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, log_slow, 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: ArcSwapOption<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 : /// Begin Hadron: If true, the pageserver has likely detected data corruption in the timeline.
401 : /// We need to feed this information back to the Safekeeper and postgres for them to take the
402 : /// appropriate action.
403 : corruption_detected: AtomicBool,
404 :
405 : /// Notifies the tenant compaction loop that there is pending L0 compaction work.
406 : l0_compaction_trigger: Arc<Notify>,
407 :
408 : /// Make sure we only have one running gc at a time.
409 : ///
410 : /// Must only be taken in two places:
411 : /// - [`Timeline::gc`] (this file)
412 : /// - [`delete::delete_local_timeline_directory`]
413 : ///
414 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
415 : gc_lock: tokio::sync::Mutex<()>,
416 :
417 : /// Cloned from [`super::TenantShard::pagestream_throttle`] on construction.
418 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
419 :
420 : /// Size estimator for aux file v2
421 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
422 :
423 : /// Some test cases directly place keys into the timeline without actually modifying the directory
424 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
425 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
426 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
427 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
428 : #[cfg(test)]
429 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
430 :
431 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
432 :
433 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
434 :
435 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
436 :
437 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
438 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
439 :
440 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
441 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
442 :
443 : pub(super) previous_heatmap: ArcSwapOption<PreviousHeatmap>,
444 :
445 : /// May host a background Tokio task which downloads all the layers from the current
446 : /// heatmap on demand.
447 : heatmap_layers_downloader: Mutex<Option<heatmap_layers_downloader::HeatmapLayersDownloader>>,
448 :
449 : pub(crate) rel_size_v2_status: ArcSwap<(Option<RelSizeMigration>, Option<Lsn>)>,
450 :
451 : wait_lsn_log_slow: tokio::sync::Semaphore,
452 :
453 : /// A channel to send async requests to prepare a basebackup for the basebackup cache.
454 : basebackup_cache: Arc<BasebackupCache>,
455 :
456 : #[expect(dead_code)]
457 : feature_resolver: Arc<TenantFeatureResolver>,
458 :
459 : /// Basebackup will collect the count and store it here. Used for reldirv2 rollout.
460 : pub(crate) db_rel_count: ArcSwapOption<(usize, usize)>,
461 : }
462 :
463 : pub(crate) enum PreviousHeatmap {
464 : Active {
465 : heatmap: HeatMapTimeline,
466 : read_at: std::time::Instant,
467 : // End LSN covered by the heatmap if known
468 : end_lsn: Option<Lsn>,
469 : },
470 : Obsolete,
471 : }
472 :
473 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
474 :
475 : pub struct WalReceiverInfo {
476 : pub wal_source_connconf: PgConnectionConfig,
477 : pub last_received_msg_lsn: Lsn,
478 : pub last_received_msg_ts: u128,
479 : }
480 :
481 : /// Information about how much history needs to be retained, needed by
482 : /// Garbage Collection.
483 : #[derive(Default)]
484 : pub(crate) struct GcInfo {
485 : /// Specific LSNs that are needed.
486 : ///
487 : /// Currently, this includes all points where child branches have
488 : /// been forked off from. In the future, could also include
489 : /// explicit user-defined snapshot points.
490 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
491 :
492 : /// The cutoff coordinates, which are combined by selecting the minimum.
493 : pub(crate) cutoffs: GcCutoffs,
494 :
495 : /// Leases granted to particular LSNs.
496 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
497 :
498 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
499 : pub(crate) within_ancestor_pitr: bool,
500 : }
501 :
502 : impl GcInfo {
503 154 : pub(crate) fn min_cutoff(&self) -> Lsn {
504 154 : self.cutoffs.select_min()
505 154 : }
506 :
507 119 : pub(super) fn insert_child(
508 119 : &mut self,
509 119 : child_id: TimelineId,
510 119 : child_lsn: Lsn,
511 119 : is_offloaded: MaybeOffloaded,
512 119 : ) {
513 119 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
514 119 : self.retain_lsns.sort_by_key(|i| i.0);
515 119 : }
516 :
517 2 : pub(super) fn remove_child_maybe_offloaded(
518 2 : &mut self,
519 2 : child_id: TimelineId,
520 2 : maybe_offloaded: MaybeOffloaded,
521 2 : ) -> bool {
522 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
523 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
524 2 : let mut removed = false;
525 3 : self.retain_lsns.retain(|i| {
526 3 : if removed {
527 1 : return true;
528 2 : }
529 2 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
530 2 : removed |= remove;
531 2 : !remove
532 3 : });
533 2 : removed
534 2 : }
535 :
536 2 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
537 2 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
538 2 : }
539 :
540 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
541 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
542 0 : }
543 119 : pub(crate) fn lsn_covered_by_lease(&self, lsn: Lsn) -> bool {
544 119 : self.leases.contains_key(&lsn)
545 119 : }
546 : }
547 :
548 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
549 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
550 : /// between time-based and space-based retention for observability and consumption metrics purposes.
551 : #[derive(Clone, Debug, Default)]
552 : pub(crate) struct GcCutoffs {
553 : /// Calculated from the [`pageserver_api::models::TenantConfig::gc_horizon`], this LSN indicates how much
554 : /// history we must keep to retain a specified number of bytes of WAL.
555 : pub(crate) space: Lsn,
556 :
557 : /// Calculated from [`pageserver_api::models::TenantConfig::pitr_interval`], this LSN indicates
558 : /// how much history we must keep to enable reading back at least the PITR interval duration.
559 : ///
560 : /// None indicates that the PITR cutoff has not been computed. A PITR interval of 0 will yield
561 : /// Some(last_record_lsn).
562 : pub(crate) time: Option<Lsn>,
563 : }
564 :
565 : impl GcCutoffs {
566 154 : fn select_min(&self) -> Lsn {
567 : // NB: if we haven't computed the PITR cutoff yet, we can't GC anything.
568 154 : self.space.min(self.time.unwrap_or_default())
569 154 : }
570 : }
571 :
572 : pub(crate) struct TimelineVisitOutcome {
573 : completed_keyspace: KeySpace,
574 : image_covered_keyspace: KeySpace,
575 : }
576 :
577 : /// An error happened in a get() operation.
578 : #[derive(thiserror::Error, Debug)]
579 : pub(crate) enum PageReconstructError {
580 : #[error(transparent)]
581 : Other(anyhow::Error),
582 :
583 : #[error("Ancestor LSN wait error: {0}")]
584 : AncestorLsnTimeout(WaitLsnError),
585 :
586 : #[error("timeline shutting down")]
587 : Cancelled,
588 :
589 : /// An error happened replaying WAL records
590 : #[error(transparent)]
591 : WalRedo(anyhow::Error),
592 :
593 : #[error("{0}")]
594 : MissingKey(Box<MissingKeyError>),
595 : }
596 :
597 : impl PageReconstructError {
598 0 : pub(crate) fn is_cancel(&self) -> bool {
599 0 : match self {
600 0 : PageReconstructError::Other(_) => false,
601 0 : PageReconstructError::AncestorLsnTimeout(e) => e.is_cancel(),
602 0 : PageReconstructError::Cancelled => true,
603 0 : PageReconstructError::WalRedo(_) => false,
604 0 : PageReconstructError::MissingKey(_) => false,
605 : }
606 0 : }
607 : #[allow(dead_code)] // we use the is_cancel + into_anyhow pattern in quite a few places, this one will follow soon enough
608 0 : pub(crate) fn into_anyhow(self) -> anyhow::Error {
609 0 : match self {
610 0 : PageReconstructError::Other(e) => e,
611 0 : PageReconstructError::AncestorLsnTimeout(e) => e.into_anyhow(),
612 0 : PageReconstructError::Cancelled => anyhow::Error::new(self),
613 0 : PageReconstructError::WalRedo(e) => e,
614 0 : PageReconstructError::MissingKey(_) => anyhow::Error::new(self),
615 : }
616 0 : }
617 : }
618 :
619 : impl From<anyhow::Error> for PageReconstructError {
620 1 : fn from(value: anyhow::Error) -> Self {
621 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
622 1 : match value.downcast::<PageReconstructError>() {
623 0 : Ok(pre) => pre,
624 1 : Err(other) => PageReconstructError::Other(other),
625 : }
626 1 : }
627 : }
628 :
629 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
630 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
631 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
632 0 : }
633 : }
634 :
635 : impl From<layer_manager::Shutdown> for PageReconstructError {
636 0 : fn from(_: layer_manager::Shutdown) -> Self {
637 0 : PageReconstructError::Cancelled
638 0 : }
639 : }
640 :
641 : impl GetVectoredError {
642 : #[cfg(test)]
643 3 : pub(crate) fn is_missing_key_error(&self) -> bool {
644 3 : matches!(self, Self::MissingKey(_))
645 3 : }
646 : }
647 :
648 : impl From<layer_manager::Shutdown> for GetVectoredError {
649 0 : fn from(_: layer_manager::Shutdown) -> Self {
650 0 : GetVectoredError::Cancelled
651 0 : }
652 : }
653 :
654 : /// A layer identifier when used in the [`ReadPath`] structure. This enum is for observability purposes
655 : /// only and not used by the "real read path".
656 : pub enum ReadPathLayerId {
657 : PersistentLayer(PersistentLayerKey),
658 : InMemoryLayer(Range<Lsn>),
659 : }
660 :
661 : impl std::fmt::Display for ReadPathLayerId {
662 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
663 0 : match self {
664 0 : ReadPathLayerId::PersistentLayer(key) => write!(f, "{key}"),
665 0 : ReadPathLayerId::InMemoryLayer(range) => {
666 0 : write!(f, "in-mem {}..{}", range.start, range.end)
667 : }
668 : }
669 0 : }
670 : }
671 : pub struct ReadPath {
672 : keyspace: KeySpace,
673 : lsn: Lsn,
674 : path: Vec<(ReadPathLayerId, KeySpace, Range<Lsn>)>,
675 : }
676 :
677 : impl ReadPath {
678 312430 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
679 312430 : Self {
680 312430 : keyspace,
681 312430 : lsn,
682 312430 : path: Vec::new(),
683 312430 : }
684 312430 : }
685 :
686 446935 : pub fn record_layer_visit(
687 446935 : &mut self,
688 446935 : layer_to_read: &ReadableLayer,
689 446935 : keyspace_to_read: &KeySpace,
690 446935 : lsn_range: &Range<Lsn>,
691 446935 : ) {
692 446935 : let id = match layer_to_read {
693 139661 : ReadableLayer::PersistentLayer(layer) => {
694 139661 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
695 : }
696 307274 : ReadableLayer::InMemoryLayer(layer) => {
697 307274 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
698 : }
699 : };
700 446935 : self.path
701 446935 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
702 446935 : }
703 : }
704 :
705 : impl std::fmt::Display for ReadPath {
706 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
707 0 : writeln!(f, "Read path for {} at lsn {}:", self.keyspace, self.lsn)?;
708 0 : for (idx, (layer_id, keyspace, lsn_range)) in self.path.iter().enumerate() {
709 0 : writeln!(
710 0 : f,
711 0 : "{}: {} {}..{} {}",
712 : idx, layer_id, lsn_range.start, lsn_range.end, keyspace
713 0 : )?;
714 : }
715 0 : Ok(())
716 0 : }
717 : }
718 :
719 : #[derive(thiserror::Error)]
720 : pub struct MissingKeyError {
721 : keyspace: KeySpace,
722 : shard: ShardNumber,
723 : query: Option<VersionedKeySpaceQuery>,
724 : // This is largest request LSN from the get page request batch
725 : original_hwm_lsn: Lsn,
726 : ancestor_lsn: Option<Lsn>,
727 : /// Debug information about the read path if there's an error
728 : read_path: Option<ReadPath>,
729 : backtrace: Option<std::backtrace::Backtrace>,
730 : }
731 :
732 : impl MissingKeyError {
733 7 : fn enrich(&mut self, query: VersionedKeySpaceQuery) {
734 7 : self.query = Some(query);
735 7 : }
736 : }
737 :
738 : impl std::fmt::Debug for MissingKeyError {
739 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
740 0 : write!(f, "{self}")
741 0 : }
742 : }
743 :
744 : impl std::fmt::Display for MissingKeyError {
745 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
746 0 : write!(
747 0 : f,
748 0 : "could not find data for key {} (shard {:?}), original HWM LSN {}",
749 : self.keyspace, self.shard, self.original_hwm_lsn
750 0 : )?;
751 :
752 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
753 0 : write!(f, ", ancestor {ancestor_lsn}")?;
754 0 : }
755 :
756 0 : if let Some(ref query) = self.query {
757 0 : write!(f, ", query {query}")?;
758 0 : }
759 :
760 0 : if let Some(ref read_path) = self.read_path {
761 0 : write!(f, "\n{read_path}")?;
762 0 : }
763 :
764 0 : if let Some(ref backtrace) = self.backtrace {
765 0 : write!(f, "\n{backtrace}")?;
766 0 : }
767 :
768 0 : Ok(())
769 0 : }
770 : }
771 :
772 : #[derive(thiserror::Error, Debug)]
773 : pub(crate) enum CreateImageLayersError {
774 : #[error("timeline shutting down")]
775 : Cancelled,
776 :
777 : #[error("read failed")]
778 : GetVectoredError(#[source] GetVectoredError),
779 :
780 : #[error("reconstruction failed")]
781 : PageReconstructError(#[source] PageReconstructError),
782 :
783 : #[error(transparent)]
784 : Other(anyhow::Error),
785 : }
786 :
787 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
788 0 : fn from(_: layer_manager::Shutdown) -> Self {
789 0 : CreateImageLayersError::Cancelled
790 0 : }
791 : }
792 :
793 : #[derive(thiserror::Error, Debug, Clone)]
794 : pub(crate) enum FlushLayerError {
795 : /// Timeline cancellation token was cancelled
796 : #[error("timeline shutting down")]
797 : Cancelled,
798 :
799 : /// We tried to flush a layer while the Timeline is in an unexpected state
800 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
801 : NotRunning(FlushLoopState),
802 :
803 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
804 : // loop via a watch channel, where we can only borrow it.
805 : #[error("create image layers (shared)")]
806 : CreateImageLayersError(Arc<CreateImageLayersError>),
807 :
808 : #[error("other (shared)")]
809 : Other(#[from] Arc<anyhow::Error>),
810 : }
811 :
812 : impl FlushLayerError {
813 : // When crossing from generic anyhow errors to this error type, we explicitly check
814 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
815 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
816 0 : let cancelled = timeline.cancel.is_cancelled()
817 : // The upload queue might have been shut down before the official cancellation of the timeline.
818 0 : || err
819 0 : .downcast_ref::<NotInitialized>()
820 0 : .map(NotInitialized::is_stopping)
821 0 : .unwrap_or_default();
822 0 : if cancelled {
823 0 : Self::Cancelled
824 : } else {
825 0 : Self::Other(Arc::new(err))
826 : }
827 0 : }
828 : }
829 :
830 : impl From<layer_manager::Shutdown> for FlushLayerError {
831 0 : fn from(_: layer_manager::Shutdown) -> Self {
832 0 : FlushLayerError::Cancelled
833 0 : }
834 : }
835 :
836 : #[derive(thiserror::Error, Debug)]
837 : pub enum GetVectoredError {
838 : #[error("timeline shutting down")]
839 : Cancelled,
840 :
841 : #[error("requested too many keys: {0} > {1}")]
842 : Oversized(u64, u64),
843 :
844 : #[error("requested at invalid LSN: {0}")]
845 : InvalidLsn(Lsn),
846 :
847 : #[error("requested key not found: {0}")]
848 : MissingKey(Box<MissingKeyError>),
849 :
850 : #[error("ancestry walk")]
851 : GetReadyAncestorError(#[source] GetReadyAncestorError),
852 :
853 : #[error(transparent)]
854 : Other(#[from] anyhow::Error),
855 : }
856 :
857 : impl From<GetReadyAncestorError> for GetVectoredError {
858 1 : fn from(value: GetReadyAncestorError) -> Self {
859 : use GetReadyAncestorError::*;
860 1 : match value {
861 0 : Cancelled => GetVectoredError::Cancelled,
862 : AncestorLsnTimeout(_) | BadState { .. } => {
863 1 : GetVectoredError::GetReadyAncestorError(value)
864 : }
865 : }
866 1 : }
867 : }
868 :
869 : #[derive(thiserror::Error, Debug)]
870 : pub enum GetReadyAncestorError {
871 : #[error("ancestor LSN wait error")]
872 : AncestorLsnTimeout(#[from] WaitLsnError),
873 :
874 : #[error("bad state on timeline {timeline_id}: {state:?}")]
875 : BadState {
876 : timeline_id: TimelineId,
877 : state: TimelineState,
878 : },
879 :
880 : #[error("cancelled")]
881 : Cancelled,
882 : }
883 :
884 : #[derive(Clone, Copy)]
885 : pub enum LogicalSizeCalculationCause {
886 : Initial,
887 : ConsumptionMetricsSyntheticSize,
888 : EvictionTaskImitation,
889 : TenantSizeHandler,
890 : }
891 :
892 : pub enum GetLogicalSizePriority {
893 : User,
894 : Background,
895 : }
896 :
897 : #[derive(Debug, enumset::EnumSetType)]
898 : pub(crate) enum CompactFlags {
899 : ForceRepartition,
900 : ForceImageLayerCreation,
901 : ForceL0Compaction,
902 : OnlyL0Compaction,
903 : EnhancedGcBottomMostCompaction,
904 : DryRun,
905 : /// Makes image compaction yield if there's pending L0 compaction. This should always be used in
906 : /// the background compaction task, since we want to aggressively compact down L0 to bound
907 : /// read amplification.
908 : ///
909 : /// It only makes sense to use this when `compaction_l0_first` is enabled (such that we yield to
910 : /// an L0 compaction pass), and without `OnlyL0Compaction` (L0 compaction shouldn't yield for L0
911 : /// compaction).
912 : YieldForL0,
913 : }
914 :
915 : #[serde_with::serde_as]
916 0 : #[derive(Debug, Clone, serde::Deserialize)]
917 : pub(crate) struct CompactRequest {
918 : pub compact_key_range: Option<CompactKeyRange>,
919 : pub compact_lsn_range: Option<CompactLsnRange>,
920 : /// Whether the compaction job should be scheduled.
921 : #[serde(default)]
922 : pub scheduled: bool,
923 : /// Whether the compaction job should be split across key ranges.
924 : #[serde(default)]
925 : pub sub_compaction: bool,
926 : /// Max job size for each subcompaction job.
927 : pub sub_compaction_max_job_size_mb: Option<u64>,
928 : }
929 :
930 0 : #[derive(Debug, Clone, serde::Deserialize)]
931 : pub(crate) struct MarkInvisibleRequest {
932 : #[serde(default)]
933 : pub is_visible: Option<bool>,
934 : }
935 :
936 : #[derive(Debug, Clone, Default)]
937 : pub(crate) struct CompactOptions {
938 : pub flags: EnumSet<CompactFlags>,
939 : /// If set, the compaction will only compact the key range specified by this option.
940 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
941 : pub compact_key_range: Option<CompactKeyRange>,
942 : /// If set, the compaction will only compact the LSN within this value.
943 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
944 : pub compact_lsn_range: Option<CompactLsnRange>,
945 : /// Enable sub-compaction (split compaction job across key ranges).
946 : /// This option is only used by GC compaction.
947 : pub sub_compaction: bool,
948 : /// Set job size for the GC compaction.
949 : /// This option is only used by GC compaction.
950 : pub sub_compaction_max_job_size_mb: Option<u64>,
951 : /// Only for GC compaction.
952 : /// If set, the compaction will compact the metadata layers. Should be only set to true in unit tests
953 : /// because metadata compaction is not fully supported yet.
954 : pub gc_compaction_do_metadata_compaction: bool,
955 : }
956 :
957 : impl CompactOptions {
958 : #[cfg(test)]
959 28 : pub fn default_for_gc_compaction_unit_tests() -> Self {
960 28 : Self {
961 28 : gc_compaction_do_metadata_compaction: true,
962 28 : ..Default::default()
963 28 : }
964 28 : }
965 : }
966 :
967 : impl std::fmt::Debug for Timeline {
968 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
969 0 : write!(f, "Timeline<{}>", self.timeline_id)
970 0 : }
971 : }
972 :
973 : #[derive(thiserror::Error, Debug, Clone)]
974 : pub enum WaitLsnError {
975 : // Called on a timeline which is shutting down
976 : #[error("Shutdown")]
977 : Shutdown,
978 :
979 : // Called on an timeline not in active state or shutting down
980 : #[error("Bad timeline state: {0:?}")]
981 : BadState(TimelineState),
982 :
983 : // Timeout expired while waiting for LSN to catch up with goal.
984 : #[error("{0}")]
985 : Timeout(String),
986 : }
987 :
988 : impl WaitLsnError {
989 0 : pub(crate) fn is_cancel(&self) -> bool {
990 0 : match self {
991 0 : WaitLsnError::Shutdown => true,
992 0 : WaitLsnError::BadState(timeline_state) => match timeline_state {
993 0 : TimelineState::Loading => false,
994 0 : TimelineState::Active => false,
995 0 : TimelineState::Stopping => true,
996 0 : TimelineState::Broken { .. } => false,
997 : },
998 0 : WaitLsnError::Timeout(_) => false,
999 : }
1000 0 : }
1001 0 : pub(crate) fn into_anyhow(self) -> anyhow::Error {
1002 0 : match self {
1003 0 : WaitLsnError::Shutdown => anyhow::Error::new(self),
1004 0 : WaitLsnError::BadState(_) => anyhow::Error::new(self),
1005 0 : WaitLsnError::Timeout(_) => anyhow::Error::new(self),
1006 : }
1007 0 : }
1008 : }
1009 :
1010 : impl From<WaitLsnError> for tonic::Status {
1011 0 : fn from(err: WaitLsnError) -> Self {
1012 : use tonic::Code;
1013 0 : let code = if err.is_cancel() {
1014 0 : Code::Unavailable
1015 : } else {
1016 0 : Code::Internal
1017 : };
1018 0 : tonic::Status::new(code, err.to_string())
1019 0 : }
1020 : }
1021 :
1022 : // The impls below achieve cancellation mapping for errors.
1023 : // Perhaps there's a way of achieving this with less cruft.
1024 :
1025 : impl From<CreateImageLayersError> for CompactionError {
1026 0 : fn from(e: CreateImageLayersError) -> Self {
1027 0 : match e {
1028 0 : CreateImageLayersError::Cancelled => CompactionError::new_cancelled(),
1029 0 : CreateImageLayersError::Other(e) => {
1030 0 : CompactionError::Other(e.context("create image layers"))
1031 : }
1032 0 : _ => CompactionError::Other(e.into()),
1033 : }
1034 0 : }
1035 : }
1036 :
1037 : impl From<CreateImageLayersError> for FlushLayerError {
1038 0 : fn from(e: CreateImageLayersError) -> Self {
1039 0 : match e {
1040 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
1041 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
1042 : }
1043 0 : }
1044 : }
1045 :
1046 : impl From<PageReconstructError> for CreateImageLayersError {
1047 0 : fn from(e: PageReconstructError) -> Self {
1048 0 : match e {
1049 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
1050 0 : _ => CreateImageLayersError::PageReconstructError(e),
1051 : }
1052 0 : }
1053 : }
1054 :
1055 : impl From<super::storage_layer::errors::PutError> for CreateImageLayersError {
1056 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
1057 0 : if e.is_cancel() {
1058 0 : CreateImageLayersError::Cancelled
1059 : } else {
1060 0 : CreateImageLayersError::Other(e.into_anyhow())
1061 : }
1062 0 : }
1063 : }
1064 :
1065 : impl From<GetVectoredError> for CreateImageLayersError {
1066 0 : fn from(e: GetVectoredError) -> Self {
1067 0 : match e {
1068 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
1069 0 : _ => CreateImageLayersError::GetVectoredError(e),
1070 : }
1071 0 : }
1072 : }
1073 :
1074 : impl From<GetVectoredError> for PageReconstructError {
1075 3 : fn from(e: GetVectoredError) -> Self {
1076 3 : match e {
1077 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
1078 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
1079 0 : err @ GetVectoredError::Oversized(_, _) => PageReconstructError::Other(err.into()),
1080 2 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
1081 1 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
1082 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
1083 : }
1084 3 : }
1085 : }
1086 :
1087 : impl From<GetReadyAncestorError> for PageReconstructError {
1088 1 : fn from(e: GetReadyAncestorError) -> Self {
1089 : use GetReadyAncestorError::*;
1090 1 : match e {
1091 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
1092 1 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
1093 0 : Cancelled => PageReconstructError::Cancelled,
1094 : }
1095 1 : }
1096 : }
1097 :
1098 : pub(crate) enum WaitLsnTimeout {
1099 : Custom(Duration),
1100 : // Use the [`PageServerConf::wait_lsn_timeout`] default
1101 : Default,
1102 : }
1103 :
1104 : pub(crate) enum WaitLsnWaiter<'a> {
1105 : Timeline(&'a Timeline),
1106 : Tenant,
1107 : PageService,
1108 : HttpEndpoint,
1109 : BaseBackupCache,
1110 : }
1111 :
1112 : /// Argument to [`Timeline::shutdown`].
1113 : #[derive(Debug, Clone, Copy)]
1114 : pub(crate) enum ShutdownMode {
1115 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk. This method can
1116 : /// take multiple seconds for a busy timeline.
1117 : ///
1118 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1119 : /// the call to [`Timeline::shutdown`].
1120 : FreezeAndFlush,
1121 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1122 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1123 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1124 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1125 : Reload,
1126 : /// Shut down immediately, without waiting for any open layers to flush.
1127 : Hard,
1128 : }
1129 :
1130 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1131 : enum ImageLayerCreationOutcome {
1132 : /// We generated an image layer
1133 : Generated {
1134 : unfinished_image_layer: ImageLayerWriter,
1135 : },
1136 : /// The key range is empty
1137 : Empty,
1138 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1139 : /// the image layer creation.
1140 : Skip,
1141 : }
1142 :
1143 : enum RepartitionError {
1144 : Other(anyhow::Error),
1145 : CollectKeyspace(CollectKeySpaceError),
1146 : }
1147 :
1148 : impl RepartitionError {
1149 0 : fn is_cancel(&self) -> bool {
1150 0 : match self {
1151 0 : RepartitionError::Other(_) => false,
1152 0 : RepartitionError::CollectKeyspace(e) => e.is_cancel(),
1153 : }
1154 0 : }
1155 0 : fn into_anyhow(self) -> anyhow::Error {
1156 0 : match self {
1157 0 : RepartitionError::Other(e) => e,
1158 0 : RepartitionError::CollectKeyspace(e) => e.into_anyhow(),
1159 : }
1160 0 : }
1161 : }
1162 :
1163 : /// Public interface functions
1164 : impl Timeline {
1165 : /// Get the LSN where this branch was created
1166 80 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1167 80 : self.ancestor_lsn
1168 80 : }
1169 :
1170 : /// Get the ancestor's timeline id
1171 416 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1172 416 : self.ancestor_timeline
1173 416 : .as_ref()
1174 416 : .map(|ancestor| ancestor.timeline_id)
1175 416 : }
1176 :
1177 : /// Get the ancestor timeline
1178 1 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1179 1 : self.ancestor_timeline.as_ref()
1180 1 : }
1181 :
1182 : /// Get the bytes written since the PITR cutoff on this branch, and
1183 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1184 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1185 : // TODO: for backwards compatibility, we return the full history back to 0 when the PITR
1186 : // cutoff has not yet been initialized. This should return None instead, but this is exposed
1187 : // in external HTTP APIs and callers may not handle a null value.
1188 0 : let gc_info = self.gc_info.read().unwrap();
1189 0 : let history = self
1190 0 : .get_last_record_lsn()
1191 0 : .checked_sub(gc_info.cutoffs.time.unwrap_or_default())
1192 0 : .unwrap_or_default()
1193 0 : .0;
1194 0 : (history, gc_info.within_ancestor_pitr)
1195 0 : }
1196 :
1197 : /// Read timeline's GC cutoff: this is the LSN at which GC has started to happen
1198 427473 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1199 427473 : self.applied_gc_cutoff_lsn.read()
1200 427473 : }
1201 :
1202 : /// Read timeline's planned GC cutoff: this is the logical end of history that users are allowed
1203 : /// to read (based on configured PITR), even if physically we have more history. Returns None
1204 : /// if the PITR cutoff has not yet been initialized.
1205 0 : pub(crate) fn get_gc_cutoff_lsn(&self) -> Option<Lsn> {
1206 0 : self.gc_info.read().unwrap().cutoffs.time
1207 0 : }
1208 :
1209 : /// Look up given page version.
1210 : ///
1211 : /// If a remote layer file is needed, it is downloaded as part of this
1212 : /// call.
1213 : ///
1214 : /// This method enforces [`Self::pagestream_throttle`] internally.
1215 : ///
1216 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1217 : /// abstraction above this needs to store suitable metadata to track what
1218 : /// data exists with what keys, in separate metadata entries. If a
1219 : /// non-existent key is requested, we may incorrectly return a value from
1220 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1221 : /// non-existing key.
1222 : ///
1223 : /// # Cancel-Safety
1224 : ///
1225 : /// This method is cancellation-safe.
1226 : #[inline(always)]
1227 301292 : pub(crate) async fn get(
1228 301292 : &self,
1229 301292 : key: Key,
1230 301292 : lsn: Lsn,
1231 301292 : ctx: &RequestContext,
1232 301292 : ) -> Result<Bytes, PageReconstructError> {
1233 301292 : if !lsn.is_valid() {
1234 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1235 301292 : }
1236 :
1237 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1238 : // already checked the key against the shard_identity when looking up the Timeline from
1239 : // page_service.
1240 301292 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1241 :
1242 301292 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1243 :
1244 301292 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
1245 :
1246 301292 : let vectored_res = self
1247 301292 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
1248 301292 : .await;
1249 :
1250 301292 : let key_value = vectored_res?.pop_first();
1251 301289 : match key_value {
1252 301283 : Some((got_key, value)) => {
1253 301283 : if got_key != key {
1254 0 : error!(
1255 0 : "Expected {}, but singular vectored get returned {}",
1256 : key, got_key
1257 : );
1258 0 : Err(PageReconstructError::Other(anyhow!(
1259 0 : "Singular vectored get returned wrong key"
1260 0 : )))
1261 : } else {
1262 301283 : value
1263 : }
1264 : }
1265 6 : None => Err(PageReconstructError::MissingKey(Box::new(
1266 6 : MissingKeyError {
1267 6 : keyspace: KeySpace::single(key..key.next()),
1268 6 : shard: self.shard_identity.get_shard_number(&key),
1269 6 : original_hwm_lsn: lsn,
1270 6 : ancestor_lsn: None,
1271 6 : backtrace: None,
1272 6 : read_path: None,
1273 6 : query: None,
1274 6 : },
1275 6 : ))),
1276 : }
1277 301292 : }
1278 :
1279 : #[inline(always)]
1280 0 : pub(crate) async fn debug_get(
1281 0 : &self,
1282 0 : key: Key,
1283 0 : lsn: Lsn,
1284 0 : ctx: &RequestContext,
1285 0 : reconstruct_state: &mut ValuesReconstructState,
1286 0 : ) -> Result<Bytes, PageReconstructError> {
1287 0 : if !lsn.is_valid() {
1288 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1289 0 : }
1290 :
1291 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1292 : // already checked the key against the shard_identity when looking up the Timeline from
1293 : // page_service.
