(GetVectoredError::Cancelled);
4511 446449 : }
4512 :
4513 446449 : if let Some(ref mut read_path) = reconstruct_state.read_path {
4514 446449 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
4515 446449 : }
4516 :
4517 : // Visit the layer and plan IOs for it
4518 446449 : let next_cont_lsn = lsn_range.start;
4519 446449 : layer_to_read
4520 446449 : .get_values_reconstruct_data(
4521 446449 : keyspace_to_read.clone(),
4522 446449 : lsn_range,
4523 446449 : reconstruct_state,
4524 446449 : ctx,
4525 446449 : )
4526 446449 : .await?;
4527 :
4528 446449 : let mut unmapped_keyspace = keyspace_to_read;
4529 446449 : let cont_lsn = next_cont_lsn;
4530 :
4531 446449 : reconstruct_state.on_layer_visited(&layer_to_read);
4532 :
4533 446449 : let (keys_done_last_step, keys_with_image_coverage) =
4534 446449 : reconstruct_state.consume_done_keys();
4535 446449 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
4536 446449 : completed_keyspace.merge(&keys_done_last_step);
4537 446449 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
4538 14120 : unmapped_keyspace
4539 14120 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
4540 14120 : image_covered_keyspace.add_range(keys_with_image_coverage);
4541 432329 : }
4542 :
4543 : // Query the layer map for the next layers to read.
4544 : //
4545 : // Do not descent any further if the last layer we visited
4546 : // completed all keys in the keyspace it inspected. This is not
4547 : // required for correctness, but avoids visiting extra layers
4548 : // which turns out to be a perf bottleneck in some cases.
4549 446449 : if !unmapped_keyspace.is_empty() {
4550 129725 : let guard = timeline.layers.read(LayerManagerLockHolder::GetPage).await;
4551 129725 : guard.update_search_fringe(&unmapped_keyspace, cont_lsn, &mut fringe)?;
4552 :
4553 : // It's safe to drop the layer map lock after planning the next round of reads.
4554 : // The fringe keeps readable handles for the layers which are safe to read even
4555 : // if layers were compacted or flushed.
4556 : //
4557 : // The more interesting consideration is: "Why is the read algorithm still correct
4558 : // if the layer map changes while it is operating?". Doing a vectored read on a
4559 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
4560 : // covered by the read. The layer map tells us how to move the lsn downwards for a
4561 : // range at *a particular point in time*. It is fine for the answer to be different
4562 : // at two different time points.
4563 129725 : drop(guard);
4564 316724 : }
4565 : }
4566 :
4567 427092 : Ok(TimelineVisitOutcome {
4568 427092 : completed_keyspace,
4569 427092 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
4570 427092 : })
4571 427092 : }
4572 :
4573 114637 : async fn get_ready_ancestor_timeline(
4574 114637 : &self,
4575 114637 : ancestor: &Arc<Timeline>,
4576 114637 : ctx: &RequestContext,
4577 114637 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
4578 : // It's possible that the ancestor timeline isn't active yet, or
4579 : // is active but hasn't yet caught up to the branch point. Wait
4580 : // for it.
4581 : //
4582 : // This cannot happen while the pageserver is running normally,
4583 : // because you cannot create a branch from a point that isn't
4584 : // present in the pageserver yet. However, we don't wait for the
4585 : // branch point to be uploaded to cloud storage before creating
4586 : // a branch. I.e., the branch LSN need not be remote consistent
4587 : // for the branching operation to succeed.
4588 : //
4589 : // Hence, if we try to load a tenant in such a state where
4590 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
4591 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
4592 : // then we will need to wait for the ancestor timeline to
4593 : // re-stream WAL up to branch_lsn before we access it.
4594 : //
4595 : // How can a tenant get in such a state?
4596 : // - ungraceful pageserver process exit
4597 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
4598 : //
4599 : // NB: this could be avoided by requiring
4600 : // branch_lsn >= remote_consistent_lsn
4601 : // during branch creation.
4602 114637 : match ancestor.wait_to_become_active(ctx).await {
4603 114636 : Ok(()) => {}
4604 : Err(TimelineState::Stopping) => {
4605 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
4606 0 : return Err(GetReadyAncestorError::Cancelled);
4607 : }
4608 1 : Err(state) => {
4609 1 : return Err(GetReadyAncestorError::BadState {
4610 1 : timeline_id: ancestor.timeline_id,
4611 1 : state,
4612 1 : });
4613 : }
4614 : }
4615 114636 : ancestor
4616 114636 : .wait_lsn(
4617 114636 : self.ancestor_lsn,
4618 114636 : WaitLsnWaiter::Timeline(self),
4619 114636 : WaitLsnTimeout::Default,
4620 114636 : ctx,
4621 114636 : )
4622 114636 : .await
4623 114636 : .map_err(|e| match e {
4624 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
4625 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
4626 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
4627 0 : timeline_id: ancestor.timeline_id,
4628 0 : state,
4629 0 : },
4630 0 : })?;
4631 :
4632 114636 : Ok(ancestor.clone())
4633 114637 : }
4634 :
4635 148592 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
4636 148592 : &self.shard_identity
4637 148592 : }
4638 :
4639 : #[inline(always)]
4640 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
4641 0 : ShardTimelineId {
4642 0 : shard_index: ShardIndex {
4643 0 : shard_number: self.shard_identity.number,
4644 0 : shard_count: self.shard_identity.count,
4645 0 : },
4646 0 : timeline_id: self.timeline_id,
4647 0 : }
4648 0 : }
4649 :
4650 : /// Returns a non-frozen open in-memory layer for ingestion.
4651 : ///
4652 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
4653 : /// this function without holding the mutex.
4654 660 : async fn get_layer_for_write(
4655 660 : &self,
4656 660 : lsn: Lsn,
4657 660 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4658 660 : ctx: &RequestContext,
4659 660 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
4660 660 : let mut guard = self
4661 660 : .layers
4662 660 : .write(LayerManagerLockHolder::GetLayerForWrite)
4663 660 : .await;
4664 :
4665 660 : let last_record_lsn = self.get_last_record_lsn();
4666 660 : ensure!(
4667 660 : lsn > last_record_lsn,
4668 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
4669 : lsn,
4670 : last_record_lsn,
4671 : );
4672 :
4673 660 : let layer = guard
4674 660 : .open_mut()?
4675 660 : .get_layer_for_write(
4676 660 : lsn,
4677 660 : self.conf,
4678 660 : self.timeline_id,
4679 660 : self.tenant_shard_id,
4680 660 : &self.gate,
4681 660 : &self.cancel,
4682 660 : ctx,
4683 : )
4684 660 : .await?;
4685 660 : Ok(layer)
4686 660 : }
4687 :
4688 2639560 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4689 2639560 : assert!(new_lsn.is_aligned());
4690 :
4691 2639560 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
4692 2639560 : self.last_record_lsn.advance(new_lsn);
4693 2639560 : }
4694 :
4695 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
4696 : ///
4697 : /// Unconditional flush loop notification is given because in sharded cases we will want to
4698 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
4699 611 : async fn freeze_inmem_layer_at(
4700 611 : &self,
4701 611 : at: Lsn,
4702 611 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4703 611 : ) -> Result<u64, FlushLayerError> {
4704 611 : let frozen = {
4705 611 : let mut guard = self
4706 611 : .layers
4707 611 : .write(LayerManagerLockHolder::TryFreezeLayer)
4708 611 : .await;
4709 611 : guard
4710 611 : .open_mut()?
4711 611 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
4712 611 : .await
4713 : };
4714 :
4715 611 : if frozen {
4716 597 : let now = Instant::now();
4717 597 : *(self.last_freeze_ts.write().unwrap()) = now;
4718 597 : }
4719 :
4720 : // Increment the flush cycle counter and wake up the flush task.
4721 : // Remember the new value, so that when we listen for the flush
4722 : // to finish, we know when the flush that we initiated has
4723 : // finished, instead of some other flush that was started earlier.
4724 611 : let mut my_flush_request = 0;
4725 :
4726 611 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
4727 611 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
4728 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
4729 611 : }
4730 :
4731 611 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
4732 611 : my_flush_request = *counter + 1;
4733 611 : *counter = my_flush_request;
4734 611 : *lsn = std::cmp::max(at, *lsn);
4735 611 : });
4736 :
4737 611 : assert_ne!(my_flush_request, 0);
4738 :
4739 611 : Ok(my_flush_request)
4740 611 : }
4741 :
4742 : /// Layer flusher task's main loop.
4743 232 : async fn flush_loop(
4744 232 : self: &Arc<Self>,
4745 232 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4746 232 : ctx: &RequestContext,
4747 232 : ) {
4748 : // Always notify waiters about the flush loop exiting since the loop might stop
4749 : // when the timeline hasn't been cancelled.
4750 232 : let scopeguard_rx = layer_flush_start_rx.clone();
4751 232 : scopeguard::defer! {
4752 : let (flush_counter, _) = *scopeguard_rx.borrow();
4753 : let _ = self
4754 : .layer_flush_done_tx
4755 : .send_replace((flush_counter, Err(FlushLayerError::Cancelled)));
4756 : }
4757 :
4758 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4759 232 : let mut watch_l0 = match self
4760 232 : .layers
4761 232 : .read(LayerManagerLockHolder::FlushLoop)
4762 232 : .await
4763 232 : .layer_map()
4764 : {
4765 232 : Ok(lm) => lm.watch_level0_deltas(),
4766 0 : Err(Shutdown) => return,
4767 : };
4768 :
4769 232 : info!("started flush loop");
4770 : loop {
4771 833 : tokio::select! {
4772 833 : _ = self.cancel.cancelled() => {
4773 5 : info!("shutting down layer flush task due to Timeline::cancel");
4774 5 : break;
4775 : },
4776 833 : _ = layer_flush_start_rx.changed() => {}
4777 : }
4778 601 : trace!("waking up");
4779 601 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4780 :
4781 : // The highest LSN to which we flushed in the loop over frozen layers
4782 601 : let mut flushed_to_lsn = Lsn(0);
4783 :
4784 601 : let result = loop {
4785 1197 : if self.cancel.is_cancelled() {
4786 0 : info!("dropping out of flush loop for timeline shutdown");
4787 0 : return;
4788 1197 : }
4789 :
4790 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4791 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4792 1197 : if flushed_to_lsn >= frozen_to_lsn {
4793 586 : break Ok(());
4794 611 : }
4795 :
4796 : // Fetch the next layer to flush, if any.
4797 611 : let (layer, l0_count, frozen_count, frozen_size, open_layer_size) = {
4798 611 : let layers = self.layers.read(LayerManagerLockHolder::FlushLoop).await;
4799 611 : let Ok(lm) = layers.layer_map() else {
4800 0 : info!("dropping out of flush loop for timeline shutdown");
4801 0 : return;
4802 : };
4803 611 : let l0_count = lm.level0_deltas().len();
4804 611 : let frozen_count = lm.frozen_layers.len();
4805 611 : let frozen_size: u64 = lm
4806 611 : .frozen_layers
4807 611 : .iter()
4808 611 : .map(|l| l.estimated_in_mem_size())
4809 611 : .sum();
4810 611 : let open_layer_size: u64 = lm
4811 611 : .open_layer
4812 611 : .as_ref()
4813 611 : .map(|l| l.estimated_in_mem_size())
4814 611 : .unwrap_or(0);
4815 611 : let layer = lm.frozen_layers.front().cloned();
4816 611 : (layer, l0_count, frozen_count, frozen_size, open_layer_size)
4817 : // drop 'layers' lock
4818 : };
4819 611 : let Some(layer) = layer else {
4820 14 : break Ok(());
4821 : };
4822 :
4823 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
4824 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
4825 597 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
4826 0 : if l0_count >= stall_threshold {
4827 0 : warn!(
4828 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
4829 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
4830 : );
4831 0 : let stall_timer = self
4832 0 : .metrics
4833 0 : .flush_delay_histo
4834 0 : .start_timer()
4835 0 : .record_on_drop();
4836 0 : tokio::select! {
4837 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
4838 0 : if let Ok(l0) = result.as_deref() {
4839 0 : let delay = stall_timer.elapsed().as_secs_f64();
4840 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
4841 0 : }
4842 : },
4843 0 : _ = self.cancel.cancelled() => {},
4844 : }
4845 0 : continue; // check again
4846 0 : }
4847 597 : }
4848 :
4849 : // Flush the layer.
4850 597 : let flush_timer = self.metrics.flush_time_histo.start_timer();
4851 597 : match self.flush_frozen_layer(layer, ctx).await {
4852 596 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
4853 : Err(FlushLayerError::Cancelled) => {
4854 0 : info!("dropping out of flush loop for timeline shutdown");
4855 0 : return;
4856 : }
4857 1 : err @ Err(
4858 : FlushLayerError::NotRunning(_)
4859 : | FlushLayerError::Other(_)
4860 : | FlushLayerError::CreateImageLayersError(_),
4861 : ) => {
4862 1 : error!("could not flush frozen layer: {err:?}");
4863 1 : break err.map(|_| ());
4864 : }
4865 : }
4866 596 : let flush_duration = flush_timer.stop_and_record();
4867 :
4868 : // Notify the tenant compaction loop if L0 compaction is needed.
4869 596 : let l0_count = *watch_l0.borrow();
4870 596 : if l0_count >= self.get_compaction_threshold() {
4871 238 : self.l0_compaction_trigger.notify_one();
4872 358 : }
4873 :
4874 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
4875 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
4876 : // ingestion by having ephemeral layer rolls wait for flushes.
4877 596 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
4878 4 : if l0_count >= delay_threshold {
4879 0 : let delay = flush_duration.as_secs_f64();
4880 0 : info!(
4881 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
4882 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
4883 : );
4884 0 : let _delay_timer = self
4885 0 : .metrics
4886 0 : .flush_delay_histo
4887 0 : .start_timer()
4888 0 : .record_on_drop();
4889 0 : tokio::select! {
4890 0 : _ = tokio::time::sleep(flush_duration) => {},
4891 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
4892 0 : _ = self.cancel.cancelled() => {},
4893 : }
4894 4 : }
4895 592 : }
4896 : };
4897 :
4898 : // Unsharded tenants should never advance their LSN beyond the end of the
4899 : // highest layer they write: such gaps between layer data and the frozen LSN
4900 : // are only legal on sharded tenants.
4901 601 : debug_assert!(
4902 601 : self.shard_identity.count.count() > 1
4903 594 : || flushed_to_lsn >= frozen_to_lsn
4904 14 : || !flushed_to_lsn.is_valid()
4905 : );
4906 :
4907 601 : if flushed_to_lsn < frozen_to_lsn
4908 15 : && self.shard_identity.count.count() > 1
4909 1 : && result.is_ok()
4910 : {
4911 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
4912 : // to us via layer_flush_start_rx, then advance it here.
4913 : //
4914 : // This path is only taken for tenants with multiple shards: single sharded tenants should
4915 : // never encounter a gap in the wal.
4916 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
4917 0 : tracing::debug!(
4918 0 : "Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}"
4919 : );
4920 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
4921 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
4922 0 : tracing::warn!(
4923 0 : "Failed to schedule metadata upload after updating disk_consistent_lsn: {e}"
4924 : );
4925 0 : }
4926 0 : }
4927 601 : }
4928 :
4929 : // Notify any listeners that we're done
4930 601 : let _ = self
4931 601 : .layer_flush_done_tx
4932 601 : .send_replace((flush_counter, result));
4933 : }
4934 5 : }
4935 :
4936 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
4937 571 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
4938 571 : let mut rx = self.layer_flush_done_tx.subscribe();
4939 : loop {
4940 : {
4941 1161 : let (last_result_counter, last_result) = &*rx.borrow();
4942 1161 : if *last_result_counter >= request {
4943 571 : if let Err(err) = last_result {
4944 : // We already logged the original error in
4945 : // flush_loop. We cannot propagate it to the caller
4946 : // here, because it might not be Cloneable
4947 1 : return Err(err.clone());
4948 : } else {
4949 570 : return Ok(());
4950 : }
4951 590 : }
4952 : }
4953 590 : trace!("waiting for flush to complete");
4954 590 : tokio::select! {
4955 590 : rx_e = rx.changed() => {
4956 590 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
4957 : },
4958 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
4959 : // the notification from [`flush_loop`] that it completed.