1294 0 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1295 :
1296 0 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
1297 0 : let vectored_res = self
1298 0 : .debug_get_vectored_impl(query, reconstruct_state, ctx)
1299 0 : .await;
1300 :
1301 0 : let key_value = vectored_res?.pop_first();
1302 0 : match key_value {
1303 0 : Some((got_key, value)) => {
1304 0 : if got_key != key {
1305 0 : error!(
1306 0 : "Expected {}, but singular vectored get returned {}",
1307 : key, got_key
1308 : );
1309 0 : Err(PageReconstructError::Other(anyhow!(
1310 0 : "Singular vectored get returned wrong key"
1311 0 : )))
1312 : } else {
1313 0 : value
1314 : }
1315 : }
1316 0 : None => Err(PageReconstructError::MissingKey(Box::new(
1317 0 : MissingKeyError {
1318 0 : keyspace: KeySpace::single(key..key.next()),
1319 0 : shard: self.shard_identity.get_shard_number(&key),
1320 0 : original_hwm_lsn: lsn,
1321 0 : ancestor_lsn: None,
1322 0 : backtrace: None,
1323 0 : read_path: None,
1324 0 : query: None,
1325 0 : },
1326 0 : ))),
1327 : }
1328 0 : }
1329 :
1330 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1331 :
1332 : /// Look up multiple page versions at a given LSN
1333 : ///
1334 : /// This naive implementation will be replaced with a more efficient one
1335 : /// which actually vectorizes the read path.
1336 : ///
1337 : /// NB: the read path must be cancellation-safe. The Tonic gRPC service will drop the future
1338 : /// if the client goes away (e.g. due to timeout or cancellation).
1339 10894 : pub(crate) async fn get_vectored(
1340 10894 : &self,
1341 10894 : query: VersionedKeySpaceQuery,
1342 10894 : io_concurrency: super::storage_layer::IoConcurrency,
1343 10894 : ctx: &RequestContext,
1344 10894 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1345 10894 : let total_keyspace = query.total_keyspace();
1346 :
1347 10894 : let key_count = total_keyspace.total_raw_size();
1348 10894 : if key_count > self.conf.max_get_vectored_keys.get() {
1349 0 : return Err(GetVectoredError::Oversized(
1350 0 : key_count as u64,
1351 0 : self.conf.max_get_vectored_keys.get() as u64,
1352 0 : ));
1353 10894 : }
1354 :
1355 34113 : for range in &total_keyspace.ranges {
1356 23219 : let mut key = range.start;
1357 65248 : while key != range.end {
1358 42029 : assert!(!self.shard_identity.is_key_disposable(&key));
1359 42029 : key = key.next();
1360 : }
1361 : }
1362 :
1363 10894 : trace!(
1364 0 : "get vectored query {} from task kind {:?}",
1365 : query,
1366 0 : ctx.task_kind(),
1367 : );
1368 :
1369 10894 : let start = crate::metrics::GET_VECTORED_LATENCY
1370 10894 : .for_task_kind(ctx.task_kind())
1371 10894 : .map(|metric| (metric, Instant::now()));
1372 :
1373 10894 : let res = self
1374 10894 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1375 10894 : .await;
1376 :
1377 10894 : if let Some((metric, start)) = start {
1378 0 : let elapsed = start.elapsed();
1379 0 : metric.observe(elapsed.as_secs_f64());
1380 10894 : }
1381 :
1382 10894 : res
1383 10894 : }
1384 :
1385 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1386 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1387 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1388 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1389 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1390 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1391 : /// the scan operation will not cause OOM in the future.
1392 8 : pub(crate) async fn scan(
1393 8 : &self,
1394 8 : keyspace: KeySpace,
1395 8 : lsn: Lsn,
1396 8 : ctx: &RequestContext,
1397 8 : io_concurrency: super::storage_layer::IoConcurrency,
1398 8 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1399 8 : if !lsn.is_valid() {
1400 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1401 8 : }
1402 :
1403 8 : trace!(
1404 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1405 : keyspace,
1406 : lsn,
1407 0 : ctx.task_kind()
1408 : );
1409 :
1410 : // We should generalize this into Keyspace::contains in the future.
1411 16 : for range in &keyspace.ranges {
1412 8 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1413 8 : || range.end.field1 > METADATA_KEY_END_PREFIX
1414 : {
1415 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1416 0 : "only metadata keyspace can be scanned"
1417 0 : )));
1418 8 : }
1419 : }
1420 :
1421 8 : let start = crate::metrics::SCAN_LATENCY
1422 8 : .for_task_kind(ctx.task_kind())
1423 8 : .map(ScanLatencyOngoingRecording::start_recording);
1424 :
1425 8 : let query = VersionedKeySpaceQuery::uniform(keyspace, lsn);
1426 :
1427 8 : let vectored_res = self
1428 8 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1429 8 : .await;
1430 :
1431 8 : if let Some(recording) = start {
1432 0 : recording.observe();
1433 8 : }
1434 :
1435 8 : vectored_res
1436 8 : }
1437 :
1438 312430 : pub(super) async fn get_vectored_impl(
1439 312430 : &self,
1440 312430 : query: VersionedKeySpaceQuery,
1441 312430 : reconstruct_state: &mut ValuesReconstructState,
1442 312430 : ctx: &RequestContext,
1443 312430 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1444 312430 : if query.is_empty() {
1445 0 : return Ok(BTreeMap::default());
1446 312430 : }
1447 :
1448 312430 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1449 312430 : Some(ReadPath::new(
1450 312430 : query.total_keyspace(),
1451 312430 : query.high_watermark_lsn()?,
1452 : ))
1453 : } else {
1454 0 : None
1455 : };
1456 :
1457 312430 : reconstruct_state.read_path = read_path;
1458 :
1459 312430 : let redo_attempt_type = if ctx.task_kind() == TaskKind::Compaction {
1460 0 : RedoAttemptType::LegacyCompaction
1461 : } else {
1462 312430 : RedoAttemptType::ReadPage
1463 : };
1464 :
1465 312430 : let traversal_res: Result<(), _> = {
1466 312430 : let ctx = RequestContextBuilder::from(ctx)
1467 312430 : .perf_span(|crnt_perf_span| {
1468 0 : info_span!(
1469 : target: PERF_TRACE_TARGET,
1470 0 : parent: crnt_perf_span,
1471 : "PLAN_IO",
1472 : )
1473 0 : })
1474 312430 : .attached_child();
1475 :
1476 312430 : self.get_vectored_reconstruct_data(query.clone(), reconstruct_state, &ctx)
1477 312430 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1478 312430 : .await
1479 : };
1480 :
1481 312430 : if let Err(err) = traversal_res {
1482 : // Wait for all the spawned IOs to complete.
1483 : // See comments on `spawn_io` inside `storage_layer` for more details.
1484 8 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1485 8 : .into_values()
1486 8 : .map(|state| state.collect_pending_ios())
1487 8 : .collect::<FuturesUnordered<_>>();
1488 8 : while collect_futs.next().await.is_some() {}
1489 :
1490 : // Enrich the missing key error with the original query.
1491 8 : if let GetVectoredError::MissingKey(mut missing_err) = err {
1492 7 : missing_err.enrich(query.clone());
1493 7 : return Err(GetVectoredError::MissingKey(missing_err));
1494 1 : }
1495 :
1496 1 : return Err(err);
1497 312422 : };
1498 :
1499 312422 : let layers_visited = reconstruct_state.get_layers_visited();
1500 :
1501 312422 : let ctx = RequestContextBuilder::from(ctx)
1502 312422 : .perf_span(|crnt_perf_span| {
1503 0 : info_span!(
1504 : target: PERF_TRACE_TARGET,
1505 0 : parent: crnt_perf_span,
1506 : "RECONSTRUCT",
1507 : )
1508 0 : })
1509 312422 : .attached_child();
1510 :
1511 312422 : let futs = FuturesUnordered::new();
1512 363551 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1513 363551 : let req_lsn_for_key = query.map_key_to_lsn(&key);
1514 :
1515 363551 : futs.push({
1516 363551 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1517 363551 : let ctx = RequestContextBuilder::from(&ctx)
1518 363551 : .perf_span(|crnt_perf_span| {
1519 0 : info_span!(
1520 : target: PERF_TRACE_TARGET,
1521 0 : parent: crnt_perf_span,
1522 : "RECONSTRUCT_KEY",
1523 : key = %key,
1524 : )
1525 0 : })
1526 363551 : .attached_child();
1527 :
1528 363551 : async move {
1529 363551 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1530 :
1531 363551 : let res = state
1532 363551 : .collect_pending_ios()
1533 363551 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1534 0 : info_span!(
1535 : target: PERF_TRACE_TARGET,
1536 0 : parent: crnt_perf_span,
1537 : "WAIT_FOR_IO_COMPLETIONS",
1538 : )
1539 0 : })
1540 363551 : .await;
1541 :
1542 363551 : let converted = match res {
1543 363551 : Ok(ok) => ok,
1544 0 : Err(err) => {
1545 0 : return (key, Err(err));
1546 : }
1547 : };
1548 363551 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1549 :
1550 : // The walredo module expects the records to be descending in terms of Lsn.
1551 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1552 363551 : debug_assert!(
1553 363551 : converted
1554 363551 : .records
1555 1401588 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1556 0 : "{converted:?}"
1557 : );
1558 :
1559 363551 : let walredo_deltas = converted.num_deltas();
1560 363551 : let walredo_res = walredo_self
1561 363551 : .reconstruct_value(key, req_lsn_for_key, converted, redo_attempt_type)
1562 363551 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1563 0 : info_span!(
1564 : target: PERF_TRACE_TARGET,
1565 0 : parent: crnt_perf_span,
1566 : "WALREDO",
1567 : deltas = %walredo_deltas,
1568 : )
1569 0 : })
1570 363551 : .await;
1571 :
1572 363551 : (key, walredo_res)
1573 363551 : }
1574 : });
1575 : }
1576 :
1577 312422 : let results = futs
1578 312422 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1579 312422 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1580 312422 : .await;
1581 :
1582 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1583 : // when they're missing. Instead they are omitted from the resulting btree
1584 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1585 : // to avoid infinite results.
1586 312422 : if !results.is_empty() {
1587 312243 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1588 0 : let total_keyspace = query.total_keyspace();
1589 0 : let max_request_lsn = query.high_watermark_lsn().expect("Validated previously");
1590 :
1591 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1592 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1593 0 : LOG_PACER.lock().unwrap().call(|| {
1594 0 : let num_keys = total_keyspace.total_raw_size();
1595 0 : let num_pages = results.len();
1596 0 : tracing::info!(
1597 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1598 : lsn = %max_request_lsn,
1599 0 : "Vectored read for {total_keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1600 : );
1601 0 : });
1602 312243 : }
1603 :
1604 : // Records the number of layers visited in a few different ways:
1605 : //
1606 : // * LAYERS_PER_READ: all layers count towards every read in the batch, because each
1607 : // layer directly affects its observed latency.
1608 : //
1609 : // * LAYERS_PER_READ_BATCH: all layers count towards each batch, to get the per-batch
1610 : // layer visits and access cost.
1611 : //
1612 : // * LAYERS_PER_READ_AMORTIZED: the average layer count per read, to get the amortized
1613 : // read amplification after batching.
1614 312243 : let layers_visited = layers_visited as f64;
1615 312243 : let avg_layers_visited = layers_visited / results.len() as f64;
1616 312243 : LAYERS_PER_READ_BATCH_GLOBAL.observe(layers_visited);
1617 675794 : for _ in &results {
1618 363551 : self.metrics.layers_per_read.observe(layers_visited);
1619 363551 : LAYERS_PER_READ_GLOBAL.observe(layers_visited);
1620 363551 : LAYERS_PER_READ_AMORTIZED_GLOBAL.observe(avg_layers_visited);
1621 363551 : }
1622 179 : }
1623 :
1624 312422 : Ok(results)
1625 312430 : }
1626 :
1627 : // A copy of the get_vectored_impl method except that we store the image and wal records into `reconstruct_state`.
1628 : // This is only used in the http getpage call for debugging purpose.
1629 0 : pub(super) async fn debug_get_vectored_impl(
1630 0 : &self,
1631 0 : query: VersionedKeySpaceQuery,
1632 0 : reconstruct_state: &mut ValuesReconstructState,
1633 0 : ctx: &RequestContext,
1634 0 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1635 0 : if query.is_empty() {
1636 0 : return Ok(BTreeMap::default());
1637 0 : }
1638 :
1639 0 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1640 0 : Some(ReadPath::new(
1641 0 : query.total_keyspace(),
1642 0 : query.high_watermark_lsn()?,
1643 : ))
1644 : } else {
1645 0 : None
1646 : };
1647 :
1648 0 : reconstruct_state.read_path = read_path;
1649 :
1650 0 : let traversal_res: Result<(), _> = self
1651 0 : .get_vectored_reconstruct_data(query.clone(), reconstruct_state, ctx)
1652 0 : .await;
1653 :
1654 0 : if let Err(err) = traversal_res {
1655 : // Wait for all the spawned IOs to complete.
1656 : // See comments on `spawn_io` inside `storage_layer` for more details.
1657 0 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1658 0 : .into_values()
1659 0 : .map(|state| state.collect_pending_ios())
1660 0 : .collect::<FuturesUnordered<_>>();
1661 0 : while collect_futs.next().await.is_some() {}
1662 0 : return Err(err);
1663 0 : };
1664 :
1665 0 : let reconstruct_state = Arc::new(Mutex::new(reconstruct_state));
1666 0 : let futs = FuturesUnordered::new();
1667 :
1668 0 : for (key, state) in std::mem::take(&mut reconstruct_state.lock().unwrap().keys) {
1669 0 : let req_lsn_for_key = query.map_key_to_lsn(&key);
1670 0 : futs.push({
1671 0 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1672 0 : let rc_clone = Arc::clone(&reconstruct_state);
1673 :
1674 0 : async move {
1675 0 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1676 :
1677 0 : let converted = match state.collect_pending_ios().await {
1678 0 : Ok(ok) => ok,
1679 0 : Err(err) => {
1680 0 : return (key, Err(err));
1681 : }
1682 : };
1683 0 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1684 :
1685 : // The walredo module expects the records to be descending in terms of Lsn.
1686 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1687 0 : debug_assert!(
1688 0 : converted
1689 0 : .records
1690 0 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1691 0 : "{converted:?}"
1692 : );
1693 0 : {
1694 0 : let mut guard = rc_clone.lock().unwrap();
1695 0 : guard.set_debug_state(&converted);
1696 0 : }
1697 : (
1698 0 : key,
1699 0 : walredo_self
1700 0 : .reconstruct_value(
1701 0 : key,
1702 0 : req_lsn_for_key,
1703 0 : converted,
1704 0 : RedoAttemptType::ReadPage,
1705 0 : )
1706 0 : .await,
1707 : )
1708 0 : }
1709 : });
1710 : }
1711 :
1712 0 : let results = futs
1713 0 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1714 0 : .await;
1715 :
1716 0 : Ok(results)
1717 0 : }
1718 :
1719 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1720 138202 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1721 138202 : self.last_record_lsn.load().last
1722 138202 : }
1723 :
1724 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1725 0 : self.last_record_lsn.load().prev
1726 0 : }
1727 :
1728 : /// Atomically get both last and prev.
1729 117 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1730 117 : self.last_record_lsn.load()
1731 117 : }
1732 :
1733 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1734 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1735 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1736 0 : self.last_record_lsn.status_receiver()
1737 0 : }
1738 :
1739 610 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1740 610 : self.disk_consistent_lsn.load()
1741 610 : }
1742 :
1743 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1744 : /// not validated with control plane yet.
1745 : /// See [`Self::get_remote_consistent_lsn_visible`].
1746 2 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1747 2 : self.remote_client.remote_consistent_lsn_projected()
1748 2 : }
1749 :
1750 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1751 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1752 : /// generation validation in the deletion queue.
1753 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1754 0 : self.remote_client.remote_consistent_lsn_visible()
1755 0 : }
1756 :
1757 : /// The sum of the file size of all historic layers in the layer map.
1758 : /// This method makes no distinction between local and remote layers.
1759 : /// Hence, the result **does not represent local filesystem usage**.
1760 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1761 0 : let guard = self
1762 0 : .layers
1763 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
1764 0 : .await;
1765 0 : guard.layer_size_sum()
1766 0 : }
1767 :
1768 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1769 0 : self.metrics.resident_physical_size_get()
1770 0 : }
1771 :
1772 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1773 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1774 0 : }
1775 :
1776 : ///
1777 : /// Wait until WAL has been received and processed up to this LSN.
1778 : ///
1779 : /// You should call this before any of the other get_* or list_* functions. Calling
1780 : /// those functions with an LSN that has been processed yet is an error.
1781 : ///
1782 114505 : pub(crate) async fn wait_lsn(
1783 114505 : &self,
1784 114505 : lsn: Lsn,
1785 114505 : who_is_waiting: WaitLsnWaiter<'_>,
1786 114505 : timeout: WaitLsnTimeout,
1787 114505 : ctx: &RequestContext, /* Prepare for use by cancellation */
1788 114505 : ) -> Result<(), WaitLsnError> {
1789 114505 : let state = self.current_state();
1790 114505 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1791 0 : return Err(WaitLsnError::Shutdown);
1792 114505 : } else if !matches!(state, TimelineState::Active) {
1793 0 : return Err(WaitLsnError::BadState(state));
1794 114505 : }
1795 :
1796 114505 : if cfg!(debug_assertions) {
1797 114505 : match ctx.task_kind() {
1798 : TaskKind::WalReceiverManager
1799 : | TaskKind::WalReceiverConnectionHandler
1800 : | TaskKind::WalReceiverConnectionPoller => {
1801 0 : let is_myself = match who_is_waiting {
1802 0 : WaitLsnWaiter::Timeline(waiter) => {
1803 0 : Weak::ptr_eq(&waiter.myself, &self.myself)
1804 : }
1805 : WaitLsnWaiter::Tenant
1806 : | WaitLsnWaiter::PageService
1807 : | WaitLsnWaiter::HttpEndpoint
1808 0 : | WaitLsnWaiter::BaseBackupCache => unreachable!(
1809 : "tenant or page_service context are not expected to have task kind {:?}",
1810 0 : ctx.task_kind()
1811 : ),
1812 : };
1813 0 : if is_myself {
1814 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1815 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1816 0 : panic!(
1817 0 : "this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock"
1818 : );
1819 0 : }
1820 0 : } else {
1821 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1822 0 : // our walreceiver task can make progress independent of theirs
1823 0 : }
1824 : }
1825 114505 : _ => {}
1826 : }
1827 0 : }
1828 :
1829 114505 : let timeout = match timeout {
1830 0 : WaitLsnTimeout::Custom(t) => t,
1831 114505 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1832 : };
1833 :
1834 114505 : let timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1835 114505 : let start_finish_counterpair_guard = self.metrics.wait_lsn_start_finish_counterpair.guard();
1836 :
1837 114505 : let wait_for_timeout = self.last_record_lsn.wait_for_timeout(lsn, timeout);
1838 114505 : let wait_for_timeout = std::pin::pin!(wait_for_timeout);
1839 : // Use threshold of 1 because even 1 second of wait for ingest is very much abnormal.
1840 114505 : let log_slow_threshold = Duration::from_secs(1);
1841 : // Use period of 10 to avoid flooding logs during an outage that affects all timelines.
1842 114505 : let log_slow_period = Duration::from_secs(10);
1843 114505 : let mut logging_permit = None;
1844 114505 : let wait_for_timeout = monitor_slow_future(
1845 114505 : log_slow_threshold,
1846 114505 : log_slow_period,
1847 114505 : wait_for_timeout,
1848 : |MonitorSlowFutureCallback {
1849 : ready,
1850 : is_slow,
1851 : elapsed_total,
1852 : elapsed_since_last_callback,
1853 114505 : }| {
1854 114505 : self.metrics
1855 114505 : .wait_lsn_in_progress_micros
1856 114505 : .inc_by(u64::try_from(elapsed_since_last_callback.as_micros()).unwrap());
1857 114505 : if !is_slow {
1858 114505 : return;
1859 0 : }
1860 : // It's slow, see if we should log it.
1861 : // (We limit the logging to one per invocation per timeline to avoid excessive
1862 : // logging during an extended broker / networking outage that affects all timelines.)
1863 0 : if logging_permit.is_none() {
1864 0 : logging_permit = self.wait_lsn_log_slow.try_acquire().ok();
1865 0 : }
1866 0 : if logging_permit.is_none() {
1867 0 : return;
1868 0 : }
1869 : // We log it.
1870 0 : if ready {
1871 0 : info!(
1872 0 : "slow wait_lsn completed after {:.3}s",
1873 0 : elapsed_total.as_secs_f64()
1874 : );
1875 : } else {
1876 0 : info!(
1877 0 : "slow wait_lsn still running for {:.3}s",
1878 0 : elapsed_total.as_secs_f64()
1879 : );
1880 : }
1881 114505 : },
1882 : );
1883 114505 : let res = wait_for_timeout.await;
1884 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1885 114505 : drop(logging_permit);
1886 114505 : drop(start_finish_counterpair_guard);
1887 114505 : drop(timer);
1888 114505 : match res {
1889 114505 : Ok(()) => Ok(()),
1890 0 : Err(e) => {
1891 : use utils::seqwait::SeqWaitError::*;
1892 0 : match e {
1893 0 : Shutdown => Err(WaitLsnError::Shutdown),
1894 : Timeout => {
1895 0 : let walreceiver_status = self.walreceiver_status();
1896 0 : Err(WaitLsnError::Timeout(format!(
1897 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1898 0 : lsn,
1899 0 : self.get_last_record_lsn(),
1900 0 : self.get_disk_consistent_lsn(),
1901 0 : walreceiver_status,
1902 0 : )))
1903 : }
1904 : }
1905 : }
1906 : }
1907 114505 : }
1908 :
1909 0 : pub(crate) fn walreceiver_status(&self) -> String {
1910 0 : match &*self.walreceiver.lock().unwrap() {
1911 0 : None => "stopping or stopped".to_string(),
1912 0 : Some(walreceiver) => match walreceiver.status() {
1913 0 : Some(status) => status.to_human_readable_string(),
1914 0 : None => "Not active".to_string(),
1915 : },
1916 : }
1917 0 : }
1918 :
1919 : /// Check that it is valid to request operations with that lsn.
1920 119 : pub(crate) fn check_lsn_is_in_scope(
1921 119 : &self,
1922 119 : lsn: Lsn,
1923 119 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1924 119 : ) -> anyhow::Result<()> {
1925 119 : ensure!(
1926 119 : lsn >= **latest_gc_cutoff_lsn,
1927 2 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1928 : lsn,
1929 2 : **latest_gc_cutoff_lsn,
1930 : );
1931 117 : Ok(())
1932 119 : }
1933 :
1934 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1935 5 : pub(crate) fn init_lsn_lease(
1936 5 : &self,
1937 5 : lsn: Lsn,
1938 5 : length: Duration,
1939 5 : ctx: &RequestContext,
1940 5 : ) -> anyhow::Result<LsnLease> {
1941 5 : self.make_lsn_lease(lsn, length, true, ctx)
1942 5 : }
1943 :
1944 : /// 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.
1945 2 : pub(crate) fn renew_lsn_lease(
1946 2 : &self,
1947 2 : lsn: Lsn,
1948 2 : length: Duration,
1949 2 : ctx: &RequestContext,
1950 2 : ) -> anyhow::Result<LsnLease> {
1951 2 : self.make_lsn_lease(lsn, length, false, ctx)
1952 2 : }
1953 :
1954 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1955 : ///
1956 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1957 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1958 : ///
1959 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1960 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1961 7 : fn make_lsn_lease(
1962 7 : &self,
1963 7 : lsn: Lsn,
1964 7 : length: Duration,
1965 7 : init: bool,
1966 7 : _ctx: &RequestContext,
1967 7 : ) -> anyhow::Result<LsnLease> {
1968 6 : let lease = {
1969 : // Normalize the requested LSN to be aligned, and move to the first record
1970 : // if it points to the beginning of the page (header).
1971 7 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1972 :
1973 7 : let mut gc_info = self.gc_info.write().unwrap();
1974 7 : let planned_cutoff = gc_info.min_cutoff();
1975 :
1976 7 : let valid_until = SystemTime::now() + length;
1977 :
1978 7 : let entry = gc_info.leases.entry(lsn);
1979 :
1980 7 : match entry {
1981 3 : Entry::Occupied(mut occupied) => {
1982 3 : let existing_lease = occupied.get_mut();
1983 3 : if valid_until > existing_lease.valid_until {
1984 1 : existing_lease.valid_until = valid_until;
1985 1 : let dt: DateTime<Utc> = valid_until.into();
1986 1 : info!("lease extended to {}", dt);
1987 : } else {
1988 2 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1989 2 : info!("existing lease covers greater length, valid until {}", dt);
1990 : }
1991 :
1992 3 : existing_lease.clone()
1993 : }
1994 4 : Entry::Vacant(vacant) => {
1995 : // Never allow a lease to be requested for an LSN below the applied GC cutoff. The data could have been deleted.
1996 4 : let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
1997 4 : if lsn < *latest_gc_cutoff_lsn {
1998 1 : bail!(
1999 1 : "tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}",
2000 : lsn,
2001 1 : *latest_gc_cutoff_lsn
2002 : );
2003 3 : }
2004 :
2005 : // We allow create lease for those below the planned gc cutoff if we are still within the grace period
2006 : // of GC blocking.
2007 3 : let validate = {
2008 3 : let conf = self.tenant_conf.load();
2009 3 : !conf.is_gc_blocked_by_lsn_lease_deadline()
2010 : };
2011 :
2012 : // Do not allow initial lease creation to be below the planned gc cutoff. The client (compute_ctl) determines
2013 : // whether it is a initial lease creation or a renewal.
2014 3 : if (init || validate) && lsn < planned_cutoff {
2015 0 : bail!(
2016 0 : "tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}",
2017 : lsn,
2018 : planned_cutoff
2019 : );
2020 3 : }
2021 :
2022 3 : let dt: DateTime<Utc> = valid_until.into();
2023 3 : info!("lease created, valid until {}", dt);
2024 3 : vacant.insert(LsnLease { valid_until }).clone()
2025 : }
2026 : }
2027 : };
2028 :
2029 6 : Ok(lease)
2030 7 : }
2031 :
2032 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
2033 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
2034 : #[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))]
2035 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
2036 : self.freeze0().await
2037 : }
2038 :
2039 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
2040 : #[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))]
2041 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
2042 : self.freeze_and_flush0().await
2043 : }
2044 :
2045 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
2046 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
2047 571 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
2048 571 : let mut g = self.write_lock.lock().await;
2049 571 : let to_lsn = self.get_last_record_lsn();
2050 571 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
2051 571 : }
2052 :
2053 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
2054 : // polluting the span hierarchy.
2055 571 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
2056 571 : let token = self.freeze0().await?;
2057 571 : self.wait_flush_completion(token).await
2058 571 : }
2059 :
2060 : // Check if an open ephemeral layer should be closed: this provides
2061 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
2062 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
2063 : // ephemeral layer bytes has been breached.
2064 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
2065 0 : debug_assert_current_span_has_tenant_and_timeline_id();
2066 :
2067 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
2068 : // If the write lock is held, there is an active wal receiver: rolling open layers
2069 : // is their responsibility while they hold this lock.
2070 0 : return;
2071 : };
2072 :
2073 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
2074 : // time, and this was missed.
2075 : // if write_guard.is_none() { return; }
2076 :
2077 0 : let Ok(layers_guard) = self.layers.try_read(LayerManagerLockHolder::TryFreezeLayer) else {
2078 : // Don't block if the layer lock is busy
2079 0 : return;
2080 : };
2081 :
2082 0 : let Ok(lm) = layers_guard.layer_map() else {
2083 0 : return;
2084 : };
2085 :
2086 0 : let Some(open_layer) = &lm.open_layer else {
2087 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
2088 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
2089 : // that didn't result in writes to this shard.
2090 :
2091 : // Must not hold the layers lock while waiting for a flush.
2092 0 : drop(layers_guard);
2093 :
2094 0 : let last_record_lsn = self.get_last_record_lsn();
2095 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
2096 0 : if last_record_lsn > disk_consistent_lsn {
2097 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
2098 : // we are a sharded tenant and have skipped some WAL
2099 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
2100 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
2101 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
2102 : // without any data ingested (yet) doesn't write a remote index as soon as it
2103 : // sees its LSN advance: we only do this if we've been layer-less
2104 : // for some time.
2105 0 : tracing::debug!(
2106 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
2107 : disk_consistent_lsn,
2108 : last_record_lsn
2109 : );
2110 :
2111 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
2112 : // We know there is no open layer, so we can request freezing without actually
2113 : // freezing anything. This is true even if we have dropped the layers_guard, we
2114 : // still hold the write_guard.
2115 0 : let _ = async {
2116 0 : let token = self
2117 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
2118 0 : .await?;
2119 0 : self.wait_flush_completion(token).await
2120 0 : }
2121 0 : .await;
2122 0 : }
2123 0 : }
2124 :
2125 0 : return;
2126 : };
2127 :
2128 0 : let current_size = open_layer.len();
2129 :
2130 0 : let current_lsn = self.get_last_record_lsn();
2131 :
2132 0 : let checkpoint_distance_override = open_layer.tick();
2133 :
2134 0 : if let Some(size_override) = checkpoint_distance_override {
2135 0 : if current_size > size_override {
2136 : // This is not harmful, but it only happens in relatively rare cases where
2137 : // time-based checkpoints are not happening fast enough to keep the amount of
2138 : // ephemeral data within configured limits. It's a sign of stress on the system.
2139 0 : tracing::info!(
2140 0 : "Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure"
2141 : );
2142 0 : }
2143 0 : }
2144 :
2145 0 : let checkpoint_distance =
2146 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
2147 :
2148 0 : if self.should_roll(
2149 0 : current_size,
2150 0 : current_size,
2151 0 : checkpoint_distance,
2152 0 : self.get_last_record_lsn(),
2153 0 : self.last_freeze_at.load(),
2154 0 : open_layer.get_opened_at(),
2155 : ) {
2156 0 : match open_layer.info() {
2157 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
2158 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
2159 : // happens asynchronously in the background.
2160 0 : tracing::debug!(
2161 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
2162 : );
2163 : }
2164 : InMemoryLayerInfo::Open { .. } => {
2165 : // Upgrade to a write lock and freeze the layer
2166 0 : drop(layers_guard);
2167 0 : let res = self
2168 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
2169 0 : .await;
2170 :
2171 0 : if let Err(e) = res {
2172 0 : tracing::info!(
2173 0 : "failed to flush frozen layer after background freeze: {e:#}"
2174 : );
2175 0 : }
2176 : }
2177 : }
2178 0 : }
2179 0 : }
2180 :
2181 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
2182 : ///
2183 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
2184 : /// child timelines that are not offloaded yet.
2185 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
2186 0 : if self.remote_client.is_archived() != Some(true) {
2187 0 : return (false, "the timeline is not archived");
2188 0 : }
2189 0 : if !self.remote_client.no_pending_work() {
2190 : // if the remote client is still processing some work, we can't offload
2191 0 : return (false, "the upload queue is not drained yet");
2192 0 : }
2193 :
2194 0 : (true, "ok")
2195 0 : }
2196 :
2197 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
2198 : /// compaction tasks.
2199 192 : pub(crate) async fn compact(
2200 192 : self: &Arc<Self>,
2201 192 : cancel: &CancellationToken,
2202 192 : flags: EnumSet<CompactFlags>,
2203 192 : ctx: &RequestContext,
2204 192 : ) -> Result<CompactionOutcome, CompactionError> {
2205 192 : let res = self
2206 192 : .compact_with_options(
2207 192 : cancel,
2208 192 : CompactOptions {
2209 192 : flags,
2210 192 : compact_key_range: None,
2211 192 : compact_lsn_range: None,
2212 192 : sub_compaction: false,
2213 192 : sub_compaction_max_job_size_mb: None,
2214 192 : gc_compaction_do_metadata_compaction: false,
2215 192 : },
2216 192 : ctx,
2217 192 : )
2218 192 : .await;
2219 192 : if let Err(err) = &res {
2220 0 : log_compaction_error(err, None, cancel.is_cancelled(), false);
2221 192 : }
2222 192 : res
2223 192 : }
2224 :
2225 : /// Outermost timeline compaction operation; downloads needed layers.
2226 : ///
2227 : /// NB: the cancellation token is usually from a background task, but can also come from a
2228 : /// request task.
2229 192 : pub(crate) async fn compact_with_options(
2230 192 : self: &Arc<Self>,
2231 192 : cancel: &CancellationToken,
2232 192 : options: CompactOptions,
2233 192 : ctx: &RequestContext,
2234 192 : ) -> Result<CompactionOutcome, CompactionError> {
2235 : // Acquire the compaction lock and task semaphore.
2236 : //
2237 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
2238 : // out by other background tasks (including image compaction). We request this via
2239 : // `BackgroundLoopKind::L0Compaction`.
2240 : //
2241 : // Yield for pending L0 compaction while waiting for the semaphore.
2242 192 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
2243 192 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
2244 0 : true => BackgroundLoopKind::L0Compaction,
2245 192 : false => BackgroundLoopKind::Compaction,
2246 : };
2247 192 : let yield_for_l0 = options.flags.contains(CompactFlags::YieldForL0);
2248 192 : if yield_for_l0 {
2249 : // If this is an L0 pass, it doesn't make sense to yield for L0.
2250 0 : debug_assert!(!is_l0_only, "YieldForL0 during L0 pass");
2251 : // If `compaction_l0_first` is disabled, there's no point yielding.
2252 0 : debug_assert!(self.get_compaction_l0_first(), "YieldForL0 without L0 pass");
2253 192 : }
2254 :
2255 192 : let acquire = async move {
2256 192 : let guard = self.compaction_lock.lock().await;
2257 192 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
2258 192 : (guard, permit)
2259 192 : };
2260 :
2261 192 : let (_guard, _permit) = tokio::select! {
2262 192 : (guard, permit) = acquire => (guard, permit),
2263 192 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
2264 0 : return Ok(CompactionOutcome::YieldForL0);
2265 : }
2266 192 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2267 192 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2268 : };
2269 :
2270 192 : let last_record_lsn = self.get_last_record_lsn();
2271 :
2272 : // Last record Lsn could be zero in case the timeline was just created
2273 192 : if !last_record_lsn.is_valid() {
2274 0 : warn!(
2275 0 : "Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}"
2276 : );
2277 0 : return Ok(CompactionOutcome::Skipped);
2278 192 : }
2279 :
2280 192 : let result = match self.get_compaction_algorithm_settings().kind {
2281 : CompactionAlgorithm::Tiered => {
2282 0 : self.compact_tiered(cancel, ctx).await?;
2283 0 : Ok(CompactionOutcome::Done)
2284 : }
2285 192 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
2286 : };
2287 :
2288 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
2289 0 : match &result {
2290 192 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
2291 0 : Err(e) if e.is_cancel() => {}
2292 0 : Err(_) => self.compaction_failed.store(true, AtomicOrdering::Relaxed),
2293 : };
2294 :
2295 192 : result
2296 192 : }
2297 :
2298 : /// Mutate the timeline with a [`TimelineWriter`].
2299 2566609 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
2300 : TimelineWriter {
2301 2566609 : tl: self,
2302 2566609 : write_guard: self.write_lock.lock().await,
2303 : }
2304 2566609 : }
2305 :
2306 0 : pub(crate) fn activate(
2307 0 : self: &Arc<Self>,
2308 0 : parent: Arc<crate::tenant::TenantShard>,
2309 0 : broker_client: BrokerClientChannel,
2310 0 : background_jobs_can_start: Option<&completion::Barrier>,
2311 0 : ctx: &RequestContext,
2312 0 : ) {
2313 0 : if self.tenant_shard_id.is_shard_zero() {
2314 0 : // Logical size is only maintained accurately on shard zero.
2315 0 : self.spawn_initial_logical_size_computation_task(ctx);
2316 0 : }
2317 0 : self.launch_wal_receiver(ctx, broker_client);
2318 0 : self.set_state(TimelineState::Active);
2319 0 : self.launch_eviction_task(parent, background_jobs_can_start);
2320 0 : }
2321 :
2322 : /// After this function returns, there are no timeline-scoped tasks are left running.
2323 : ///
2324 : /// The preferred pattern for is:
2325 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
2326 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
2327 : /// go the extra mile and keep track of JoinHandles
2328 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
2329 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
2330 : ///
2331 : /// For legacy reasons, we still have multiple tasks spawned using
2332 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
2333 : /// We refer to these as "timeline-scoped task_mgr tasks".
2334 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
2335 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
2336 : /// or [`task_mgr::shutdown_watcher`].
2337 : /// We want to gradually convert the code base away from these.
2338 : ///
2339 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
2340 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
2341 : /// ones that aren't mentioned here):
2342 : /// - [`TaskKind::TimelineDeletionWorker`]
2343 : /// - NB: also used for tenant deletion
2344 : /// - [`TaskKind::RemoteUploadTask`]`
2345 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
2346 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
2347 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
2348 : /// - [`TaskKind::Eviction`]
2349 : /// - [`TaskKind::LayerFlushTask`]
2350 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
2351 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
2352 5 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
2353 5 : debug_assert_current_span_has_tenant_and_timeline_id();
2354 :
2355 : // Regardless of whether we're going to try_freeze_and_flush
2356 : // cancel walreceiver to stop ingesting more data asap.
2357 : //
2358 : // Note that we're accepting a race condition here where we may
2359 : // do the final flush below, before walreceiver observes the
2360 : // cancellation and exits.
2361 : // This means we may open a new InMemoryLayer after the final flush below.
2362 : // Flush loop is also still running for a short while, so, in theory, it
2363 : // could also make its way into the upload queue.
2364 : //
2365 : // If we wait for the shutdown of the walreceiver before moving on to the
2366 : // flush, then that would be avoided. But we don't do it because the
2367 : // walreceiver entertains reads internally, which means that it possibly
2368 : // depends on the download of layers. Layer download is only sensitive to
2369 : // the cancellation of the entire timeline, so cancelling the walreceiver
2370 : // will have no effect on the individual get requests.
2371 : // This would cause problems when there is a lot of ongoing downloads or
2372 : // there is S3 unavailabilities, i.e. detach, deletion, etc would hang,
2373 : // and we can't deallocate resources of the timeline, etc.
2374 5 : let walreceiver = self.walreceiver.lock().unwrap().take();
2375 5 : tracing::debug!(
2376 0 : is_some = walreceiver.is_some(),
2377 0 : "Waiting for WalReceiverManager..."
2378 : );
2379 5 : if let Some(walreceiver) = walreceiver {
2380 0 : walreceiver.cancel().await;
2381 5 : }
2382 : // ... and inform any waiters for newer LSNs that there won't be any.
2383 5 : self.last_record_lsn.shutdown();
2384 :
2385 5 : if let ShutdownMode::FreezeAndFlush = mode {
2386 3 : let do_flush = if let Some((open, frozen)) = self
2387 3 : .layers
2388 3 : .read(LayerManagerLockHolder::Shutdown)
2389 3 : .await
2390 3 : .layer_map()
2391 3 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
2392 3 : .ok()
2393 3 : .filter(|(open, frozen)| *open || *frozen > 0)
2394 : {
2395 0 : if self.remote_client.is_archived() == Some(true) {
2396 : // No point flushing on shutdown for an archived timeline: it is not important
2397 : // to have it nice and fresh after our restart, and trying to flush here might
2398 : // race with trying to offload it (which also stops the flush loop)
2399 0 : false
2400 : } else {
2401 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
2402 0 : true
2403 : }
2404 : } else {
2405 : // this is double-shutdown, it'll be a no-op
2406 3 : true
2407 : };
2408 :
2409 : // we shut down walreceiver above, so, we won't add anything more
2410 : // to the InMemoryLayer; freeze it and wait for all frozen layers
2411 : // to reach the disk & upload queue, then shut the upload queue and
2412 : // wait for it to drain.
2413 3 : if do_flush {
2414 3 : match self.freeze_and_flush().await {
2415 : Ok(_) => {
2416 : // drain the upload queue
2417 : // if we did not wait for completion here, it might be our shutdown process
2418 : // didn't wait for remote uploads to complete at all, as new tasks can forever
2419 : // be spawned.
2420 : //
2421 : // what is problematic is the shutting down of RemoteTimelineClient, because
2422 : // obviously it does not make sense to stop while we wait for it, but what
2423 : // about corner cases like s3 suddenly hanging up?
2424 3 : self.remote_client.shutdown().await;
2425 : }
2426 : Err(FlushLayerError::Cancelled) => {
2427 : // this is likely the second shutdown, ignore silently.
2428 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2429 0 : debug_assert!(self.cancel.is_cancelled());
2430 : }
2431 0 : Err(e) => {
2432 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
2433 : // we have some extra WAL replay to do next time the timeline starts.
2434 0 : warn!("failed to freeze and flush: {e:#}");
2435 : }
2436 : }
2437 :
2438 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2439 : // we also do a final check here to ensure that the queue is empty.
2440 3 : if !self.remote_client.no_pending_work() {
2441 0 : warn!(
2442 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2443 : );
2444 3 : }
2445 0 : }
2446 2 : }
2447 :
2448 5 : if let ShutdownMode::Reload = mode {
2449 : // drain the upload queue
2450 1 : self.remote_client.shutdown().await;
2451 1 : if !self.remote_client.no_pending_work() {
2452 0 : warn!(
2453 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2454 : );
2455 1 : }
2456 4 : }
2457 :
2458 : // Signal any subscribers to our cancellation token to drop out
2459 5 : tracing::debug!("Cancelling CancellationToken");
2460 5 : self.cancel.cancel();
2461 :
2462 : // If we have a background task downloading heatmap layers stop it.
2463 : // The background downloads are sensitive to timeline cancellation (done above),
2464 : // so the drain will be immediate.
2465 5 : self.stop_and_drain_heatmap_layers_download().await;
2466 :
2467 : // Ensure Prevent new page service requests from starting.
2468 5 : self.handles.shutdown();
2469 :
2470 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2471 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2472 5 : self.remote_client.stop();
2473 :
2474 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2475 : // to shut down the upload queue tasks.
2476 : // TODO: fix that, task management should be encapsulated inside remote_client.
2477 5 : task_mgr::shutdown_tasks(
2478 5 : Some(TaskKind::RemoteUploadTask),
2479 5 : Some(self.tenant_shard_id),
2480 5 : Some(self.timeline_id),
2481 5 : )
2482 5 : .await;
2483 :
2484 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2485 5 : tracing::debug!("Waiting for tasks...");
2486 5 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2487 :
2488 : {
2489 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2490 : // open.
2491 5 : let mut write_guard = self.write_lock.lock().await;
2492 5 : self.layers
2493 5 : .write(LayerManagerLockHolder::Shutdown)
2494 5 : .await
2495 5 : .shutdown(&mut write_guard);
2496 : }
2497 :
2498 : // Finally wait until any gate-holders are complete.
2499 : //
2500 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2501 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2502 5 : self.gate.close().await;
2503 :
2504 5 : self.metrics.shutdown();
2505 5 : }
2506 :
2507 236 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2508 236 : match (self.current_state(), new_state) {
2509 236 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2510 1 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2511 : }
2512 0 : (st, TimelineState::Loading) => {
2513 0 : error!("ignoring transition from {st:?} into Loading state");
2514 : }
2515 0 : (TimelineState::Broken { .. }, new_state) => {
2516 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2517 : }
2518 : (TimelineState::Stopping, TimelineState::Active) => {
2519 0 : error!("Not activating a Stopping timeline");
2520 : }
2521 235 : (_, new_state) => {
2522 235 : self.state.send_replace(new_state);
2523 235 : }
2524 : }
2525 236 : }
2526 :
2527 1 : pub(crate) fn set_broken(&self, reason: String) {
2528 1 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2529 1 : let broken_state = TimelineState::Broken {
2530 1 : reason,
2531 1 : backtrace: backtrace_str,
2532 1 : };
2533 1 : self.set_state(broken_state);
2534 :
2535 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2536 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2537 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2538 1 : self.cancel.cancel();
2539 1 : }
2540 :
2541 115436 : pub(crate) fn current_state(&self) -> TimelineState {
2542 115436 : self.state.borrow().clone()
2543 115436 : }
2544 :
2545 3 : pub(crate) fn is_broken(&self) -> bool {
2546 3 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2547 3 : }
2548 :
2549 126 : pub(crate) fn is_active(&self) -> bool {
2550 126 : self.current_state() == TimelineState::Active
2551 126 : }
2552 :
2553 8 : pub(crate) fn is_archived(&self) -> Option<bool> {
2554 8 : self.remote_client.is_archived()
2555 8 : }
2556 :
2557 8 : pub(crate) fn is_invisible(&self) -> Option<bool> {
2558 8 : self.remote_client.is_invisible()
2559 8 : }
2560 :
2561 569 : pub(crate) fn is_stopping(&self) -> bool {
2562 569 : self.current_state() == TimelineState::Stopping
2563 569 : }
2564 :
2565 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2566 0 : self.state.subscribe()
2567 0 : }
2568 :
2569 114506 : pub(crate) async fn wait_to_become_active(
2570 114506 : &self,
2571 114506 : _ctx: &RequestContext, // Prepare for use by cancellation
2572 114506 : ) -> Result<(), TimelineState> {
2573 114506 : let mut receiver = self.state.subscribe();
2574 : loop {
2575 114506 : let current_state = receiver.borrow().clone();
2576 114506 : match current_state {
2577 : TimelineState::Loading => {
2578 0 : receiver
2579 0 : .changed()
2580 0 : .await
2581 0 : .expect("holding a reference to self");
2582 : }
2583 : TimelineState::Active => {
2584 114505 : return Ok(());
2585 : }
2586 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2587 : // There's no chance the timeline can transition back into ::Active
2588 1 : return Err(current_state);
2589 : }
2590 : }
2591 : }
2592 114506 : }
2593 :
2594 0 : pub(crate) async fn layer_map_info(
2595 0 : &self,
2596 0 : reset: LayerAccessStatsReset,
2597 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2598 0 : let guard = self
2599 0 : .layers
2600 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
2601 0 : .await;
2602 0 : let layer_map = guard.layer_map()?;
2603 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2604 0 : if let Some(open_layer) = &layer_map.open_layer {
2605 0 : in_memory_layers.push(open_layer.info());
2606 0 : }
2607 0 : for frozen_layer in &layer_map.frozen_layers {
2608 0 : in_memory_layers.push(frozen_layer.info());
2609 0 : }
2610 :
2611 0 : let historic_layers = layer_map
2612 0 : .iter_historic_layers()
2613 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2614 0 : .collect();
2615 :
2616 0 : Ok(LayerMapInfo {
2617 0 : in_memory_layers,
2618 0 : historic_layers,
2619 0 : })
2620 0 : }
2621 :
2622 : #[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))]
2623 : pub(crate) async fn download_layer(
2624 : &self,
2625 : layer_file_name: &LayerName,
2626 : ctx: &RequestContext,
2627 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2628 : let Some(layer) = self
2629 : .find_layer(layer_file_name)
2630 : .await
2631 0 : .map_err(|e| match e {
2632 : layer_manager::Shutdown => {
2633 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2634 : }
2635 0 : })?
2636 : else {
2637 : return Ok(None);
2638 : };
2639 :
2640 : layer.download(ctx).await?;
2641 :
2642 : Ok(Some(true))
2643 : }
2644 :
2645 : /// Evict just one layer.
2646 : ///
2647 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2648 0 : pub(crate) async fn evict_layer(
2649 0 : &self,
2650 0 : layer_file_name: &LayerName,
2651 0 : ) -> anyhow::Result<Option<bool>> {
2652 0 : let _gate = self
2653 0 : .gate
2654 0 : .enter()
2655 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2656 :
2657 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2658 0 : return Ok(None);
2659 : };
2660 :
2661 : // curl has this by default
2662 0 : let timeout = std::time::Duration::from_secs(120);
2663 :
2664 0 : match local_layer.evict_and_wait(timeout).await {
2665 0 : Ok(()) => Ok(Some(true)),
2666 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2667 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2668 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2669 : }
2670 0 : }
2671 :
2672 2401506 : fn should_roll(
2673 2401506 : &self,
2674 2401506 : layer_size: u64,
2675 2401506 : projected_layer_size: u64,
2676 2401506 : checkpoint_distance: u64,
2677 2401506 : projected_lsn: Lsn,
2678 2401506 : last_freeze_at: Lsn,
2679 2401506 : opened_at: Instant,
2680 2401506 : ) -> bool {
2681 2401506 : let distance = projected_lsn.widening_sub(last_freeze_at);
2682 :
2683 : // Rolling the open layer can be triggered by:
2684 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2685 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2686 : // account for how writes are distributed across shards: we expect each node to consume
2687 : // 1/count of the LSN on average.
2688 : // 2. The size of the currently open layer.
2689 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2690 : // up and suspend activity.
2691 2401506 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2692 0 : info!(
2693 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2694 : projected_lsn, layer_size, distance
2695 : );
2696 :
2697 0 : true
2698 2401506 : } else if projected_layer_size >= checkpoint_distance {
2699 : // NB: this check is relied upon by:
2700 40 : let _ = IndexEntry::validate_checkpoint_distance;
2701 40 : info!(
2702 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2703 : projected_lsn, layer_size, projected_layer_size
2704 : );
2705 :
2706 40 : true
2707 2401466 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2708 0 : info!(
2709 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2710 : projected_lsn,
2711 : layer_size,
2712 0 : opened_at.elapsed()
2713 : );
2714 :
2715 0 : true
2716 : } else {
2717 2401466 : false
2718 : }
2719 2401506 : }
2720 :
2721 1 : pub(crate) fn is_basebackup_cache_enabled(&self) -> bool {
2722 1 : let tenant_conf = self.tenant_conf.load();
2723 1 : tenant_conf
2724 1 : .tenant_conf
2725 1 : .basebackup_cache_enabled
2726 1 : .unwrap_or(self.conf.default_tenant_conf.basebackup_cache_enabled)
2727 1 : }
2728 :
2729 : /// Try to get a basebackup from the on-disk cache.
2730 0 : pub(crate) async fn get_cached_basebackup(&self, lsn: Lsn) -> Option<tokio::fs::File> {
2731 0 : self.basebackup_cache
2732 0 : .get(self.tenant_shard_id.tenant_id, self.timeline_id, lsn)
2733 0 : .await
2734 0 : }
2735 :
2736 : /// Convenience method to attempt fetching a basebackup for the timeline if enabled and safe for
2737 : /// the given request parameters.
2738 : ///
2739 : /// TODO: consider moving this onto GrpcPageServiceHandler once the libpq handler is gone.
2740 0 : pub async fn get_cached_basebackup_if_enabled(
2741 0 : &self,
2742 0 : lsn: Option<Lsn>,
2743 0 : prev_lsn: Option<Lsn>,
2744 0 : full: bool,
2745 0 : replica: bool,
2746 0 : gzip: bool,
2747 0 : ) -> Option<tokio::fs::File> {
2748 0 : if !self.is_basebackup_cache_enabled() || !self.basebackup_cache.is_enabled() {
2749 0 : return None;
2750 0 : }
2751 : // We have to know which LSN to fetch the basebackup for.
2752 0 : let lsn = lsn?;
2753 : // We only cache gzipped, non-full basebackups for primary computes with automatic prev_lsn.
2754 0 : if prev_lsn.is_some() || full || replica || !gzip {
2755 0 : return None;
2756 0 : }
2757 0 : self.get_cached_basebackup(lsn).await
2758 0 : }
2759 :
2760 : /// Prepare basebackup for the given LSN and store it in the basebackup cache.
2761 : /// The method is asynchronous and returns immediately.
2762 : /// The actual basebackup preparation is performed in the background
2763 : /// by the basebackup cache on a best-effort basis.
2764 1 : pub(crate) fn prepare_basebackup(&self, lsn: Lsn) {
2765 1 : if !self.is_basebackup_cache_enabled() {
2766 1 : return;
2767 0 : }
2768 0 : if !self.tenant_shard_id.is_shard_zero() {
2769 : // In theory we should never get here, but just in case check it.
2770 : // Preparing basebackup doesn't make sense for shards other than shard zero.
2771 0 : return;
2772 0 : }
2773 0 : if !self.is_active() {
2774 : // May happen during initial timeline creation.
2775 : // Such timeline is not in the global timeline map yet,
2776 : // so basebackup cache will not be able to find it.
2777 : // TODO(diko): We can prepare such timelines in finish_creation().
2778 0 : return;
2779 0 : }
2780 :
2781 0 : self.basebackup_cache
2782 0 : .send_prepare(self.tenant_shard_id, self.timeline_id, lsn);
2783 1 : }
2784 : }
2785 :
2786 : /// Number of times we will compute partition within a checkpoint distance.
2787 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2788 :
2789 : // Private functions
2790 : impl Timeline {
2791 6 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2792 6 : let tenant_conf = self.tenant_conf.load();
2793 6 : tenant_conf
2794 6 : .tenant_conf
2795 6 : .lsn_lease_length
2796 6 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2797 6 : }
2798 :
2799 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2800 0 : let tenant_conf = self.tenant_conf.load();
2801 0 : tenant_conf
2802 0 : .tenant_conf
2803 0 : .lsn_lease_length_for_ts
2804 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2805 0 : }
2806 :
2807 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2808 0 : let tenant_conf = self.tenant_conf.load();
2809 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2810 0 : }
2811 :
2812 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2813 0 : let tenant_conf = self.tenant_conf.load();
2814 0 : tenant_conf
2815 0 : .tenant_conf
2816 0 : .lazy_slru_download
2817 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2818 0 : }
2819 :
2820 : /// Checks if a get page request should get perf tracing
2821 : ///
2822 : /// The configuration priority is: tenant config override, default tenant config,
2823 : /// pageserver config.
2824 0 : pub(crate) fn is_get_page_request_sampled(&self) -> bool {
2825 0 : let tenant_conf = self.tenant_conf.load();
2826 0 : let ratio = tenant_conf
2827 0 : .tenant_conf
2828 0 : .sampling_ratio
2829 0 : .flatten()
2830 0 : .or(self.conf.default_tenant_conf.sampling_ratio)
2831 0 : .or(self.conf.tracing.as_ref().map(|t| t.sampling_ratio));
2832 :
2833 0 : match ratio {
2834 0 : Some(r) => {
2835 0 : if r.numerator == 0 {
2836 0 : false
2837 : } else {
2838 0 : rand::rng().random_range(0..r.denominator) < r.numerator
2839 : }
2840 : }
2841 0 : None => false,
2842 : }
2843 0 : }
2844 :
2845 2402187 : fn get_checkpoint_distance(&self) -> u64 {
2846 2402187 : let tenant_conf = self.tenant_conf.load();
2847 2402187 : tenant_conf
2848 2402187 : .tenant_conf
2849 2402187 : .checkpoint_distance
2850 2402187 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2851 2402187 : }
2852 :
2853 2401522 : fn get_checkpoint_timeout(&self) -> Duration {
2854 2401522 : let tenant_conf = self.tenant_conf.load();
2855 2401522 : tenant_conf
2856 2401522 : .tenant_conf
2857 2401522 : .checkpoint_timeout
2858 2401522 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2859 2401522 : }
2860 :
2861 1 : pub(crate) fn get_pitr_interval(&self) -> Duration {
2862 1 : let tenant_conf = &self.tenant_conf.load().tenant_conf;
2863 1 : tenant_conf
2864 1 : .pitr_interval
2865 1 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2866 1 : }
2867 :
2868 1273 : fn get_compaction_period(&self) -> Duration {
2869 1273 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2870 1273 : tenant_conf
2871 1273 : .compaction_period
2872 1273 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2873 1273 : }
2874 :
2875 352 : fn get_compaction_target_size(&self) -> u64 {
2876 352 : let tenant_conf = self.tenant_conf.load();
2877 352 : tenant_conf
2878 352 : .tenant_conf
2879 352 : .compaction_target_size
2880 352 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2881 352 : }
2882 :
2883 820 : fn get_compaction_threshold(&self) -> usize {
2884 820 : let tenant_conf = self.tenant_conf.load();
2885 820 : tenant_conf
2886 820 : .tenant_conf
2887 820 : .compaction_threshold
2888 820 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2889 820 : }
2890 :
2891 : /// Returns `true` if the rel_size_v2 config is enabled. NOTE: the write path and read path
2892 : /// should look at `get_rel_size_v2_status()` to get the actual status of the timeline. It is
2893 : /// possible that the index part persists the state while the config doesn't get persisted.
2894 973 : pub(crate) fn get_rel_size_v2_expected_state(&self) -> RelSizeMigration {
2895 973 : let tenant_conf = self.tenant_conf.load();
2896 973 : let v2_enabled = tenant_conf
2897 973 : .tenant_conf
2898 973 : .rel_size_v2_enabled
2899 973 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled);
2900 973 : let v1_access_disabled = tenant_conf
2901 973 : .tenant_conf
2902 973 : .rel_size_v1_access_disabled
2903 973 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v1_access_disabled);
2904 :
2905 973 : match (v2_enabled, v1_access_disabled) {
2906 0 : (true, false) => RelSizeMigration::Migrating,
2907 0 : (true, true) => RelSizeMigration::Migrated,
2908 0 : (false, true) => RelSizeMigration::Legacy, // This should never happen
2909 973 : (false, false) => RelSizeMigration::Legacy,
2910 : }
2911 973 : }
2912 :
2913 1099 : pub(crate) fn get_rel_size_v2_status(&self) -> (RelSizeMigration, Option<Lsn>) {
2914 1099 : let (status, migrated_at) = self.rel_size_v2_status.load().as_ref().clone();
2915 1099 : (status.unwrap_or(RelSizeMigration::Legacy), migrated_at)
2916 1099 : }
2917 :
2918 23 : fn get_compaction_upper_limit(&self) -> usize {
2919 23 : let tenant_conf = self.tenant_conf.load();
2920 23 : tenant_conf
2921 23 : .tenant_conf
2922 23 : .compaction_upper_limit
2923 23 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2924 23 : }
2925 :
2926 0 : pub fn get_compaction_l0_first(&self) -> bool {
2927 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2928 0 : tenant_conf
2929 0 : .compaction_l0_first
2930 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2931 0 : }
2932 :
2933 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2934 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2935 0 : tenant_conf
2936 0 : .compaction_l0_semaphore
2937 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2938 0 : }
2939 :
2940 636 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2941 : // By default, delay L0 flushes at 3x the compaction threshold. The compaction threshold
2942 : // defaults to 10, and L0 compaction is generally able to keep L0 counts below 30.
2943 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 3;
2944 :
2945 : // If compaction is disabled, don't delay.
2946 636 : if self.get_compaction_period() == Duration::ZERO {
2947 632 : return None;
2948 4 : }
2949 :
2950 4 : let compaction_threshold = self.get_compaction_threshold();
2951 4 : let tenant_conf = self.tenant_conf.load();
2952 4 : let l0_flush_delay_threshold = tenant_conf
2953 4 : .tenant_conf
2954 4 : .l0_flush_delay_threshold
2955 4 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2956 4 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2957 :
2958 : // 0 disables backpressure.
2959 4 : if l0_flush_delay_threshold == 0 {
2960 0 : return None;
2961 4 : }
2962 :
2963 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2964 : // backpressure flushes below this.
2965 : // TODO: the tenant config should have validation to prevent this instead.
2966 4 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2967 4 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2968 636 : }
2969 :
2970 637 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2971 : // Disable L0 stalls by default. Stalling can cause unavailability if L0 compaction isn't
2972 : // responsive, and it can e.g. block on other compaction via the compaction semaphore or
2973 : // sibling timelines. We need more confidence before enabling this.
2974 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2975 :
2976 : // If compaction is disabled, don't stall.
2977 637 : if self.get_compaction_period() == Duration::ZERO {
2978 632 : return None;
2979 5 : }
2980 :
2981 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2982 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2983 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2984 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2985 5 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2986 0 : return None;
2987 5 : }
2988 :
2989 5 : let compaction_threshold = self.get_compaction_threshold();
2990 5 : let tenant_conf = self.tenant_conf.load();
2991 5 : let l0_flush_stall_threshold = tenant_conf
2992 5 : .tenant_conf
2993 5 : .l0_flush_stall_threshold
2994 5 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2995 :
2996 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2997 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2998 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2999 5 : if cfg!(feature = "testing")
3000 5 : && compaction_threshold == 1
3001 0 : && l0_flush_stall_threshold.is_none()
3002 : {
3003 0 : return None;
3004 5 : }
3005 :
3006 5 : let l0_flush_stall_threshold = l0_flush_stall_threshold
3007 5 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
3008 :
3009 : // 0 disables backpressure.
3010 5 : if l0_flush_stall_threshold == 0 {
3011 5 : return None;
3012 0 : }
3013 :
3014 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
3015 : // backpressure flushes below this.
3016 : // TODO: the tenant config should have validation to prevent this instead.