4960 590 : _ = self.cancel.cancelled() => {
4961 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
4962 0 : return Ok(())
4963 : }
4964 : };
4965 590 : trace!("done")
4966 : }
4967 571 : }
4968 :
4969 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
4970 : ///
4971 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
4972 : #[instrument(skip_all, fields(layer=%frozen_layer))]
4973 : async fn flush_frozen_layer(
4974 : self: &Arc<Self>,
4975 : frozen_layer: Arc<InMemoryLayer>,
4976 : ctx: &RequestContext,
4977 : ) -> Result<Lsn, FlushLayerError> {
4978 : debug_assert_current_span_has_tenant_and_timeline_id();
4979 :
4980 : // As a special case, when we have just imported an image into the repository,
4981 : // instead of writing out a L0 delta layer, we directly write out image layer
4982 : // files instead. This is possible as long as *all* the data imported into the
4983 : // repository have the same LSN.
4984 : let lsn_range = frozen_layer.get_lsn_range();
4985 :
4986 : // Whether to directly create image layers for this flush, or flush them as delta layers
4987 : let create_image_layer =
4988 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
4989 :
4990 : #[cfg(test)]
4991 : {
4992 : match &mut *self.flush_loop_state.lock().unwrap() {
4993 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
4994 : panic!("flush loop not running")
4995 : }
4996 : FlushLoopState::Running {
4997 : expect_initdb_optimization,
4998 : initdb_optimization_count,
4999 : ..
5000 : } => {
5001 : if create_image_layer {
5002 : *initdb_optimization_count += 1;
5003 : } else {
5004 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
5005 : }
5006 : }
5007 : }
5008 : }
5009 :
5010 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
5011 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
5012 : // require downloading anything during initial import.
5013 : let ((rel_partition, metadata_partition), _lsn) = self
5014 : .repartition(
5015 : self.initdb_lsn,
5016 : self.get_compaction_target_size(),
5017 : EnumSet::empty(),
5018 : ctx,
5019 : )
5020 : .await
5021 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into_anyhow()))?;
5022 :
5023 : if self.cancel.is_cancelled() {
5024 : return Err(FlushLayerError::Cancelled);
5025 : }
5026 :
5027 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
5028 : // So that the key ranges don't overlap.
5029 : let mut partitions = KeyPartitioning::default();
5030 : partitions.parts.extend(rel_partition.parts);
5031 : if !metadata_partition.parts.is_empty() {
5032 : assert_eq!(
5033 : metadata_partition.parts.len(),
5034 : 1,
5035 : "currently sparse keyspace should only contain a single metadata keyspace"
5036 : );
5037 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
5038 : // every single key within the keyspace, and therefore, it's safe to force converting it
5039 : // into a dense keyspace before calling this function.
5040 : partitions
5041 : .parts
5042 : .extend(metadata_partition.into_dense().parts);
5043 : }
5044 :
5045 : let mut layers_to_upload = Vec::new();
5046 : let (generated_image_layers, is_complete) = self
5047 : .create_image_layers(
5048 : &partitions,
5049 : self.initdb_lsn,
5050 : None,
5051 : ImageLayerCreationMode::Initial,
5052 : ctx,
5053 : LastImageLayerCreationStatus::Initial,
5054 : false, // don't yield for L0, we're flushing L0
5055 : )
5056 : .instrument(info_span!("create_image_layers", mode = %ImageLayerCreationMode::Initial, partition_mode = "initial", lsn = %self.initdb_lsn))
5057 : .await?;
5058 : debug_assert!(
5059 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
5060 : "init image generation mode must fully cover the keyspace"
5061 : );
5062 : layers_to_upload.extend(generated_image_layers);
5063 :
5064 : (layers_to_upload, None)
5065 : } else {
5066 : // Normal case, write out a L0 delta layer file.
5067 : // `create_delta_layer` will not modify the layer map.
5068 : // We will remove frozen layer and add delta layer in one atomic operation later.
5069 : let Some(layer) = self
5070 : .create_delta_layer(&frozen_layer, None, ctx)
5071 : .await
5072 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
5073 : else {
5074 : panic!("delta layer cannot be empty if no filter is applied");
5075 : };
5076 : (
5077 : // FIXME: even though we have a single image and single delta layer assumption
5078 : // we push them to vec
5079 : vec![layer.clone()],
5080 : Some(layer),
5081 : )
5082 : };
5083 :
5084 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
5085 :
5086 : if self.cancel.is_cancelled() {
5087 : return Err(FlushLayerError::Cancelled);
5088 : }
5089 :
5090 1 : fail_point!("flush-layer-before-update-remote-consistent-lsn", |_| {
5091 1 : Err(FlushLayerError::Other(anyhow!("failpoint").into()))
5092 1 : });
5093 :
5094 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
5095 :
5096 : // The new on-disk layers are now in the layer map. We can remove the
5097 : // in-memory layer from the map now. The flushed layer is stored in
5098 : // the mapping in `create_delta_layer`.
5099 : {
5100 : let mut guard = self
5101 : .layers
5102 : .write(LayerManagerLockHolder::FlushFrozenLayer)
5103 : .await;
5104 :
5105 : guard.open_mut()?.finish_flush_l0_layer(
5106 : delta_layer_to_add.as_ref(),
5107 : &frozen_layer,
5108 : &self.metrics,
5109 : );
5110 :
5111 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
5112 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
5113 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
5114 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
5115 : }
5116 : // release lock on 'layers'
5117 : };
5118 :
5119 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
5120 : // a compaction can delete the file and then it won't be available for uploads any more.
5121 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
5122 : // race situation.
5123 : // See https://github.com/neondatabase/neon/issues/4526
5124 : pausable_failpoint!("flush-frozen-pausable");
5125 :
5126 : // This failpoint is used by another test case `test_pageserver_recovery`.
5127 : fail_point!("flush-frozen-exit");
5128 :
5129 : Ok(Lsn(lsn_range.end.0 - 1))
5130 : }
5131 :
5132 : /// Return true if the value changed
5133 : ///
5134 : /// This function must only be used from the layer flush task.
5135 596 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
5136 596 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
5137 596 : assert!(
5138 596 : new_value >= old_value,
5139 0 : "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}"
5140 : );
5141 :
5142 596 : self.metrics
5143 596 : .disk_consistent_lsn_gauge
5144 596 : .set(new_value.0 as i64);
5145 596 : new_value != old_value
5146 596 : }
5147 :
5148 : /// Update metadata file
5149 623 : fn schedule_uploads(
5150 623 : &self,
5151 623 : disk_consistent_lsn: Lsn,
5152 623 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
5153 623 : ) -> anyhow::Result<()> {
5154 : // We can only save a valid 'prev_record_lsn' value on disk if we
5155 : // flushed *all* in-memory changes to disk. We only track
5156 : // 'prev_record_lsn' in memory for the latest processed record, so we
5157 : // don't remember what the correct value that corresponds to some old
5158 : // LSN is. But if we flush everything, then the value corresponding
5159 : // current 'last_record_lsn' is correct and we can store it on disk.
5160 : let RecordLsn {
5161 623 : last: last_record_lsn,
5162 623 : prev: prev_record_lsn,
5163 623 : } = self.last_record_lsn.load();
5164 623 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
5165 560 : Some(prev_record_lsn)
5166 : } else {
5167 63 : None
5168 : };
5169 :
5170 623 : let update = crate::tenant::metadata::MetadataUpdate::new(
5171 623 : disk_consistent_lsn,
5172 623 : ondisk_prev_record_lsn,
5173 623 : *self.applied_gc_cutoff_lsn.read(),
5174 : );
5175 :
5176 623 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
5177 0 : "{}",
5178 0 : x.unwrap()
5179 : ));
5180 :
5181 1225 : for layer in layers_to_upload {
5182 602 : self.remote_client.schedule_layer_file_upload(layer)?;
5183 : }
5184 623 : self.remote_client
5185 623 : .schedule_index_upload_for_metadata_update(&update)?;
5186 :
5187 623 : Ok(())
5188 623 : }
5189 :
5190 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
5191 0 : self.remote_client
5192 0 : .preserve_initdb_archive(
5193 0 : &self.tenant_shard_id.tenant_id,
5194 0 : &self.timeline_id,
5195 0 : &self.cancel,
5196 0 : )
5197 0 : .await
5198 0 : }
5199 :
5200 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
5201 : // in layer map immediately. The caller is responsible to put it into the layer map.
5202 486 : async fn create_delta_layer(
5203 486 : self: &Arc<Self>,
5204 486 : frozen_layer: &Arc<InMemoryLayer>,
5205 486 : key_range: Option<Range<Key>>,
5206 486 : ctx: &RequestContext,
5207 486 : ) -> anyhow::Result<Option<ResidentLayer>> {
5208 486 : let self_clone = Arc::clone(self);
5209 486 : let frozen_layer = Arc::clone(frozen_layer);
5210 486 : let ctx = ctx.attached_child();
5211 486 : let work = async move {
5212 486 : let Some((desc, path)) = frozen_layer
5213 486 : .write_to_disk(
5214 486 : &ctx,
5215 486 : key_range,
5216 486 : self_clone.l0_flush_global_state.inner(),
5217 486 : &self_clone.gate,
5218 486 : self_clone.cancel.clone(),
5219 486 : )
5220 486 : .await?
5221 : else {
5222 0 : return Ok(None);
5223 : };
5224 486 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
5225 :
5226 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
5227 : // We just need to fsync the directory in which these inodes are linked,
5228 : // which we know to be the timeline directory.
5229 : //
5230 : // We use fatal_err() below because the after write_to_disk returns with success,
5231 : // the in-memory state of the filesystem already has the layer file in its final place,
5232 : // and subsequent pageserver code could think it's durable while it really isn't.
5233 486 : let timeline_dir = VirtualFile::open(
5234 486 : &self_clone
5235 486 : .conf
5236 486 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
5237 486 : &ctx,
5238 486 : )
5239 486 : .await
5240 486 : .fatal_err("VirtualFile::open for timeline dir fsync");
5241 486 : timeline_dir
5242 486 : .sync_all()
5243 486 : .await
5244 486 : .fatal_err("VirtualFile::sync_all timeline dir");
5245 486 : anyhow::Ok(Some(new_delta))
5246 486 : };
5247 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
5248 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
5249 : use crate::virtual_file::io_engine::IoEngine;
5250 486 : match crate::virtual_file::io_engine::get() {
5251 0 : IoEngine::NotSet => panic!("io engine not set"),
5252 : IoEngine::StdFs => {
5253 0 : let span = tracing::info_span!("blocking");
5254 0 : tokio::task::spawn_blocking({
5255 0 : move || Handle::current().block_on(work.instrument(span))
5256 : })
5257 0 : .await
5258 0 : .context("spawn_blocking")
5259 0 : .and_then(|x| x)
5260 : }
5261 : #[cfg(target_os = "linux")]
5262 486 : IoEngine::TokioEpollUring => work.await,
5263 : }
5264 486 : }
5265 :
5266 303 : async fn repartition(
5267 303 : &self,
5268 303 : lsn: Lsn,
5269 303 : partition_size: u64,
5270 303 : flags: EnumSet<CompactFlags>,
5271 303 : ctx: &RequestContext,
5272 303 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), RepartitionError> {
5273 303 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
5274 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
5275 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
5276 : // and hence before the compaction task starts.
5277 0 : return Err(RepartitionError::Other(anyhow!(
5278 0 : "repartition() called concurrently"
5279 0 : )));
5280 : };
5281 303 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
5282 303 : if lsn < *partition_lsn {
5283 0 : return Err(RepartitionError::Other(anyhow!(
5284 0 : "repartition() called with LSN going backwards, this should not happen"
5285 0 : )));
5286 303 : }
5287 :
5288 303 : let distance = lsn.0 - partition_lsn.0;
5289 303 : if *partition_lsn != Lsn(0)
5290 141 : && distance <= self.repartition_threshold
5291 141 : && !flags.contains(CompactFlags::ForceRepartition)
5292 : {
5293 134 : debug!(
5294 : distance,
5295 : threshold = self.repartition_threshold,
5296 0 : "no repartitioning needed"
5297 : );
5298 134 : return Ok((
5299 134 : (dense_partition.clone(), sparse_partition.clone()),
5300 134 : *partition_lsn,
5301 134 : ));
5302 169 : }
5303 :
5304 169 : let (dense_ks, sparse_ks) = self
5305 169 : .collect_keyspace(lsn, ctx)
5306 169 : .await
5307 169 : .map_err(RepartitionError::CollectKeyspace)?;
5308 169 : let dense_partitioning = dense_ks.partition(
5309 169 : &self.shard_identity,
5310 169 : partition_size,
5311 169 : postgres_ffi::BLCKSZ as u64,
5312 : );
5313 169 : let sparse_partitioning = SparseKeyPartitioning {
5314 169 : parts: vec![sparse_ks],
5315 169 : }; // no partitioning for metadata keys for now
5316 169 : let result = ((dense_partitioning, sparse_partitioning), lsn);
5317 169 : guard.write(result.clone());
5318 169 : Ok(result)
5319 303 : }
5320 :
5321 : // Is it time to create a new image layer for the given partition? True if we want to generate.
5322 57 : async fn time_for_new_image_layer(
5323 57 : &self,
5324 57 : partition: &KeySpace,
5325 57 : lsn: Lsn,
5326 57 : force_image_creation_lsn: Option<Lsn>,
5327 57 : ) -> bool {
5328 57 : let threshold = self.get_image_creation_threshold();
5329 :
5330 57 : let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
5331 57 : let Ok(layers) = guard.layer_map() else {
5332 0 : return false;
5333 : };
5334 57 : let mut min_image_lsn: Lsn = Lsn::MAX;
5335 57 : let mut max_deltas = 0;
5336 364 : for part_range in &partition.ranges {
5337 307 : let image_coverage = layers.image_coverage(part_range, lsn);
5338 614 : for (img_range, last_img) in image_coverage {
5339 307 : let img_lsn = if let Some(last_img) = last_img {
5340 0 : last_img.get_lsn_range().end
5341 : } else {
5342 307 : Lsn(0)
5343 : };
5344 : // Let's consider an example:
5345 : //
5346 : // delta layer with LSN range 71-81
5347 : // delta layer with LSN range 81-91
5348 : // delta layer with LSN range 91-101
5349 : // image layer at LSN 100
5350 : //
5351 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
5352 : // there's no need to create a new one. We check this case explicitly, to avoid passing
5353 : // a bogus range to count_deltas below, with start > end. It's even possible that there
5354 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
5355 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
5356 307 : if img_lsn < lsn {
5357 307 : let num_deltas =
5358 307 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
5359 :
5360 307 : max_deltas = max_deltas.max(num_deltas);
5361 307 : if num_deltas >= threshold {
5362 0 : debug!(
5363 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
5364 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
5365 : );
5366 0 : return true;
5367 307 : }
5368 0 : }
5369 307 : min_image_lsn = min(min_image_lsn, img_lsn);
5370 : }
5371 : }
5372 :
5373 : // HADRON
5374 : // for child timelines, we consider all pages up to ancestor_LSN are redone successfully by the parent timeline
5375 57 : min_image_lsn = min_image_lsn.max(self.get_ancestor_lsn());
5376 57 : if min_image_lsn < force_image_creation_lsn.unwrap_or(Lsn(0)) && max_deltas > 0 {
5377 0 : info!(
5378 0 : "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}, num deltas: {}",
5379 0 : partition.ranges[0].start,
5380 0 : partition.ranges[0].end,
5381 : min_image_lsn,
5382 0 : force_image_creation_lsn.unwrap(),
5383 : max_deltas
5384 : );
5385 0 : return true;
5386 57 : }
5387 :
5388 57 : debug!(
5389 : max_deltas,
5390 0 : "none of the partitioned ranges had >= {threshold} deltas"
5391 : );
5392 57 : false
5393 57 : }
5394 :
5395 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
5396 : /// so that at most one image layer will be produced from this function.
5397 : #[allow(clippy::too_many_arguments)]
5398 124 : async fn create_image_layer_for_rel_blocks(
5399 124 : self: &Arc<Self>,
5400 124 : partition: &KeySpace,
5401 124 : mut image_layer_writer: ImageLayerWriter,
5402 124 : lsn: Lsn,
5403 124 : ctx: &RequestContext,
5404 124 : img_range: Range<Key>,
5405 124 : io_concurrency: IoConcurrency,
5406 124 : progress: Option<(usize, usize)>,
5407 124 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5408 124 : let mut wrote_keys = false;
5409 :
5410 124 : let mut key_request_accum = KeySpaceAccum::new();
5411 824 : for range in &partition.ranges {
5412 700 : let mut key = range.start;
5413 1517 : while key < range.end {
5414 : // Decide whether to retain this key: usually we do, but sharded tenants may
5415 : // need to drop keys that don't belong to them. If we retain the key, add it
5416 : // to `key_request_accum` for later issuing a vectored get
5417 817 : if self.shard_identity.is_key_disposable(&key) {
5418 0 : debug!(
5419 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
5420 : key,
5421 0 : self.shard_identity.get_shard_number(&key)
5422 : );
5423 817 : } else {
5424 817 : key_request_accum.add_key(key);
5425 817 : }
5426 :
5427 817 : let last_key_in_range = key.next() == range.end;
5428 817 : key = key.next();
5429 :
5430 : // Maybe flush `key_rest_accum`
5431 817 : if key_request_accum.raw_size() >= self.conf.max_get_vectored_keys.get() as u64
5432 817 : || (last_key_in_range && key_request_accum.raw_size() > 0)
5433 : {
5434 700 : let query =
5435 700 : VersionedKeySpaceQuery::uniform(key_request_accum.consume_keyspace(), lsn);
5436 :
5437 700 : let results = self
5438 700 : .get_vectored(query, io_concurrency.clone(), ctx)
5439 700 : .await?;
5440 :
5441 700 : if self.cancel.is_cancelled() {
5442 0 : return Err(CreateImageLayersError::Cancelled);
5443 700 : }
5444 :
5445 1517 : for (img_key, img) in results {
5446 817 : let img = match img {
5447 817 : Ok(img) => img,
5448 0 : Err(err) => {
5449 : // If we fail to reconstruct a VM or FSM page, we can zero the
5450 : // page without losing any actual user data. That seems better
5451 : // than failing repeatedly and getting stuck.
5452 : //
5453 : // We had a bug at one point, where we truncated the FSM and VM
5454 : // in the pageserver, but the Postgres didn't know about that
5455 : // and continued to generate incremental WAL records for pages
5456 : // that didn't exist in the pageserver. Trying to replay those
5457 : // WAL records failed to find the previous image of the page.
5458 : // This special case allows us to recover from that situation.
5459 : // See https://github.com/neondatabase/neon/issues/2601.
5460 : //
5461 : // Unfortunately we cannot do this for the main fork, or for
5462 : // any metadata keys, keys, as that would lead to actual data
5463 : // loss.
5464 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
5465 0 : warn!(
5466 0 : "could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}"
5467 : );
5468 0 : ZERO_PAGE.clone()
5469 : } else {
5470 0 : return Err(CreateImageLayersError::from(err));
5471 : }
5472 : }
5473 : };
5474 :
5475 : // Write all the keys we just read into our new image layer.
5476 817 : image_layer_writer.put_image(img_key, img, ctx).await?;
5477 817 : wrote_keys = true;
5478 : }
5479 117 : }
5480 : }
5481 : }
5482 :
5483 124 : let progress_report = progress
5484 124 : .map(|(idx, total)| format!("({idx}/{total}) "))
5485 124 : .unwrap_or_default();
5486 124 : if wrote_keys {
5487 : // Normal path: we have written some data into the new image layer for this
5488 : // partition, so flush it to disk.
5489 124 : info!(
5490 0 : "{} produced image layer for rel {}",
5491 : progress_report,
5492 0 : ImageLayerName {
5493 0 : key_range: img_range.clone(),
5494 0 : lsn
5495 0 : },
5496 : );
5497 124 : Ok(ImageLayerCreationOutcome::Generated {
5498 124 : unfinished_image_layer: image_layer_writer,
5499 124 : })
5500 : } else {
5501 0 : tracing::debug!(
5502 0 : "{} no data in range {}-{}",
5503 : progress_report,
5504 : img_range.start,
5505 : img_range.end
5506 : );
5507 0 : Ok(ImageLayerCreationOutcome::Empty)
5508 : }
5509 124 : }
5510 :
5511 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
5512 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
5513 : /// would not be too large to fit in a single image layer.
5514 : ///
5515 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
5516 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
5517 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
5518 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
5519 : #[allow(clippy::too_many_arguments)]
5520 169 : async fn create_image_layer_for_metadata_keys(
5521 169 : self: &Arc<Self>,
5522 169 : partition: &KeySpace,
5523 169 : mut image_layer_writer: ImageLayerWriter,
5524 169 : lsn: Lsn,
5525 169 : ctx: &RequestContext,
5526 169 : img_range: Range<Key>,
5527 169 : mode: ImageLayerCreationMode,
5528 169 : io_concurrency: IoConcurrency,
5529 169 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5530 : // Metadata keys image layer creation.
5531 169 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
5532 169 : let begin = Instant::now();
5533 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
5534 : // not contain too many keys, otherwise this takes a lot of memory.
5535 169 : let data = self
5536 169 : .get_vectored_impl(
5537 169 : VersionedKeySpaceQuery::uniform(partition.clone(), lsn),
5538 169 : &mut reconstruct_state,
5539 169 : ctx,
5540 169 : )
5541 169 : .await?;
5542 169 : let (data, total_kb_retrieved, total_keys_retrieved) = {
5543 169 : let mut new_data = BTreeMap::new();
5544 169 : let mut total_kb_retrieved = 0;
5545 169 : let mut total_keys_retrieved = 0;
5546 5175 : for (k, v) in data {
5547 5006 : let v = v?;
5548 5006 : total_kb_retrieved += KEY_SIZE + v.len();
5549 5006 : total_keys_retrieved += 1;
5550 5006 : new_data.insert(k, v);
5551 : }
5552 169 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
5553 : };
5554 169 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
5555 169 : let elapsed = begin.elapsed();
5556 :
5557 169 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
5558 169 : info!(
5559 0 : "metadata key compaction: trigger_generation={trigger_generation}, delta_files_accessed={delta_files_accessed}, total_kb_retrieved={total_kb_retrieved}, total_keys_retrieved={total_keys_retrieved}, read_time={}s",
5560 0 : elapsed.as_secs_f64()
5561 : );
5562 :
5563 169 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
5564 16 : return Ok(ImageLayerCreationOutcome::Skip);
5565 153 : }
5566 153 : if self.cancel.is_cancelled() {
5567 0 : return Err(CreateImageLayersError::Cancelled);
5568 153 : }
5569 153 : let mut wrote_any_image = false;
5570 5159 : for (k, v) in data {
5571 5006 : if v.is_empty() {
5572 : // the key has been deleted, it does not need an image
5573 : // in metadata keyspace, an empty image == tombstone
5574 4 : continue;
5575 5002 : }
5576 5002 : wrote_any_image = true;
5577 :
5578 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
5579 :
5580 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
5581 : // on the normal data path either.
5582 5002 : image_layer_writer.put_image(k, v, ctx).await?;
5583 : }
5584 :
5585 153 : if wrote_any_image {
5586 : // Normal path: we have written some data into the new image layer for this
5587 : // partition, so flush it to disk.
5588 6 : info!(
5589 0 : "created image layer for metadata {}",
5590 0 : ImageLayerName {
5591 0 : key_range: img_range.clone(),
5592 0 : lsn
5593 0 : }
5594 : );
5595 6 : Ok(ImageLayerCreationOutcome::Generated {
5596 6 : unfinished_image_layer: image_layer_writer,
5597 6 : })
5598 : } else {
5599 147 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
5600 147 : Ok(ImageLayerCreationOutcome::Empty)
5601 : }
5602 169 : }
5603 :
5604 : /// Predicate function which indicates whether we should check if new image layers
5605 : /// are required. Since checking if new image layers are required is expensive in
5606 : /// terms of CPU, we only do it in the following cases:
5607 : /// 1. If the timeline has ingested sufficient WAL to justify the cost or ...
5608 : /// 2. If enough time has passed since the last check:
5609 : /// 1. For large tenants, we wish to perform the check more often since they
5610 : /// suffer from the lack of image layers. Note that we assume sharded tenants
5611 : /// to be large since non-zero shards do not track the logical size.
5612 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
5613 191 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
5614 191 : let large_timeline_threshold = self.conf.image_layer_generation_large_timeline_threshold;
5615 :
5616 191 : let last_checks_at = self.last_image_layer_creation_check_at.load();
5617 191 : let distance = lsn
5618 191 : .checked_sub(last_checks_at)
5619 191 : .expect("Attempt to compact with LSN going backwards");
5620 191 : let min_distance =
5621 191 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
5622 :
5623 191 : let distance_based_decision = distance.0 >= min_distance;
5624 :
5625 191 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
5626 191 : let check_required_after = (|| {
5627 191 : if self.shard_identity.is_unsharded() {
5628 135 : if let CurrentLogicalSize::Exact(logical_size) =
5629 186 : self.current_logical_size.current_size()
5630 : {
5631 135 : if Some(Into::<u64>::into(&logical_size)) < large_timeline_threshold {
5632 135 : return Duration::from_secs(3600 * 48);
5633 0 : }
5634 51 : }
5635 5 : }
5636 :
5637 56 : self.get_checkpoint_timeout()
5638 : })();
5639 :
5640 191 : let time_based_decision = match *last_check_instant {
5641 29 : Some(last_check) => {
5642 29 : let elapsed = last_check.elapsed();
5643 29 : elapsed >= check_required_after
5644 : }
5645 162 : None => true,
5646 : };
5647 :
5648 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
5649 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
5650 : // check.
5651 191 : let decision = distance_based_decision || time_based_decision;
5652 191 : tracing::info!(
5653 0 : "Decided to check image layers: {}. Distance-based decision: {}, time-based decision: {}",
5654 : decision,
5655 : distance_based_decision,
5656 : time_based_decision
5657 : );
5658 191 : if decision {
5659 162 : self.last_image_layer_creation_check_at.store(lsn);
5660 162 : *last_check_instant = Some(Instant::now());
5661 162 : }
5662 :
5663 191 : decision
5664 191 : }
5665 :
5666 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
5667 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
5668 : ///
5669 : /// `partition_mode` is only for logging purpose and is not used anywhere in this function.
5670 : #[allow(clippy::too_many_arguments)]
5671 191 : async fn create_image_layers(
5672 191 : self: &Arc<Timeline>,
5673 191 : partitioning: &KeyPartitioning,
5674 191 : lsn: Lsn,
5675 191 : force_image_creation_lsn: Option<Lsn>,
5676 191 : mode: ImageLayerCreationMode,
5677 191 : ctx: &RequestContext,
5678 191 : last_status: LastImageLayerCreationStatus,
5679 191 : yield_for_l0: bool,
5680 191 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
5681 191 : let timer = self.metrics.create_images_time_histo.start_timer();
5682 :
5683 191 : if partitioning.parts.is_empty() {
5684 0 : warn!("no partitions to create image layers for");
5685 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5686 191 : }
5687 :
5688 : // We need to avoid holes between generated image layers.
5689 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
5690 : // image layer with hole between them. In this case such layer can not be utilized by GC.
5691 : //
5692 : // How such hole between partitions can appear?
5693 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
5694 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
5695 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
5696 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
5697 191 : let mut start = Key::MIN;
5698 :
5699 191 : let check_for_image_layers =
5700 191 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5701 0 : info!(
5702 0 : "resuming image layer creation: last_status=incomplete, continue from {}",
5703 : last_key
5704 : );
5705 0 : true
5706 : } else {
5707 191 : self.should_check_if_image_layers_required(lsn)
5708 : };
5709 :
5710 191 : let mut batch_image_writer = BatchLayerWriter::new(self.conf);
5711 :
5712 191 : let mut all_generated = true;
5713 :
5714 191 : let mut partition_processed = 0;
5715 191 : let mut total_partitions = partitioning.parts.len();
5716 191 : let mut last_partition_processed = None;
5717 191 : let mut partition_parts = partitioning.parts.clone();
5718 :
5719 191 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5720 : // We need to skip the partitions that have already been processed.
5721 0 : let mut found = false;
5722 0 : for (i, partition) in partition_parts.iter().enumerate() {
5723 0 : if last_key <= partition.end().unwrap() {
5724 : // ```plain
5725 : // |------|--------|----------|------|
5726 : // ^last_key
5727 : // ^start from this partition
5728 : // ```
5729 : // Why `i+1` instead of `i`?
5730 : // It is possible that the user did some writes after the previous image layer creation attempt so that
5731 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
5732 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
5733 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
5734 0 : if i + 1 >= total_partitions {
5735 : // In general, this case should not happen -- if last_key is on the last partition, the previous
5736 : // iteration of image layer creation should return a complete status.
5737 0 : break; // with found=false
5738 0 : }
5739 0 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
5740 0 : total_partitions = partition_parts.len();
5741 : // Update the start key to the partition start.
5742 0 : start = partition_parts[0].start().unwrap();
5743 0 : found = true;
5744 0 : break;
5745 0 : }
5746 : }
5747 0 : if !found {
5748 : // Last key is within the last partition, or larger than all partitions.
5749 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5750 0 : }
5751 191 : }
5752 :
5753 191 : let total = partition_parts.len();
5754 394 : for (idx, partition) in partition_parts.iter().enumerate() {
5755 394 : if self.cancel.is_cancelled() {
5756 0 : return Err(CreateImageLayersError::Cancelled);
5757 394 : }
5758 394 : partition_processed += 1;
5759 394 : let img_range = start..partition.ranges.last().unwrap().end;
5760 394 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
5761 394 : if compact_metadata {
5762 764 : for range in &partition.ranges {
5763 573 : assert!(
5764 573 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
5765 573 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
5766 0 : "metadata keys must be partitioned separately"
5767 : );
5768 : }
5769 191 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
5770 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
5771 : // might mess up with evictions.
5772 22 : start = img_range.end;
5773 22 : continue;
5774 169 : }
5775 : // For initial and force modes, we always generate image layers for metadata keys.
5776 203 : } else if let ImageLayerCreationMode::Try = mode {
5777 : // check_for_image_layers = false -> skip
5778 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
5779 79 : if !check_for_image_layers
5780 57 : || !self
5781 57 : .time_for_new_image_layer(partition, lsn, force_image_creation_lsn)
5782 57 : .await
5783 : {
5784 79 : start = img_range.end;
5785 79 : continue;
5786 0 : }
5787 124 : }
5788 293 : if let ImageLayerCreationMode::Force = mode {
5789 : // When forced to create image layers, we might try and create them where they already
5790 : // exist. This mode is only used in tests/debug.
5791 14 : let layers = self.layers.read(LayerManagerLockHolder::Compaction).await;
5792 14 : if layers.contains_key(&PersistentLayerKey {
5793 14 : key_range: img_range.clone(),
5794 14 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
5795 14 : is_delta: false,
5796 14 : }) {
5797 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
5798 : // in the future.
5799 0 : tracing::info!(
5800 0 : "Skipping image layer at {lsn} {}..{}, already exists",
5801 : img_range.start,
5802 : img_range.end
5803 : );
5804 0 : start = img_range.end;
5805 0 : continue;
5806 14 : }
5807 279 : }
5808 :
5809 293 : let image_layer_writer = ImageLayerWriter::new(
5810 293 : self.conf,
5811 293 : self.timeline_id,
5812 293 : self.tenant_shard_id,
5813 293 : &img_range,
5814 293 : lsn,
5815 293 : &self.gate,
5816 293 : self.cancel.clone(),
5817 293 : ctx,
5818 293 : )
5819 293 : .await
5820 293 : .map_err(CreateImageLayersError::Other)?;
5821 :
5822 293 : fail_point!("image-layer-writer-fail-before-finish", |_| {
5823 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
5824 0 : "failpoint image-layer-writer-fail-before-finish"
5825 0 : )))
5826 0 : });
5827 :
5828 293 : let io_concurrency = IoConcurrency::spawn_from_conf(
5829 293 : self.conf.get_vectored_concurrent_io,
5830 293 : self.gate
5831 293 : .enter()
5832 293 : .map_err(|_| CreateImageLayersError::Cancelled)?,
5833 : );
5834 :
5835 293 : let outcome = if !compact_metadata {
5836 124 : self.create_image_layer_for_rel_blocks(
5837 124 : partition,
5838 124 : image_layer_writer,
5839 124 : lsn,
5840 124 : ctx,
5841 124 : img_range.clone(),
5842 124 : io_concurrency,
5843 124 : Some((idx, total)),
5844 124 : )
5845 124 : .await?