3017 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
3018 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
3019 637 : }
3020 :
3021 57 : fn get_image_creation_threshold(&self) -> usize {
3022 57 : let tenant_conf = self.tenant_conf.load();
3023 57 : tenant_conf
3024 57 : .tenant_conf
3025 57 : .image_creation_threshold
3026 57 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3027 57 : }
3028 :
3029 : // HADRON
3030 193 : fn get_image_layer_force_creation_period(&self) -> Option<Duration> {
3031 193 : let tenant_conf = self.tenant_conf.load();
3032 193 : tenant_conf
3033 193 : .tenant_conf
3034 193 : .image_layer_force_creation_period
3035 193 : .or(self
3036 193 : .conf
3037 193 : .default_tenant_conf
3038 193 : .image_layer_force_creation_period)
3039 193 : }
3040 :
3041 192 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
3042 192 : let tenant_conf = &self.tenant_conf.load();
3043 192 : tenant_conf
3044 192 : .tenant_conf
3045 192 : .compaction_algorithm
3046 192 : .as_ref()
3047 192 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
3048 192 : .clone()
3049 192 : }
3050 :
3051 192 : pub fn get_compaction_shard_ancestor(&self) -> bool {
3052 192 : let tenant_conf = self.tenant_conf.load();
3053 192 : tenant_conf
3054 192 : .tenant_conf
3055 192 : .compaction_shard_ancestor
3056 192 : .unwrap_or(self.conf.default_tenant_conf.compaction_shard_ancestor)
3057 192 : }
3058 :
3059 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
3060 0 : let tenant_conf = self.tenant_conf.load();
3061 0 : tenant_conf
3062 0 : .tenant_conf
3063 0 : .eviction_policy
3064 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
3065 0 : }
3066 :
3067 235 : fn get_evictions_low_residence_duration_metric_threshold(
3068 235 : tenant_conf: &pageserver_api::models::TenantConfig,
3069 235 : default_tenant_conf: &pageserver_api::config::TenantConfigToml,
3070 235 : ) -> Duration {
3071 235 : tenant_conf
3072 235 : .evictions_low_residence_duration_metric_threshold
3073 235 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
3074 235 : }
3075 :
3076 191 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
3077 191 : let tenant_conf = self.tenant_conf.load();
3078 191 : tenant_conf
3079 191 : .tenant_conf
3080 191 : .image_layer_creation_check_threshold
3081 191 : .unwrap_or(
3082 191 : self.conf
3083 191 : .default_tenant_conf
3084 191 : .image_layer_creation_check_threshold,
3085 : )
3086 191 : }
3087 :
3088 27 : fn get_gc_compaction_settings(&self) -> GcCompactionCombinedSettings {
3089 27 : let tenant_conf = &self.tenant_conf.load();
3090 27 : let gc_compaction_enabled = tenant_conf
3091 27 : .tenant_conf
3092 27 : .gc_compaction_enabled
3093 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_enabled);
3094 27 : let gc_compaction_verification = tenant_conf
3095 27 : .tenant_conf
3096 27 : .gc_compaction_verification
3097 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_verification);
3098 27 : let gc_compaction_initial_threshold_kb = tenant_conf
3099 27 : .tenant_conf
3100 27 : .gc_compaction_initial_threshold_kb
3101 27 : .unwrap_or(
3102 27 : self.conf
3103 27 : .default_tenant_conf
3104 27 : .gc_compaction_initial_threshold_kb,
3105 : );
3106 27 : let gc_compaction_ratio_percent = tenant_conf
3107 27 : .tenant_conf
3108 27 : .gc_compaction_ratio_percent
3109 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_ratio_percent);
3110 27 : GcCompactionCombinedSettings {
3111 27 : gc_compaction_enabled,
3112 27 : gc_compaction_verification,
3113 27 : gc_compaction_initial_threshold_kb,
3114 27 : gc_compaction_ratio_percent,
3115 27 : }
3116 27 : }
3117 :
3118 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
3119 0 : let tenant_conf = self.tenant_conf.load();
3120 0 : tenant_conf
3121 0 : .tenant_conf
3122 0 : .image_creation_preempt_threshold
3123 0 : .unwrap_or(
3124 0 : self.conf
3125 0 : .default_tenant_conf
3126 0 : .image_creation_preempt_threshold,
3127 : )
3128 0 : }
3129 :
3130 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
3131 : // NB: Most tenant conf options are read by background loops, so,
3132 : // changes will automatically be picked up.
3133 :
3134 : // The threshold is embedded in the metric. So, we need to update it.
3135 : {
3136 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
3137 0 : &new_conf.tenant_conf,
3138 0 : &self.conf.default_tenant_conf,
3139 : );
3140 :
3141 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
3142 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
3143 :
3144 0 : let timeline_id_str = self.timeline_id.to_string();
3145 :
3146 0 : self.remote_client.update_config(&new_conf.location);
3147 :
3148 0 : let mut rel_size_cache = self.rel_size_snapshot_cache.lock().unwrap();
3149 0 : if let Some(new_capacity) = new_conf.tenant_conf.relsize_snapshot_cache_capacity {
3150 0 : if new_capacity != rel_size_cache.capacity() {
3151 0 : rel_size_cache.set_capacity(new_capacity);
3152 0 : }
3153 0 : }
3154 :
3155 0 : self.metrics
3156 0 : .evictions_with_low_residence_duration
3157 0 : .write()
3158 0 : .unwrap()
3159 0 : .change_threshold(
3160 0 : &tenant_id_str,
3161 0 : &shard_id_str,
3162 0 : &timeline_id_str,
3163 0 : new_threshold,
3164 0 : );
3165 : }
3166 0 : }
3167 :
3168 : /// Open a Timeline handle.
3169 : ///
3170 : /// Loads the metadata for the timeline into memory, but not the layer map.
3171 : #[allow(clippy::too_many_arguments)]
3172 235 : pub(super) fn new(
3173 235 : conf: &'static PageServerConf,
3174 235 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
3175 235 : metadata: &TimelineMetadata,
3176 235 : previous_heatmap: Option<PreviousHeatmap>,
3177 235 : ancestor: Option<Arc<Timeline>>,
3178 235 : timeline_id: TimelineId,
3179 235 : tenant_shard_id: TenantShardId,
3180 235 : generation: Generation,
3181 235 : shard_identity: ShardIdentity,
3182 235 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
3183 235 : resources: TimelineResources,
3184 235 : pg_version: PgMajorVersion,
3185 235 : state: TimelineState,
3186 235 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
3187 235 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
3188 235 : gc_compaction_state: Option<GcCompactionState>,
3189 235 : rel_size_v2_status: Option<RelSizeMigration>,
3190 235 : rel_size_migrated_at: Option<Lsn>,
3191 235 : cancel: CancellationToken,
3192 235 : ) -> Arc<Self> {
3193 235 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
3194 235 : let (state, _) = watch::channel(state);
3195 :
3196 235 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
3197 235 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
3198 :
3199 235 : let evictions_low_residence_duration_metric_threshold = {
3200 235 : let loaded_tenant_conf = tenant_conf.load();
3201 235 : Self::get_evictions_low_residence_duration_metric_threshold(
3202 235 : &loaded_tenant_conf.tenant_conf,
3203 235 : &conf.default_tenant_conf,
3204 : )
3205 : };
3206 :
3207 235 : if let Some(ancestor) = &ancestor {
3208 118 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
3209 118 : // If we construct an explicit timeline object, it's obviously not offloaded
3210 118 : let is_offloaded = MaybeOffloaded::No;
3211 118 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
3212 118 : }
3213 :
3214 235 : let relsize_snapshot_cache_capacity = {
3215 235 : let loaded_tenant_conf = tenant_conf.load();
3216 235 : loaded_tenant_conf
3217 235 : .tenant_conf
3218 235 : .relsize_snapshot_cache_capacity
3219 235 : .unwrap_or(conf.default_tenant_conf.relsize_snapshot_cache_capacity)
3220 : };
3221 :
3222 235 : Arc::new_cyclic(|myself| {
3223 235 : let metrics = Arc::new(TimelineMetrics::new(
3224 235 : &tenant_shard_id,
3225 235 : &timeline_id,
3226 235 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
3227 : "mtime",
3228 235 : evictions_low_residence_duration_metric_threshold,
3229 : ),
3230 : ));
3231 235 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
3232 :
3233 235 : let mut result = Timeline {
3234 235 : conf,
3235 235 : tenant_conf,
3236 235 : myself: myself.clone(),
3237 235 : timeline_id,
3238 235 : tenant_shard_id,
3239 235 : generation,
3240 235 : shard_identity,
3241 235 : pg_version,
3242 235 : layers: Default::default(),
3243 235 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
3244 :
3245 235 : walredo_mgr,
3246 235 : walreceiver: Mutex::new(None),
3247 :
3248 235 : remote_client: Arc::new(resources.remote_client),
3249 :
3250 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
3251 235 : last_record_lsn: SeqWait::new(RecordLsn {
3252 235 : last: disk_consistent_lsn,
3253 235 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
3254 235 : }),
3255 235 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
3256 :
3257 235 : gc_compaction_state: ArcSwapOption::from_pointee(gc_compaction_state),
3258 :
3259 235 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
3260 235 : last_freeze_ts: RwLock::new(Instant::now()),
3261 :
3262 235 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
3263 :
3264 235 : ancestor_timeline: ancestor,
3265 235 : ancestor_lsn: metadata.ancestor_lsn(),
3266 :
3267 235 : metrics,
3268 :
3269 235 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
3270 235 : &tenant_shard_id,
3271 235 : &timeline_id,
3272 235 : resources.pagestream_throttle_metrics,
3273 : ),
3274 :
3275 1880 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
3276 1880 : directory_metrics_inited: array::from_fn(|_| AtomicBool::new(false)),
3277 :
3278 235 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
3279 :
3280 235 : layer_flush_start_tx,
3281 235 : layer_flush_done_tx,
3282 :
3283 235 : write_lock: tokio::sync::Mutex::new(None),
3284 :
3285 235 : gc_info: std::sync::RwLock::new(GcInfo::default()),
3286 :
3287 235 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
3288 235 : LastImageLayerCreationStatus::default(),
3289 : )),
3290 :
3291 235 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
3292 235 : initdb_lsn: metadata.initdb_lsn(),
3293 :
3294 235 : current_logical_size: if disk_consistent_lsn.is_valid() {
3295 : // we're creating timeline data with some layer files existing locally,
3296 : // need to recalculate timeline's logical size based on data in the layers.
3297 120 : LogicalSize::deferred_initial(disk_consistent_lsn)
3298 : } else {
3299 : // we're creating timeline data without any layers existing locally,
3300 : // initial logical size is 0.
3301 115 : LogicalSize::empty_initial()
3302 : },
3303 :
3304 235 : partitioning: GuardArcSwap::new((
3305 235 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
3306 235 : Lsn(0),
3307 235 : )),
3308 : repartition_threshold: 0,
3309 235 : last_image_layer_creation_check_at: AtomicLsn::new(0),
3310 235 : last_image_layer_creation_check_instant: Mutex::new(None),
3311 235 : last_received_wal: Mutex::new(None),
3312 235 : rel_size_latest_cache: RwLock::new(HashMap::new()),
3313 235 : rel_size_snapshot_cache: Mutex::new(LruCache::new(relsize_snapshot_cache_capacity)),
3314 :
3315 235 : download_all_remote_layers_task_info: RwLock::new(None),
3316 :
3317 235 : state,
3318 :
3319 235 : eviction_task_timeline_state: tokio::sync::Mutex::new(
3320 235 : EvictionTaskTimelineState::default(),
3321 : ),
3322 235 : delete_progress: TimelineDeleteProgress::default(),
3323 :
3324 235 : cancel,
3325 235 : gate: Gate::default(),
3326 :
3327 235 : compaction_lock: tokio::sync::Mutex::default(),
3328 235 : compaction_failed: AtomicBool::default(),
3329 235 : corruption_detected: AtomicBool::default(),
3330 235 : l0_compaction_trigger: resources.l0_compaction_trigger,
3331 235 : gc_lock: tokio::sync::Mutex::default(),
3332 :
3333 235 : standby_horizon: AtomicLsn::new(0),
3334 :
3335 235 : pagestream_throttle: resources.pagestream_throttle,
3336 :
3337 235 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
3338 :
3339 : #[cfg(test)]
3340 235 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
3341 :
3342 235 : l0_flush_global_state: resources.l0_flush_global_state,
3343 :
3344 235 : handles: Default::default(),
3345 :
3346 235 : attach_wal_lag_cooldown,
3347 :
3348 235 : create_idempotency,
3349 :
3350 235 : page_trace: Default::default(),
3351 :
3352 235 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
3353 :
3354 235 : heatmap_layers_downloader: Mutex::new(None),
3355 :
3356 235 : rel_size_v2_status: ArcSwap::from_pointee((
3357 235 : rel_size_v2_status,
3358 235 : rel_size_migrated_at,
3359 235 : )),
3360 :
3361 235 : wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
3362 :
3363 235 : basebackup_cache: resources.basebackup_cache,
3364 :
3365 235 : feature_resolver: resources.feature_resolver.clone(),
3366 :
3367 235 : db_rel_count: ArcSwapOption::from_pointee(None),
3368 : };
3369 :
3370 235 : result.repartition_threshold =
3371 235 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
3372 :
3373 235 : result
3374 235 : .metrics
3375 235 : .last_record_lsn_gauge
3376 235 : .set(disk_consistent_lsn.0 as i64);
3377 235 : result
3378 235 : })
3379 235 : }
3380 :
3381 343 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
3382 343 : let Ok(guard) = self.gate.enter() else {
3383 0 : info!("cannot start flush loop when the timeline gate has already been closed");
3384 0 : return;
3385 : };
3386 343 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
3387 343 : match *flush_loop_state {
3388 232 : FlushLoopState::NotStarted => (),
3389 : FlushLoopState::Running { .. } => {
3390 111 : info!(
3391 0 : "skipping attempt to start flush_loop twice {}/{}",
3392 0 : self.tenant_shard_id, self.timeline_id
3393 : );
3394 111 : return;
3395 : }
3396 : FlushLoopState::Exited => {
3397 0 : info!(
3398 0 : "ignoring attempt to restart exited flush_loop {}/{}",
3399 0 : self.tenant_shard_id, self.timeline_id
3400 : );
3401 0 : return;
3402 : }
3403 : }
3404 :
3405 232 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
3406 232 : let self_clone = Arc::clone(self);
3407 :
3408 232 : debug!("spawning flush loop");
3409 232 : *flush_loop_state = FlushLoopState::Running {
3410 232 : #[cfg(test)]
3411 232 : expect_initdb_optimization: false,
3412 232 : #[cfg(test)]
3413 232 : initdb_optimization_count: 0,
3414 232 : };
3415 232 : task_mgr::spawn(
3416 232 : task_mgr::BACKGROUND_RUNTIME.handle(),
3417 232 : task_mgr::TaskKind::LayerFlushTask,
3418 232 : self.tenant_shard_id,
3419 232 : Some(self.timeline_id),
3420 232 : "layer flush task",
3421 232 : async move {
3422 232 : let _guard = guard;
3423 232 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error).with_scope_timeline(&self_clone);
3424 232 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
3425 5 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
3426 5 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
3427 5 : *flush_loop_state = FlushLoopState::Exited;
3428 5 : Ok(())
3429 5 : }
3430 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))
3431 : );
3432 343 : }
3433 :
3434 0 : pub(crate) fn update_gc_compaction_state(
3435 0 : &self,
3436 0 : gc_compaction_state: GcCompactionState,
3437 0 : ) -> anyhow::Result<()> {
3438 0 : self.gc_compaction_state
3439 0 : .store(Some(Arc::new(gc_compaction_state.clone())));
3440 0 : self.remote_client
3441 0 : .schedule_index_upload_for_gc_compaction_state_update(gc_compaction_state)
3442 0 : }
3443 :
3444 0 : pub(crate) fn update_rel_size_v2_status(
3445 0 : &self,
3446 0 : rel_size_v2_status: RelSizeMigration,
3447 0 : rel_size_migrated_at: Option<Lsn>,
3448 0 : ) -> anyhow::Result<()> {
3449 0 : self.rel_size_v2_status.store(Arc::new((
3450 0 : Some(rel_size_v2_status.clone()),
3451 0 : rel_size_migrated_at,
3452 0 : )));
3453 0 : self.remote_client
3454 0 : .schedule_index_upload_for_rel_size_v2_status_update(
3455 0 : rel_size_v2_status,
3456 0 : rel_size_migrated_at,
3457 : )
3458 0 : }
3459 :
3460 0 : pub(crate) fn get_gc_compaction_state(&self) -> Option<GcCompactionState> {
3461 0 : self.gc_compaction_state
3462 0 : .load()
3463 0 : .as_ref()
3464 0 : .map(|x| x.as_ref().clone())
3465 0 : }
3466 :
3467 : /// Creates and starts the wal receiver.
3468 : ///
3469 : /// This function is expected to be called at most once per Timeline's lifecycle
3470 : /// when the timeline is activated.
3471 0 : fn launch_wal_receiver(
3472 0 : self: &Arc<Self>,
3473 0 : ctx: &RequestContext,
3474 0 : broker_client: BrokerClientChannel,
3475 0 : ) {
3476 0 : info!(
3477 0 : "launching WAL receiver for timeline {} of tenant {}",
3478 0 : self.timeline_id, self.tenant_shard_id
3479 : );
3480 :
3481 0 : let tenant_conf = self.tenant_conf.load();
3482 0 : let wal_connect_timeout = tenant_conf
3483 0 : .tenant_conf
3484 0 : .walreceiver_connect_timeout
3485 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
3486 0 : let lagging_wal_timeout = tenant_conf
3487 0 : .tenant_conf
3488 0 : .lagging_wal_timeout
3489 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
3490 0 : let max_lsn_wal_lag = tenant_conf
3491 0 : .tenant_conf
3492 0 : .max_lsn_wal_lag
3493 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
3494 :
3495 0 : let mut guard = self.walreceiver.lock().unwrap();
3496 0 : assert!(
3497 0 : guard.is_none(),
3498 0 : "multiple launches / re-launches of WAL receiver are not supported"
3499 : );
3500 :
3501 0 : let protocol = PostgresClientProtocol::Interpreted {
3502 0 : format: utils::postgres_client::InterpretedFormat::Protobuf,
3503 0 : compression: Some(utils::postgres_client::Compression::Zstd { level: 1 }),
3504 0 : };
3505 :
3506 0 : *guard = Some(WalReceiver::start(
3507 0 : Arc::clone(self),
3508 0 : WalReceiverConf {
3509 0 : protocol,
3510 0 : wal_connect_timeout,
3511 0 : lagging_wal_timeout,
3512 0 : max_lsn_wal_lag,
3513 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
3514 0 : availability_zone: self.conf.availability_zone.clone(),
3515 0 : ingest_batch_size: self.conf.ingest_batch_size,
3516 0 : validate_wal_contiguity: self.conf.validate_wal_contiguity,
3517 0 : },
3518 0 : broker_client,
3519 0 : ctx,
3520 0 : ));
3521 0 : }
3522 :
3523 : /// Initialize with an empty layer map. Used when creating a new timeline.
3524 232 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
3525 232 : let mut layers = self.layers.try_write(LayerManagerLockHolder::Init).expect(
3526 232 : "in the context where we call this function, no other task has access to the object",
3527 : );
3528 232 : layers
3529 232 : .open_mut()
3530 232 : .expect("in this context the LayerManager must still be open")
3531 232 : .initialize_empty(Lsn(start_lsn.0));
3532 232 : }
3533 :
3534 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
3535 : /// files.
3536 3 : pub(super) async fn load_layer_map(
3537 3 : &self,
3538 3 : disk_consistent_lsn: Lsn,
3539 3 : index_part: IndexPart,
3540 3 : ) -> anyhow::Result<()> {
3541 : use LayerName::*;
3542 : use init::Decision::*;
3543 : use init::{Discovered, DismissedLayer};
3544 :
3545 3 : let mut guard = self
3546 3 : .layers
3547 3 : .write(LayerManagerLockHolder::LoadLayerMap)
3548 3 : .await;
3549 :
3550 3 : let timer = self.metrics.load_layer_map_histo.start_timer();
3551 :
3552 : // Scan timeline directory and create ImageLayerName and DeltaFilename
3553 : // structs representing all files on disk
3554 3 : let timeline_path = self
3555 3 : .conf
3556 3 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
3557 3 : let conf = self.conf;
3558 3 : let span = tracing::Span::current();
3559 :
3560 : // Copy to move into the task we're about to spawn
3561 3 : let this = self.myself.upgrade().expect("&self method holds the arc");
3562 :
3563 3 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
3564 3 : move || {
3565 3 : let _g = span.entered();
3566 3 : let discovered = init::scan_timeline_dir(&timeline_path)?;
3567 3 : let mut discovered_layers = Vec::with_capacity(discovered.len());
3568 3 : let mut unrecognized_files = Vec::new();
3569 :
3570 3 : let mut path = timeline_path;
3571 :
3572 11 : for discovered in discovered {
3573 8 : let (name, kind) = match discovered {
3574 8 : Discovered::Layer(layer_file_name, local_metadata) => {
3575 8 : discovered_layers.push((layer_file_name, local_metadata));
3576 8 : continue;
3577 : }
3578 0 : Discovered::IgnoredBackup(path) => {
3579 0 : std::fs::remove_file(path)
3580 0 : .or_else(fs_ext::ignore_not_found)
3581 0 : .fatal_err("Removing .old file");
3582 0 : continue;
3583 : }
3584 0 : Discovered::Unknown(file_name) => {
3585 : // we will later error if there are any
3586 0 : unrecognized_files.push(file_name);
3587 0 : continue;
3588 : }
3589 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
3590 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
3591 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
3592 : };
3593 0 : path.push(Utf8Path::new(&name));
3594 0 : init::cleanup(&path, kind)?;
3595 0 : path.pop();
3596 : }
3597 :
3598 3 : if !unrecognized_files.is_empty() {
3599 : // assume that if there are any there are many many.
3600 0 : let n = unrecognized_files.len();
3601 0 : let first = &unrecognized_files[..n.min(10)];
3602 0 : anyhow::bail!(
3603 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
3604 : );
3605 3 : }
3606 :
3607 3 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
3608 :
3609 3 : let mut loaded_layers = Vec::new();
3610 3 : let mut needs_cleanup = Vec::new();
3611 3 : let mut total_physical_size = 0;
3612 :
3613 11 : for (name, decision) in decided {
3614 8 : let decision = match decision {
3615 8 : Ok(decision) => decision,
3616 0 : Err(DismissedLayer::Future { local }) => {
3617 0 : if let Some(local) = local {
3618 0 : init::cleanup_future_layer(
3619 0 : &local.local_path,
3620 0 : &name,
3621 0 : disk_consistent_lsn,
3622 0 : )?;
3623 0 : }
3624 0 : needs_cleanup.push(name);
3625 0 : continue;
3626 : }
3627 0 : Err(DismissedLayer::LocalOnly(local)) => {
3628 0 : init::cleanup_local_only_file(&name, &local)?;
3629 : // this file never existed remotely, we will have to do rework
3630 0 : continue;
3631 : }
3632 0 : Err(DismissedLayer::BadMetadata(local)) => {
3633 0 : init::cleanup_local_file_for_remote(&local)?;
3634 : // this file never existed remotely, we will have to do rework
3635 0 : continue;
3636 : }
3637 : };
3638 :
3639 8 : match &name {
3640 6 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
3641 2 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
3642 : }
3643 :
3644 8 : tracing::debug!(layer=%name, ?decision, "applied");
3645 :
3646 8 : let layer = match decision {
3647 8 : Resident { local, remote } => {
3648 8 : total_physical_size += local.file_size;
3649 8 : Layer::for_resident(conf, &this, local.local_path, name, remote)
3650 8 : .drop_eviction_guard()
3651 : }
3652 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
3653 : };
3654 :
3655 8 : loaded_layers.push(layer);
3656 : }
3657 3 : Ok((loaded_layers, needs_cleanup, total_physical_size))
3658 3 : }
3659 : })
3660 3 : .await
3661 3 : .map_err(anyhow::Error::new)
3662 3 : .and_then(|x| x)?;
3663 :
3664 3 : let num_layers = loaded_layers.len();
3665 :
3666 3 : guard
3667 3 : .open_mut()
3668 3 : .expect("layermanager must be open during init")
3669 3 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3670 :
3671 3 : self.remote_client
3672 3 : .schedule_layer_file_deletion(&needs_cleanup)?;
3673 3 : self.remote_client
3674 3 : .schedule_index_upload_for_file_changes()?;
3675 : // This barrier orders above DELETEs before any later operations.
3676 : // This is critical because code executing after the barrier might
3677 : // create again objects with the same key that we just scheduled for deletion.
3678 : // For example, if we just scheduled deletion of an image layer "from the future",
3679 : // later compaction might run again and re-create the same image layer.
3680 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3681 : // "same" here means same key range and LSN.
3682 : //
3683 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3684 : // the upload queue may execute the PUT first, then the DELETE.
3685 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3686 : //
3687 : // 1. a future image layer is created and uploaded
3688 : // 2. ps restart
3689 : // 3. the future layer from (1) is deleted during load layer map
3690 : // 4. image layer is re-created and uploaded
3691 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3692 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3693 : //
3694 : // See https://github.com/neondatabase/neon/issues/5878
3695 : //
3696 : // NB: generation numbers naturally protect against this because they disambiguate
3697 : // (1) and (4)
3698 : // TODO: this is basically a no-op now, should we remove it?
3699 3 : self.remote_client.schedule_barrier()?;
3700 : // TenantShard::create_timeline will wait for these uploads to happen before returning, or
3701 : // on retry.
3702 :
3703 3 : info!(
3704 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3705 : num_layers, disk_consistent_lsn, total_physical_size
3706 : );
3707 :
3708 3 : timer.stop_and_record();
3709 3 : Ok(())
3710 3 : }
3711 :
3712 : /// Retrieve current logical size of the timeline.
3713 : ///
3714 : /// The size could be lagging behind the actual number, in case
3715 : /// the initial size calculation has not been run (gets triggered on the first size access).
3716 : ///
3717 : /// return size and boolean flag that shows if the size is exact
3718 0 : pub(crate) fn get_current_logical_size(
3719 0 : self: &Arc<Self>,
3720 0 : priority: GetLogicalSizePriority,
3721 0 : ctx: &RequestContext,
3722 0 : ) -> logical_size::CurrentLogicalSize {
3723 0 : if !self.tenant_shard_id.is_shard_zero() {
3724 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3725 : // when HTTP API is serving a GET for timeline zero, return zero
3726 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3727 0 : }
3728 :
3729 0 : let current_size = self.current_logical_size.current_size();
3730 0 : debug!("Current size: {current_size:?}");
3731 :
3732 0 : match (current_size.accuracy(), priority) {
3733 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3734 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3735 0 : // background task will eventually deliver an exact value, we're in no rush
3736 0 : }
3737 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3738 : // background task is not ready, but user is asking for it now;
3739 : // => make the background task skip the line
3740 : // (The alternative would be to calculate the size here, but,
3741 : // it can actually take a long time if the user has a lot of rels.
3742 : // And we'll inevitable need it again; So, let the background task do the work.)
3743 0 : match self
3744 0 : .current_logical_size
3745 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3746 0 : .get()
3747 : {
3748 0 : Some(cancel) => cancel.cancel(),
3749 : None => {
3750 0 : match self.current_state() {
3751 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3752 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3753 0 : // Don't make noise.
3754 0 : }
3755 : TimelineState::Loading => {
3756 : // Import does not return an activated timeline.
3757 0 : info!(
3758 0 : "discarding priority boost for logical size calculation because timeline is not yet active"
3759 : );
3760 : }
3761 : TimelineState::Active => {
3762 : // activation should be setting the once cell
3763 0 : warn!(
3764 0 : "unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work"
3765 : );
3766 0 : debug_assert!(false);
3767 : }
3768 : }
3769 : }
3770 : }
3771 : }
3772 : }
3773 :
3774 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3775 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3776 0 : let first = self
3777 0 : .current_logical_size
3778 0 : .did_return_approximate_to_walreceiver
3779 0 : .compare_exchange(
3780 0 : false,
3781 0 : true,
3782 0 : AtomicOrdering::Relaxed,
3783 0 : AtomicOrdering::Relaxed,
3784 0 : )
3785 0 : .is_ok();
3786 0 : if first {
3787 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3788 0 : }
3789 0 : }
3790 0 : }
3791 :
3792 0 : current_size
3793 0 : }
3794 :
3795 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3796 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3797 : // nothing to do for freshly created timelines;
3798 0 : assert_eq!(
3799 0 : self.current_logical_size.current_size().accuracy(),
3800 : logical_size::Accuracy::Exact,
3801 : );
3802 0 : self.current_logical_size.initialized.add_permits(1);
3803 0 : return;
3804 : };
3805 :
3806 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3807 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3808 0 : self.current_logical_size
3809 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3810 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3811 :
3812 0 : let self_clone = Arc::clone(self);
3813 0 : let background_ctx = ctx.detached_child(
3814 0 : TaskKind::InitialLogicalSizeCalculation,
3815 0 : DownloadBehavior::Download,
3816 : );
3817 0 : task_mgr::spawn(
3818 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3819 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3820 0 : self.tenant_shard_id,
3821 0 : Some(self.timeline_id),
3822 0 : "initial size calculation",
3823 : // NB: don't log errors here, task_mgr will do that.
3824 0 : async move {
3825 0 : self_clone
3826 0 : .initial_logical_size_calculation_task(
3827 0 : initial_part_end,
3828 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3829 0 : background_ctx,
3830 0 : )
3831 0 : .await;
3832 0 : Ok(())
3833 0 : }
3834 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)),
3835 : );
3836 0 : }
3837 :
3838 : /// # Cancellation
3839 : ///
3840 : /// This method is sensitive to `Timeline::cancel`.
3841 : ///
3842 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3843 : ///
3844 : /// # Cancel-Safety
3845 : ///
3846 : /// It does Timeline IO, hence this should be polled to completion because
3847 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3848 : /// to reason about.
3849 0 : async fn initial_logical_size_calculation_task(
3850 0 : self: Arc<Self>,
3851 0 : initial_part_end: Lsn,
3852 0 : skip_concurrency_limiter: CancellationToken,
3853 0 : background_ctx: RequestContext,
3854 0 : ) {
3855 0 : scopeguard::defer! {
3856 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3857 : self.current_logical_size.initialized.add_permits(1);
3858 : }
3859 :
3860 0 : let try_once = |attempt: usize| {
3861 0 : let background_ctx = &background_ctx;
3862 0 : let self_ref = &self;
3863 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3864 0 : async move {
3865 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3866 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3867 0 : background_ctx,
3868 : );
3869 :
3870 : use crate::metrics::initial_logical_size::StartCircumstances;
3871 0 : let (_maybe_permit, circumstances) = tokio::select! {
3872 0 : permit = wait_for_permit => {
3873 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3874 : }
3875 0 : _ = self_ref.cancel.cancelled() => {
3876 0 : return Err(CalculateLogicalSizeError::Cancelled);
3877 : }
3878 0 : () = skip_concurrency_limiter.cancelled() => {
3879 : // Some action that is part of a end user interaction requested logical size
3880 : // => break out of the rate limit
3881 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3882 : // but then again what happens if they cancel; also, we should just be using
3883 : // one runtime across the entire process, so, let's leave this for now.