5846 : } else {
5847 169 : self.create_image_layer_for_metadata_keys(
5848 169 : partition,
5849 169 : image_layer_writer,
5850 169 : lsn,
5851 169 : ctx,
5852 169 : img_range.clone(),
5853 169 : mode,
5854 169 : io_concurrency,
5855 169 : )
5856 169 : .await?
5857 : };
5858 293 : match outcome {
5859 147 : ImageLayerCreationOutcome::Empty => {
5860 147 : // No data in this partition, so we don't need to create an image layer (for now).
5861 147 : // The next image layer should cover this key range, so we don't advance the `start`
5862 147 : // key.
5863 147 : }
5864 : ImageLayerCreationOutcome::Generated {
5865 130 : unfinished_image_layer,
5866 130 : } => {
5867 130 : batch_image_writer.add_unfinished_image_writer(
5868 130 : unfinished_image_layer,
5869 130 : img_range.clone(),
5870 130 : lsn,
5871 130 : );
5872 130 : // The next image layer should be generated right after this one.
5873 130 : start = img_range.end;
5874 130 : }
5875 16 : ImageLayerCreationOutcome::Skip => {
5876 16 : // We don't need to create an image layer for this partition.
5877 16 : // The next image layer should NOT cover this range, otherwise
5878 16 : // the keyspace becomes empty (reads don't go past image layers).
5879 16 : start = img_range.end;
5880 16 : }
5881 : }
5882 :
5883 293 : if let ImageLayerCreationMode::Try = mode {
5884 : // We have at least made some progress
5885 51 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
5886 : // The `Try` mode is currently only used on the compaction path. We want to avoid
5887 : // image layer generation taking too long time and blocking L0 compaction. So in this
5888 : // mode, we also inspect the current number of L0 layers and skip image layer generation
5889 : // if there are too many of them.
5890 0 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
5891 0 : * self.get_compaction_threshold();
5892 : // TODO: currently we do not respect `get_image_creation_preempt_threshold` and always yield
5893 : // when there is a single timeline with more than L0 threshold L0 layers. As long as the
5894 : // `get_image_creation_preempt_threshold` is set to a value greater than 0, we will yield for L0 compaction.
5895 0 : if image_preempt_threshold != 0 {
5896 0 : let should_yield = self
5897 0 : .l0_compaction_trigger
5898 0 : .notified()
5899 0 : .now_or_never()
5900 0 : .is_some();
5901 0 : if should_yield {
5902 0 : tracing::info!(
5903 0 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers",
5904 0 : partition.start().unwrap(),
5905 0 : partition.end().unwrap()
5906 : );
5907 0 : last_partition_processed = Some(partition.clone());
5908 0 : all_generated = false;
5909 0 : break;
5910 0 : }
5911 0 : }
5912 51 : }
5913 242 : }
5914 : }
5915 :
5916 191 : let image_layers = batch_image_writer
5917 191 : .finish(self, ctx)
5918 191 : .await
5919 191 : .map_err(CreateImageLayersError::Other)?;
5920 :
5921 191 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
5922 :
5923 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
5924 : // now they are being scheduled outside of write lock; current way is inconsistent with
5925 : // compaction lock order.
5926 191 : guard
5927 191 : .open_mut()?
5928 191 : .track_new_image_layers(&image_layers, &self.metrics);
5929 191 : drop_layer_manager_wlock(guard);
5930 191 : let duration = timer.stop_and_record();
5931 :
5932 : // Creating image layers may have caused some previously visible layers to be covered
5933 191 : if !image_layers.is_empty() {
5934 118 : self.update_layer_visibility().await?;
5935 73 : }
5936 :
5937 191 : let total_layer_size = image_layers
5938 191 : .iter()
5939 191 : .map(|l| l.metadata().file_size)
5940 191 : .sum::<u64>();
5941 :
5942 191 : if !image_layers.is_empty() {
5943 118 : info!(
5944 0 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
5945 0 : image_layers.len(),
5946 : total_layer_size,
5947 0 : duration.as_secs_f64(),
5948 : partition_processed,
5949 : total_partitions
5950 : );
5951 73 : }
5952 :
5953 : Ok((
5954 191 : image_layers,
5955 191 : if all_generated {
5956 191 : LastImageLayerCreationStatus::Complete
5957 : } else {
5958 : LastImageLayerCreationStatus::Incomplete {
5959 0 : last_key: if let Some(last_partition_processed) = last_partition_processed {
5960 0 : last_partition_processed.end().unwrap_or(Key::MIN)
5961 : } else {
5962 : // This branch should be unreachable, but in case it happens, we can just return the start key.
5963 0 : Key::MIN
5964 : },
5965 : }
5966 : },
5967 : ))
5968 191 : }
5969 :
5970 : /// Wait until the background initial logical size calculation is complete, or
5971 : /// this Timeline is shut down. Calling this function will cause the initial
5972 : /// logical size calculation to skip waiting for the background jobs barrier.
5973 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
5974 0 : if !self.shard_identity.is_shard_zero() {
5975 : // We don't populate logical size on shard >0: skip waiting for it.
5976 0 : return;
5977 0 : }
5978 :
5979 0 : if self.remote_client.is_deleting() {
5980 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
5981 0 : return;
5982 0 : }
5983 :
5984 0 : if self.current_logical_size.current_size().is_exact() {
5985 : // root timelines are initialized with exact count, but never start the background
5986 : // calculation
5987 0 : return;
5988 0 : }
5989 :
5990 0 : if self.cancel.is_cancelled() {
5991 : // We already requested stopping the tenant, so we cannot wait for the logical size
5992 : // calculation to complete given the task might have been already cancelled.
5993 0 : return;
5994 0 : }
5995 :
5996 0 : if let Some(await_bg_cancel) = self
5997 0 : .current_logical_size
5998 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
5999 0 : .get()
6000 0 : {
6001 0 : await_bg_cancel.cancel();
6002 0 : } else {
6003 : // We should not wait if we were not able to explicitly instruct
6004 : // the logical size cancellation to skip the concurrency limit semaphore.
6005 : // TODO: this is an unexpected case. We should restructure so that it
6006 : // can't happen.
6007 0 : tracing::warn!(
6008 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
6009 : );
6010 0 : debug_assert!(false);
6011 : }
6012 :
6013 0 : tokio::select!(
6014 0 : _ = self.current_logical_size.initialized.acquire() => {},
6015 0 : _ = self.cancel.cancelled() => {}
6016 : )
6017 0 : }
6018 :
6019 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
6020 : /// Timelines layers up to the ancestor_lsn.
6021 : ///
6022 : /// Requires a timeline that:
6023 : /// - has an ancestor to detach from
6024 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
6025 : /// a technical requirement
6026 : ///
6027 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
6028 : /// polled again until completion.
6029 : ///
6030 : /// During the operation all timelines sharing the data with this timeline will be reparented
6031 : /// from our ancestor to be branches of this timeline.
6032 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
6033 0 : self: &Arc<Timeline>,
6034 0 : tenant: &crate::tenant::TenantShard,
6035 0 : options: detach_ancestor::Options,
6036 0 : behavior: DetachBehavior,
6037 0 : ctx: &RequestContext,
6038 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
6039 0 : detach_ancestor::prepare(self, tenant, behavior, options, ctx).await
6040 0 : }
6041 :
6042 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
6043 : /// reparents any reparentable children of previous ancestor.
6044 : ///
6045 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
6046 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
6047 : /// successfully, tenant must be reloaded.
6048 : ///
6049 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
6050 : /// resetting the tenant.
6051 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
6052 0 : self: &Arc<Timeline>,
6053 0 : tenant: &crate::tenant::TenantShard,
6054 0 : prepared: detach_ancestor::PreparedTimelineDetach,
6055 0 : ancestor_timeline_id: TimelineId,
6056 0 : ancestor_lsn: Lsn,
6057 0 : behavior: DetachBehavior,
6058 0 : ctx: &RequestContext,
6059 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
6060 0 : detach_ancestor::detach_and_reparent(
6061 0 : self,
6062 0 : tenant,
6063 0 : prepared,
6064 0 : ancestor_timeline_id,
6065 0 : ancestor_lsn,
6066 0 : behavior,
6067 0 : ctx,
6068 0 : )
6069 0 : .await
6070 0 : }
6071 :
6072 : /// Final step which unblocks the GC.
6073 : ///
6074 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
6075 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
6076 0 : self: &Arc<Timeline>,
6077 0 : tenant: &crate::tenant::TenantShard,
6078 0 : attempt: detach_ancestor::Attempt,
6079 0 : ctx: &RequestContext,
6080 0 : ) -> Result<(), detach_ancestor::Error> {
6081 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
6082 0 : }
6083 : }
6084 :
6085 : impl Drop for Timeline {
6086 5 : fn drop(&mut self) {
6087 5 : if let Some(ancestor) = &self.ancestor_timeline {
6088 : // This lock should never be poisoned, but in case it is we do a .map() instead of
6089 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
6090 2 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
6091 2 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
6092 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
6093 0 : "Couldn't remove retain_lsn entry from timeline's parent on drop: already removed");
6094 2 : }
6095 0 : }
6096 3 : }
6097 5 : info!(
6098 0 : "Timeline {} for tenant {} is being dropped",
6099 : self.timeline_id, self.tenant_shard_id.tenant_id
6100 : );
6101 5 : }
6102 : }
6103 :
6104 : pub(crate) use compaction_error::CompactionError;
6105 : /// In a private mod to enforce that [`CompactionError::is_cancel`] is used
6106 : /// instead of `match`ing on [`CompactionError::ShuttingDown`].
6107 : mod compaction_error {
6108 : use utils::sync::gate::GateError;
6109 :
6110 : use crate::{
6111 : pgdatadir_mapping::CollectKeySpaceError,
6112 : tenant::{PageReconstructError, blob_io::WriteBlobError, upload_queue::NotInitialized},
6113 : virtual_file::owned_buffers_io::write::FlushTaskError,
6114 : };
6115 :
6116 : /// Top-level failure to compact. Use [`Self::is_cancel`].
6117 : #[derive(Debug, thiserror::Error)]
6118 : pub(crate) enum CompactionError {
6119 : /// Use [`Self::is_cancel`] instead of checking for this variant.
6120 : #[error("The timeline or pageserver is shutting down")]
6121 : #[allow(private_interfaces)]
6122 : ShuttingDown(ForbidMatching), // private ForbidMatching enforces use of [`Self::is_cancel`].
6123 : #[error(transparent)]
6124 : Other(anyhow::Error),
6125 : }
6126 :
6127 : #[derive(Debug)]
6128 : struct ForbidMatching;
6129 :
6130 : impl CompactionError {
6131 0 : pub fn new_cancelled() -> Self {
6132 0 : Self::ShuttingDown(ForbidMatching)
6133 0 : }
6134 : /// Errors that can be ignored, i.e., cancel and shutdown.
6135 0 : pub fn is_cancel(&self) -> bool {
6136 0 : let other = match self {
6137 0 : CompactionError::ShuttingDown(_) => return true,
6138 0 : CompactionError::Other(other) => other,
6139 : };
6140 :
6141 : // The write path of compaction in particular often lacks differentiated
6142 : // handling errors stemming from cancellation from other errors.
6143 : // So, if requested, we also check the ::Other variant by downcasting.
6144 : // The list below has been found empirically from flaky tests and production logs.
6145 : // The process is simple: on ::Other(), compaction will print the enclosed
6146 : // anyhow::Error in debug mode, i.e., with backtrace. That backtrace contains the
6147 : // line where the write path / compaction code does undifferentiated error handling
6148 : // from a non-anyhow type to an anyhow type. Add the type to the list of downcasts
6149 : // below, following the same is_cancel() pattern.
6150 :
6151 0 : let root_cause = other.root_cause();
6152 :
6153 0 : let upload_queue = root_cause
6154 0 : .downcast_ref::<NotInitialized>()
6155 0 : .is_some_and(|e| e.is_stopping());
6156 0 : let timeline = root_cause
6157 0 : .downcast_ref::<PageReconstructError>()
6158 0 : .is_some_and(|e| e.is_cancel());
6159 0 : let buffered_writer_flush_task_canelled = root_cause
6160 0 : .downcast_ref::<FlushTaskError>()
6161 0 : .is_some_and(|e| e.is_cancel());
6162 0 : let write_blob_cancelled = root_cause
6163 0 : .downcast_ref::<WriteBlobError>()
6164 0 : .is_some_and(|e| e.is_cancel());
6165 0 : let gate_closed = root_cause
6166 0 : .downcast_ref::<GateError>()
6167 0 : .is_some_and(|e| e.is_cancel());
6168 0 : upload_queue
6169 0 : || timeline
6170 0 : || buffered_writer_flush_task_canelled
6171 0 : || write_blob_cancelled
6172 0 : || gate_closed
6173 0 : }
6174 0 : pub fn into_anyhow(self) -> anyhow::Error {
6175 0 : match self {
6176 0 : CompactionError::ShuttingDown(ForbidMatching) => anyhow::Error::new(self),
6177 0 : CompactionError::Other(e) => e,
6178 : }
6179 0 : }
6180 0 : pub fn from_collect_keyspace(err: CollectKeySpaceError) -> Self {
6181 0 : if err.is_cancel() {
6182 0 : Self::new_cancelled()
6183 : } else {
6184 0 : Self::Other(err.into_anyhow())
6185 : }
6186 0 : }
6187 : }
6188 : }
6189 :
6190 : impl From<super::upload_queue::NotInitialized> for CompactionError {
6191 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
6192 0 : match value {
6193 : super::upload_queue::NotInitialized::Uninitialized => {
6194 0 : CompactionError::Other(anyhow::anyhow!(value))
6195 : }
6196 : super::upload_queue::NotInitialized::ShuttingDown
6197 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::new_cancelled(),
6198 : }
6199 0 : }
6200 : }
6201 :
6202 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
6203 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
6204 0 : match e {
6205 : super::storage_layer::layer::DownloadError::TimelineShutdown
6206 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
6207 0 : CompactionError::new_cancelled()
6208 : }
6209 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
6210 : | super::storage_layer::layer::DownloadError::DownloadRequired
6211 : | super::storage_layer::layer::DownloadError::NotFile(_)
6212 : | super::storage_layer::layer::DownloadError::DownloadFailed
6213 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
6214 0 : CompactionError::Other(anyhow::anyhow!(e))
6215 : }
6216 : #[cfg(test)]
6217 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
6218 0 : CompactionError::Other(anyhow::anyhow!(e))
6219 : }
6220 : }
6221 0 : }
6222 : }
6223 :
6224 : impl From<layer_manager::Shutdown> for CompactionError {
6225 0 : fn from(_: layer_manager::Shutdown) -> Self {
6226 0 : CompactionError::new_cancelled()
6227 0 : }
6228 : }
6229 :
6230 : impl From<super::storage_layer::errors::PutError> for CompactionError {
6231 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
6232 0 : if e.is_cancel() {
6233 0 : CompactionError::new_cancelled()
6234 : } else {
6235 0 : CompactionError::Other(e.into_anyhow())
6236 : }
6237 0 : }
6238 : }
6239 :
6240 : #[serde_as]
6241 : #[derive(serde::Serialize)]
6242 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
6243 :
6244 : #[derive(Default)]
6245 : enum DurationRecorder {
6246 : #[default]
6247 : NotStarted,
6248 : Recorded(RecordedDuration, tokio::time::Instant),
6249 : }
6250 :
6251 : impl DurationRecorder {
6252 115 : fn till_now(&self) -> DurationRecorder {
6253 115 : match self {
6254 : DurationRecorder::NotStarted => {
6255 0 : panic!("must only call on recorded measurements")
6256 : }
6257 115 : DurationRecorder::Recorded(_, ended) => {
6258 115 : let now = tokio::time::Instant::now();
6259 115 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
6260 : }
6261 : }
6262 115 : }
6263 138 : fn into_recorded(self) -> Option<RecordedDuration> {
6264 138 : match self {
6265 0 : DurationRecorder::NotStarted => None,
6266 138 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
6267 : }
6268 138 : }
6269 : }
6270 :
6271 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
6272 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
6273 : /// the layer descriptor requires the user to provide the ranges, which should cover all
6274 : /// keys specified in the `data` field.