3884 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3885 : }
3886 : };
3887 :
3888 0 : let metrics_guard = if attempt == 1 {
3889 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3890 : } else {
3891 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3892 : };
3893 :
3894 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3895 0 : self_ref.conf.get_vectored_concurrent_io,
3896 0 : self_ref
3897 0 : .gate
3898 0 : .enter()
3899 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3900 : );
3901 :
3902 0 : let calculated_size = self_ref
3903 0 : .logical_size_calculation_task(
3904 0 : initial_part_end,
3905 0 : LogicalSizeCalculationCause::Initial,
3906 0 : background_ctx,
3907 0 : )
3908 0 : .await?;
3909 :
3910 0 : self_ref
3911 0 : .trigger_aux_file_size_computation(
3912 0 : initial_part_end,
3913 0 : background_ctx,
3914 0 : io_concurrency,
3915 0 : )
3916 0 : .await?;
3917 :
3918 : // TODO: add aux file size to logical size
3919 :
3920 0 : Ok((calculated_size, metrics_guard))
3921 0 : }
3922 0 : };
3923 :
3924 0 : let retrying = async {
3925 0 : let mut attempt = 0;
3926 : loop {
3927 0 : attempt += 1;
3928 :
3929 0 : match try_once(attempt).await {
3930 0 : Ok(res) => return ControlFlow::Continue(res),
3931 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3932 : Err(
3933 0 : e @ (CalculateLogicalSizeError::Decode(_)
3934 : | CalculateLogicalSizeError::PageRead(_)),
3935 : ) => {
3936 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3937 : // exponential back-off doesn't make sense at these long intervals;
3938 : // use fixed retry interval with generous jitter instead
3939 0 : let sleep_duration = Duration::from_secs(
3940 0 : u64::try_from(
3941 : // 1hour base
3942 0 : (60_i64 * 60_i64)
3943 0 : // 10min jitter
3944 0 : + rand::rng().random_range(-10 * 60..10 * 60),
3945 : )
3946 0 : .expect("10min < 1hour"),
3947 : );
3948 0 : tokio::select! {
3949 0 : _ = tokio::time::sleep(sleep_duration) => {}
3950 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3951 : }
3952 : }
3953 : }
3954 : }
3955 0 : };
3956 :
3957 0 : let (calculated_size, metrics_guard) = match retrying.await {
3958 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3959 0 : ControlFlow::Break(()) => return,
3960 : };
3961 :
3962 : // we cannot query current_logical_size.current_size() to know the current
3963 : // *negative* value, only truncated to u64.
3964 0 : let added = self
3965 0 : .current_logical_size
3966 0 : .size_added_after_initial
3967 0 : .load(AtomicOrdering::Relaxed);
3968 :
3969 0 : let sum = calculated_size.saturating_add_signed(added);
3970 :
3971 : // set the gauge value before it can be set in `update_current_logical_size`.
3972 0 : self.metrics.current_logical_size_gauge.set(sum);
3973 :
3974 0 : self.current_logical_size
3975 0 : .initial_logical_size
3976 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3977 0 : .ok()
3978 0 : .expect("only this task sets it");
3979 0 : }
3980 :
3981 7 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3982 7 : self: &Arc<Self>,
3983 7 : lsn: Lsn,
3984 7 : cause: LogicalSizeCalculationCause,
3985 7 : ctx: RequestContext,
3986 7 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3987 7 : let (sender, receiver) = oneshot::channel();
3988 7 : let self_clone = Arc::clone(self);
3989 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3990 : // we should stop the size calculation work and return an error.
3991 : // That would require restructuring this function's API to
3992 : // return the result directly, instead of a Receiver for the result.
3993 7 : let ctx = ctx.detached_child(
3994 7 : TaskKind::OndemandLogicalSizeCalculation,
3995 7 : DownloadBehavior::Download,
3996 : );
3997 7 : task_mgr::spawn(
3998 7 : task_mgr::BACKGROUND_RUNTIME.handle(),
3999 7 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
4000 7 : self.tenant_shard_id,
4001 7 : Some(self.timeline_id),
4002 7 : "ondemand logical size calculation",
4003 7 : async move {
4004 7 : let res = self_clone
4005 7 : .logical_size_calculation_task(lsn, cause, &ctx)
4006 7 : .await;
4007 7 : let _ = sender.send(res).ok();
4008 7 : Ok(()) // Receiver is responsible for handling errors
4009 7 : }
4010 7 : .in_current_span(),
4011 : );
4012 7 : receiver
4013 7 : }
4014 :
4015 : #[instrument(skip_all)]
4016 : async fn logical_size_calculation_task(
4017 : self: &Arc<Self>,
4018 : lsn: Lsn,
4019 : cause: LogicalSizeCalculationCause,
4020 : ctx: &RequestContext,
4021 : ) -> Result<u64, CalculateLogicalSizeError> {
4022 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
4023 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
4024 : // accurate relation sizes, and they do not emit consumption metrics.
4025 : debug_assert!(self.tenant_shard_id.is_shard_zero());
4026 :
4027 : let guard = self
4028 : .gate
4029 : .enter()
4030 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
4031 :
4032 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
4033 : }
4034 :
4035 : /// Calculate the logical size of the database at the latest LSN.
4036 : ///
4037 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
4038 : /// especially if we need to download remote layers.
4039 7 : async fn calculate_logical_size(
4040 7 : &self,
4041 7 : up_to_lsn: Lsn,
4042 7 : cause: LogicalSizeCalculationCause,
4043 7 : _guard: &GateGuard,
4044 7 : ctx: &RequestContext,
4045 7 : ) -> Result<u64, CalculateLogicalSizeError> {
4046 7 : info!(
4047 0 : "Calculating logical size for timeline {} at {}",
4048 : self.timeline_id, up_to_lsn
4049 : );
4050 :
4051 7 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
4052 : {
4053 0 : return Err(CalculateLogicalSizeError::Cancelled);
4054 7 : }
4055 :
4056 : // See if we've already done the work for initial size calculation.
4057 : // This is a short-cut for timelines that are mostly unused.
4058 7 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
4059 0 : return Ok(size);
4060 7 : }
4061 7 : let storage_time_metrics = match cause {
4062 : LogicalSizeCalculationCause::Initial
4063 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
4064 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
4065 : LogicalSizeCalculationCause::EvictionTaskImitation => {
4066 7 : &self.metrics.imitate_logical_size_histo
4067 : }
4068 : };
4069 7 : let timer = storage_time_metrics.start_timer();
4070 7 : let logical_size = self
4071 7 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
4072 7 : .await?;
4073 7 : debug!("calculated logical size: {logical_size}");
4074 7 : timer.stop_and_record();
4075 7 : Ok(logical_size)
4076 7 : }
4077 :
4078 : /// Update current logical size, adding `delta' to the old value.
4079 135285 : fn update_current_logical_size(&self, delta: i64) {
4080 135285 : let logical_size = &self.current_logical_size;
4081 135285 : logical_size.increment_size(delta);
4082 :
4083 : // Also set the value in the prometheus gauge. Note that
4084 : // there is a race condition here: if this is is called by two
4085 : // threads concurrently, the prometheus gauge might be set to
4086 : // one value while current_logical_size is set to the
4087 : // other.
4088 135285 : match logical_size.current_size() {
4089 135285 : CurrentLogicalSize::Exact(ref new_current_size) => self
4090 135285 : .metrics
4091 135285 : .current_logical_size_gauge
4092 135285 : .set(new_current_size.into()),
4093 0 : CurrentLogicalSize::Approximate(_) => {
4094 0 : // don't update the gauge yet, this allows us not to update the gauge back and
4095 0 : // forth between the initial size calculation task.
4096 0 : }
4097 : }
4098 135285 : }
4099 :
4100 1527 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: MetricsUpdate) {
4101 : // TODO: this directory metrics is not correct -- we could have multiple reldirs in the system
4102 : // for each of the database, but we only store one value, and therefore each pgdirmodification
4103 : // would overwrite the previous value if they modify different databases.
4104 :
4105 1527 : match count {
4106 566 : MetricsUpdate::Set(count) => {
4107 566 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
4108 566 : self.directory_metrics_inited[kind.offset()].store(true, AtomicOrdering::Relaxed);
4109 566 : }
4110 960 : MetricsUpdate::Add(count) => {
4111 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
4112 : // it's fine.
4113 960 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
4114 960 : // The metrics has been initialized with `MetricsUpdate::Set` before, so we can add/sub
4115 960 : // the value reliably.
4116 960 : self.directory_metrics[kind.offset()].fetch_add(count, AtomicOrdering::Relaxed);
4117 960 : }
4118 : // Otherwise, ignore this update
4119 : }
4120 1 : MetricsUpdate::Sub(count) => {
4121 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
4122 : // it's fine.
4123 1 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
4124 1 : // The metrics has been initialized with `MetricsUpdate::Set` before.
4125 1 : // The operation could overflow so we need to normalize the value.
4126 1 : let prev_val =
4127 1 : self.directory_metrics[kind.offset()].load(AtomicOrdering::Relaxed);
4128 1 : let res = prev_val.saturating_sub(count);
4129 1 : self.directory_metrics[kind.offset()].store(res, AtomicOrdering::Relaxed);
4130 1 : }
4131 : // Otherwise, ignore this update
4132 : }
4133 : };
4134 :
4135 : // TODO: remove this, there's no place in the code that updates this aux metrics.
4136 1527 : let aux_metric =
4137 1527 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
4138 :
4139 1527 : let sum_of_entries = self
4140 1527 : .directory_metrics
4141 1527 : .iter()
4142 12216 : .map(|v| v.load(AtomicOrdering::Relaxed))
4143 1527 : .sum();
4144 : // Set a high general threshold and a lower threshold for the auxiliary files,
4145 : // as we can have large numbers of relations in the db directory.
4146 : const SUM_THRESHOLD: u64 = 5000;
4147 : const AUX_THRESHOLD: u64 = 1000;
4148 1527 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
4149 0 : self.metrics
4150 0 : .directory_entries_count_gauge
4151 0 : .set(sum_of_entries);
4152 1527 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
4153 0 : metric.set(sum_of_entries);
4154 1527 : }
4155 1527 : }
4156 :
4157 0 : async fn find_layer(
4158 0 : &self,
4159 0 : layer_name: &LayerName,
4160 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
4161 0 : let guard = self
4162 0 : .layers
4163 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
4164 0 : .await;
4165 0 : let layer = guard
4166 0 : .layer_map()?
4167 0 : .iter_historic_layers()
4168 0 : .find(|l| &l.layer_name() == layer_name)
4169 0 : .map(|found| guard.get_from_desc(&found));
4170 0 : Ok(layer)
4171 0 : }
4172 :
4173 0 : pub(super) fn should_keep_previous_heatmap(&self, new_heatmap_end_lsn: Lsn) -> bool {
4174 0 : let crnt = self.previous_heatmap.load();
4175 0 : match crnt.as_deref() {
4176 0 : Some(PreviousHeatmap::Active { end_lsn, .. }) => match end_lsn {
4177 0 : Some(crnt_end_lsn) => *crnt_end_lsn > new_heatmap_end_lsn,
4178 0 : None => true,
4179 : },
4180 0 : Some(PreviousHeatmap::Obsolete) => false,
4181 0 : None => false,
4182 : }
4183 0 : }
4184 :
4185 : /// The timeline heatmap is a hint to secondary locations from the primary location,
4186 : /// indicating which layers are currently on-disk on the primary.
4187 : ///
4188 : /// None is returned if the Timeline is in a state where uploading a heatmap
4189 : /// doesn't make sense, such as shutting down or initializing. The caller
4190 : /// should treat this as a cue to simply skip doing any heatmap uploading
4191 : /// for this timeline.
4192 8 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
4193 8 : if !self.is_active() {
4194 0 : return None;
4195 8 : }
4196 :
4197 8 : let guard = self
4198 8 : .layers
4199 8 : .read(LayerManagerLockHolder::GenerateHeatmap)
4200 8 : .await;
4201 :
4202 : // Firstly, if there's any heatmap left over from when this location
4203 : // was a secondary, take that into account. Keep layers that are:
4204 : // * present in the layer map
4205 : // * visible
4206 : // * non-resident
4207 : // * not evicted since we read the heatmap
4208 : //
4209 : // Without this, a new cold, attached location would clobber the previous
4210 : // heatamp.
4211 8 : let previous_heatmap = self.previous_heatmap.load();
4212 8 : let visible_non_resident = match previous_heatmap.as_deref() {
4213 : Some(PreviousHeatmap::Active {
4214 6 : heatmap, read_at, ..
4215 23 : }) => Some(heatmap.all_layers().filter_map(|hl| {
4216 23 : let desc: PersistentLayerDesc = hl.name.clone().into();
4217 23 : let layer = guard.try_get_from_key(&desc.key())?;
4218 :
4219 23 : if layer.visibility() == LayerVisibilityHint::Covered {
4220 0 : return None;
4221 23 : }
4222 :
4223 23 : if layer.is_likely_resident() {
4224 10 : return None;
4225 13 : }
4226 :
4227 13 : if layer.last_evicted_at().happened_after(*read_at) {
4228 3 : return None;
4229 10 : }
4230 :
4231 10 : Some((desc, hl.metadata.clone(), hl.access_time, hl.cold))
4232 23 : })),
4233 0 : Some(PreviousHeatmap::Obsolete) => None,
4234 2 : None => None,
4235 : };
4236 :
4237 : // Secondly, all currently visible, resident layers are included.
4238 18 : let resident = guard.likely_resident_layers().filter_map(|layer| {
4239 18 : match layer.visibility() {
4240 : LayerVisibilityHint::Visible => {
4241 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
4242 17 : let last_activity_ts = layer.latest_activity();
4243 17 : Some((
4244 17 : layer.layer_desc().clone(),
4245 17 : layer.metadata(),
4246 17 : last_activity_ts,
4247 17 : false, // these layers are not cold
4248 17 : ))
4249 : }
4250 : LayerVisibilityHint::Covered => {
4251 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
4252 1 : None
4253 : }
4254 : }
4255 18 : });
4256 :
4257 8 : let mut layers = match visible_non_resident {
4258 6 : Some(non_resident) => {
4259 6 : let mut non_resident = non_resident.peekable();
4260 6 : if non_resident.peek().is_none() {
4261 2 : tracing::info!(timeline_id=%self.timeline_id, "Previous heatmap now obsolete");
4262 2 : self.previous_heatmap
4263 2 : .store(Some(PreviousHeatmap::Obsolete.into()));
4264 4 : }
4265 :
4266 6 : non_resident.chain(resident).collect::<Vec<_>>()
4267 : }
4268 2 : None => resident.collect::<Vec<_>>(),
4269 : };
4270 :
4271 : // Sort layers in order of which to download first. For a large set of layers to download, we
4272 : // want to prioritize those layers which are most likely to still be in the resident many minutes
4273 : // or hours later:
4274 : // - Cold layers go last for convenience when a human inspects the heatmap.
4275 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
4276 : // only exist for a few minutes before being compacted into L1s.
4277 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
4278 : // the layer is likely to be covered by an image layer during compaction.
4279 60 : layers.sort_by_key(|(desc, _meta, _atime, cold)| {
4280 60 : std::cmp::Reverse((
4281 60 : *cold,
4282 60 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
4283 60 : desc.lsn_range.end,
4284 60 : ))
4285 60 : });
4286 :
4287 8 : let layers = layers
4288 8 : .into_iter()
4289 27 : .map(|(desc, meta, atime, cold)| {
4290 27 : HeatMapLayer::new(desc.layer_name(), meta, atime, cold)
4291 27 : })
4292 8 : .collect();
4293 :
4294 8 : Some(HeatMapTimeline::new(self.timeline_id, layers))
4295 8 : }
4296 :
4297 0 : pub(super) async fn generate_unarchival_heatmap(&self, end_lsn: Lsn) -> PreviousHeatmap {
4298 0 : let guard = self
4299 0 : .layers
4300 0 : .read(LayerManagerLockHolder::GenerateHeatmap)
4301 0 : .await;
4302 :
4303 0 : let now = SystemTime::now();
4304 0 : let mut heatmap_layers = Vec::default();
4305 0 : for vl in guard.visible_layers() {
4306 0 : if vl.layer_desc().get_lsn_range().start >= end_lsn {
4307 0 : continue;
4308 0 : }
4309 :
4310 0 : let hl = HeatMapLayer {
4311 0 : name: vl.layer_desc().layer_name(),
4312 0 : metadata: vl.metadata(),
4313 0 : access_time: now,
4314 0 : cold: true,
4315 0 : };
4316 0 : heatmap_layers.push(hl);
4317 : }
4318 :
4319 0 : tracing::info!(
4320 0 : "Generating unarchival heatmap with {} layers",
4321 0 : heatmap_layers.len()
4322 : );
4323 :
4324 0 : let heatmap = HeatMapTimeline::new(self.timeline_id, heatmap_layers);
4325 0 : PreviousHeatmap::Active {
4326 0 : heatmap,
4327 0 : read_at: Instant::now(),
4328 0 : end_lsn: Some(end_lsn),
4329 0 : }
4330 0 : }
4331 :
4332 : /// Returns true if the given lsn is or was an ancestor branchpoint.
4333 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
4334 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
4335 : // branchpoint in the value in IndexPart::lineage
4336 0 : self.ancestor_lsn == lsn
4337 0 : || (self.ancestor_lsn == Lsn::INVALID
4338 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
4339 0 : }
4340 : }
4341 :
4342 : #[derive(Clone)]
4343 : /// Type representing a query in the ([`Lsn`], [`Key`]) space.
4344 : /// In other words, a set of segments in a 2D space.
4345 : ///
4346 : /// This representation has the advatange of avoiding hash map
4347 : /// allocations for uniform queries.
4348 : pub(crate) enum VersionedKeySpaceQuery {
4349 : /// Variant for queries at a single [`Lsn`]
4350 : Uniform { keyspace: KeySpace, lsn: Lsn },
4351 : /// Variant for queries at multiple [`Lsn`]s
4352 : Scattered {
4353 : keyspaces_at_lsn: Vec<(Lsn, KeySpace)>,
4354 : },
4355 : }
4356 :
4357 : impl VersionedKeySpaceQuery {
4358 302238 : pub(crate) fn uniform(keyspace: KeySpace, lsn: Lsn) -> Self {
4359 302238 : Self::Uniform { keyspace, lsn }
4360 302238 : }
4361 :
4362 10192 : pub(crate) fn scattered(keyspaces_at_lsn: Vec<(Lsn, KeySpace)>) -> Self {
4363 10192 : Self::Scattered { keyspaces_at_lsn }
4364 10192 : }
4365 :
4366 : /// Returns the most recent (largest) LSN included in the query.
4367 : /// If any of the LSNs included in the query are invalid, returns
4368 : /// an error instead.
4369 624860 : fn high_watermark_lsn(&self) -> Result<Lsn, GetVectoredError> {
4370 624860 : match self {
4371 604476 : Self::Uniform { lsn, .. } => {
4372 604476 : if !lsn.is_valid() {
4373 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4374 604476 : }
4375 :
4376 604476 : Ok(*lsn)
4377 : }
4378 20384 : Self::Scattered { keyspaces_at_lsn } => {
4379 20384 : let mut max_lsn = None;
4380 42216 : for (lsn, _keyspace) in keyspaces_at_lsn.iter() {
4381 42216 : if !lsn.is_valid() {
4382 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4383 42216 : }
4384 42216 : max_lsn = std::cmp::max(max_lsn, Some(lsn));
4385 : }
4386 :
4387 20384 : if let Some(computed) = max_lsn {
4388 20384 : Ok(*computed)
4389 : } else {
4390 0 : Err(GetVectoredError::Other(anyhow!("empty input")))
4391 : }
4392 : }
4393 : }
4394 624860 : }
4395 :
4396 : /// Returns the total keyspace being queried: the result of projecting
4397 : /// everything in the key dimensions onto the key axis.
4398 323507 : fn total_keyspace(&self) -> KeySpace {
4399 323507 : match self {
4400 303123 : Self::Uniform { keyspace, .. } => keyspace.clone(),
4401 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4402 20384 : .iter()
4403 20384 : .map(|(_lsn, keyspace)| keyspace)
4404 42216 : .fold(KeySpace::default(), |mut acc, v| {
4405 42216 : acc.merge(v);
4406 42216 : acc
4407 42216 : }),
4408 : }
4409 323507 : }
4410 :
4411 : /// Returns LSN for a specific key.
4412 : ///
4413 : /// Invariant: requested key must be part of [`Self::total_keyspace`]
4414 395551 : pub(super) fn map_key_to_lsn(&self, key: &Key) -> Lsn {
4415 395551 : match self {
4416 322359 : Self::Uniform { lsn, .. } => *lsn,
4417 73192 : Self::Scattered { keyspaces_at_lsn } => {
4418 73192 : keyspaces_at_lsn
4419 73192 : .iter()
4420 418356 : .find(|(_lsn, keyspace)| keyspace.contains(key))
4421 73192 : .expect("Returned key was requested")
4422 : .0
4423 : }
4424 : }
4425 395551 : }
4426 :
4427 : /// Remove any parts of the query (segments) which overlap with the provided
4428 : /// key space (also segments).
4429 968376 : fn remove_overlapping_with(&mut self, to_remove: &KeySpace) -> KeySpace {
4430 968376 : match self {
4431 947992 : Self::Uniform { keyspace, .. } => keyspace.remove_overlapping_with(to_remove),
4432 20384 : Self::Scattered { keyspaces_at_lsn } => {
4433 20384 : let mut removed_accum = KeySpaceRandomAccum::new();
4434 42216 : keyspaces_at_lsn.iter_mut().for_each(|(_lsn, keyspace)| {
4435 42216 : let removed = keyspace.remove_overlapping_with(to_remove);
4436 42216 : removed_accum.add_keyspace(removed);
4437 42216 : });
4438 :
4439 20384 : removed_accum.to_keyspace()
4440 : }
4441 : }
4442 968376 : }
4443 :
4444 739365 : fn is_empty(&self) -> bool {
4445 739365 : match self {
4446 718981 : Self::Uniform { keyspace, .. } => keyspace.is_empty(),
4447 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4448 20384 : .iter()
4449 31300 : .all(|(_lsn, keyspace)| keyspace.is_empty()),
4450 : }
4451 739365 : }
4452 :
4453 : /// "Lower" the query on the LSN dimension
4454 114506 : fn lower(&mut self, to: Lsn) {
4455 114506 : match self {
4456 114506 : Self::Uniform { lsn, .. } => {
4457 114506 : // If the originally requested LSN is smaller than the starting
4458 114506 : // LSN of the ancestor we are descending into, we need to respect that.
4459 114506 : // Hence the min.
4460 114506 : *lsn = std::cmp::min(*lsn, to);
4461 114506 : }
4462 0 : Self::Scattered { keyspaces_at_lsn } => {
4463 0 : keyspaces_at_lsn.iter_mut().for_each(|(lsn, _keyspace)| {
4464 0 : *lsn = std::cmp::min(*lsn, to);
4465 0 : });
4466 : }
4467 : }
4468 114506 : }
4469 : }
4470 :
4471 : impl std::fmt::Display for VersionedKeySpaceQuery {
4472 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
4473 0 : write!(f, "[")?;
4474 :
4475 0 : match self {
4476 0 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4477 0 : write!(f, "{keyspace} @ {lsn}")?;
4478 : }
4479 0 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4480 0 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4481 0 : write!(f, "{keyspace} @ {lsn},")?;
4482 : }
4483 : }
4484 : }
4485 :
4486 0 : write!(f, "]")
4487 0 : }
4488 : }
4489 :
4490 : impl Timeline {
4491 : #[allow(clippy::doc_lazy_continuation)]
4492 : /// Get the data needed to reconstruct all keys in the provided keyspace
4493 : ///
4494 : /// The algorithm is as follows:
4495 : /// 1. While some keys are still not done and there's a timeline to visit:
4496 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
4497 : /// 2.1: Build the fringe for the current keyspace
4498 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
4499 : /// intersects
4500 : /// 2.3. Pop the timeline from the fringe
4501 : /// 2.4. If the fringe is empty, go back to 1
4502 312430 : async fn get_vectored_reconstruct_data(
4503 312430 : &self,
4504 312430 : mut query: VersionedKeySpaceQuery,
4505 312430 : reconstruct_state: &mut ValuesReconstructState,
4506 312430 : ctx: &RequestContext,
4507 312430 : ) -> Result<(), GetVectoredError> {
4508 312430 : let original_hwm_lsn = query.high_watermark_lsn().unwrap();
4509 :
4510 : let mut timeline_owned: Arc<Timeline>;
4511 312430 : let mut timeline = self;
4512 :
4513 312429 : let missing_keyspace = loop {
4514 426935 : if self.cancel.is_cancelled() {
4515 0 : return Err(GetVectoredError::Cancelled);
4516 426935 : }
4517 :
4518 : let TimelineVisitOutcome {
4519 426935 : completed_keyspace: completed,
4520 426935 : image_covered_keyspace,
4521 : } = {
4522 426935 : let ctx = RequestContextBuilder::from(ctx)
4523 426935 : .perf_span(|crnt_perf_span| {
4524 0 : info_span!(
4525 : target: PERF_TRACE_TARGET,
4526 0 : parent: crnt_perf_span,
4527 : "PLAN_IO_TIMELINE",
4528 : timeline = %timeline.timeline_id,
4529 0 : high_watermark_lsn = %query.high_watermark_lsn().unwrap(),
4530 : )
4531 0 : })
4532 426935 : .attached_child();
4533 :
4534 426935 : Self::get_vectored_reconstruct_data_timeline(
4535 426935 : timeline,
4536 426935 : &query,
4537 426935 : reconstruct_state,
4538 426935 : &self.cancel,
4539 426935 : &ctx,
4540 : )
4541 426935 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4542 426935 : .await?
4543 : };
4544 :
4545 426935 : query.remove_overlapping_with(&completed);
4546 :
4547 : // Do not descend into the ancestor timeline for aux files.
4548 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
4549 : // stalling compaction.
4550 426935 : query.remove_overlapping_with(&KeySpace {
4551 426935 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
4552 426935 : });
4553 :
4554 : // Keyspace is fully retrieved
4555 426935 : if query.is_empty() {
4556 312246 : break None;
4557 114689 : }
4558 :
4559 114689 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
4560 : // Not fully retrieved but no ancestor timeline.
4561 183 : break Some(query.total_keyspace());
4562 : };
4563 :
4564 : // Now we see if there are keys covered by the image layer but does not exist in the
4565 : // image layer, which means that the key does not exist.
4566 :
4567 : // The block below will stop the vectored search if any of the keys encountered an image layer
4568 : // which did not contain a snapshot for said key. Since we have already removed all completed
4569 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
4570 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
4571 : // and stop the search as a result of that.
4572 114506 : let mut removed = query.remove_overlapping_with(&image_covered_keyspace);
4573 : // Do not fire missing key error and end early for sparse keys. Note that we hava already removed
4574 : // non-inherited keyspaces before, so we can safely do a full `SPARSE_RANGE` remove instead of
4575 : // figuring out what is the inherited key range and do a fine-grained pruning.
4576 114506 : removed.remove_overlapping_with(&KeySpace {
4577 114506 : ranges: vec![SPARSE_RANGE],
4578 114506 : });
4579 114506 : if !removed.is_empty() {
4580 0 : break Some(removed);
4581 114506 : }
4582 :
4583 : // Each key range in the original query is at some point in the LSN space.
4584 : // When descending into the ancestor, lower all ranges in the LSN space
4585 : // such that new changes on the parent timeline are not visible.
4586 114506 : query.lower(timeline.ancestor_lsn);
4587 :
4588 114506 : let ctx = RequestContextBuilder::from(ctx)
4589 114506 : .perf_span(|crnt_perf_span| {
4590 0 : info_span!(
4591 : target: PERF_TRACE_TARGET,
4592 0 : parent: crnt_perf_span,
4593 : "GET_ANCESTOR",
4594 : timeline = %timeline.timeline_id,
4595 0 : ancestor = %ancestor_timeline.timeline_id,
4596 : ancestor_lsn = %timeline.ancestor_lsn
4597 : )
4598 0 : })
4599 114506 : .attached_child();
4600 :
4601 114506 : timeline_owned = timeline
4602 114506 : .get_ready_ancestor_timeline(ancestor_timeline, &ctx)
4603 114506 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4604 114506 : .await?;
4605 114505 : timeline = &*timeline_owned;
4606 : };
4607 :
4608 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
4609 312429 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
4610 183 : let mut missing_keyspace = missing_keyspace;
4611 183 : missing_keyspace.remove_overlapping_with(&KeySpace {
4612 183 : ranges: vec![SPARSE_RANGE],
4613 183 : });
4614 183 : if missing_keyspace.is_empty() {
4615 176 : None
4616 : } else {
4617 7 : Some(missing_keyspace)
4618 : }
4619 : } else {
4620 312246 : None
4621 : };
4622 :
4623 312429 : if let Some(missing_keyspace) = missing_keyspace {
4624 7 : return Err(GetVectoredError::MissingKey(Box::new(MissingKeyError {
4625 7 : keyspace: missing_keyspace, /* better if we can store the full keyspace */
4626 7 : shard: self.shard_identity.number,
4627 7 : original_hwm_lsn,
4628 7 : ancestor_lsn: Some(timeline.ancestor_lsn),
4629 7 : backtrace: None,
4630 7 : read_path: std::mem::take(&mut reconstruct_state.read_path),
4631 7 : query: None,
4632 7 : })));
4633 312422 : }
4634 :
4635 312422 : Ok(())
4636 312430 : }
4637 :
4638 426935 : async fn get_vectored_init_fringe(
4639 426935 : &self,
4640 426935 : query: &VersionedKeySpaceQuery,
4641 426935 : ) -> Result<LayerFringe, GetVectoredError> {
4642 426935 : let mut fringe = LayerFringe::new();
4643 426935 : let guard = self.layers.read(LayerManagerLockHolder::GetPage).await;
4644 :
4645 426935 : match query {
4646 416743 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4647 : // LSNs requested by the compute or determined by the pageserver
4648 : // are inclusive. Queries to the layer map use exclusive LSNs.
4649 : // Hence, bump the value before the query - same in the other
4650 : // match arm.
4651 416743 : let cont_lsn = Lsn(lsn.0 + 1);
4652 416743 : guard.update_search_fringe(keyspace, cont_lsn, &mut fringe)?;
4653 : }
4654 10192 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4655 21108 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4656 21108 : let cont_lsn_for_keyspace = Lsn(lsn.0 + 1);
4657 21108 : guard.update_search_fringe(keyspace, cont_lsn_for_keyspace, &mut fringe)?;
4658 : }
4659 : }
4660 : }
4661 :
4662 426935 : Ok(fringe)
4663 426935 : }
4664 :
4665 : /// Collect the reconstruct data for a keyspace from the specified timeline.
4666 : ///
4667 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
4668 : /// the current keyspace. The current keyspace of the search at any given timeline
4669 : /// is the original keyspace minus all the keys that have been completed minus
4670 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
4671 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
4672 : ///
4673 : /// This is basically a depth-first search visitor implementation where a vertex
4674 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
4675 : ///
4676 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
4677 : /// and get all the required reconstruct data from the layer in one go.
4678 : ///
4679 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
4680 : /// decides how to deal with these two keyspaces.
4681 426935 : async fn get_vectored_reconstruct_data_timeline(
4682 426935 : timeline: &Timeline,
4683 426935 : query: &VersionedKeySpaceQuery,
4684 426935 : reconstruct_state: &mut ValuesReconstructState,
4685 426935 : cancel: &CancellationToken,
4686 426935 : ctx: &RequestContext,
4687 426935 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
4688 : // Prevent GC from progressing while visiting the current timeline.
4689 : // If we are GC-ing because a new image layer was added while traversing
4690 : // the timeline, then it will remove layers that are required for fulfilling
4691 : // the current get request (read-path cannot "look back" and notice the new
4692 : // image layer).
4693 426935 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
4694 :
4695 : // See `compaction::compact_with_gc` for why we need this.
4696 426935 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
4697 :
4698 : // Initialize the fringe
4699 426935 : let mut fringe = timeline.get_vectored_init_fringe(query).await?;
4700 :
4701 426935 : let mut completed_keyspace = KeySpace::default();
4702 426935 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
4703 :
4704 873870 : while let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
4705 446935 : if cancel.is_cancelled() {
4706 0 : return Err(GetVectoredError::Cancelled);
4707 446935 : }
4708 :
4709 446935 : if let Some(ref mut read_path) = reconstruct_state.read_path {
4710 446935 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
4711 446935 : }
4712 :
4713 : // Visit the layer and plan IOs for it
4714 446935 : let next_cont_lsn = lsn_range.start;
4715 446935 : layer_to_read
4716 446935 : .get_values_reconstruct_data(
4717 446935 : keyspace_to_read.clone(),
4718 446935 : lsn_range,
4719 446935 : reconstruct_state,
4720 446935 : ctx,
4721 446935 : )
4722 446935 : .await?;
4723 :
4724 446935 : let mut unmapped_keyspace = keyspace_to_read;
4725 446935 : let cont_lsn = next_cont_lsn;
4726 :
4727 446935 : reconstruct_state.on_layer_visited(&layer_to_read);
4728 :
4729 446935 : let (keys_done_last_step, keys_with_image_coverage) =
4730 446935 : reconstruct_state.consume_done_keys();
4731 446935 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
4732 446935 : completed_keyspace.merge(&keys_done_last_step);
4733 446935 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
4734 14229 : unmapped_keyspace
4735 14229 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
4736 14229 : image_covered_keyspace.add_range(keys_with_image_coverage);
4737 432706 : }
4738 :
4739 : // Query the layer map for the next layers to read.
4740 : //
4741 : // Do not descent any further if the last layer we visited
4742 : // completed all keys in the keyspace it inspected. This is not
4743 : // required for correctness, but avoids visiting extra layers
4744 : // which turns out to be a perf bottleneck in some cases.
4745 446935 : if !unmapped_keyspace.is_empty() {
4746 130175 : let guard = timeline.layers.read(LayerManagerLockHolder::GetPage).await;
4747 130175 : guard.update_search_fringe(&unmapped_keyspace, cont_lsn, &mut fringe)?;
4748 :
4749 : // It's safe to drop the layer map lock after planning the next round of reads.
4750 : // The fringe keeps readable handles for the layers which are safe to read even
4751 : // if layers were compacted or flushed.
4752 : //
4753 : // The more interesting consideration is: "Why is the read algorithm still correct
4754 : // if the layer map changes while it is operating?". Doing a vectored read on a
4755 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
4756 : // covered by the read. The layer map tells us how to move the lsn downwards for a
4757 : // range at *a particular point in time*. It is fine for the answer to be different
4758 : // at two different time points.
4759 130175 : drop(guard);
4760 316760 : }
4761 : }
4762 :
4763 426935 : Ok(TimelineVisitOutcome {
4764 426935 : completed_keyspace,
4765 426935 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
4766 426935 : })
4767 426935 : }
4768 :
4769 114506 : async fn get_ready_ancestor_timeline(
4770 114506 : &self,
4771 114506 : ancestor: &Arc<Timeline>,
4772 114506 : ctx: &RequestContext,
4773 114506 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
4774 : // It's possible that the ancestor timeline isn't active yet, or
4775 : // is active but hasn't yet caught up to the branch point. Wait
4776 : // for it.
4777 : //
4778 : // This cannot happen while the pageserver is running normally,
4779 : // because you cannot create a branch from a point that isn't
4780 : // present in the pageserver yet. However, we don't wait for the
4781 : // branch point to be uploaded to cloud storage before creating
4782 : // a branch. I.e., the branch LSN need not be remote consistent
4783 : // for the branching operation to succeed.
4784 : //
4785 : // Hence, if we try to load a tenant in such a state where
4786 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
4787 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
4788 : // then we will need to wait for the ancestor timeline to
4789 : // re-stream WAL up to branch_lsn before we access it.
4790 : //
4791 : // How can a tenant get in such a state?
4792 : // - ungraceful pageserver process exit
4793 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
4794 : //
4795 : // NB: this could be avoided by requiring
4796 : // branch_lsn >= remote_consistent_lsn
4797 : // during branch creation.
4798 114506 : match ancestor.wait_to_become_active(ctx).await {
4799 114505 : Ok(()) => {}
4800 : Err(TimelineState::Stopping) => {
4801 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
4802 0 : return Err(GetReadyAncestorError::Cancelled);
4803 : }
4804 1 : Err(state) => {
4805 1 : return Err(GetReadyAncestorError::BadState {
4806 1 : timeline_id: ancestor.timeline_id,
4807 1 : state,
4808 1 : });
4809 : }
4810 : }
4811 114505 : ancestor
4812 114505 : .wait_lsn(
4813 114505 : self.ancestor_lsn,
4814 114505 : WaitLsnWaiter::Timeline(self),
4815 114505 : WaitLsnTimeout::Default,
4816 114505 : ctx,
4817 114505 : )
4818 114505 : .await
4819 114505 : .map_err(|e| match e {
4820 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
4821 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
4822 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
4823 0 : timeline_id: ancestor.timeline_id,
4824 0 : state,
4825 0 : },
4826 0 : })?;
4827 :
4828 114505 : Ok(ancestor.clone())
4829 114506 : }
4830 :
4831 148592 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
4832 148592 : &self.shard_identity
4833 148592 : }
4834 :
4835 : #[inline(always)]
4836 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
4837 0 : ShardTimelineId {
4838 0 : shard_index: ShardIndex {
4839 0 : shard_number: self.shard_identity.number,
4840 0 : shard_count: self.shard_identity.count,
4841 0 : },
4842 0 : timeline_id: self.timeline_id,
4843 0 : }
4844 0 : }
4845 :
4846 : /// Returns a non-frozen open in-memory layer for ingestion.
4847 : ///
4848 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
4849 : /// this function without holding the mutex.
4850 660 : async fn get_layer_for_write(
4851 660 : &self,
4852 660 : lsn: Lsn,
4853 660 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4854 660 : ctx: &RequestContext,
4855 660 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
4856 660 : let mut guard = self
4857 660 : .layers
4858 660 : .write(LayerManagerLockHolder::GetLayerForWrite)
4859 660 : .await;
4860 :
4861 660 : let last_record_lsn = self.get_last_record_lsn();
4862 660 : ensure!(
4863 660 : lsn > last_record_lsn,
4864 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
4865 : lsn,
4866 : last_record_lsn,
4867 : );
4868 :
4869 660 : let layer = guard
4870 660 : .open_mut()?
4871 660 : .get_layer_for_write(
4872 660 : lsn,
4873 660 : self.conf,
4874 660 : self.timeline_id,
4875 660 : self.tenant_shard_id,
4876 660 : &self.gate,
4877 660 : &self.cancel,
4878 660 : ctx,
4879 : )
4880 660 : .await?;
4881 660 : Ok(layer)
4882 660 : }
4883 :
4884 2639560 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4885 2639560 : assert!(new_lsn.is_aligned());
4886 :
4887 2639560 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
4888 2639560 : self.last_record_lsn.advance(new_lsn);
4889 2639560 : }
4890 :
4891 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
4892 : ///
4893 : /// Unconditional flush loop notification is given because in sharded cases we will want to
4894 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
4895 611 : async fn freeze_inmem_layer_at(
4896 611 : &self,
4897 611 : at: Lsn,
4898 611 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4899 611 : ) -> Result<u64, FlushLayerError> {
4900 611 : let frozen = {
4901 611 : let mut guard = self
4902 611 : .layers
4903 611 : .write(LayerManagerLockHolder::TryFreezeLayer)
4904 611 : .await;
4905 611 : guard
4906 611 : .open_mut()?
4907 611 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
4908 611 : .await
4909 : };
4910 :
4911 611 : if frozen {
4912 597 : let now = Instant::now();
4913 597 : *(self.last_freeze_ts.write().unwrap()) = now;
4914 597 : }
4915 :
4916 : // Increment the flush cycle counter and wake up the flush task.
4917 : // Remember the new value, so that when we listen for the flush
4918 : // to finish, we know when the flush that we initiated has
4919 : // finished, instead of some other flush that was started earlier.
4920 611 : let mut my_flush_request = 0;
4921 :
4922 611 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
4923 611 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
4924 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
4925 611 : }
4926 :
4927 611 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
4928 611 : my_flush_request = *counter + 1;
4929 611 : *counter = my_flush_request;
4930 611 : *lsn = std::cmp::max(at, *lsn);
4931 611 : });
4932 :
4933 611 : assert_ne!(my_flush_request, 0);
4934 :
4935 611 : Ok(my_flush_request)
4936 611 : }
4937 :
4938 : /// Layer flusher task's main loop.
4939 232 : async fn flush_loop(
4940 232 : self: &Arc<Self>,
4941 232 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4942 232 : ctx: &RequestContext,
4943 232 : ) {
4944 : // Always notify waiters about the flush loop exiting since the loop might stop
4945 : // when the timeline hasn't been cancelled.
4946 232 : let scopeguard_rx = layer_flush_start_rx.clone();
4947 232 : scopeguard::defer! {
4948 : let (flush_counter, _) = *scopeguard_rx.borrow();
4949 : let _ = self
4950 : .layer_flush_done_tx
4951 : .send_replace((flush_counter, Err(FlushLayerError::Cancelled)));
4952 : }
4953 :
4954 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4955 232 : let mut watch_l0 = match self
4956 232 : .layers
4957 232 : .read(LayerManagerLockHolder::FlushLoop)
4958 232 : .await
4959 232 : .layer_map()
4960 : {
4961 232 : Ok(lm) => lm.watch_level0_deltas(),
4962 0 : Err(Shutdown) => return,
4963 : };
4964 :
4965 232 : info!("started flush loop");
4966 : loop {
4967 836 : tokio::select! {
4968 836 : _ = self.cancel.cancelled() => {
4969 5 : info!("shutting down layer flush task due to Timeline::cancel");
4970 5 : break;
4971 : },
4972 836 : _ = layer_flush_start_rx.changed() => {}
4973 : }
4974 604 : trace!("waking up");
4975 604 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4976 :
4977 : // The highest LSN to which we flushed in the loop over frozen layers
4978 604 : let mut flushed_to_lsn = Lsn(0);
4979 :
4980 604 : let result = loop {
4981 1200 : if self.cancel.is_cancelled() {
4982 0 : info!("dropping out of flush loop for timeline shutdown");
4983 0 : return;
4984 1200 : }
4985 :
4986 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4987 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4988 1200 : if flushed_to_lsn >= frozen_to_lsn {
4989 589 : break Ok(());
4990 611 : }
4991 :
4992 : // Fetch the next layer to flush, if any.
4993 611 : let (layer, l0_count, frozen_count, frozen_size, open_layer_size) = {
4994 611 : let layers = self.layers.read(LayerManagerLockHolder::FlushLoop).await;
4995 611 : let Ok(lm) = layers.layer_map() else {
4996 0 : info!("dropping out of flush loop for timeline shutdown");
4997 0 : return;
4998 : };
4999 611 : let l0_count = lm.level0_deltas().len();
5000 611 : let frozen_count = lm.frozen_layers.len();
5001 611 : let frozen_size: u64 = lm
5002 611 : .frozen_layers
5003 611 : .iter()
5004 611 : .map(|l| l.estimated_in_mem_size())
5005 611 : .sum();
5006 611 : let open_layer_size: u64 = lm
5007 611 : .open_layer
5008 611 : .as_ref()
5009 611 : .map(|l| l.estimated_in_mem_size())
5010 611 : .unwrap_or(0);
5011 611 : let layer = lm.frozen_layers.front().cloned();
5012 611 : (layer, l0_count, frozen_count, frozen_size, open_layer_size)
5013 : // drop 'layers' lock
5014 : };
5015 611 : let Some(layer) = layer else {
5016 14 : break Ok(());
5017 : };
5018 :
5019 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
5020 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
5021 597 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
5022 0 : if l0_count >= stall_threshold {
5023 0 : warn!(
5024 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
5025 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
5026 : );
5027 0 : let stall_timer = self
5028 0 : .metrics
5029 0 : .flush_delay_histo
5030 0 : .start_timer()
5031 0 : .record_on_drop();
5032 0 : tokio::select! {
5033 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
5034 0 : if let Ok(l0) = result.as_deref() {
5035 0 : let delay = stall_timer.elapsed().as_secs_f64();
5036 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
5037 0 : }
5038 : },
5039 0 : _ = self.cancel.cancelled() => {},
5040 : }
5041 0 : continue; // check again
5042 0 : }
5043 597 : }
5044 :
5045 : // Flush the layer.
5046 597 : let flush_timer = self.metrics.flush_time_histo.start_timer();
5047 597 : match self.flush_frozen_layer(layer, ctx).await {
5048 596 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
5049 : Err(FlushLayerError::Cancelled) => {
5050 0 : info!("dropping out of flush loop for timeline shutdown");
5051 0 : return;
5052 : }
5053 1 : err @ Err(
5054 : FlushLayerError::NotRunning(_)
5055 : | FlushLayerError::Other(_)
5056 : | FlushLayerError::CreateImageLayersError(_),
5057 : ) => {
5058 1 : error!("could not flush frozen layer: {err:?}");
5059 1 : break err.map(|_| ());
5060 : }
5061 : }
5062 596 : let flush_duration = flush_timer.stop_and_record();
5063 :
5064 : // Notify the tenant compaction loop if L0 compaction is needed.
5065 596 : let l0_count = *watch_l0.borrow();
5066 596 : if l0_count >= self.get_compaction_threshold() {
5067 238 : self.l0_compaction_trigger.notify_one();
5068 358 : }
5069 :
5070 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
5071 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
5072 : // ingestion by having ephemeral layer rolls wait for flushes.
5073 596 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
5074 4 : if l0_count >= delay_threshold {
5075 0 : let delay = flush_duration.as_secs_f64();
5076 0 : info!(
5077 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
5078 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
5079 : );
5080 0 : let _delay_timer = self
5081 0 : .metrics
5082 0 : .flush_delay_histo
5083 0 : .start_timer()
5084 0 : .record_on_drop();
5085 0 : tokio::select! {
5086 0 : _ = tokio::time::sleep(flush_duration) => {},
5087 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
5088 0 : _ = self.cancel.cancelled() => {},
5089 : }
5090 4 : }
5091 592 : }
5092 : };
5093 :
5094 : // Unsharded tenants should never advance their LSN beyond the end of the
5095 : // highest layer they write: such gaps between layer data and the frozen LSN
5096 : // are only legal on sharded tenants.
5097 604 : debug_assert!(
5098 604 : self.shard_identity.count.count() > 1
5099 597 : || flushed_to_lsn >= frozen_to_lsn
5100 14 : || !flushed_to_lsn.is_valid()
5101 : );
5102 :
5103 604 : if flushed_to_lsn < frozen_to_lsn
5104 15 : && self.shard_identity.count.count() > 1
5105 1 : && result.is_ok()
5106 : {
5107 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
5108 : // to us via layer_flush_start_rx, then advance it here.
5109 : //
5110 : // This path is only taken for tenants with multiple shards: single sharded tenants should
5111 : // never encounter a gap in the wal.
5112 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
5113 0 : tracing::debug!(
5114 0 : "Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}"
5115 : );
5116 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
5117 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
5118 0 : tracing::warn!(
5119 0 : "Failed to schedule metadata upload after updating disk_consistent_lsn: {e}"
5120 : );
5121 0 : }
5122 0 : }
5123 604 : }
5124 :
5125 : // Notify any listeners that we're done
5126 604 : let _ = self
5127 604 : .layer_flush_done_tx
5128 604 : .send_replace((flush_counter, result));
5129 : }
5130 5 : }
5131 :
5132 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
5133 571 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
5134 571 : let mut rx = self.layer_flush_done_tx.subscribe();
5135 : loop {
5136 : {
5137 1157 : let (last_result_counter, last_result) = &*rx.borrow();
5138 1157 : if *last_result_counter >= request {
5139 571 : if let Err(err) = last_result {
5140 : // We already logged the original error in
5141 : // flush_loop. We cannot propagate it to the caller
5142 : // here, because it might not be Cloneable
5143 1 : return Err(err.clone());
5144 : } else {
5145 570 : return Ok(());
5146 : }
5147 586 : }
5148 : }
5149 586 : trace!("waiting for flush to complete");
5150 586 : tokio::select! {
5151 586 : rx_e = rx.changed() => {
5152 586 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
5153 : },
5154 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
5155 : // the notification from [`flush_loop`] that it completed.
5156 586 : _ = self.cancel.cancelled() => {
5157 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
5158 0 : return Ok(())
5159 : }
5160 : };
5161 586 : trace!("done")
5162 : }
5163 571 : }
5164 :
5165 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
5166 : ///
5167 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
5168 : #[instrument(skip_all, fields(layer=%frozen_layer))]
5169 : async fn flush_frozen_layer(
5170 : self: &Arc<Self>,
5171 : frozen_layer: Arc<InMemoryLayer>,
5172 : ctx: &RequestContext,
5173 : ) -> Result<Lsn, FlushLayerError> {
5174 : debug_assert_current_span_has_tenant_and_timeline_id();
5175 :
5176 : // As a special case, when we have just imported an image into the repository,
5177 : // instead of writing out a L0 delta layer, we directly write out image layer
5178 : // files instead. This is possible as long as *all* the data imported into the
5179 : // repository have the same LSN.
5180 : let lsn_range = frozen_layer.get_lsn_range();
5181 :
5182 : // Whether to directly create image layers for this flush, or flush them as delta layers
5183 : let create_image_layer =
5184 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
5185 :
5186 : #[cfg(test)]
5187 : {
5188 : match &mut *self.flush_loop_state.lock().unwrap() {
5189 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
5190 : panic!("flush loop not running")
5191 : }
5192 : FlushLoopState::Running {
5193 : expect_initdb_optimization,
5194 : initdb_optimization_count,
5195 : ..
5196 : } => {
5197 : if create_image_layer {
5198 : *initdb_optimization_count += 1;
5199 : } else {
5200 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
5201 : }
5202 : }
5203 : }
5204 : }
5205 :
5206 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
5207 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
5208 : // require downloading anything during initial import.
5209 : let ((rel_partition, metadata_partition), _lsn) = self
5210 : .repartition(
5211 : self.initdb_lsn,
5212 : self.get_compaction_target_size(),
5213 : EnumSet::empty(),
5214 : ctx,
5215 : )
5216 : .await
5217 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into_anyhow()))?;
5218 :
5219 : if self.cancel.is_cancelled() {
5220 : return Err(FlushLayerError::Cancelled);
5221 : }
5222 :
5223 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
5224 : // So that the key ranges don't overlap.
5225 : let mut partitions = KeyPartitioning::default();
5226 : partitions.parts.extend(rel_partition.parts);
5227 : if !metadata_partition.parts.is_empty() {
5228 : assert_eq!(
5229 : metadata_partition.parts.len(),
5230 : 1,
5231 : "currently sparse keyspace should only contain a single metadata keyspace"
5232 : );
5233 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
5234 : // every single key within the keyspace, and therefore, it's safe to force converting it
5235 : // into a dense keyspace before calling this function.
5236 : partitions
5237 : .parts
5238 : .extend(metadata_partition.into_dense().parts);
5239 : }
5240 :
5241 : let mut layers_to_upload = Vec::new();
5242 : let (generated_image_layers, is_complete) = self
5243 : .create_image_layers(
5244 : &partitions,
5245 : self.initdb_lsn,
5246 : None,
5247 : ImageLayerCreationMode::Initial,
5248 : ctx,
5249 : LastImageLayerCreationStatus::Initial,
5250 : false, // don't yield for L0, we're flushing L0
5251 : )
5252 : .instrument(info_span!("create_image_layers", mode = %ImageLayerCreationMode::Initial, partition_mode = "initial", lsn = %self.initdb_lsn))
5253 : .await?;
5254 : debug_assert!(
5255 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
5256 : "init image generation mode must fully cover the keyspace"
5257 : );
5258 : layers_to_upload.extend(generated_image_layers);
5259 :
5260 : (layers_to_upload, None)
5261 : } else {
5262 : // Normal case, write out a L0 delta layer file.
5263 : // `create_delta_layer` will not modify the layer map.
5264 : // We will remove frozen layer and add delta layer in one atomic operation later.
5265 : let Some(layer) = self
5266 : .create_delta_layer(&frozen_layer, None, ctx)
5267 : .await
5268 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
5269 : else {
5270 : panic!("delta layer cannot be empty if no filter is applied");
5271 : };
5272 : (
5273 : // FIXME: even though we have a single image and single delta layer assumption
5274 : // we push them to vec
5275 : vec![layer.clone()],
5276 : Some(layer),
5277 : )
5278 : };
5279 :
5280 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
5281 :
5282 : if self.cancel.is_cancelled() {
5283 : return Err(FlushLayerError::Cancelled);
5284 : }
5285 :
5286 1 : fail_point!("flush-layer-before-update-remote-consistent-lsn", |_| {
5287 1 : Err(FlushLayerError::Other(anyhow!("failpoint").into()))
5288 1 : });
5289 :
5290 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
5291 :
5292 : // The new on-disk layers are now in the layer map. We can remove the
5293 : // in-memory layer from the map now. The flushed layer is stored in
5294 : // the mapping in `create_delta_layer`.
5295 : {
5296 : let mut guard = self
5297 : .layers
5298 : .write(LayerManagerLockHolder::FlushFrozenLayer)
5299 : .await;
5300 :
5301 : guard.open_mut()?.finish_flush_l0_layer(
5302 : delta_layer_to_add.as_ref(),
5303 : &frozen_layer,
5304 : &self.metrics,
5305 : );
5306 :
5307 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
5308 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
5309 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
5310 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
5311 : }
5312 : // release lock on 'layers'
5313 : };
5314 :
5315 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
5316 : // a compaction can delete the file and then it won't be available for uploads any more.
5317 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
5318 : // race situation.
5319 : // See https://github.com/neondatabase/neon/issues/4526
5320 : pausable_failpoint!("flush-frozen-pausable");
5321 :
5322 : // This failpoint is used by another test case `test_pageserver_recovery`.
5323 : fail_point!("flush-frozen-exit");
5324 :
5325 : Ok(Lsn(lsn_range.end.0 - 1))
5326 : }
5327 :
5328 : /// Return true if the value changed
5329 : ///
5330 : /// This function must only be used from the layer flush task.
5331 596 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
5332 596 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
5333 596 : assert!(
5334 596 : new_value >= old_value,
5335 0 : "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}"
5336 : );
5337 :
5338 596 : self.metrics
5339 596 : .disk_consistent_lsn_gauge
5340 596 : .set(new_value.0 as i64);
5341 596 : new_value != old_value
5342 596 : }
5343 :
5344 : /// Update metadata file
5345 623 : fn schedule_uploads(
5346 623 : &self,
5347 623 : disk_consistent_lsn: Lsn,
5348 623 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
5349 623 : ) -> anyhow::Result<()> {
5350 : // We can only save a valid 'prev_record_lsn' value on disk if we
5351 : // flushed *all* in-memory changes to disk. We only track
5352 : // 'prev_record_lsn' in memory for the latest processed record, so we
5353 : // don't remember what the correct value that corresponds to some old
5354 : // LSN is. But if we flush everything, then the value corresponding
5355 : // current 'last_record_lsn' is correct and we can store it on disk.
5356 : let RecordLsn {
5357 623 : last: last_record_lsn,
5358 623 : prev: prev_record_lsn,
5359 623 : } = self.last_record_lsn.load();
5360 623 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
5361 561 : Some(prev_record_lsn)
5362 : } else {
5363 62 : None
5364 : };
5365 :
5366 623 : let update = crate::tenant::metadata::MetadataUpdate::new(
5367 623 : disk_consistent_lsn,
5368 623 : ondisk_prev_record_lsn,
5369 623 : *self.applied_gc_cutoff_lsn.read(),
5370 : );
5371 :
5372 623 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
5373 0 : "{}",
5374 0 : x.unwrap()
5375 : ));
5376 :
5377 1225 : for layer in layers_to_upload {
5378 602 : self.remote_client.schedule_layer_file_upload(layer)?;
5379 : }
5380 623 : self.remote_client
5381 623 : .schedule_index_upload_for_metadata_update(&update)?;
5382 :
5383 623 : Ok(())
5384 623 : }
5385 :
5386 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
5387 0 : self.remote_client
5388 0 : .preserve_initdb_archive(
5389 0 : &self.tenant_shard_id.tenant_id,
5390 0 : &self.timeline_id,
5391 0 : &self.cancel,
5392 0 : )
5393 0 : .await
5394 0 : }
5395 :
5396 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
5397 : // in layer map immediately. The caller is responsible to put it into the layer map.
5398 486 : async fn create_delta_layer(
5399 486 : self: &Arc<Self>,
5400 486 : frozen_layer: &Arc<InMemoryLayer>,
5401 486 : key_range: Option<Range<Key>>,
5402 486 : ctx: &RequestContext,
5403 486 : ) -> anyhow::Result<Option<ResidentLayer>> {
5404 486 : let self_clone = Arc::clone(self);
5405 486 : let frozen_layer = Arc::clone(frozen_layer);
5406 486 : let ctx = ctx.attached_child();
5407 486 : let work = async move {
5408 486 : let Some((desc, path)) = frozen_layer
5409 486 : .write_to_disk(
5410 486 : &ctx,
5411 486 : key_range,
5412 486 : self_clone.l0_flush_global_state.inner(),
5413 486 : &self_clone.gate,
5414 486 : self_clone.cancel.clone(),
5415 486 : )
5416 486 : .await?
5417 : else {
5418 0 : return Ok(None);
5419 : };
5420 486 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
5421 :
5422 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
5423 : // We just need to fsync the directory in which these inodes are linked,
5424 : // which we know to be the timeline directory.
5425 : //
5426 : // We use fatal_err() below because the after write_to_disk returns with success,
5427 : // the in-memory state of the filesystem already has the layer file in its final place,
5428 : // and subsequent pageserver code could think it's durable while it really isn't.
5429 486 : let timeline_dir = VirtualFile::open(
5430 486 : &self_clone
5431 486 : .conf
5432 486 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
5433 486 : &ctx,
5434 486 : )
5435 486 : .await
5436 486 : .fatal_err("VirtualFile::open for timeline dir fsync");
5437 486 : timeline_dir
5438 486 : .sync_all()
5439 486 : .await
5440 486 : .fatal_err("VirtualFile::sync_all timeline dir");
5441 486 : anyhow::Ok(Some(new_delta))
5442 486 : };
5443 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
5444 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
5445 : use crate::virtual_file::io_engine::IoEngine;
5446 486 : match crate::virtual_file::io_engine::get() {
5447 0 : IoEngine::NotSet => panic!("io engine not set"),
5448 : IoEngine::StdFs => {
5449 0 : let span = tracing::info_span!("blocking");
5450 0 : tokio::task::spawn_blocking({
5451 0 : move || Handle::current().block_on(work.instrument(span))
5452 : })
5453 0 : .await
5454 0 : .context("spawn_blocking")
5455 0 : .and_then(|x| x)
5456 : }
5457 : #[cfg(target_os = "linux")]
5458 486 : IoEngine::TokioEpollUring => work.await,
5459 : }
5460 486 : }
5461 :
5462 303 : async fn repartition(
5463 303 : &self,
5464 303 : lsn: Lsn,
5465 303 : partition_size: u64,
5466 303 : flags: EnumSet<CompactFlags>,
5467 303 : ctx: &RequestContext,
5468 303 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), RepartitionError> {
5469 303 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
5470 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
5471 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
5472 : // and hence before the compaction task starts.
5473 0 : return Err(RepartitionError::Other(anyhow!(
5474 0 : "repartition() called concurrently"
5475 0 : )));
5476 : };
5477 303 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
5478 303 : if lsn < *partition_lsn {
5479 0 : return Err(RepartitionError::Other(anyhow!(
5480 0 : "repartition() called with LSN going backwards, this should not happen"
5481 0 : )));
5482 303 : }
5483 :
5484 303 : let distance = lsn.0 - partition_lsn.0;
5485 303 : if *partition_lsn != Lsn(0)
5486 141 : && distance <= self.repartition_threshold
5487 141 : && !flags.contains(CompactFlags::ForceRepartition)
5488 : {
5489 134 : debug!(
5490 : distance,
5491 : threshold = self.repartition_threshold,
5492 0 : "no repartitioning needed"
5493 : );
5494 134 : return Ok((
5495 134 : (dense_partition.clone(), sparse_partition.clone()),
5496 134 : *partition_lsn,
5497 134 : ));
5498 169 : }
5499 :
5500 169 : let (dense_ks, sparse_ks) = self
5501 169 : .collect_keyspace(lsn, ctx)
5502 169 : .await
5503 169 : .map_err(RepartitionError::CollectKeyspace)?;
5504 169 : let dense_partitioning = dense_ks.partition(
5505 169 : &self.shard_identity,
5506 169 : partition_size,
5507 169 : postgres_ffi::BLCKSZ as u64,
5508 : );
5509 169 : let sparse_partitioning = SparseKeyPartitioning {
5510 169 : parts: vec![sparse_ks],
5511 169 : }; // no partitioning for metadata keys for now
5512 169 : let result = ((dense_partitioning, sparse_partitioning), lsn);
5513 169 : guard.write(result.clone());
5514 169 : Ok(result)
5515 303 : }
5516 :
5517 : // Is it time to create a new image layer for the given partition? True if we want to generate.
5518 57 : async fn time_for_new_image_layer(
5519 57 : &self,
5520 57 : partition: &KeySpace,
5521 57 : lsn: Lsn,
5522 57 : force_image_creation_lsn: Option<Lsn>,
5523 57 : ) -> bool {
5524 57 : let threshold = self.get_image_creation_threshold();
5525 :
5526 57 : let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
5527 57 : let Ok(layers) = guard.layer_map() else {
5528 0 : return false;
5529 : };
5530 57 : let mut min_image_lsn: Lsn = Lsn::MAX;
5531 57 : let mut max_deltas = 0;
5532 364 : for part_range in &partition.ranges {
5533 307 : let image_coverage = layers.image_coverage(part_range, lsn);
5534 614 : for (img_range, last_img) in image_coverage {
5535 307 : let img_lsn = if let Some(last_img) = last_img {
5536 0 : last_img.get_lsn_range().end
5537 : } else {
5538 307 : Lsn(0)
5539 : };
5540 : // Let's consider an example:
5541 : //
5542 : // delta layer with LSN range 71-81
5543 : // delta layer with LSN range 81-91
5544 : // delta layer with LSN range 91-101
5545 : // image layer at LSN 100
5546 : //
5547 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
5548 : // there's no need to create a new one. We check this case explicitly, to avoid passing
5549 : // a bogus range to count_deltas below, with start > end. It's even possible that there
5550 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
5551 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
5552 307 : if img_lsn < lsn {
5553 307 : let num_deltas =
5554 307 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
5555 :
5556 307 : max_deltas = max_deltas.max(num_deltas);
5557 307 : if num_deltas >= threshold {
5558 0 : debug!(
5559 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
5560 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
5561 : );
5562 0 : return true;
5563 307 : }
5564 0 : }
5565 307 : min_image_lsn = min(min_image_lsn, img_lsn);
5566 : }
5567 : }
5568 :
5569 : // HADRON
5570 : // for child timelines, we consider all pages up to ancestor_LSN are redone successfully by the parent timeline
5571 57 : min_image_lsn = min_image_lsn.max(self.get_ancestor_lsn());
5572 57 : if min_image_lsn < force_image_creation_lsn.unwrap_or(Lsn(0)) && max_deltas > 0 {
5573 0 : info!(
5574 0 : "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}, num deltas: {}",
5575 0 : partition.ranges[0].start,
5576 0 : partition.ranges[0].end,
5577 : min_image_lsn,
5578 0 : force_image_creation_lsn.unwrap(),
5579 : max_deltas
5580 : );
5581 0 : return true;
5582 57 : }
5583 :
5584 57 : debug!(
5585 : max_deltas,
5586 0 : "none of the partitioned ranges had >= {threshold} deltas"
5587 : );
5588 57 : false
5589 57 : }
5590 :
5591 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
5592 : /// so that at most one image layer will be produced from this function.