6275 : #[cfg(test)]
6276 : #[derive(Clone)]
6277 : pub struct DeltaLayerTestDesc {
6278 : pub lsn_range: Range<Lsn>,
6279 : pub key_range: Range<Key>,
6280 : pub data: Vec<(Key, Lsn, Value)>,
6281 : }
6282 :
6283 : #[cfg(test)]
6284 : #[derive(Clone)]
6285 : pub struct InMemoryLayerTestDesc {
6286 : pub lsn_range: Range<Lsn>,
6287 : pub data: Vec<(Key, Lsn, Value)>,
6288 : pub is_open: bool,
6289 : }
6290 :
6291 : #[cfg(test)]
6292 : impl DeltaLayerTestDesc {
6293 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
6294 2 : Self {
6295 2 : lsn_range,
6296 2 : key_range,
6297 2 : data,
6298 2 : }
6299 2 : }
6300 :
6301 45 : pub fn new_with_inferred_key_range(
6302 45 : lsn_range: Range<Lsn>,
6303 45 : data: Vec<(Key, Lsn, Value)>,
6304 45 : ) -> Self {
6305 45 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
6306 45 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
6307 45 : Self {
6308 45 : key_range: (*key_min)..(key_max.next()),
6309 45 : lsn_range,
6310 45 : data,
6311 45 : }
6312 45 : }
6313 :
6314 5 : pub(crate) fn layer_name(&self) -> LayerName {
6315 5 : LayerName::Delta(super::storage_layer::DeltaLayerName {
6316 5 : key_range: self.key_range.clone(),
6317 5 : lsn_range: self.lsn_range.clone(),
6318 5 : })
6319 5 : }
6320 : }
6321 :
6322 : impl Timeline {
6323 23 : async fn finish_compact_batch(
6324 23 : self: &Arc<Self>,
6325 23 : new_deltas: &[ResidentLayer],
6326 23 : new_images: &[ResidentLayer],
6327 23 : layers_to_remove: &[Layer],
6328 23 : ) -> Result<(), CompactionError> {
6329 23 : let mut guard = tokio::select! {
6330 23 : guard = self.layers.write(LayerManagerLockHolder::Compaction) => guard,
6331 23 : _ = self.cancel.cancelled() => {
6332 0 : return Err(CompactionError::new_cancelled());
6333 : }
6334 : };
6335 :
6336 23 : let mut duplicated_layers = HashSet::new();
6337 :
6338 23 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
6339 :
6340 186 : for l in new_deltas {
6341 163 : if guard.contains(l.as_ref()) {
6342 : // expected in tests
6343 0 : tracing::error!(layer=%l, "duplicated L1 layer");
6344 :
6345 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
6346 : // `guard` on self.layers. as of writing this, there are no error returns except
6347 : // for compact_level0_phase1 creating an L0, which does not happen in practice
6348 : // because we have not implemented L0 => L0 compaction.
6349 0 : duplicated_layers.insert(l.layer_desc().key());
6350 163 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
6351 0 : return Err(CompactionError::Other(anyhow::anyhow!(
6352 0 : "compaction generates a L0 layer file as output, which will cause infinite compaction."
6353 0 : )));
6354 163 : } else {
6355 163 : insert_layers.push(l.clone());
6356 163 : }
6357 : }
6358 :
6359 : // only remove those inputs which were not outputs
6360 23 : let remove_layers: Vec<Layer> = layers_to_remove
6361 23 : .iter()
6362 201 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
6363 23 : .cloned()
6364 23 : .collect();
6365 :
6366 23 : if !new_images.is_empty() {
6367 0 : guard
6368 0 : .open_mut()?
6369 0 : .track_new_image_layers(new_images, &self.metrics);
6370 23 : }
6371 :
6372 23 : guard
6373 23 : .open_mut()?
6374 23 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
6375 :
6376 23 : self.remote_client
6377 23 : .schedule_compaction_update(&remove_layers, new_deltas)?;
6378 :
6379 23 : drop_layer_manager_wlock(guard);
6380 :
6381 23 : Ok(())
6382 23 : }
6383 :
6384 0 : async fn rewrite_layers(
6385 0 : self: &Arc<Self>,
6386 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
6387 0 : mut drop_layers: Vec<Layer>,
6388 0 : ) -> Result<(), CompactionError> {
6389 0 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
6390 :
6391 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
6392 : // to avoid double-removing, and avoid rewriting something that was removed.
6393 0 : replace_layers.retain(|(l, _)| guard.contains(l));
6394 0 : drop_layers.retain(|l| guard.contains(l));
6395 :
6396 0 : guard
6397 0 : .open_mut()?
6398 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
6399 :
6400 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
6401 :
6402 0 : self.remote_client
6403 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
6404 :
6405 0 : Ok(())
6406 0 : }
6407 :
6408 : /// Schedules the uploads of the given image layers
6409 80 : fn upload_new_image_layers(
6410 80 : self: &Arc<Self>,
6411 80 : new_images: impl IntoIterator<Item = ResidentLayer>,
6412 80 : ) -> Result<(), super::upload_queue::NotInitialized> {
6413 93 : for layer in new_images {
6414 13 : self.remote_client.schedule_layer_file_upload(layer)?;
6415 : }
6416 : // should any new image layer been created, not uploading index_part will
6417 : // result in a mismatch between remote_physical_size and layermap calculated
6418 : // size, which will fail some tests, but should not be an issue otherwise.
6419 80 : self.remote_client
6420 80 : .schedule_index_upload_for_file_changes()?;
6421 80 : Ok(())
6422 80 : }
6423 :
6424 0 : async fn find_gc_time_cutoff(
6425 0 : &self,
6426 0 : now: SystemTime,
6427 0 : pitr: Duration,
6428 0 : cancel: &CancellationToken,
6429 0 : ctx: &RequestContext,
6430 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
6431 0 : debug_assert_current_span_has_tenant_and_timeline_id();
6432 0 : if self.shard_identity.is_shard_zero() {
6433 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
6434 0 : let time_range = if pitr == Duration::ZERO {
6435 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
6436 : } else {
6437 0 : pitr
6438 : };
6439 :
6440 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
6441 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
6442 0 : let timestamp = to_pg_timestamp(time_cutoff);
6443 :
6444 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
6445 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
6446 0 : LsnForTimestamp::Future(lsn) => {
6447 : // The timestamp is in the future. That sounds impossible,
6448 : // but what it really means is that there hasn't been
6449 : // any commits since the cutoff timestamp.
6450 : //
6451 : // In this case we should use the LSN of the most recent commit,
6452 : // which is implicitly the last LSN in the log.
6453 0 : debug!("future({})", lsn);
6454 0 : Some(self.get_last_record_lsn())
6455 : }
6456 0 : LsnForTimestamp::Past(lsn) => {
6457 0 : debug!("past({})", lsn);
6458 0 : None
6459 : }
6460 0 : LsnForTimestamp::NoData(lsn) => {
6461 0 : debug!("nodata({})", lsn);
6462 0 : None
6463 : }
6464 : };
6465 0 : Ok(time_cutoff)
6466 : } else {
6467 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
6468 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
6469 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
6470 : // space cutoff that we would also have respected ourselves.
6471 0 : match self
6472 0 : .remote_client
6473 0 : .download_foreign_index(ShardNumber(0), cancel)
6474 0 : .await
6475 : {
6476 0 : Ok((index_part, index_generation, _index_mtime)) => {
6477 0 : tracing::info!(
6478 0 : "GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
6479 0 : index_part.metadata.latest_gc_cutoff_lsn()
6480 : );
6481 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
6482 : }
6483 : Err(DownloadError::NotFound) => {
6484 : // This is unexpected, because during timeline creations shard zero persists to remote
6485 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
6486 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
6487 : // state, then shard zero should _eventually_ write an index when it GCs.
6488 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
6489 0 : Ok(None)
6490 : }
6491 0 : Err(e) => {
6492 : // TODO: this function should return a different error type than page reconstruct error
6493 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
6494 : }
6495 : }
6496 :
6497 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
6498 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
6499 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
6500 : // new location. This is legal in principle, but problematic in practice because it might result
6501 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
6502 : // because they have GC'd past the branch point.
6503 : }
6504 0 : }
6505 :
6506 : /// Find the Lsns above which layer files need to be retained on
6507 : /// garbage collection.
6508 : ///
6509 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
6510 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
6511 : /// the space-based retention.
6512 : ///
6513 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
6514 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
6515 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
6516 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
6517 : /// in the response as the GC cutoff point for the timeline.
6518 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
6519 : pub(super) async fn find_gc_cutoffs(
6520 : &self,
6521 : now: SystemTime,
6522 : space_cutoff: Lsn,
6523 : pitr: Duration,
6524 : cancel: &CancellationToken,
6525 : ctx: &RequestContext,
6526 : ) -> Result<GcCutoffs, PageReconstructError> {
6527 : let _timer = self
6528 : .metrics
6529 : .find_gc_cutoffs_histo
6530 : .start_timer()
6531 : .record_on_drop();
6532 :
6533 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
6534 :
6535 : if cfg!(test) && pitr == Duration::ZERO {
6536 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
6537 : return Ok(GcCutoffs {
6538 : time: Some(self.get_last_record_lsn()),
6539 : space: space_cutoff,
6540 : });
6541 : }
6542 :
6543 : // Calculate a time-based limit on how much to retain:
6544 : // - if PITR interval is set, then this is our cutoff.
6545 : // - if PITR interval is not set, then we do a lookup
6546 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
6547 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
6548 :
6549 : Ok(match (pitr, time_cutoff) {
6550 : (Duration::ZERO, Some(time_cutoff)) => {
6551 : // PITR is not set. Retain the size-based limit, or the default time retention,
6552 : // whichever requires less data.
6553 : GcCutoffs {
6554 : time: Some(self.get_last_record_lsn()),
6555 : space: std::cmp::max(time_cutoff, space_cutoff),
6556 : }
6557 : }
6558 : (Duration::ZERO, None) => {
6559 : // PITR is not set, and time lookup failed
6560 : GcCutoffs {
6561 : time: Some(self.get_last_record_lsn()),
6562 : space: space_cutoff,
6563 : }
6564 : }
6565 : (_, None) => {
6566 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
6567 : // cannot advance beyond what was already GC'd, and respect space-based retention
6568 : GcCutoffs {
6569 : time: Some(*self.get_applied_gc_cutoff_lsn()),
6570 : space: space_cutoff,
6571 : }
6572 : }
6573 : (_, Some(time_cutoff)) => {
6574 : // PITR interval is set and we looked up timestamp successfully. Ignore
6575 : // size based retention and make time cutoff authoritative
6576 : GcCutoffs {
6577 : time: Some(time_cutoff),
6578 : space: time_cutoff,
6579 : }
6580 : }
6581 : })
6582 : }
6583 :
6584 : /// Garbage collect layer files on a timeline that are no longer needed.
6585 : ///
6586 : /// Currently, we don't make any attempt at removing unneeded page versions
6587 : /// within a layer file. We can only remove the whole file if it's fully
6588 : /// obsolete.
6589 377 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
6590 : // this is most likely the background tasks, but it might be the spawned task from
6591 : // immediate_gc
6592 377 : let _g = tokio::select! {
6593 377 : guard = self.gc_lock.lock() => guard,
6594 377 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
6595 : };
6596 377 : let timer = self.metrics.garbage_collect_histo.start_timer();
6597 :
6598 377 : fail_point!("before-timeline-gc");
6599 :
6600 : // Is the timeline being deleted?
6601 377 : if self.is_stopping() {
6602 0 : return Err(GcError::TimelineCancelled);
6603 377 : }
6604 :
6605 377 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
6606 377 : let gc_info = self.gc_info.read().unwrap();
6607 :
6608 377 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
6609 377 : let time_cutoff = gc_info.cutoffs.time;
6610 377 : let retain_lsns = gc_info
6611 377 : .retain_lsns
6612 377 : .iter()
6613 377 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
6614 377 : .collect();
6615 :
6616 : // Gets the maximum LSN that holds the valid lease.
6617 : //
6618 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
6619 : // Here, we do not check for stale leases again.
6620 377 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
6621 :
6622 377 : (
6623 377 : space_cutoff,
6624 377 : time_cutoff,
6625 377 : retain_lsns,
6626 377 : max_lsn_with_valid_lease,
6627 377 : )
6628 : };
6629 :
6630 377 : let mut new_gc_cutoff = space_cutoff.min(time_cutoff.unwrap_or_default());
6631 377 : let standby_horizon = self.standby_horizon.load();
6632 : // Hold GC for the standby, but as a safety guard do it only within some
6633 : // reasonable lag.
6634 377 : if standby_horizon != Lsn::INVALID {
6635 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
6636 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
6637 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
6638 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
6639 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
6640 : } else {
6641 0 : warn!(
6642 0 : "standby is lagging for more than {}MB, not holding gc for it",
6643 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
6644 : )
6645 : }
6646 0 : }
6647 377 : }
6648 :
6649 : // Reset standby horizon to ignore it if it is not updated till next GC.
6650 : // It is an easy way to unset it when standby disappears without adding
6651 : // more conf options.
6652 377 : self.standby_horizon.store(Lsn::INVALID);
6653 377 : self.metrics
6654 377 : .standby_horizon_gauge
6655 377 : .set(Lsn::INVALID.0 as i64);
6656 :
6657 377 : let res = self
6658 377 : .gc_timeline(
6659 377 : space_cutoff,
6660 377 : time_cutoff,
6661 377 : retain_lsns,
6662 377 : max_lsn_with_valid_lease,
6663 377 : new_gc_cutoff,
6664 : )
6665 377 : .instrument(
6666 377 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
6667 : )
6668 377 : .await?;
6669 :
6670 : // only record successes
6671 377 : timer.stop_and_record();
6672 :
6673 377 : Ok(res)
6674 377 : }
6675 :
6676 377 : async fn gc_timeline(
6677 377 : &self,
6678 377 : space_cutoff: Lsn,
6679 377 : time_cutoff: Option<Lsn>, // None if uninitialized
6680 377 : retain_lsns: Vec<Lsn>,
6681 377 : max_lsn_with_valid_lease: Option<Lsn>,
6682 377 : new_gc_cutoff: Lsn,
6683 377 : ) -> Result<GcResult, GcError> {
6684 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
6685 :
6686 377 : let now = SystemTime::now();
6687 377 : let mut result: GcResult = GcResult::default();
6688 :
6689 : // Nothing to GC. Return early.
6690 377 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
6691 377 : if latest_gc_cutoff >= new_gc_cutoff {
6692 11 : info!(
6693 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
6694 : );
6695 11 : return Ok(result);
6696 366 : }
6697 :
6698 366 : let Some(time_cutoff) = time_cutoff else {
6699 : // The GC cutoff should have been computed by now, but let's be defensive.
6700 0 : info!("Nothing to GC: time_cutoff not yet computed");
6701 0 : return Ok(result);
6702 : };
6703 :
6704 : // We need to ensure that no one tries to read page versions or create
6705 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
6706 : // for details. This will block until the old value is no longer in use.
6707 : //
6708 : // The GC cutoff should only ever move forwards.
6709 366 : let waitlist = {
6710 366 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
6711 366 : if *write_guard > new_gc_cutoff {
6712 0 : return Err(GcError::BadLsn {
6713 0 : why: format!(
6714 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
6715 0 : *write_guard, new_gc_cutoff
6716 0 : ),
6717 0 : });
6718 366 : }
6719 :
6720 366 : write_guard.store_and_unlock(new_gc_cutoff)
6721 : };
6722 366 : waitlist.wait().await;
6723 :
6724 366 : info!("GC starting");
6725 :
6726 366 : debug!("retain_lsns: {:?}", retain_lsns);
6727 :
6728 366 : let max_retain_lsn = retain_lsns.iter().max();
6729 :
6730 : // Scan all layers in the timeline (remote or on-disk).
6731 : //
6732 : // Garbage collect the layer if all conditions are satisfied:
6733 : // 1. it is older than cutoff LSN;
6734 : // 2. it is older than PITR interval;
6735 : // 3. it doesn't need to be retained for 'retain_lsns';
6736 : // 4. it does not need to be kept for LSNs holding valid leases.
6737 : // 5. newer on-disk image layers cover the layer's whole key range
6738 366 : let layers_to_remove = {
6739 366 : let mut layers_to_remove = Vec::new();
6740 :
6741 366 : let guard = self
6742 366 : .layers
6743 366 : .read(LayerManagerLockHolder::GarbageCollection)
6744 366 : .await;
6745 366 : let layers = guard.layer_map()?;
6746 6209 : 'outer: for l in layers.iter_historic_layers() {
6747 6209 : result.layers_total += 1;
6748 :
6749 : // 1. Is it newer than GC horizon cutoff point?