5593 : #[allow(clippy::too_many_arguments)]
5594 124 : async fn create_image_layer_for_rel_blocks(
5595 124 : self: &Arc<Self>,
5596 124 : partition: &KeySpace,
5597 124 : mut image_layer_writer: ImageLayerWriter,
5598 124 : lsn: Lsn,
5599 124 : ctx: &RequestContext,
5600 124 : img_range: Range<Key>,
5601 124 : io_concurrency: IoConcurrency,
5602 124 : progress: Option<(usize, usize)>,
5603 124 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5604 124 : let mut wrote_keys = false;
5605 :
5606 124 : let mut key_request_accum = KeySpaceAccum::new();
5607 824 : for range in &partition.ranges {
5608 700 : let mut key = range.start;
5609 1517 : while key < range.end {
5610 : // Decide whether to retain this key: usually we do, but sharded tenants may
5611 : // need to drop keys that don't belong to them. If we retain the key, add it
5612 : // to `key_request_accum` for later issuing a vectored get
5613 817 : if self.shard_identity.is_key_disposable(&key) {
5614 0 : debug!(
5615 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
5616 : key,
5617 0 : self.shard_identity.get_shard_number(&key)
5618 : );
5619 817 : } else {
5620 817 : key_request_accum.add_key(key);
5621 817 : }
5622 :
5623 817 : let last_key_in_range = key.next() == range.end;
5624 817 : key = key.next();
5625 :
5626 : // Maybe flush `key_rest_accum`
5627 817 : if key_request_accum.raw_size() >= self.conf.max_get_vectored_keys.get() as u64
5628 817 : || (last_key_in_range && key_request_accum.raw_size() > 0)
5629 : {
5630 700 : let query =
5631 700 : VersionedKeySpaceQuery::uniform(key_request_accum.consume_keyspace(), lsn);
5632 :
5633 700 : let results = self
5634 700 : .get_vectored(query, io_concurrency.clone(), ctx)
5635 700 : .await?;
5636 :
5637 700 : if self.cancel.is_cancelled() {
5638 0 : return Err(CreateImageLayersError::Cancelled);
5639 700 : }
5640 :
5641 1517 : for (img_key, img) in results {
5642 817 : let img = match img {
5643 817 : Ok(img) => img,
5644 0 : Err(err) => {
5645 : // If we fail to reconstruct a VM or FSM page, we can zero the
5646 : // page without losing any actual user data. That seems better
5647 : // than failing repeatedly and getting stuck.
5648 : //
5649 : // We had a bug at one point, where we truncated the FSM and VM
5650 : // in the pageserver, but the Postgres didn't know about that
5651 : // and continued to generate incremental WAL records for pages
5652 : // that didn't exist in the pageserver. Trying to replay those
5653 : // WAL records failed to find the previous image of the page.
5654 : // This special case allows us to recover from that situation.
5655 : // See https://github.com/neondatabase/neon/issues/2601.
5656 : //
5657 : // Unfortunately we cannot do this for the main fork, or for
5658 : // any metadata keys, keys, as that would lead to actual data
5659 : // loss.
5660 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
5661 0 : warn!(
5662 0 : "could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}"
5663 : );
5664 0 : ZERO_PAGE.clone()
5665 : } else {
5666 0 : return Err(CreateImageLayersError::from(err));
5667 : }
5668 : }
5669 : };
5670 :
5671 : // Write all the keys we just read into our new image layer.
5672 817 : image_layer_writer.put_image(img_key, img, ctx).await?;
5673 817 : wrote_keys = true;
5674 : }
5675 117 : }
5676 : }
5677 : }
5678 :
5679 124 : let progress_report = progress
5680 124 : .map(|(idx, total)| format!("({idx}/{total}) "))
5681 124 : .unwrap_or_default();
5682 124 : if wrote_keys {
5683 : // Normal path: we have written some data into the new image layer for this
5684 : // partition, so flush it to disk.
5685 124 : info!(
5686 0 : "{} produced image layer for rel {}",
5687 : progress_report,
5688 0 : ImageLayerName {
5689 0 : key_range: img_range.clone(),
5690 0 : lsn
5691 0 : },
5692 : );
5693 124 : Ok(ImageLayerCreationOutcome::Generated {
5694 124 : unfinished_image_layer: image_layer_writer,
5695 124 : })
5696 : } else {
5697 0 : tracing::debug!(
5698 0 : "{} no data in range {}-{}",
5699 : progress_report,
5700 : img_range.start,
5701 : img_range.end
5702 : );
5703 0 : Ok(ImageLayerCreationOutcome::Empty)
5704 : }
5705 124 : }
5706 :
5707 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
5708 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
5709 : /// would not be too large to fit in a single image layer.
5710 : ///
5711 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
5712 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
5713 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
5714 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
5715 : #[allow(clippy::too_many_arguments)]
5716 169 : async fn create_image_layer_for_metadata_keys(
5717 169 : self: &Arc<Self>,
5718 169 : partition: &KeySpace,
5719 169 : mut image_layer_writer: ImageLayerWriter,
5720 169 : lsn: Lsn,
5721 169 : ctx: &RequestContext,
5722 169 : img_range: Range<Key>,
5723 169 : mode: ImageLayerCreationMode,
5724 169 : io_concurrency: IoConcurrency,
5725 169 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5726 : // Metadata keys image layer creation.
5727 169 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
5728 169 : let begin = Instant::now();
5729 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
5730 : // not contain too many keys, otherwise this takes a lot of memory.
5731 169 : let data = self
5732 169 : .get_vectored_impl(
5733 169 : VersionedKeySpaceQuery::uniform(partition.clone(), lsn),
5734 169 : &mut reconstruct_state,
5735 169 : ctx,
5736 169 : )
5737 169 : .await?;
5738 169 : let (data, total_kb_retrieved, total_keys_retrieved) = {
5739 169 : let mut new_data = BTreeMap::new();
5740 169 : let mut total_kb_retrieved = 0;
5741 169 : let mut total_keys_retrieved = 0;
5742 5175 : for (k, v) in data {
5743 5006 : let v = v?;
5744 5006 : total_kb_retrieved += KEY_SIZE + v.len();
5745 5006 : total_keys_retrieved += 1;
5746 5006 : new_data.insert(k, v);
5747 : }
5748 169 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
5749 : };
5750 169 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
5751 169 : let elapsed = begin.elapsed();
5752 :
5753 169 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
5754 169 : info!(
5755 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",
5756 0 : elapsed.as_secs_f64()
5757 : );
5758 :
5759 169 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
5760 16 : return Ok(ImageLayerCreationOutcome::Skip);
5761 153 : }
5762 153 : if self.cancel.is_cancelled() {
5763 0 : return Err(CreateImageLayersError::Cancelled);
5764 153 : }
5765 153 : let mut wrote_any_image = false;
5766 5159 : for (k, v) in data {
5767 5006 : if v.is_empty() {
5768 : // the key has been deleted, it does not need an image
5769 : // in metadata keyspace, an empty image == tombstone
5770 4 : continue;
5771 5002 : }
5772 5002 : wrote_any_image = true;
5773 :
5774 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
5775 :
5776 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
5777 : // on the normal data path either.
5778 5002 : image_layer_writer.put_image(k, v, ctx).await?;
5779 : }
5780 :
5781 153 : if wrote_any_image {
5782 : // Normal path: we have written some data into the new image layer for this
5783 : // partition, so flush it to disk.
5784 6 : info!(
5785 0 : "created image layer for metadata {}",
5786 0 : ImageLayerName {
5787 0 : key_range: img_range.clone(),
5788 0 : lsn
5789 0 : }
5790 : );
5791 6 : Ok(ImageLayerCreationOutcome::Generated {
5792 6 : unfinished_image_layer: image_layer_writer,
5793 6 : })
5794 : } else {
5795 147 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
5796 147 : Ok(ImageLayerCreationOutcome::Empty)
5797 : }
5798 169 : }
5799 :
5800 : /// Predicate function which indicates whether we should check if new image layers
5801 : /// are required. Since checking if new image layers are required is expensive in
5802 : /// terms of CPU, we only do it in the following cases:
5803 : /// 1. If the timeline has ingested sufficient WAL to justify the cost or ...
5804 : /// 2. If enough time has passed since the last check:
5805 : /// 1. For large tenants, we wish to perform the check more often since they
5806 : /// suffer from the lack of image layers. Note that we assume sharded tenants
5807 : /// to be large since non-zero shards do not track the logical size.
5808 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
5809 191 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
5810 191 : let large_timeline_threshold = self.conf.image_layer_generation_large_timeline_threshold;
5811 :
5812 191 : let last_checks_at = self.last_image_layer_creation_check_at.load();
5813 191 : let distance = lsn
5814 191 : .checked_sub(last_checks_at)
5815 191 : .expect("Attempt to compact with LSN going backwards");
5816 191 : let min_distance =
5817 191 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
5818 :
5819 191 : let distance_based_decision = distance.0 >= min_distance;
5820 :
5821 191 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
5822 191 : let check_required_after = (|| {
5823 191 : if self.shard_identity.is_unsharded() {
5824 135 : if let CurrentLogicalSize::Exact(logical_size) =
5825 186 : self.current_logical_size.current_size()
5826 : {
5827 135 : if Some(Into::<u64>::into(&logical_size)) < large_timeline_threshold {
5828 135 : return Duration::from_secs(3600 * 48);
5829 0 : }
5830 51 : }
5831 5 : }
5832 :
5833 56 : self.get_checkpoint_timeout()
5834 : })();
5835 :
5836 191 : let time_based_decision = match *last_check_instant {
5837 29 : Some(last_check) => {
5838 29 : let elapsed = last_check.elapsed();
5839 29 : elapsed >= check_required_after
5840 : }
5841 162 : None => true,
5842 : };
5843 :
5844 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
5845 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
5846 : // check.
5847 191 : let decision = distance_based_decision || time_based_decision;
5848 191 : tracing::info!(
5849 0 : "Decided to check image layers: {}. Distance-based decision: {}, time-based decision: {}",
5850 : decision,
5851 : distance_based_decision,
5852 : time_based_decision
5853 : );
5854 191 : if decision {
5855 162 : self.last_image_layer_creation_check_at.store(lsn);
5856 162 : *last_check_instant = Some(Instant::now());
5857 162 : }
5858 :
5859 191 : decision
5860 191 : }
5861 :
5862 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
5863 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
5864 : ///
5865 : /// `partition_mode` is only for logging purpose and is not used anywhere in this function.
5866 : #[allow(clippy::too_many_arguments)]
5867 191 : async fn create_image_layers(
5868 191 : self: &Arc<Timeline>,
5869 191 : partitioning: &KeyPartitioning,
5870 191 : lsn: Lsn,
5871 191 : force_image_creation_lsn: Option<Lsn>,
5872 191 : mode: ImageLayerCreationMode,
5873 191 : ctx: &RequestContext,
5874 191 : last_status: LastImageLayerCreationStatus,
5875 191 : yield_for_l0: bool,
5876 191 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
5877 191 : let timer = self.metrics.create_images_time_histo.start_timer();
5878 :
5879 191 : if partitioning.parts.is_empty() {
5880 0 : warn!("no partitions to create image layers for");
5881 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5882 191 : }
5883 :
5884 : // We need to avoid holes between generated image layers.
5885 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
5886 : // image layer with hole between them. In this case such layer can not be utilized by GC.
5887 : //
5888 : // How such hole between partitions can appear?
5889 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
5890 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
5891 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
5892 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
5893 191 : let mut start = Key::MIN;
5894 :
5895 191 : let check_for_image_layers =
5896 191 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5897 0 : info!(
5898 0 : "resuming image layer creation: last_status=incomplete, continue from {}",
5899 : last_key
5900 : );
5901 0 : true
5902 : } else {
5903 191 : self.should_check_if_image_layers_required(lsn)
5904 : };
5905 :
5906 191 : let mut batch_image_writer = BatchLayerWriter::new(self.conf);
5907 :
5908 191 : let mut all_generated = true;
5909 :
5910 191 : let mut partition_processed = 0;
5911 191 : let mut total_partitions = partitioning.parts.len();
5912 191 : let mut last_partition_processed = None;
5913 191 : let mut partition_parts = partitioning.parts.clone();
5914 :
5915 191 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5916 : // We need to skip the partitions that have already been processed.
5917 0 : let mut found = false;
5918 0 : for (i, partition) in partition_parts.iter().enumerate() {
5919 0 : if last_key <= partition.end().unwrap() {
5920 : // ```plain
5921 : // |------|--------|----------|------|
5922 : // ^last_key
5923 : // ^start from this partition
5924 : // ```
5925 : // Why `i+1` instead of `i`?
5926 : // It is possible that the user did some writes after the previous image layer creation attempt so that
5927 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
5928 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
5929 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
5930 0 : if i + 1 >= total_partitions {
5931 : // In general, this case should not happen -- if last_key is on the last partition, the previous
5932 : // iteration of image layer creation should return a complete status.
5933 0 : break; // with found=false
5934 0 : }
5935 0 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
5936 0 : total_partitions = partition_parts.len();
5937 : // Update the start key to the partition start.
5938 0 : start = partition_parts[0].start().unwrap();
5939 0 : found = true;
5940 0 : break;
5941 0 : }
5942 : }
5943 0 : if !found {
5944 : // Last key is within the last partition, or larger than all partitions.
5945 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5946 0 : }
5947 191 : }
5948 :
5949 191 : let total = partition_parts.len();
5950 394 : for (idx, partition) in partition_parts.iter().enumerate() {
5951 394 : if self.cancel.is_cancelled() {
5952 0 : return Err(CreateImageLayersError::Cancelled);
5953 394 : }
5954 394 : partition_processed += 1;
5955 394 : let img_range = start..partition.ranges.last().unwrap().end;
5956 394 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
5957 394 : if compact_metadata {
5958 764 : for range in &partition.ranges {
5959 573 : assert!(
5960 573 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
5961 573 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
5962 0 : "metadata keys must be partitioned separately"
5963 : );
5964 : }
5965 191 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
5966 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
5967 : // might mess up with evictions.
5968 22 : start = img_range.end;
5969 22 : continue;
5970 169 : }
5971 : // For initial and force modes, we always generate image layers for metadata keys.
5972 203 : } else if let ImageLayerCreationMode::Try = mode {
5973 : // check_for_image_layers = false -> skip
5974 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
5975 79 : if !check_for_image_layers
5976 57 : || !self
5977 57 : .time_for_new_image_layer(partition, lsn, force_image_creation_lsn)
5978 57 : .await
5979 : {
5980 79 : start = img_range.end;
5981 79 : continue;
5982 0 : }
5983 124 : }
5984 293 : if let ImageLayerCreationMode::Force = mode {
5985 : // When forced to create image layers, we might try and create them where they already
5986 : // exist. This mode is only used in tests/debug.
5987 14 : let layers = self.layers.read(LayerManagerLockHolder::Compaction).await;
5988 14 : if layers.contains_key(&PersistentLayerKey {
5989 14 : key_range: img_range.clone(),
5990 14 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
5991 14 : is_delta: false,
5992 14 : }) {
5993 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
5994 : // in the future.
5995 0 : tracing::info!(
5996 0 : "Skipping image layer at {lsn} {}..{}, already exists",
5997 : img_range.start,
5998 : img_range.end
5999 : );
6000 0 : start = img_range.end;
6001 0 : continue;
6002 14 : }
6003 279 : }
6004 :
6005 293 : let image_layer_writer = ImageLayerWriter::new(
6006 293 : self.conf,
6007 293 : self.timeline_id,
6008 293 : self.tenant_shard_id,
6009 293 : &img_range,
6010 293 : lsn,
6011 293 : &self.gate,
6012 293 : self.cancel.clone(),
6013 293 : ctx,
6014 293 : )
6015 293 : .await
6016 293 : .map_err(CreateImageLayersError::Other)?;
6017 :
6018 293 : fail_point!("image-layer-writer-fail-before-finish", |_| {
6019 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
6020 0 : "failpoint image-layer-writer-fail-before-finish"
6021 0 : )))
6022 0 : });
6023 :
6024 : // Begin Hadron
6025 : //
6026 293 : fail_point!("create-image-layer-fail-simulated-corruption", |_| {
6027 0 : self.corruption_detected
6028 0 : .store(true, std::sync::atomic::Ordering::Relaxed);
6029 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
6030 0 : "failpoint create-image-layer-fail-simulated-corruption"
6031 0 : )))
6032 0 : });
6033 : // End Hadron
6034 :
6035 293 : let io_concurrency = IoConcurrency::spawn_from_conf(
6036 293 : self.conf.get_vectored_concurrent_io,
6037 293 : self.gate
6038 293 : .enter()
6039 293 : .map_err(|_| CreateImageLayersError::Cancelled)?,
6040 : );
6041 :
6042 293 : let outcome = if !compact_metadata {
6043 124 : self.create_image_layer_for_rel_blocks(
6044 124 : partition,
6045 124 : image_layer_writer,
6046 124 : lsn,
6047 124 : ctx,
6048 124 : img_range.clone(),
6049 124 : io_concurrency,
6050 124 : Some((idx, total)),
6051 124 : )
6052 124 : .await?
6053 : } else {
6054 169 : self.create_image_layer_for_metadata_keys(
6055 169 : partition,
6056 169 : image_layer_writer,
6057 169 : lsn,
6058 169 : ctx,
6059 169 : img_range.clone(),
6060 169 : mode,
6061 169 : io_concurrency,
6062 169 : )
6063 169 : .await?
6064 : };
6065 293 : match outcome {
6066 147 : ImageLayerCreationOutcome::Empty => {
6067 147 : // No data in this partition, so we don't need to create an image layer (for now).
6068 147 : // The next image layer should cover this key range, so we don't advance the `start`
6069 147 : // key.
6070 147 : }
6071 : ImageLayerCreationOutcome::Generated {
6072 130 : unfinished_image_layer,
6073 130 : } => {
6074 130 : batch_image_writer.add_unfinished_image_writer(
6075 130 : unfinished_image_layer,
6076 130 : img_range.clone(),
6077 130 : lsn,
6078 130 : );
6079 130 : // The next image layer should be generated right after this one.
6080 130 : start = img_range.end;
6081 130 : }
6082 16 : ImageLayerCreationOutcome::Skip => {
6083 16 : // We don't need to create an image layer for this partition.
6084 16 : // The next image layer should NOT cover this range, otherwise
6085 16 : // the keyspace becomes empty (reads don't go past image layers).
6086 16 : start = img_range.end;
6087 16 : }
6088 : }
6089 :
6090 293 : if let ImageLayerCreationMode::Try = mode {
6091 : // We have at least made some progress
6092 51 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
6093 : // The `Try` mode is currently only used on the compaction path. We want to avoid
6094 : // image layer generation taking too long time and blocking L0 compaction. So in this
6095 : // mode, we also inspect the current number of L0 layers and skip image layer generation
6096 : // if there are too many of them.
6097 0 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
6098 0 : * self.get_compaction_threshold();
6099 : // TODO: currently we do not respect `get_image_creation_preempt_threshold` and always yield
6100 : // when there is a single timeline with more than L0 threshold L0 layers. As long as the
6101 : // `get_image_creation_preempt_threshold` is set to a value greater than 0, we will yield for L0 compaction.
6102 0 : if image_preempt_threshold != 0 {
6103 0 : let should_yield = self
6104 0 : .l0_compaction_trigger
6105 0 : .notified()
6106 0 : .now_or_never()
6107 0 : .is_some();
6108 0 : if should_yield {
6109 0 : tracing::info!(
6110 0 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers",
6111 0 : partition.start().unwrap(),
6112 0 : partition.end().unwrap()
6113 : );
6114 0 : last_partition_processed = Some(partition.clone());
6115 0 : all_generated = false;
6116 0 : break;
6117 0 : }
6118 0 : }
6119 51 : }
6120 242 : }
6121 : }
6122 :
6123 191 : let image_layers = batch_image_writer
6124 191 : .finish(self, ctx)
6125 191 : .await
6126 191 : .map_err(CreateImageLayersError::Other)?;
6127 :
6128 191 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
6129 :
6130 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
6131 : // now they are being scheduled outside of write lock; current way is inconsistent with
6132 : // compaction lock order.
6133 191 : guard
6134 191 : .open_mut()?
6135 191 : .track_new_image_layers(&image_layers, &self.metrics);
6136 191 : drop_layer_manager_wlock(guard);
6137 191 : let duration = timer.stop_and_record();
6138 :
6139 : // Creating image layers may have caused some previously visible layers to be covered
6140 191 : if !image_layers.is_empty() {
6141 118 : self.update_layer_visibility().await?;
6142 73 : }
6143 :
6144 191 : let total_layer_size = image_layers
6145 191 : .iter()
6146 191 : .map(|l| l.metadata().file_size)
6147 191 : .sum::<u64>();
6148 :
6149 191 : if !image_layers.is_empty() {
6150 118 : info!(
6151 0 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
6152 0 : image_layers.len(),
6153 : total_layer_size,
6154 0 : duration.as_secs_f64(),
6155 : partition_processed,
6156 : total_partitions
6157 : );
6158 73 : }
6159 :
6160 : Ok((
6161 191 : image_layers,
6162 191 : if all_generated {
6163 191 : LastImageLayerCreationStatus::Complete
6164 : } else {
6165 : LastImageLayerCreationStatus::Incomplete {
6166 0 : last_key: if let Some(last_partition_processed) = last_partition_processed {
6167 0 : last_partition_processed.end().unwrap_or(Key::MIN)
6168 : } else {
6169 : // This branch should be unreachable, but in case it happens, we can just return the start key.
6170 0 : Key::MIN
6171 : },
6172 : }
6173 : },
6174 : ))
6175 191 : }
6176 :
6177 : /// Wait until the background initial logical size calculation is complete, or
6178 : /// this Timeline is shut down. Calling this function will cause the initial
6179 : /// logical size calculation to skip waiting for the background jobs barrier.
6180 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
6181 0 : if !self.shard_identity.is_shard_zero() {
6182 : // We don't populate logical size on shard >0: skip waiting for it.
6183 0 : return;
6184 0 : }
6185 :
6186 0 : if self.remote_client.is_deleting() {
6187 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
6188 0 : return;
6189 0 : }
6190 :
6191 0 : if self.current_logical_size.current_size().is_exact() {
6192 : // root timelines are initialized with exact count, but never start the background
6193 : // calculation
6194 0 : return;
6195 0 : }
6196 :
6197 0 : if self.cancel.is_cancelled() {
6198 : // We already requested stopping the tenant, so we cannot wait for the logical size
6199 : // calculation to complete given the task might have been already cancelled.
6200 0 : return;
6201 0 : }
6202 :
6203 0 : if let Some(await_bg_cancel) = self
6204 0 : .current_logical_size
6205 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
6206 0 : .get()
6207 0 : {
6208 0 : await_bg_cancel.cancel();
6209 0 : } else {
6210 : // We should not wait if we were not able to explicitly instruct
6211 : // the logical size cancellation to skip the concurrency limit semaphore.
6212 : // TODO: this is an unexpected case. We should restructure so that it
6213 : // can't happen.
6214 0 : tracing::warn!(
6215 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
6216 : );
6217 0 : debug_assert!(false);
6218 : }
6219 :
6220 0 : tokio::select!(
6221 0 : _ = self.current_logical_size.initialized.acquire() => {},
6222 0 : _ = self.cancel.cancelled() => {}
6223 : )
6224 0 : }
6225 :
6226 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
6227 : /// Timelines layers up to the ancestor_lsn.
6228 : ///
6229 : /// Requires a timeline that:
6230 : /// - has an ancestor to detach from
6231 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
6232 : /// a technical requirement
6233 : ///
6234 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
6235 : /// polled again until completion.
6236 : ///
6237 : /// During the operation all timelines sharing the data with this timeline will be reparented
6238 : /// from our ancestor to be branches of this timeline.
6239 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
6240 0 : self: &Arc<Timeline>,
6241 0 : tenant: &crate::tenant::TenantShard,
6242 0 : options: detach_ancestor::Options,
6243 0 : behavior: DetachBehavior,
6244 0 : ctx: &RequestContext,
6245 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
6246 0 : detach_ancestor::prepare(self, tenant, behavior, options, ctx).await
6247 0 : }
6248 :
6249 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
6250 : /// reparents any reparentable children of previous ancestor.
6251 : ///
6252 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
6253 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
6254 : /// successfully, tenant must be reloaded.
6255 : ///
6256 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
6257 : /// resetting the tenant.
6258 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
6259 0 : self: &Arc<Timeline>,
6260 0 : tenant: &crate::tenant::TenantShard,
6261 0 : prepared: detach_ancestor::PreparedTimelineDetach,
6262 0 : ancestor_timeline_id: TimelineId,
6263 0 : ancestor_lsn: Lsn,
6264 0 : behavior: DetachBehavior,
6265 0 : ctx: &RequestContext,
6266 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
6267 0 : detach_ancestor::detach_and_reparent(
6268 0 : self,
6269 0 : tenant,
6270 0 : prepared,
6271 0 : ancestor_timeline_id,
6272 0 : ancestor_lsn,
6273 0 : behavior,
6274 0 : ctx,
6275 0 : )
6276 0 : .await
6277 0 : }
6278 :
6279 : /// Final step which unblocks the GC.
6280 : ///
6281 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
6282 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
6283 0 : self: &Arc<Timeline>,
6284 0 : tenant: &crate::tenant::TenantShard,
6285 0 : attempt: detach_ancestor::Attempt,
6286 0 : ctx: &RequestContext,
6287 0 : ) -> Result<(), detach_ancestor::Error> {
6288 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
6289 0 : }
6290 : }
6291 :
6292 : impl Drop for Timeline {
6293 5 : fn drop(&mut self) {
6294 5 : if let Some(ancestor) = &self.ancestor_timeline {
6295 : // This lock should never be poisoned, but in case it is we do a .map() instead of
6296 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
6297 2 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
6298 2 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
6299 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
6300 0 : "Couldn't remove retain_lsn entry from timeline's parent on drop: already removed");
6301 2 : }
6302 0 : }
6303 3 : }
6304 5 : info!(
6305 0 : "Timeline {} for tenant {} is being dropped",
6306 : self.timeline_id, self.tenant_shard_id.tenant_id
6307 : );
6308 5 : }
6309 : }
6310 :
6311 : pub(crate) use compaction_error::CompactionError;
6312 : /// In a private mod to enforce that [`CompactionError::is_cancel`] is used
6313 : /// instead of `match`ing on [`CompactionError::ShuttingDown`].
6314 : mod compaction_error {
6315 : use utils::sync::gate::GateError;
6316 :
6317 : use crate::{
6318 : pgdatadir_mapping::CollectKeySpaceError,
6319 : tenant::{PageReconstructError, blob_io::WriteBlobError, upload_queue::NotInitialized},
6320 : virtual_file::owned_buffers_io::write::FlushTaskError,
6321 : };
6322 :
6323 : /// Top-level failure to compact. Use [`Self::is_cancel`].
6324 : #[derive(Debug, thiserror::Error)]
6325 : pub(crate) enum CompactionError {
6326 : /// Use [`Self::is_cancel`] instead of checking for this variant.
6327 : #[error("The timeline or pageserver is shutting down")]
6328 : #[allow(private_interfaces)]
6329 : ShuttingDown(ForbidMatching), // private ForbidMatching enforces use of [`Self::is_cancel`].
6330 : #[error(transparent)]
6331 : Other(anyhow::Error),
6332 : }
6333 :
6334 : #[derive(Debug)]
6335 : struct ForbidMatching;
6336 :
6337 : impl CompactionError {
6338 0 : pub fn new_cancelled() -> Self {
6339 0 : Self::ShuttingDown(ForbidMatching)
6340 0 : }
6341 : /// Errors that can be ignored, i.e., cancel and shutdown.
6342 0 : pub fn is_cancel(&self) -> bool {
6343 0 : let other = match self {
6344 0 : CompactionError::ShuttingDown(_) => return true,
6345 0 : CompactionError::Other(other) => other,
6346 : };
6347 :
6348 : // The write path of compaction in particular often lacks differentiated
6349 : // handling errors stemming from cancellation from other errors.
6350 : // So, if requested, we also check the ::Other variant by downcasting.
6351 : // The list below has been found empirically from flaky tests and production logs.
6352 : // The process is simple: on ::Other(), compaction will print the enclosed
6353 : // anyhow::Error in debug mode, i.e., with backtrace. That backtrace contains the
6354 : // line where the write path / compaction code does undifferentiated error handling
6355 : // from a non-anyhow type to an anyhow type. Add the type to the list of downcasts
6356 : // below, following the same is_cancel() pattern.
6357 :
6358 0 : let root_cause = other.root_cause();
6359 :
6360 0 : let upload_queue = root_cause
6361 0 : .downcast_ref::<NotInitialized>()
6362 0 : .is_some_and(|e| e.is_stopping());
6363 0 : let timeline = root_cause
6364 0 : .downcast_ref::<PageReconstructError>()
6365 0 : .is_some_and(|e| e.is_cancel());
6366 0 : let buffered_writer_flush_task_canelled = root_cause
6367 0 : .downcast_ref::<FlushTaskError>()
6368 0 : .is_some_and(|e| e.is_cancel());
6369 0 : let write_blob_cancelled = root_cause
6370 0 : .downcast_ref::<WriteBlobError>()
6371 0 : .is_some_and(|e| e.is_cancel());
6372 0 : let gate_closed = root_cause
6373 0 : .downcast_ref::<GateError>()
6374 0 : .is_some_and(|e| e.is_cancel());
6375 0 : upload_queue
6376 0 : || timeline
6377 0 : || buffered_writer_flush_task_canelled
6378 0 : || write_blob_cancelled
6379 0 : || gate_closed
6380 0 : }
6381 0 : pub fn into_anyhow(self) -> anyhow::Error {
6382 0 : match self {
6383 0 : CompactionError::ShuttingDown(ForbidMatching) => anyhow::Error::new(self),
6384 0 : CompactionError::Other(e) => e,
6385 : }
6386 0 : }
6387 0 : pub fn from_collect_keyspace(err: CollectKeySpaceError) -> Self {
6388 0 : if err.is_cancel() {
6389 0 : Self::new_cancelled()
6390 : } else {
6391 0 : Self::Other(err.into_anyhow())
6392 : }
6393 0 : }
6394 : }
6395 : }
6396 :
6397 : impl From<super::upload_queue::NotInitialized> for CompactionError {
6398 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
6399 0 : match value {
6400 : super::upload_queue::NotInitialized::Uninitialized => {
6401 0 : CompactionError::Other(anyhow::anyhow!(value))
6402 : }
6403 : super::upload_queue::NotInitialized::ShuttingDown
6404 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::new_cancelled(),
6405 : }
6406 0 : }
6407 : }
6408 :
6409 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
6410 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
6411 0 : match e {
6412 : super::storage_layer::layer::DownloadError::TimelineShutdown
6413 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
6414 0 : CompactionError::new_cancelled()
6415 : }
6416 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
6417 : | super::storage_layer::layer::DownloadError::DownloadRequired
6418 : | super::storage_layer::layer::DownloadError::NotFile(_)
6419 : | super::storage_layer::layer::DownloadError::DownloadFailed
6420 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
6421 0 : CompactionError::Other(anyhow::anyhow!(e))
6422 : }
6423 : #[cfg(test)]
6424 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
6425 0 : CompactionError::Other(anyhow::anyhow!(e))
6426 : }
6427 : }
6428 0 : }
6429 : }
6430 :
6431 : impl From<layer_manager::Shutdown> for CompactionError {
6432 0 : fn from(_: layer_manager::Shutdown) -> Self {
6433 0 : CompactionError::new_cancelled()
6434 0 : }
6435 : }
6436 :
6437 : impl From<super::storage_layer::errors::PutError> for CompactionError {
6438 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
6439 0 : if e.is_cancel() {
6440 0 : CompactionError::new_cancelled()
6441 : } else {
6442 0 : CompactionError::Other(e.into_anyhow())
6443 : }
6444 0 : }
6445 : }
6446 :
6447 : #[serde_as]
6448 : #[derive(serde::Serialize)]
6449 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
6450 :
6451 : #[derive(Default)]
6452 : enum DurationRecorder {
6453 : #[default]
6454 : NotStarted,
6455 : Recorded(RecordedDuration, tokio::time::Instant),
6456 : }
6457 :
6458 : impl DurationRecorder {
6459 115 : fn till_now(&self) -> DurationRecorder {
6460 115 : match self {
6461 : DurationRecorder::NotStarted => {
6462 0 : panic!("must only call on recorded measurements")
6463 : }
6464 115 : DurationRecorder::Recorded(_, ended) => {
6465 115 : let now = tokio::time::Instant::now();
6466 115 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
6467 : }
6468 : }
6469 115 : }
6470 138 : fn into_recorded(self) -> Option<RecordedDuration> {
6471 138 : match self {
6472 0 : DurationRecorder::NotStarted => None,
6473 138 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
6474 : }
6475 138 : }
6476 : }
6477 :
6478 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
6479 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
6480 : /// the layer descriptor requires the user to provide the ranges, which should cover all
6481 : /// keys specified in the `data` field.