6750 6209 : if l.get_lsn_range().end > space_cutoff {
6751 371 : debug!(
6752 0 : "keeping {} because it's newer than space_cutoff {}",
6753 0 : l.layer_name(),
6754 : space_cutoff,
6755 : );
6756 371 : result.layers_needed_by_cutoff += 1;
6757 371 : continue 'outer;
6758 5838 : }
6759 :
6760 : // 2. It is newer than PiTR cutoff point?
6761 5838 : if l.get_lsn_range().end > time_cutoff {
6762 0 : debug!(
6763 0 : "keeping {} because it's newer than time_cutoff {}",
6764 0 : l.layer_name(),
6765 : time_cutoff,
6766 : );
6767 0 : result.layers_needed_by_pitr += 1;
6768 0 : continue 'outer;
6769 5838 : }
6770 :
6771 : // 3. Is it needed by a child branch?
6772 : // NOTE With that we would keep data that
6773 : // might be referenced by child branches forever.
6774 : // We can track this in child timeline GC and delete parent layers when
6775 : // they are no longer needed. This might be complicated with long inheritance chains.
6776 5838 : if let Some(retain_lsn) = max_retain_lsn {
6777 : // start_lsn is inclusive
6778 6 : if &l.get_lsn_range().start <= retain_lsn {
6779 6 : debug!(
6780 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
6781 0 : l.layer_name(),
6782 : retain_lsn,
6783 0 : l.is_incremental(),
6784 : );
6785 6 : result.layers_needed_by_branches += 1;
6786 6 : continue 'outer;
6787 0 : }
6788 5832 : }
6789 :
6790 : // 4. Is there a valid lease that requires us to keep this layer?
6791 5832 : if let Some(lsn) = &max_lsn_with_valid_lease {
6792 : // keep if layer start <= any of the lease
6793 9 : if &l.get_lsn_range().start <= lsn {
6794 7 : debug!(
6795 0 : "keeping {} because there is a valid lease preventing GC at {}",
6796 0 : l.layer_name(),
6797 : lsn,
6798 : );
6799 7 : result.layers_needed_by_leases += 1;
6800 7 : continue 'outer;
6801 2 : }
6802 5823 : }
6803 :
6804 : // 5. Is there a later on-disk layer for this relation?
6805 : //
6806 : // The end-LSN is exclusive, while disk_consistent_lsn is
6807 : // inclusive. For example, if disk_consistent_lsn is 100, it is
6808 : // OK for a delta layer to have end LSN 101, but if the end LSN
6809 : // is 102, then it might not have been fully flushed to disk
6810 : // before crash.
6811 : //
6812 : // For example, imagine that the following layers exist:
6813 : //
6814 : // 1000 - image (A)
6815 : // 1000-2000 - delta (B)
6816 : // 2000 - image (C)
6817 : // 2000-3000 - delta (D)
6818 : // 3000 - image (E)
6819 : //
6820 : // If GC horizon is at 2500, we can remove layers A and B, but
6821 : // we cannot remove C, even though it's older than 2500, because
6822 : // the delta layer 2000-3000 depends on it.
6823 5825 : if !layers
6824 5825 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
6825 : {
6826 5821 : debug!("keeping {} because it is the latest layer", l.layer_name());
6827 5821 : result.layers_not_updated += 1;
6828 5821 : continue 'outer;
6829 4 : }
6830 :
6831 : // We didn't find any reason to keep this file, so remove it.
6832 4 : info!(
6833 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
6834 0 : l.layer_name(),
6835 0 : l.is_incremental(),
6836 : );
6837 4 : layers_to_remove.push(l);
6838 : }
6839 :
6840 366 : layers_to_remove
6841 : };
6842 :
6843 366 : if !layers_to_remove.is_empty() {
6844 : // Persist the new GC cutoff value before we actually remove anything.
6845 : // This unconditionally schedules also an index_part.json update, even though, we will
6846 : // be doing one a bit later with the unlinked gc'd layers.
6847 3 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
6848 3 : self.schedule_uploads(disk_consistent_lsn, None)
6849 3 : .map_err(|e| {
6850 0 : if self.cancel.is_cancelled() {
6851 0 : GcError::TimelineCancelled
6852 : } else {
6853 0 : GcError::Remote(e)
6854 : }
6855 0 : })?;
6856 :
6857 3 : let mut guard = self
6858 3 : .layers
6859 3 : .write(LayerManagerLockHolder::GarbageCollection)
6860 3 : .await;
6861 :
6862 3 : let gc_layers = layers_to_remove
6863 3 : .iter()
6864 4 : .flat_map(|desc| guard.try_get_from_key(&desc.key()).cloned())
6865 3 : .collect::<Vec<Layer>>();
6866 :
6867 3 : result.layers_removed = gc_layers.len() as u64;
6868 :
6869 3 : self.remote_client.schedule_gc_update(&gc_layers)?;
6870 3 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
6871 :
6872 : #[cfg(feature = "testing")]
6873 3 : {
6874 3 : result.doomed_layers = gc_layers;
6875 3 : }
6876 363 : }
6877 :
6878 366 : info!(
6879 0 : "GC completed removing {} layers, cutoff {}",
6880 : result.layers_removed, new_gc_cutoff
6881 : );
6882 :
6883 366 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
6884 366 : Ok(result)
6885 377 : }
6886 :
6887 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
6888 364910 : pub(crate) async fn reconstruct_value(
6889 364910 : &self,
6890 364910 : key: Key,
6891 364910 : request_lsn: Lsn,
6892 364910 : mut data: ValueReconstructState,
6893 364910 : redo_attempt_type: RedoAttemptType,
6894 364910 : ) -> Result<Bytes, PageReconstructError> {
6895 : // Perform WAL redo if needed
6896 364910 : data.records.reverse();
6897 :
6898 364910 : let fire_critical_error = match redo_attempt_type {
6899 363577 : RedoAttemptType::ReadPage => true,
6900 0 : RedoAttemptType::LegacyCompaction => true,
6901 1333 : RedoAttemptType::GcCompaction => false,
6902 : };
6903 :
6904 : // If we have a page image, and no WAL, we're all set
6905 364910 : if data.records.is_empty() {
6906 338136 : if let Some((img_lsn, img)) = &data.img {
6907 338136 : trace!(
6908 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
6909 : key, img_lsn, request_lsn,
6910 : );
6911 338136 : Ok(img.clone())
6912 : } else {
6913 0 : Err(PageReconstructError::from(anyhow!(
6914 0 : "base image for {key} at {request_lsn} not found"
6915 0 : )))
6916 : }
6917 : } else {
6918 : // We need to do WAL redo.
6919 : //
6920 : // If we don't have a base image, then the oldest WAL record better initialize
6921 : // the page
6922 26774 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
6923 0 : Err(PageReconstructError::from(anyhow!(
6924 0 : "Base image for {} at {} not found, but got {} WAL records",
6925 0 : key,
6926 0 : request_lsn,
6927 0 : data.records.len()
6928 0 : )))
6929 : } else {
6930 26774 : if data.img.is_some() {
6931 13029 : trace!(
6932 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
6933 0 : data.records.len(),
6934 : key,
6935 : request_lsn
6936 : );
6937 : } else {
6938 13745 : trace!(
6939 0 : "found {} WAL records that will init the page for {} at {}, performing WAL redo",
6940 0 : data.records.len(),
6941 : key,
6942 : request_lsn
6943 : );
6944 : };
6945 26774 : let res = self
6946 26774 : .walredo_mgr
6947 26774 : .as_ref()
6948 26774 : .context("timeline has no walredo manager")
6949 26774 : .map_err(PageReconstructError::WalRedo)?
6950 26774 : .request_redo(
6951 26774 : key,
6952 26774 : request_lsn,
6953 26774 : data.img,
6954 26774 : data.records,
6955 26774 : self.pg_version,
6956 26774 : redo_attempt_type,
6957 : )
6958 26774 : .await;
6959 26773 : let img = match res {
6960 26773 : Ok(img) => img,
6961 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
6962 1 : Err(walredo::Error::Other(err)) => {
6963 1 : if fire_critical_error {
6964 0 : critical_timeline!(
6965 0 : self.tenant_shard_id,
6966 0 : self.timeline_id,
6967 0 : "walredo failure during page reconstruction: {err:?}"
6968 : );
6969 1 : }
6970 1 : return Err(PageReconstructError::WalRedo(
6971 1 : err.context("reconstruct a page image"),
6972 1 : ));
6973 : }
6974 : };
6975 26773 : Ok(img)
6976 : }
6977 : }
6978 364910 : }
6979 :
6980 0 : pub(crate) async fn spawn_download_all_remote_layers(
6981 0 : self: Arc<Self>,
6982 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6983 0 : ctx: &RequestContext,
6984 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
6985 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6986 :
6987 : // this is not really needed anymore; it has tests which really check the return value from
6988 : // http api. it would be better not to maintain this anymore.
6989 :
6990 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
6991 0 : if let Some(st) = &*status_guard {
6992 0 : match &st.state {
6993 : DownloadRemoteLayersTaskState::Running => {
6994 0 : return Err(st.clone());
6995 : }
6996 : DownloadRemoteLayersTaskState::ShutDown
6997 0 : | DownloadRemoteLayersTaskState::Completed => {
6998 0 : *status_guard = None;
6999 0 : }
7000 : }
7001 0 : }
7002 :
7003 0 : let self_clone = Arc::clone(&self);
7004 0 : let task_ctx = ctx.detached_child(
7005 0 : TaskKind::DownloadAllRemoteLayers,
7006 0 : DownloadBehavior::Download,
7007 : );
7008 0 : let task_id = task_mgr::spawn(
7009 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
7010 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
7011 0 : self.tenant_shard_id,
7012 0 : Some(self.timeline_id),
7013 0 : "download all remote layers task",
7014 0 : async move {
7015 0 : self_clone.download_all_remote_layers(request, &task_ctx).await;
7016 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
7017 0 : match &mut *status_guard {
7018 : None => {
7019 0 : warn!("tasks status is supposed to be Some(), since we are running");
7020 : }
7021 0 : Some(st) => {
7022 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
7023 0 : if st.task_id != exp_task_id {
7024 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
7025 0 : } else {
7026 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7027 0 : }
7028 : }
7029 : };
7030 0 : Ok(())
7031 0 : }
7032 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))
7033 : );
7034 :
7035 0 : let initial_info = DownloadRemoteLayersTaskInfo {
7036 0 : task_id: format!("{task_id}"),
7037 0 : state: DownloadRemoteLayersTaskState::Running,
7038 0 : total_layer_count: 0,
7039 0 : successful_download_count: 0,
7040 0 : failed_download_count: 0,
7041 0 : };
7042 0 : *status_guard = Some(initial_info.clone());
7043 :
7044 0 : Ok(initial_info)
7045 0 : }
7046 :
7047 0 : async fn download_all_remote_layers(
7048 0 : self: &Arc<Self>,
7049 0 : request: DownloadRemoteLayersTaskSpawnRequest,
7050 0 : ctx: &RequestContext,
7051 0 : ) {
7052 : use pageserver_api::models::DownloadRemoteLayersTaskState;
7053 :
7054 0 : let remaining = {
7055 0 : let guard = self
7056 0 : .layers
7057 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7058 0 : .await;
7059 0 : let Ok(lm) = guard.layer_map() else {
7060 : // technically here we could look into iterating accessible layers, but downloading
7061 : // all layers of a shutdown timeline makes no sense regardless.
7062 0 : tracing::info!("attempted to download all layers of shutdown timeline");
7063 0 : return;
7064 : };
7065 0 : lm.iter_historic_layers()
7066 0 : .map(|desc| guard.get_from_desc(&desc))
7067 0 : .collect::<Vec<_>>()
7068 : };
7069 0 : let total_layer_count = remaining.len();
7070 :
7071 : macro_rules! lock_status {
7072 : ($st:ident) => {
7073 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
7074 : let st = st
7075 : .as_mut()
7076 : .expect("this function is only called after the task has been spawned");
7077 : assert_eq!(
7078 : st.task_id,
7079 : format!(
7080 : "{}",
7081 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
7082 : )
7083 : );
7084 : let $st = st;
7085 : };
7086 : }
7087 :
7088 : {
7089 0 : lock_status!(st);
7090 0 : st.total_layer_count = total_layer_count as u64;
7091 : }
7092 :
7093 0 : let mut remaining = remaining.into_iter();
7094 0 : let mut have_remaining = true;
7095 0 : let mut js = tokio::task::JoinSet::new();
7096 :
7097 0 : let cancel = task_mgr::shutdown_token();
7098 :
7099 0 : let limit = request.max_concurrent_downloads;
7100 :
7101 : loop {
7102 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
7103 0 : let Some(next) = remaining.next() else {
7104 0 : have_remaining = false;
7105 0 : break;
7106 : };
7107 :
7108 0 : let span = tracing::info_span!("download", layer = %next);
7109 :
7110 0 : let ctx = ctx.attached_child();
7111 0 : js.spawn(
7112 0 : async move {
7113 0 : let res = next.download(&ctx).await;
7114 0 : (next, res)
7115 0 : }
7116 0 : .instrument(span),
7117 : );
7118 : }
7119 :
7120 0 : while let Some(res) = js.join_next().await {
7121 0 : match res {
7122 : Ok((_, Ok(_))) => {
7123 0 : lock_status!(st);
7124 0 : st.successful_download_count += 1;
7125 : }
7126 0 : Ok((layer, Err(e))) => {
7127 0 : tracing::error!(%layer, "download failed: {e:#}");
7128 0 : lock_status!(st);
7129 0 : st.failed_download_count += 1;
7130 : }
7131 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
7132 0 : Err(je) if je.is_panic() => {
7133 0 : lock_status!(st);
7134 0 : st.failed_download_count += 1;
7135 : }
7136 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
7137 : }
7138 : }
7139 :
7140 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
7141 0 : break;
7142 0 : }
7143 : }
7144 :
7145 : {
7146 0 : lock_status!(st);
7147 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7148 : }
7149 0 : }
7150 :
7151 0 : pub(crate) fn get_download_all_remote_layers_task_info(
7152 0 : &self,
7153 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
7154 0 : self.download_all_remote_layers_task_info
7155 0 : .read()
7156 0 : .unwrap()
7157 0 : .clone()
7158 0 : }
7159 :
7160 : /* BEGIN_HADRON */
7161 0 : pub(crate) async fn compute_image_consistent_lsn(&self) -> anyhow::Result<Lsn> {
7162 0 : let guard = self
7163 0 : .layers
7164 0 : .read(LayerManagerLockHolder::ComputeImageConsistentLsn)
7165 0 : .await;
7166 0 : let layer_map = guard.layer_map()?;
7167 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
7168 :
7169 0 : Ok(layer_map.compute_image_consistent_lsn(disk_consistent_lsn))
7170 0 : }
7171 : /* END_HADRON */
7172 : }
7173 :
7174 : impl Timeline {
7175 : /// Returns non-remote layers for eviction.
7176 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
7177 0 : let guard = self.layers.read(LayerManagerLockHolder::Eviction).await;
7178 0 : let mut max_layer_size: Option<u64> = None;
7179 :
7180 0 : let resident_layers = guard
7181 0 : .likely_resident_layers()
7182 0 : .map(|layer| {
7183 0 : let file_size = layer.layer_desc().file_size;
7184 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
7185 :
7186 0 : let last_activity_ts = layer.latest_activity();
7187 :
7188 0 : EvictionCandidate {
7189 0 : layer: layer.to_owned().into(),
7190 0 : last_activity_ts,
7191 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
7192 0 : visibility: layer.visibility(),
7193 0 : }
7194 0 : })
7195 0 : .collect();
7196 :
7197 0 : DiskUsageEvictionInfo {
7198 0 : max_layer_size,
7199 0 : resident_layers,
7200 0 : }
7201 0 : }
7202 :
7203 962 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
7204 962 : ShardIndex {
7205 962 : shard_number: self.tenant_shard_id.shard_number,
7206 962 : shard_count: self.tenant_shard_id.shard_count,
7207 962 : }
7208 962 : }
7209 :
7210 : /// Persistently blocks gc for `Manual` reason.
7211 : ///
7212 : /// Returns true if no such block existed before, false otherwise.
7213 0 : pub(crate) async fn block_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<bool> {
7214 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7215 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7216 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
7217 0 : }
7218 :
7219 : /// Persistently unblocks gc for `Manual` reason.
7220 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<()> {
7221 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7222 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7223 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
7224 0 : }
7225 :
7226 : #[cfg(test)]
7227 31 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
7228 31 : self.last_record_lsn.advance(new_lsn);
7229 31 : }
7230 :
7231 : #[cfg(test)]
7232 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
7233 2 : self.disk_consistent_lsn.store(new_value);
7234 2 : }
7235 :
7236 : /// Force create an image layer and place it into the layer map.
7237 : ///
7238 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7239 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7240 : /// placed into the layer map in one run AND be validated.
7241 : #[cfg(test)]
7242 36 : pub(super) async fn force_create_image_layer(
7243 36 : self: &Arc<Timeline>,
7244 36 : lsn: Lsn,
7245 36 : mut images: Vec<(Key, Bytes)>,
7246 36 : check_start_lsn: Option<Lsn>,
7247 36 : ctx: &RequestContext,
7248 36 : ) -> anyhow::Result<()> {
7249 36 : let last_record_lsn = self.get_last_record_lsn();
7250 36 : assert!(
7251 36 : lsn <= last_record_lsn,
7252 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
7253 : );
7254 36 : if let Some(check_start_lsn) = check_start_lsn {
7255 36 : assert!(lsn >= check_start_lsn);
7256 0 : }
7257 240 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
7258 36 : let min_key = *images.first().map(|(k, _)| k).unwrap();
7259 36 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
7260 36 : let mut image_layer_writer = ImageLayerWriter::new(
7261 36 : self.conf,
7262 36 : self.timeline_id,
7263 36 : self.tenant_shard_id,
7264 36 : &(min_key..end_key),
7265 36 : lsn,
7266 36 : &self.gate,
7267 36 : self.cancel.clone(),
7268 36 : ctx,
7269 36 : )
7270 36 : .await?;
7271 312 : for (key, img) in images {
7272 276 : image_layer_writer.put_image(key, img, ctx).await?;
7273 : }
7274 36 : let (desc, path) = image_layer_writer.finish(ctx).await?;
7275 36 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7276 36 : info!("force created image layer {}", image_layer.local_path());
7277 : {
7278 36 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7279 36 : guard
7280 36 : .open_mut()
7281 36 : .unwrap()
7282 36 : .force_insert_layer(image_layer.clone());
7283 : }
7284 :
7285 : // Update remote_timeline_client state to reflect existence of this layer
7286 36 : self.remote_client
7287 36 : .schedule_layer_file_upload(image_layer)
7288 36 : .unwrap();
7289 :
7290 36 : Ok(())
7291 36 : }
7292 :
7293 : /// Force create a delta layer and place it into the layer map.
7294 : ///
7295 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7296 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7297 : /// placed into the layer map in one run AND be validated.
7298 : #[cfg(test)]
7299 50 : pub(super) async fn force_create_delta_layer(
7300 50 : self: &Arc<Timeline>,
7301 50 : mut deltas: DeltaLayerTestDesc,
7302 50 : check_start_lsn: Option<Lsn>,
7303 50 : ctx: &RequestContext,
7304 50 : ) -> anyhow::Result<()> {
7305 50 : let last_record_lsn = self.get_last_record_lsn();
7306 50 : deltas
7307 50 : .data
7308 124364 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7309 50 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
7310 50 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
7311 10464 : for (_, lsn, _) in &deltas.data {
7312 10414 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
7313 : }
7314 50 : assert!(
7315 50 : deltas.lsn_range.end <= last_record_lsn,
7316 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
7317 : deltas.lsn_range.end,
7318 : last_record_lsn
7319 : );
7320 50 : if let Some(check_start_lsn) = check_start_lsn {
7321 50 : assert!(deltas.lsn_range.start >= check_start_lsn);
7322 0 : }
7323 50 : let mut delta_layer_writer = DeltaLayerWriter::new(
7324 50 : self.conf,
7325 50 : self.timeline_id,
7326 50 : self.tenant_shard_id,
7327 50 : deltas.key_range.start,
7328 50 : deltas.lsn_range,
7329 50 : &self.gate,
7330 50 : self.cancel.clone(),
7331 50 : ctx,
7332 50 : )
7333 50 : .await?;
7334 10464 : for (key, lsn, val) in deltas.data {
7335 10414 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
7336 : }
7337 50 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
7338 50 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7339 50 : info!("force created delta layer {}", delta_layer.local_path());
7340 : {
7341 50 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7342 50 : guard
7343 50 : .open_mut()
7344 50 : .unwrap()
7345 50 : .force_insert_layer(delta_layer.clone());
7346 : }
7347 :
7348 : // Update remote_timeline_client state to reflect existence of this layer
7349 50 : self.remote_client
7350 50 : .schedule_layer_file_upload(delta_layer)
7351 50 : .unwrap();
7352 :
7353 50 : Ok(())
7354 50 : }
7355 :
7356 : /// Force create an in-memory layer and place them into the layer map.
7357 : #[cfg(test)]
7358 4 : pub(super) async fn force_create_in_memory_layer(
7359 4 : self: &Arc<Timeline>,
7360 4 : mut in_memory: InMemoryLayerTestDesc,
7361 4 : check_start_lsn: Option<Lsn>,
7362 4 : ctx: &RequestContext,
7363 4 : ) -> anyhow::Result<()> {
7364 : use utils::bin_ser::BeSer;
7365 :
7366 : // Validate LSNs
7367 4 : if let Some(check_start_lsn) = check_start_lsn {
7368 4 : assert!(in_memory.lsn_range.start >= check_start_lsn);
7369 0 : }
7370 :
7371 4 : let last_record_lsn = self.get_last_record_lsn();
7372 4 : let layer_end_lsn = if in_memory.is_open {
7373 1 : in_memory
7374 1 : .data
7375 1 : .iter()
7376 1 : .map(|(_key, lsn, _value)| lsn)
7377 1 : .max()
7378 1 : .cloned()
7379 : } else {
7380 3 : Some(in_memory.lsn_range.end)
7381 : };
7382 :
7383 4 : if let Some(end) = layer_end_lsn {
7384 4 : assert!(
7385 4 : end <= last_record_lsn,
7386 0 : "advance last record lsn before inserting a layer, end_lsn={end}, last_record_lsn={last_record_lsn}",
7387 : );
7388 0 : }
7389 :
7390 19820 : in_memory.data.iter().for_each(|(_key, lsn, _value)| {
7391 19820 : assert!(*lsn >= in_memory.lsn_range.start);
7392 19820 : assert!(*lsn < in_memory.lsn_range.end);
7393 19820 : });
7394 :
7395 : // Build the batch
7396 4 : in_memory
7397 4 : .data
7398 273384 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7399 :
7400 4 : let data = in_memory
7401 4 : .data
7402 4 : .into_iter()
7403 19820 : .map(|(key, lsn, value)| {
7404 19820 : let value_size = value.serialized_size().unwrap() as usize;
7405 19820 : (key.to_compact(), lsn, value_size, value)
7406 19820 : })
7407 4 : .collect::<Vec<_>>();
7408 :
7409 4 : let batch = SerializedValueBatch::from_values(data);
7410 :
7411 : // Create the in-memory layer and write the batch into it
7412 4 : let layer = InMemoryLayer::create(
7413 4 : self.conf,
7414 4 : self.timeline_id,
7415 4 : self.tenant_shard_id,
7416 4 : in_memory.lsn_range.start,
7417 4 : &self.gate,
7418 4 : // TODO: if we ever use this function in production code, we need to pass the real cancellation token
7419 4 : &CancellationToken::new(),
7420 4 : ctx,
7421 4 : )
7422 4 : .await
7423 4 : .unwrap();
7424 :
7425 4 : layer.put_batch(batch, ctx).await.unwrap();
7426 4 : if !in_memory.is_open {
7427 3 : layer.freeze(in_memory.lsn_range.end).await;
7428 1 : }
7429 :
7430 4 : info!("force created in-memory layer {:?}", in_memory.lsn_range);
7431 :
7432 : // Link the layer to the layer map
7433 : {
7434 4 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7435 4 : let layer_map = guard.open_mut().unwrap();
7436 4 : layer_map.force_insert_in_memory_layer(Arc::new(layer));
7437 : }
7438 :
7439 4 : Ok(())
7440 4 : }
7441 :
7442 : /// Return all keys at the LSN in the image layers
7443 : #[cfg(test)]
7444 3 : pub(crate) async fn inspect_image_layers(
7445 3 : self: &Arc<Timeline>,
7446 3 : lsn: Lsn,
7447 3 : ctx: &RequestContext,
7448 3 : io_concurrency: IoConcurrency,
7449 3 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
7450 3 : let mut all_data = Vec::new();
7451 3 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7452 17 : for layer in guard.layer_map()?.iter_historic_layers() {
7453 17 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
7454 4 : let layer = guard.get_from_desc(&layer);
7455 4 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
7456 4 : layer
7457 4 : .get_values_reconstruct_data(
7458 4 : KeySpace::single(Key::MIN..Key::MAX),
7459 4 : lsn..Lsn(lsn.0 + 1),
7460 4 : &mut reconstruct_data,
7461 4 : ctx,
7462 4 : )
7463 4 : .await?;
7464 33 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
7465 33 : let v = v.collect_pending_ios().await?;
7466 33 : all_data.push((k, v.img.unwrap().1));
7467 : }
7468 13 : }
7469 : }
7470 3 : all_data.sort();
7471 3 : Ok(all_data)
7472 3 : }
7473 :
7474 : /// Get all historic layer descriptors in the layer map
7475 : #[cfg(test)]
7476 12 : pub(crate) async fn inspect_historic_layers(
7477 12 : self: &Arc<Timeline>,
7478 12 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
7479 12 : let mut layers = Vec::new();
7480 12 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7481 57 : for layer in guard.layer_map()?.iter_historic_layers() {
7482 57 : layers.push(layer.key());
7483 57 : }
7484 12 : Ok(layers)
7485 12 : }
7486 :
7487 : #[cfg(test)]
7488 5 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
7489 5 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
7490 5 : keyspace.merge(&ks);
7491 5 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
7492 5 : }
7493 : }
7494 :
7495 : /// Tracking writes ingestion does to a particular in-memory layer.
7496 : ///
7497 : /// Cleared upon freezing a layer.
7498 : pub(crate) struct TimelineWriterState {
7499 : open_layer: Arc<InMemoryLayer>,
7500 : current_size: u64,
7501 : // Previous Lsn which passed through
7502 : prev_lsn: Option<Lsn>,
7503 : // Largest Lsn which passed through the current writer
7504 : max_lsn: Option<Lsn>,
7505 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
7506 : cached_last_freeze_at: Lsn,
7507 : }
7508 :
7509 : impl TimelineWriterState {
7510 660 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
7511 660 : Self {
7512 660 : open_layer,
7513 660 : current_size,
7514 660 : prev_lsn: None,
7515 660 : max_lsn: None,
7516 660 : cached_last_freeze_at: last_freeze_at,
7517 660 : }
7518 660 : }
7519 : }
7520 :
7521 : /// Various functions to mutate the timeline.
7522 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
7523 : // This is probably considered a bad practice in Rust and should be fixed eventually,
7524 : // but will cause large code changes.
7525 : pub(crate) struct TimelineWriter<'a> {
7526 : tl: &'a Timeline,
7527 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
7528 : }
7529 :
7530 : impl Deref for TimelineWriter<'_> {
7531 : type Target = Timeline;
7532 :
7533 4949226 : fn deref(&self) -> &Self::Target {
7534 4949226 : self.tl
7535 4949226 : }
7536 : }
7537 :
7538 : #[derive(PartialEq)]
7539 : enum OpenLayerAction {
7540 : Roll,
7541 : Open,
7542 : None,
7543 : }
7544 :
7545 : impl TimelineWriter<'_> {
7546 2402129 : async fn handle_open_layer_action(
7547 2402129 : &mut self,
7548 2402129 : at: Lsn,
7549 2402129 : action: OpenLayerAction,
7550 2402129 : ctx: &RequestContext,
7551 2402129 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
7552 2402129 : match action {
7553 : OpenLayerAction::Roll => {
7554 40 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
7555 40 : self.roll_layer(freeze_at).await?;
7556 40 : self.open_layer(at, ctx).await?;
7557 : }
7558 620 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
7559 : OpenLayerAction::None => {
7560 2401469 : assert!(self.write_guard.is_some());
7561 : }
7562 : }
7563 :
7564 2402129 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
7565 2402129 : }
7566 :
7567 660 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
7568 660 : let layer = self
7569 660 : .tl
7570 660 : .get_layer_for_write(at, &self.write_guard, ctx)
7571 660 : .await?;
7572 660 : let initial_size = layer.len();
7573 :
7574 660 : let last_freeze_at = self.last_freeze_at.load();
7575 660 : self.write_guard.replace(TimelineWriterState::new(
7576 660 : layer,
7577 660 : initial_size,
7578 660 : last_freeze_at,
7579 660 : ));
7580 :
7581 660 : Ok(())
7582 660 : }
7583 :
7584 40 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
7585 40 : let current_size = self.write_guard.as_ref().unwrap().current_size;
7586 :
7587 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
7588 : // to propagate the backpressure up into WAL ingestion.
7589 40 : let l0_count = self
7590 40 : .tl
7591 40 : .layers
7592 40 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7593 40 : .await
7594 40 : .layer_map()?
7595 40 : .level0_deltas()
7596 40 : .len();
7597 40 : let wait_thresholds = [
7598 40 : self.get_l0_flush_delay_threshold(),
7599 40 : self.get_l0_flush_stall_threshold(),
7600 40 : ];
7601 40 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
7602 :
7603 : // self.write_guard will be taken by the freezing
7604 40 : let flush_id = self
7605 40 : .tl
7606 40 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
7607 40 : .await?;
7608 :
7609 40 : assert!(self.write_guard.is_none());
7610 :
7611 40 : if let Some(wait_threshold) = wait_threshold {
7612 0 : if l0_count >= wait_threshold {
7613 0 : debug!(
7614 0 : "layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers"
7615 : );
7616 0 : self.tl.wait_flush_completion(flush_id).await?;
7617 0 : }
7618 40 : }
7619 :
7620 40 : if current_size >= self.get_checkpoint_distance() * 2 {
7621 0 : warn!("Flushed oversized open layer with size {}", current_size)
7622 40 : }
7623 :
7624 40 : Ok(())
7625 40 : }
7626 :
7627 2402129 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
7628 2402129 : let state = &*self.write_guard;
7629 2402129 : let Some(state) = &state else {
7630 620 : return OpenLayerAction::Open;
7631 : };
7632 :
7633 : #[cfg(feature = "testing")]
7634 2401509 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
7635 : // this check and assertion are not really needed because
7636 : // LayerManager::try_freeze_in_memory_layer will always clear out the
7637 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
7638 : // is no TimelineWriterState.
7639 0 : assert!(
7640 0 : state.open_layer.end_lsn.get().is_some(),
7641 0 : "our open_layer must be outdated"
7642 : );
7643 :
7644 : // this would be a memory leak waiting to happen because the in-memory layer always has
7645 : // an index
7646 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
7647 2401509 : }
7648 :
7649 2401509 : if state.prev_lsn == Some(lsn) {
7650 : // Rolling mid LSN is not supported by [downstream code].
7651 : // Hence, only roll at LSN boundaries.
7652 : //
7653 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
7654 3 : return OpenLayerAction::None;
7655 2401506 : }
7656 :
7657 2401506 : if state.current_size == 0 {
7658 : // Don't roll empty layers
7659 0 : return OpenLayerAction::None;
7660 2401506 : }
7661 :
7662 2401506 : if self.tl.should_roll(
7663 2401506 : state.current_size,
7664 2401506 : state.current_size + new_value_size,
7665 2401506 : self.get_checkpoint_distance(),
7666 2401506 : lsn,
7667 2401506 : state.cached_last_freeze_at,
7668 2401506 : state.open_layer.get_opened_at(),
7669 : ) {
7670 40 : OpenLayerAction::Roll
7671 : } else {
7672 2401466 : OpenLayerAction::None
7673 : }
7674 2402129 : }
7675 :
7676 : /// Put a batch of keys at the specified Lsns.
7677 2402128 : pub(crate) async fn put_batch(
7678 2402128 : &mut self,
7679 2402128 : batch: SerializedValueBatch,
7680 2402128 : ctx: &RequestContext,
7681 2402128 : ) -> anyhow::Result<()> {
7682 2402128 : if !batch.has_data() {
7683 0 : return Ok(());
7684 2402128 : }
7685 :
7686 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
7687 : // We don't assert this in release builds, since key ownership policies may change over
7688 : // time. Stray keys will be removed during compaction.
7689 2402128 : if cfg!(debug_assertions) {
7690 4947529 : for metadata in &batch.metadata {
7691 2545401 : if let ValueMeta::Serialized(metadata) = metadata {
7692 2545401 : let key = Key::from_compact(metadata.key);
7693 2545401 : assert!(
7694 2545401 : self.shard_identity.is_key_local(&key)
7695 12 : || self.shard_identity.is_key_global(&key),
7696 0 : "key {key} does not belong on shard {}",
7697 0 : self.shard_identity.shard_index()
7698 : );
7699 0 : }
7700 : }
7701 0 : }
7702 :
7703 2402128 : let batch_max_lsn = batch.max_lsn;
7704 2402128 : let buf_size: u64 = batch.buffer_size() as u64;
7705 :
7706 2402128 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
7707 2402128 : let layer = self
7708 2402128 : .handle_open_layer_action(batch_max_lsn, action, ctx)
7709 2402128 : .await?;
7710 :
7711 2402128 : let res = layer.put_batch(batch, ctx).await;
7712 :
7713 2402128 : if res.is_ok() {
7714 2402128 : // Update the current size only when the entire write was ok.
7715 2402128 : // In case of failures, we may have had partial writes which
7716 2402128 : // render the size tracking out of sync. That's ok because
7717 2402128 : // the checkpoint distance should be significantly smaller
7718 2402128 : // than the S3 single shot upload limit of 5GiB.
7719 2402128 : let state = self.write_guard.as_mut().unwrap();
7720 2402128 :
7721 2402128 : state.current_size += buf_size;
7722 2402128 : state.prev_lsn = Some(batch_max_lsn);
7723 2402128 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
7724 2402128 : }
7725 :
7726 2402128 : res
7727 2402128 : }
7728 :
7729 : #[cfg(test)]
7730 : /// Test helper, for tests that would like to poke individual values without composing a batch
7731 2195079 : pub(crate) async fn put(
7732 2195079 : &mut self,
7733 2195079 : key: Key,
7734 2195079 : lsn: Lsn,
7735 2195079 : value: &Value,
7736 2195079 : ctx: &RequestContext,
7737 2195079 : ) -> anyhow::Result<()> {
7738 : use utils::bin_ser::BeSer;
7739 2195079 : if !key.is_valid_key_on_write_path() {
7740 0 : bail!(
7741 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
7742 : key
7743 : );
7744 2195079 : }
7745 2195079 : let val_ser_size = value.serialized_size().unwrap() as usize;
7746 2195079 : let batch = SerializedValueBatch::from_values(vec![(
7747 2195079 : key.to_compact(),
7748 2195079 : lsn,
7749 2195079 : val_ser_size,
7750 2195079 : value.clone(),
7751 2195079 : )]);
7752 :
7753 2195079 : self.put_batch(batch, ctx).await
7754 2195079 : }
7755 :
7756 1 : pub(crate) async fn delete_batch(
7757 1 : &mut self,
7758 1 : batch: &[(Range<Key>, Lsn)],
7759 1 : ctx: &RequestContext,
7760 1 : ) -> anyhow::Result<()> {
7761 1 : if let Some((_, lsn)) = batch.first() {
7762 1 : let action = self.get_open_layer_action(*lsn, 0);
7763 1 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
7764 1 : layer.put_tombstones(batch).await?;
7765 0 : }
7766 :
7767 1 : Ok(())
7768 1 : }
7769 :
7770 : /// Track the end of the latest digested WAL record.
7771 : /// Remember the (end of) last valid WAL record remembered in the timeline.
7772 : ///
7773 : /// Call this after you have finished writing all the WAL up to 'lsn'.
7774 : ///
7775 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
7776 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
7777 : /// the latest and can be read.
7778 2639560 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
7779 2639560 : self.tl.finish_write(new_lsn);
7780 2639560 : }
7781 :
7782 135285 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
7783 135285 : self.tl.update_current_logical_size(delta)
7784 135285 : }
7785 : }
7786 :
7787 : // We need TimelineWriter to be send in upcoming conversion of
7788 : // Timeline::layers to tokio::sync::RwLock.
7789 : #[test]
7790 1 : fn is_send() {
7791 1 : fn _assert_send<T: Send>() {}
7792 1 : _assert_send::<TimelineWriter<'_>>();
7793 1 : }
7794 :
7795 : #[cfg(test)]
7796 : mod tests {
7797 : use std::sync::Arc;
7798 :
7799 : use pageserver_api::key::Key;
7800 : use postgres_ffi::PgMajorVersion;
7801 : use std::iter::Iterator;
7802 : use tracing::Instrument;
7803 : use utils::id::TimelineId;
7804 : use utils::lsn::Lsn;
7805 : use wal_decoder::models::value::Value;
7806 :
7807 : use super::HeatMapTimeline;
7808 : use crate::context::RequestContextBuilder;
7809 : use crate::tenant::harness::{TenantHarness, test_img};
7810 : use crate::tenant::layer_map::LayerMap;
7811 : use crate::tenant::storage_layer::{Layer, LayerName, LayerVisibilityHint};
7812 : use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
7813 : use crate::tenant::timeline::{DeltaLayerTestDesc, EvictionError};
7814 : use crate::tenant::{PreviousHeatmap, Timeline};
7815 :
7816 5 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
7817 5 : assert_eq!(lhs.all_layers().count(), rhs.all_layers().count());
7818 5 : let lhs_rhs = lhs.all_layers().zip(rhs.all_layers());
7819 25 : for (l, r) in lhs_rhs {
7820 20 : assert_eq!(l.name, r.name);
7821 20 : assert_eq!(l.metadata, r.metadata);
7822 : }
7823 5 : }
7824 :
7825 : #[tokio::test]
7826 1 : async fn test_heatmap_generation() {
7827 1 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
7828 :
7829 1 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7830 1 : Lsn(0x10)..Lsn(0x20),
7831 1 : vec![(
7832 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7833 1 : Lsn(0x11),
7834 1 : Value::Image(test_img("foo")),
7835 1 : )],
7836 : );
7837 1 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7838 1 : Lsn(0x10)..Lsn(0x20),
7839 1 : vec![(
7840 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7841 1 : Lsn(0x11),
7842 1 : Value::Image(test_img("foo")),
7843 1 : )],
7844 : );
7845 1 : let l0_delta = DeltaLayerTestDesc::new(
7846 1 : Lsn(0x20)..Lsn(0x30),
7847 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7848 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7849 1 : vec![(
7850 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7851 1 : Lsn(0x25),
7852 1 : Value::Image(test_img("foo")),
7853 1 : )],
7854 : );
7855 1 : let delta_layers = vec![
7856 1 : covered_delta.clone(),
7857 1 : visible_delta.clone(),
7858 1 : l0_delta.clone(),
7859 : ];
7860 :
7861 1 : let image_layer = (
7862 1 : Lsn(0x40),
7863 1 : vec![(
7864 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7865 1 : test_img("bar"),
7866 1 : )],
7867 1 : );
7868 1 : let image_layers = vec![image_layer];
7869 :
7870 1 : let (tenant, ctx) = harness.load().await;
7871 1 : let timeline = tenant
7872 1 : .create_test_timeline_with_layers(
7873 1 : TimelineId::generate(),
7874 1 : Lsn(0x10),
7875 1 : PgMajorVersion::PG14,
7876 1 : &ctx,
7877 1 : Vec::new(), // in-memory layers
7878 1 : delta_layers,
7879 1 : image_layers,
7880 1 : Lsn(0x100),
7881 1 : )
7882 1 : .await
7883 1 : .unwrap();
7884 1 : let ctx = &ctx.with_scope_timeline(&timeline);
7885 :
7886 : // Layer visibility is an input to heatmap generation, so refresh it first
7887 1 : timeline.update_layer_visibility().await.unwrap();
7888 :
7889 1 : let heatmap = timeline
7890 1 : .generate_heatmap()
7891 1 : .await
7892 1 : .expect("Infallible while timeline is not shut down");
7893 :
7894 1 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
7895 :
7896 : // L0 should come last
7897 1 : let heatmap_layers = heatmap.all_layers().collect::<Vec<_>>();
7898 1 : assert_eq!(heatmap_layers.last().unwrap().name, l0_delta.layer_name());
7899 :
7900 1 : let mut last_lsn = Lsn::MAX;
7901 5 : for layer in heatmap_layers {
7902 : // Covered layer should be omitted
7903 4 : assert!(layer.name != covered_delta.layer_name());
7904 :
7905 4 : let layer_lsn = match &layer.name {
7906 2 : LayerName::Delta(d) => d.lsn_range.end,
7907 2 : LayerName::Image(i) => i.lsn,
7908 : };
7909 :
7910 : // Apart from L0s, newest Layers should come first
7911 4 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
7912 3 : assert!(layer_lsn <= last_lsn);
7913 3 : last_lsn = layer_lsn;
7914 1 : }
7915 : }
7916 :
7917 : // Evict all the layers and stash the old heatmap in the timeline.
7918 : // This simulates a migration to a cold secondary location.
7919 :
7920 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
7921 1 : let mut all_layers = Vec::new();
7922 1 : let forever = std::time::Duration::from_secs(120);
7923 5 : for layer in guard.likely_resident_layers() {
7924 5 : all_layers.push(layer.clone());
7925 5 : layer.evict_and_wait(forever).await.unwrap();
7926 : }
7927 1 : drop(guard);
7928 :
7929 1 : timeline
7930 1 : .previous_heatmap
7931 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
7932 1 : heatmap: heatmap.clone(),
7933 1 : read_at: std::time::Instant::now(),
7934 1 : end_lsn: None,
7935 1 : })));
7936 :
7937 : // Generate a new heatmap and assert that it contains the same layers as the old one.
7938 1 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
7939 1 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
7940 :
7941 : // Download each layer one by one. Generate the heatmap at each step and check
7942 : // that it's stable.
7943 6 : for layer in all_layers {
7944 5 : if layer.visibility() == LayerVisibilityHint::Covered {
7945 1 : continue;
7946 4 : }
7947 1 :
7948 4 : eprintln!("Downloading {layer} and re-generating heatmap");
7949 1 :
7950 4 : let ctx = &RequestContextBuilder::from(ctx)
7951 4 : .download_behavior(crate::context::DownloadBehavior::Download)
7952 4 : .attached_child();
7953 1 :
7954 4 : let _resident = layer
7955 4 : .download_and_keep_resident(ctx)
7956 4 : .instrument(tracing::info_span!(
7957 4 : parent: None,
7958 1 : "download_layer",
7959 1 : tenant_id = %timeline.tenant_shard_id.tenant_id,
7960 1 : shard_id = %timeline.tenant_shard_id.shard_slug(),
7961 1 : timeline_id = %timeline.timeline_id
7962 1 : ))
7963 4 : .await
7964 4 : .unwrap();
7965 1 :
7966 4 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
7967 4 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
7968 1 : }
7969 1 :
7970 1 : // Everything from the post-migration heatmap is now resident.
7971 1 : // Check that we drop it from memory.
7972 1 : assert!(matches!(
7973 1 : timeline.previous_heatmap.load().as_deref(),
7974 1 : Some(PreviousHeatmap::Obsolete)
7975 1 : ));
7976 1 : }
7977 :
7978 : #[tokio::test]
7979 1 : async fn test_previous_heatmap_obsoletion() {
7980 1 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
7981 1 : .await
7982 1 : .unwrap();
7983 :
7984 1 : let l0_delta = DeltaLayerTestDesc::new(
7985 1 : Lsn(0x20)..Lsn(0x30),
7986 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7987 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7988 1 : vec![(
7989 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7990 1 : Lsn(0x25),
7991 1 : Value::Image(test_img("foo")),
7992 1 : )],
7993 : );
7994 :
7995 1 : let image_layer = (
7996 1 : Lsn(0x40),
7997 1 : vec![(
7998 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7999 1 : test_img("bar"),
8000 1 : )],
8001 1 : );
8002 :
8003 1 : let delta_layers = vec![l0_delta];
8004 1 : let image_layers = vec![image_layer];
8005 :
8006 1 : let (tenant, ctx) = harness.load().await;
8007 1 : let timeline = tenant
8008 1 : .create_test_timeline_with_layers(
8009 1 : TimelineId::generate(),
8010 1 : Lsn(0x10),
8011 1 : PgMajorVersion::PG14,
8012 1 : &ctx,
8013 1 : Vec::new(), // in-memory layers
8014 1 : delta_layers,
8015 1 : image_layers,
8016 1 : Lsn(0x100),
8017 1 : )
8018 1 : .await
8019 1 : .unwrap();
8020 :
8021 : // Layer visibility is an input to heatmap generation, so refresh it first
8022 1 : timeline.update_layer_visibility().await.unwrap();
8023 :
8024 1 : let heatmap = timeline
8025 1 : .generate_heatmap()
8026 1 : .await
8027 1 : .expect("Infallible while timeline is not shut down");
8028 :
8029 : // Both layers should be in the heatmap
8030 1 : assert!(heatmap.all_layers().count() > 0);
8031 :
8032 : // Now simulate a migration.
8033 1 : timeline
8034 1 : .previous_heatmap
8035 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
8036 1 : heatmap: heatmap.clone(),
8037 1 : read_at: std::time::Instant::now(),
8038 1 : end_lsn: None,
8039 1 : })));
8040 :
8041 : // Evict all the layers in the previous heatmap
8042 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
8043 1 : let forever = std::time::Duration::from_secs(120);
8044 3 : for layer in guard.likely_resident_layers() {
8045 3 : layer.evict_and_wait(forever).await.unwrap();
8046 : }
8047 1 : drop(guard);
8048 :
8049 : // Generate a new heatmap and check that the previous heatmap
8050 : // has been marked obsolete.
8051 1 : let post_eviction_heatmap = timeline
8052 1 : .generate_heatmap()
8053 1 : .await
8054 1 : .expect("Infallible while timeline is not shut down");
8055 :
8056 1 : assert_eq!(post_eviction_heatmap.all_layers().count(), 0);
8057 1 : assert!(matches!(
8058 1 : timeline.previous_heatmap.load().as_deref(),
8059 1 : Some(PreviousHeatmap::Obsolete)
8060 1 : ));
8061 1 : }
8062 :
8063 : #[tokio::test]
8064 1 : async fn two_layer_eviction_attempts_at_the_same_time() {
8065 1 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
8066 1 : .await
8067 1 : .unwrap();
8068 :
8069 1 : let (tenant, ctx) = harness.load().await;
8070 1 : let timeline = tenant
8071 1 : .create_test_timeline(
8072 1 : TimelineId::generate(),
8073 1 : Lsn(0x10),
8074 1 : PgMajorVersion::PG14,
8075 1 : &ctx,
8076 1 : )
8077 1 : .await
8078 1 : .unwrap();
8079 :
8080 1 : let layer = find_some_layer(&timeline).await;
8081 1 : let layer = layer
8082 1 : .keep_resident()
8083 1 : .await
8084 1 : .expect("no download => no downloading errors")
8085 1 : .drop_eviction_guard();
8086 :
8087 1 : let forever = std::time::Duration::from_secs(120);
8088 :
8089 1 : let first = layer.evict_and_wait(forever);
8090 1 : let second = layer.evict_and_wait(forever);
8091 :
8092 1 : let (first, second) = tokio::join!(first, second);
8093 :
8094 1 : let res = layer.keep_resident().await;
8095 1 : assert!(res.is_none(), "{res:?}");
8096 :
8097 1 : match (first, second) {
8098 1 : (Ok(()), Ok(())) => {
8099 1 : // because there are no more timeline locks being taken on eviction path, we can
8100 1 : // witness all three outcomes here.
8101 1 : }
8102 1 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
8103 0 : // if one completes before the other, this is fine just as well.
8104 0 : }
8105 1 : other => unreachable!("unexpected {:?}", other),
8106 1 : }
8107 1 : }
8108 :
8109 1 : async fn find_some_layer(timeline: &Timeline) -> Layer {
8110 1 : let layers = timeline
8111 1 : .layers
8112 1 : .read(LayerManagerLockHolder::GetLayerMapInfo)
8113 1 : .await;
8114 1 : let desc = layers
8115 1 : .layer_map()
8116 1 : .unwrap()
8117 1 : .iter_historic_layers()
8118 1 : .next()
8119 1 : .expect("must find one layer to evict");
8120 :
8121 1 : layers.get_from_desc(&desc)
8122 1 : }
8123 : }
|