6482 : #[cfg(test)]
6483 : #[derive(Clone)]
6484 : pub struct DeltaLayerTestDesc {
6485 : pub lsn_range: Range<Lsn>,
6486 : pub key_range: Range<Key>,
6487 : pub data: Vec<(Key, Lsn, Value)>,
6488 : }
6489 :
6490 : #[cfg(test)]
6491 : #[derive(Clone)]
6492 : pub struct InMemoryLayerTestDesc {
6493 : pub lsn_range: Range<Lsn>,
6494 : pub data: Vec<(Key, Lsn, Value)>,
6495 : pub is_open: bool,
6496 : }
6497 :
6498 : #[cfg(test)]
6499 : impl DeltaLayerTestDesc {
6500 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
6501 2 : Self {
6502 2 : lsn_range,
6503 2 : key_range,
6504 2 : data,
6505 2 : }
6506 2 : }
6507 :
6508 45 : pub fn new_with_inferred_key_range(
6509 45 : lsn_range: Range<Lsn>,
6510 45 : data: Vec<(Key, Lsn, Value)>,
6511 45 : ) -> Self {
6512 45 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
6513 45 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
6514 45 : Self {
6515 45 : key_range: (*key_min)..(key_max.next()),
6516 45 : lsn_range,
6517 45 : data,
6518 45 : }
6519 45 : }
6520 :
6521 5 : pub(crate) fn layer_name(&self) -> LayerName {
6522 5 : LayerName::Delta(super::storage_layer::DeltaLayerName {
6523 5 : key_range: self.key_range.clone(),
6524 5 : lsn_range: self.lsn_range.clone(),
6525 5 : })
6526 5 : }
6527 : }
6528 :
6529 : impl Timeline {
6530 23 : async fn finish_compact_batch(
6531 23 : self: &Arc<Self>,
6532 23 : new_deltas: &[ResidentLayer],
6533 23 : new_images: &[ResidentLayer],
6534 23 : layers_to_remove: &[Layer],
6535 23 : ) -> Result<(), CompactionError> {
6536 23 : let mut guard = tokio::select! {
6537 23 : guard = self.layers.write(LayerManagerLockHolder::Compaction) => guard,
6538 23 : _ = self.cancel.cancelled() => {
6539 0 : return Err(CompactionError::new_cancelled());
6540 : }
6541 : };
6542 :
6543 23 : let mut duplicated_layers = HashSet::new();
6544 :
6545 23 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
6546 :
6547 186 : for l in new_deltas {
6548 163 : if guard.contains(l.as_ref()) {
6549 : // expected in tests
6550 0 : tracing::error!(layer=%l, "duplicated L1 layer");
6551 :
6552 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
6553 : // `guard` on self.layers. as of writing this, there are no error returns except
6554 : // for compact_level0_phase1 creating an L0, which does not happen in practice
6555 : // because we have not implemented L0 => L0 compaction.
6556 0 : duplicated_layers.insert(l.layer_desc().key());
6557 163 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
6558 0 : return Err(CompactionError::Other(anyhow::anyhow!(
6559 0 : "compaction generates a L0 layer file as output, which will cause infinite compaction."
6560 0 : )));
6561 163 : } else {
6562 163 : insert_layers.push(l.clone());
6563 163 : }
6564 : }
6565 :
6566 : // only remove those inputs which were not outputs
6567 23 : let remove_layers: Vec<Layer> = layers_to_remove
6568 23 : .iter()
6569 201 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
6570 23 : .cloned()
6571 23 : .collect();
6572 :
6573 23 : if !new_images.is_empty() {
6574 0 : guard
6575 0 : .open_mut()?
6576 0 : .track_new_image_layers(new_images, &self.metrics);
6577 23 : }
6578 :
6579 23 : guard
6580 23 : .open_mut()?
6581 23 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
6582 :
6583 23 : self.remote_client
6584 23 : .schedule_compaction_update(&remove_layers, new_deltas)?;
6585 :
6586 23 : drop_layer_manager_wlock(guard);
6587 :
6588 23 : Ok(())
6589 23 : }
6590 :
6591 0 : async fn rewrite_layers(
6592 0 : self: &Arc<Self>,
6593 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
6594 0 : mut drop_layers: Vec<Layer>,
6595 0 : ) -> Result<(), CompactionError> {
6596 0 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
6597 :
6598 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
6599 : // to avoid double-removing, and avoid rewriting something that was removed.
6600 0 : replace_layers.retain(|(l, _)| guard.contains(l));
6601 0 : drop_layers.retain(|l| guard.contains(l));
6602 :
6603 0 : guard
6604 0 : .open_mut()?
6605 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
6606 :
6607 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
6608 :
6609 0 : self.remote_client
6610 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
6611 :
6612 0 : Ok(())
6613 0 : }
6614 :
6615 : /// Schedules the uploads of the given image layers
6616 80 : fn upload_new_image_layers(
6617 80 : self: &Arc<Self>,
6618 80 : new_images: impl IntoIterator<Item = ResidentLayer>,
6619 80 : ) -> Result<(), super::upload_queue::NotInitialized> {
6620 93 : for layer in new_images {
6621 13 : self.remote_client.schedule_layer_file_upload(layer)?;
6622 : }
6623 : // should any new image layer been created, not uploading index_part will
6624 : // result in a mismatch between remote_physical_size and layermap calculated
6625 : // size, which will fail some tests, but should not be an issue otherwise.
6626 80 : self.remote_client
6627 80 : .schedule_index_upload_for_file_changes()?;
6628 80 : Ok(())
6629 80 : }
6630 :
6631 0 : async fn find_gc_time_cutoff(
6632 0 : &self,
6633 0 : now: SystemTime,
6634 0 : pitr: Duration,
6635 0 : cancel: &CancellationToken,
6636 0 : ctx: &RequestContext,
6637 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
6638 0 : debug_assert_current_span_has_tenant_and_timeline_id();
6639 0 : if self.shard_identity.is_shard_zero() {
6640 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
6641 0 : let time_range = if pitr == Duration::ZERO {
6642 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
6643 : } else {
6644 0 : pitr
6645 : };
6646 :
6647 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
6648 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
6649 0 : let timestamp = to_pg_timestamp(time_cutoff);
6650 :
6651 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
6652 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
6653 0 : LsnForTimestamp::Future(lsn) => {
6654 : // The timestamp is in the future. That sounds impossible,
6655 : // but what it really means is that there hasn't been
6656 : // any commits since the cutoff timestamp.
6657 : //
6658 : // In this case we should use the LSN of the most recent commit,
6659 : // which is implicitly the last LSN in the log.
6660 0 : debug!("future({})", lsn);
6661 0 : Some(self.get_last_record_lsn())
6662 : }
6663 0 : LsnForTimestamp::Past(lsn) => {
6664 0 : debug!("past({})", lsn);
6665 0 : None
6666 : }
6667 0 : LsnForTimestamp::NoData(lsn) => {
6668 0 : debug!("nodata({})", lsn);
6669 0 : None
6670 : }
6671 : };
6672 0 : Ok(time_cutoff)
6673 : } else {
6674 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
6675 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
6676 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
6677 : // space cutoff that we would also have respected ourselves.
6678 0 : match self
6679 0 : .remote_client
6680 0 : .download_foreign_index(ShardNumber(0), cancel)
6681 0 : .await
6682 : {
6683 0 : Ok((index_part, index_generation, _index_mtime)) => {
6684 0 : tracing::info!(
6685 0 : "GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
6686 0 : index_part.metadata.latest_gc_cutoff_lsn()
6687 : );
6688 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
6689 : }
6690 : Err(DownloadError::NotFound) => {
6691 : // This is unexpected, because during timeline creations shard zero persists to remote
6692 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
6693 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
6694 : // state, then shard zero should _eventually_ write an index when it GCs.
6695 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
6696 0 : Ok(None)
6697 : }
6698 0 : Err(e) => {
6699 : // TODO: this function should return a different error type than page reconstruct error
6700 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
6701 : }
6702 : }
6703 :
6704 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
6705 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
6706 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
6707 : // new location. This is legal in principle, but problematic in practice because it might result
6708 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
6709 : // because they have GC'd past the branch point.
6710 : }
6711 0 : }
6712 :
6713 : /// Find the Lsns above which layer files need to be retained on
6714 : /// garbage collection.
6715 : ///
6716 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
6717 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
6718 : /// the space-based retention.
6719 : ///
6720 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
6721 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
6722 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
6723 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
6724 : /// in the response as the GC cutoff point for the timeline.
6725 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
6726 : pub(super) async fn find_gc_cutoffs(
6727 : &self,
6728 : now: SystemTime,
6729 : space_cutoff: Lsn,
6730 : pitr: Duration,
6731 : cancel: &CancellationToken,
6732 : ctx: &RequestContext,
6733 : ) -> Result<GcCutoffs, PageReconstructError> {
6734 : let _timer = self
6735 : .metrics
6736 : .find_gc_cutoffs_histo
6737 : .start_timer()
6738 : .record_on_drop();
6739 :
6740 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
6741 :
6742 : if cfg!(test) && pitr == Duration::ZERO {
6743 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
6744 : return Ok(GcCutoffs {
6745 : time: Some(self.get_last_record_lsn()),
6746 : space: space_cutoff,
6747 : });
6748 : }
6749 :
6750 : // Calculate a time-based limit on how much to retain:
6751 : // - if PITR interval is set, then this is our cutoff.
6752 : // - if PITR interval is not set, then we do a lookup
6753 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
6754 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
6755 :
6756 : Ok(match (pitr, time_cutoff) {
6757 : (Duration::ZERO, Some(time_cutoff)) => {
6758 : // PITR is not set. Retain the size-based limit, or the default time retention,
6759 : // whichever requires less data.
6760 : GcCutoffs {
6761 : time: Some(self.get_last_record_lsn()),
6762 : space: std::cmp::max(time_cutoff, space_cutoff),
6763 : }
6764 : }
6765 : (Duration::ZERO, None) => {
6766 : // PITR is not set, and time lookup failed
6767 : GcCutoffs {
6768 : time: Some(self.get_last_record_lsn()),
6769 : space: space_cutoff,
6770 : }
6771 : }
6772 : (_, None) => {
6773 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
6774 : // cannot advance beyond what was already GC'd, and respect space-based retention
6775 : GcCutoffs {
6776 : time: Some(*self.get_applied_gc_cutoff_lsn()),
6777 : space: space_cutoff,
6778 : }
6779 : }
6780 : (_, Some(time_cutoff)) => {
6781 : // PITR interval is set and we looked up timestamp successfully. Ignore
6782 : // size based retention and make time cutoff authoritative
6783 : GcCutoffs {
6784 : time: Some(time_cutoff),
6785 : space: time_cutoff,
6786 : }
6787 : }
6788 : })
6789 : }
6790 :
6791 : /// Garbage collect layer files on a timeline that are no longer needed.
6792 : ///
6793 : /// Currently, we don't make any attempt at removing unneeded page versions
6794 : /// within a layer file. We can only remove the whole file if it's fully
6795 : /// obsolete.
6796 377 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
6797 : // this is most likely the background tasks, but it might be the spawned task from
6798 : // immediate_gc
6799 377 : let _g = tokio::select! {
6800 377 : guard = self.gc_lock.lock() => guard,
6801 377 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
6802 : };
6803 377 : let timer = self.metrics.garbage_collect_histo.start_timer();
6804 :
6805 377 : fail_point!("before-timeline-gc");
6806 :
6807 : // Is the timeline being deleted?
6808 377 : if self.is_stopping() {
6809 0 : return Err(GcError::TimelineCancelled);
6810 377 : }
6811 :
6812 377 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
6813 377 : let gc_info = self.gc_info.read().unwrap();
6814 :
6815 377 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
6816 377 : let time_cutoff = gc_info.cutoffs.time;
6817 377 : let retain_lsns = gc_info
6818 377 : .retain_lsns
6819 377 : .iter()
6820 377 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
6821 377 : .collect();
6822 :
6823 : // Gets the maximum LSN that holds the valid lease.
6824 : //
6825 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
6826 : // Here, we do not check for stale leases again.
6827 377 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
6828 :
6829 377 : (
6830 377 : space_cutoff,
6831 377 : time_cutoff,
6832 377 : retain_lsns,
6833 377 : max_lsn_with_valid_lease,
6834 377 : )
6835 : };
6836 :
6837 377 : let mut new_gc_cutoff = space_cutoff.min(time_cutoff.unwrap_or_default());
6838 377 : let standby_horizon = self.standby_horizon.load();
6839 : // Hold GC for the standby, but as a safety guard do it only within some
6840 : // reasonable lag.
6841 377 : if standby_horizon != Lsn::INVALID {
6842 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
6843 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
6844 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
6845 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
6846 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
6847 : } else {
6848 0 : warn!(
6849 0 : "standby is lagging for more than {}MB, not holding gc for it",
6850 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
6851 : )
6852 : }
6853 0 : }
6854 377 : }
6855 :
6856 : // Reset standby horizon to ignore it if it is not updated till next GC.
6857 : // It is an easy way to unset it when standby disappears without adding
6858 : // more conf options.
6859 377 : self.standby_horizon.store(Lsn::INVALID);
6860 377 : self.metrics
6861 377 : .standby_horizon_gauge
6862 377 : .set(Lsn::INVALID.0 as i64);
6863 :
6864 377 : let res = self
6865 377 : .gc_timeline(
6866 377 : space_cutoff,
6867 377 : time_cutoff,
6868 377 : retain_lsns,
6869 377 : max_lsn_with_valid_lease,
6870 377 : new_gc_cutoff,
6871 : )
6872 377 : .instrument(
6873 377 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
6874 : )
6875 377 : .await?;
6876 :
6877 : // only record successes
6878 377 : timer.stop_and_record();
6879 :
6880 377 : Ok(res)
6881 377 : }
6882 :
6883 377 : async fn gc_timeline(
6884 377 : &self,
6885 377 : space_cutoff: Lsn,
6886 377 : time_cutoff: Option<Lsn>, // None if uninitialized
6887 377 : retain_lsns: Vec<Lsn>,
6888 377 : max_lsn_with_valid_lease: Option<Lsn>,
6889 377 : new_gc_cutoff: Lsn,
6890 377 : ) -> Result<GcResult, GcError> {
6891 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
6892 :
6893 377 : let now = SystemTime::now();
6894 377 : let mut result: GcResult = GcResult::default();
6895 :
6896 : // Nothing to GC. Return early.
6897 377 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
6898 377 : if latest_gc_cutoff >= new_gc_cutoff {
6899 11 : info!(
6900 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
6901 : );
6902 11 : return Ok(result);
6903 366 : }
6904 :
6905 366 : let Some(time_cutoff) = time_cutoff else {
6906 : // The GC cutoff should have been computed by now, but let's be defensive.
6907 0 : info!("Nothing to GC: time_cutoff not yet computed");
6908 0 : return Ok(result);
6909 : };
6910 :
6911 : // We need to ensure that no one tries to read page versions or create
6912 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
6913 : // for details. This will block until the old value is no longer in use.
6914 : //
6915 : // The GC cutoff should only ever move forwards.
6916 366 : let waitlist = {
6917 366 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
6918 366 : if *write_guard > new_gc_cutoff {
6919 0 : return Err(GcError::BadLsn {
6920 0 : why: format!(
6921 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
6922 0 : *write_guard, new_gc_cutoff
6923 0 : ),
6924 0 : });
6925 366 : }
6926 :
6927 366 : write_guard.store_and_unlock(new_gc_cutoff)
6928 : };
6929 366 : let waitlist_wait_fut = std::pin::pin!(waitlist.wait());
6930 366 : log_slow(
6931 366 : "applied_gc_cutoff waitlist wait",
6932 366 : Duration::from_secs(30),
6933 366 : waitlist_wait_fut,
6934 366 : )
6935 366 : .await;
6936 :
6937 366 : info!("GC starting");
6938 :
6939 366 : debug!("retain_lsns: {:?}", retain_lsns);
6940 :
6941 366 : let max_retain_lsn = retain_lsns.iter().max();
6942 :
6943 : // Scan all layers in the timeline (remote or on-disk).
6944 : //
6945 : // Garbage collect the layer if all conditions are satisfied:
6946 : // 1. it is older than cutoff LSN;
6947 : // 2. it is older than PITR interval;
6948 : // 3. it doesn't need to be retained for 'retain_lsns';
6949 : // 4. it does not need to be kept for LSNs holding valid leases.
6950 : // 5. newer on-disk image layers cover the layer's whole key range
6951 366 : let layers_to_remove = {
6952 366 : let mut layers_to_remove = Vec::new();
6953 :
6954 366 : let guard = self
6955 366 : .layers
6956 366 : .read(LayerManagerLockHolder::GarbageCollection)
6957 366 : .await;
6958 366 : let layers = guard.layer_map()?;
6959 6209 : 'outer: for l in layers.iter_historic_layers() {
6960 6209 : result.layers_total += 1;
6961 :
6962 : // 1. Is it newer than GC horizon cutoff point?
6963 6209 : if l.get_lsn_range().end > space_cutoff {
6964 371 : debug!(
6965 0 : "keeping {} because it's newer than space_cutoff {}",
6966 0 : l.layer_name(),
6967 : space_cutoff,
6968 : );
6969 371 : result.layers_needed_by_cutoff += 1;
6970 371 : continue 'outer;
6971 5838 : }
6972 :
6973 : // 2. It is newer than PiTR cutoff point?
6974 5838 : if l.get_lsn_range().end > time_cutoff {
6975 0 : debug!(
6976 0 : "keeping {} because it's newer than time_cutoff {}",
6977 0 : l.layer_name(),
6978 : time_cutoff,
6979 : );
6980 0 : result.layers_needed_by_pitr += 1;
6981 0 : continue 'outer;
6982 5838 : }
6983 :
6984 : // 3. Is it needed by a child branch?
6985 : // NOTE With that we would keep data that
6986 : // might be referenced by child branches forever.
6987 : // We can track this in child timeline GC and delete parent layers when
6988 : // they are no longer needed. This might be complicated with long inheritance chains.
6989 5838 : if let Some(retain_lsn) = max_retain_lsn {
6990 : // start_lsn is inclusive
6991 6 : if &l.get_lsn_range().start <= retain_lsn {
6992 6 : debug!(
6993 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
6994 0 : l.layer_name(),
6995 : retain_lsn,
6996 0 : l.is_incremental(),
6997 : );
6998 6 : result.layers_needed_by_branches += 1;
6999 6 : continue 'outer;
7000 0 : }
7001 5832 : }
7002 :
7003 : // 4. Is there a valid lease that requires us to keep this layer?
7004 5832 : if let Some(lsn) = &max_lsn_with_valid_lease {
7005 : // keep if layer start <= any of the lease
7006 9 : if &l.get_lsn_range().start <= lsn {
7007 7 : debug!(
7008 0 : "keeping {} because there is a valid lease preventing GC at {}",
7009 0 : l.layer_name(),
7010 : lsn,
7011 : );
7012 7 : result.layers_needed_by_leases += 1;
7013 7 : continue 'outer;
7014 2 : }
7015 5823 : }
7016 :
7017 : // 5. Is there a later on-disk layer for this relation?
7018 : //
7019 : // The end-LSN is exclusive, while disk_consistent_lsn is
7020 : // inclusive. For example, if disk_consistent_lsn is 100, it is
7021 : // OK for a delta layer to have end LSN 101, but if the end LSN
7022 : // is 102, then it might not have been fully flushed to disk
7023 : // before crash.
7024 : //
7025 : // For example, imagine that the following layers exist:
7026 : //
7027 : // 1000 - image (A)
7028 : // 1000-2000 - delta (B)
7029 : // 2000 - image (C)
7030 : // 2000-3000 - delta (D)
7031 : // 3000 - image (E)
7032 : //
7033 : // If GC horizon is at 2500, we can remove layers A and B, but
7034 : // we cannot remove C, even though it's older than 2500, because
7035 : // the delta layer 2000-3000 depends on it.
7036 5825 : if !layers
7037 5825 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
7038 : {
7039 5821 : debug!("keeping {} because it is the latest layer", l.layer_name());
7040 5821 : result.layers_not_updated += 1;
7041 5821 : continue 'outer;
7042 4 : }
7043 :
7044 : // We didn't find any reason to keep this file, so remove it.
7045 4 : info!(
7046 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
7047 0 : l.layer_name(),
7048 0 : l.is_incremental(),
7049 : );
7050 4 : layers_to_remove.push(l);
7051 : }
7052 :
7053 366 : layers_to_remove
7054 : };
7055 :
7056 366 : if !layers_to_remove.is_empty() {
7057 : // Persist the new GC cutoff value before we actually remove anything.
7058 : // This unconditionally schedules also an index_part.json update, even though, we will
7059 : // be doing one a bit later with the unlinked gc'd layers.
7060 3 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
7061 3 : self.schedule_uploads(disk_consistent_lsn, None)
7062 3 : .map_err(|e| {
7063 0 : if self.cancel.is_cancelled() {
7064 0 : GcError::TimelineCancelled
7065 : } else {
7066 0 : GcError::Remote(e)
7067 : }
7068 0 : })?;
7069 :
7070 3 : let mut guard = self
7071 3 : .layers
7072 3 : .write(LayerManagerLockHolder::GarbageCollection)
7073 3 : .await;
7074 :
7075 3 : let gc_layers = layers_to_remove
7076 3 : .iter()
7077 4 : .flat_map(|desc| guard.try_get_from_key(&desc.key()).cloned())
7078 3 : .collect::<Vec<Layer>>();
7079 :
7080 3 : result.layers_removed = gc_layers.len() as u64;
7081 :
7082 3 : self.remote_client.schedule_gc_update(&gc_layers)?;
7083 3 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
7084 :
7085 : #[cfg(feature = "testing")]
7086 3 : {
7087 3 : result.doomed_layers = gc_layers;
7088 3 : }
7089 363 : }
7090 :
7091 366 : info!(
7092 0 : "GC completed removing {} layers, cutoff {}",
7093 : result.layers_removed, new_gc_cutoff
7094 : );
7095 :
7096 366 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
7097 366 : Ok(result)
7098 377 : }
7099 :
7100 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
7101 364884 : pub(crate) async fn reconstruct_value(
7102 364884 : &self,
7103 364884 : key: Key,
7104 364884 : request_lsn: Lsn,
7105 364884 : mut data: ValueReconstructState,
7106 364884 : redo_attempt_type: RedoAttemptType,
7107 364884 : ) -> Result<Bytes, PageReconstructError> {
7108 : // Perform WAL redo if needed
7109 364884 : data.records.reverse();
7110 :
7111 364884 : let fire_critical_error = match redo_attempt_type {
7112 363551 : RedoAttemptType::ReadPage => true,
7113 0 : RedoAttemptType::LegacyCompaction => true,
7114 1333 : RedoAttemptType::GcCompaction => false,
7115 : };
7116 :
7117 : // If we have a page image, and no WAL, we're all set
7118 364884 : if data.records.is_empty() {
7119 338235 : if let Some((img_lsn, img)) = &data.img {
7120 338235 : trace!(
7121 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
7122 : key, img_lsn, request_lsn,
7123 : );
7124 338235 : Ok(img.clone())
7125 : } else {
7126 0 : Err(PageReconstructError::from(anyhow!(
7127 0 : "base image for {key} at {request_lsn} not found"
7128 0 : )))
7129 : }
7130 : } else {
7131 : // We need to do WAL redo.
7132 : //
7133 : // If we don't have a base image, then the oldest WAL record better initialize
7134 : // the page
7135 26649 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
7136 0 : Err(PageReconstructError::from(anyhow!(
7137 0 : "Base image for {} at {} not found, but got {} WAL records",
7138 0 : key,
7139 0 : request_lsn,
7140 0 : data.records.len()
7141 0 : )))
7142 : } else {
7143 26649 : if data.img.is_some() {
7144 12781 : trace!(
7145 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
7146 0 : data.records.len(),
7147 : key,
7148 : request_lsn
7149 : );
7150 : } else {
7151 13868 : trace!(
7152 0 : "found {} WAL records that will init the page for {} at {}, performing WAL redo",
7153 0 : data.records.len(),
7154 : key,
7155 : request_lsn
7156 : );
7157 : };
7158 26649 : let res = self
7159 26649 : .walredo_mgr
7160 26649 : .as_ref()
7161 26649 : .context("timeline has no walredo manager")
7162 26649 : .map_err(PageReconstructError::WalRedo)?
7163 26649 : .request_redo(
7164 26649 : key,
7165 26649 : request_lsn,
7166 26649 : data.img,
7167 26649 : data.records,
7168 26649 : self.pg_version,
7169 26649 : redo_attempt_type,
7170 : )
7171 26649 : .await;
7172 26648 : let img = match res {
7173 26648 : Ok(img) => img,
7174 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
7175 1 : Err(walredo::Error::Other(err)) => {
7176 1 : if fire_critical_error {
7177 0 : critical_timeline!(
7178 0 : self.tenant_shard_id,
7179 0 : self.timeline_id,
7180 0 : Some(&self.corruption_detected),
7181 0 : "walredo failure during page reconstruction: {err:?}"
7182 : );
7183 1 : }
7184 1 : return Err(PageReconstructError::WalRedo(
7185 1 : err.context("reconstruct a page image"),
7186 1 : ));
7187 : }
7188 : };
7189 26648 : Ok(img)
7190 : }
7191 : }
7192 364884 : }
7193 :
7194 0 : pub(crate) async fn spawn_download_all_remote_layers(
7195 0 : self: Arc<Self>,
7196 0 : request: DownloadRemoteLayersTaskSpawnRequest,
7197 0 : ctx: &RequestContext,
7198 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
7199 : use pageserver_api::models::DownloadRemoteLayersTaskState;
7200 :
7201 : // this is not really needed anymore; it has tests which really check the return value from
7202 : // http api. it would be better not to maintain this anymore.
7203 :
7204 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
7205 0 : if let Some(st) = &*status_guard {
7206 0 : match &st.state {
7207 : DownloadRemoteLayersTaskState::Running => {
7208 0 : return Err(st.clone());
7209 : }
7210 : DownloadRemoteLayersTaskState::ShutDown
7211 0 : | DownloadRemoteLayersTaskState::Completed => {
7212 0 : *status_guard = None;
7213 0 : }
7214 : }
7215 0 : }
7216 :
7217 0 : let self_clone = Arc::clone(&self);
7218 0 : let task_ctx = ctx.detached_child(
7219 0 : TaskKind::DownloadAllRemoteLayers,
7220 0 : DownloadBehavior::Download,
7221 : );
7222 0 : let task_id = task_mgr::spawn(
7223 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
7224 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
7225 0 : self.tenant_shard_id,
7226 0 : Some(self.timeline_id),
7227 0 : "download all remote layers task",
7228 0 : async move {
7229 0 : self_clone.download_all_remote_layers(request, &task_ctx).await;
7230 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
7231 0 : match &mut *status_guard {
7232 : None => {
7233 0 : warn!("tasks status is supposed to be Some(), since we are running");
7234 : }
7235 0 : Some(st) => {
7236 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
7237 0 : if st.task_id != exp_task_id {
7238 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
7239 0 : } else {
7240 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7241 0 : }
7242 : }
7243 : };
7244 0 : Ok(())
7245 0 : }
7246 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))
7247 : );
7248 :
7249 0 : let initial_info = DownloadRemoteLayersTaskInfo {
7250 0 : task_id: format!("{task_id}"),
7251 0 : state: DownloadRemoteLayersTaskState::Running,
7252 0 : total_layer_count: 0,
7253 0 : successful_download_count: 0,
7254 0 : failed_download_count: 0,
7255 0 : };
7256 0 : *status_guard = Some(initial_info.clone());
7257 :
7258 0 : Ok(initial_info)
7259 0 : }
7260 :
7261 0 : async fn download_all_remote_layers(
7262 0 : self: &Arc<Self>,
7263 0 : request: DownloadRemoteLayersTaskSpawnRequest,
7264 0 : ctx: &RequestContext,
7265 0 : ) {
7266 : use pageserver_api::models::DownloadRemoteLayersTaskState;
7267 :
7268 0 : let remaining = {
7269 0 : let guard = self
7270 0 : .layers
7271 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7272 0 : .await;
7273 0 : let Ok(lm) = guard.layer_map() else {
7274 : // technically here we could look into iterating accessible layers, but downloading
7275 : // all layers of a shutdown timeline makes no sense regardless.
7276 0 : tracing::info!("attempted to download all layers of shutdown timeline");
7277 0 : return;
7278 : };
7279 0 : lm.iter_historic_layers()
7280 0 : .map(|desc| guard.get_from_desc(&desc))
7281 0 : .collect::<Vec<_>>()
7282 : };
7283 0 : let total_layer_count = remaining.len();
7284 :
7285 : macro_rules! lock_status {
7286 : ($st:ident) => {
7287 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
7288 : let st = st
7289 : .as_mut()
7290 : .expect("this function is only called after the task has been spawned");
7291 : assert_eq!(
7292 : st.task_id,
7293 : format!(
7294 : "{}",
7295 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
7296 : )
7297 : );
7298 : let $st = st;
7299 : };
7300 : }
7301 :
7302 : {
7303 0 : lock_status!(st);
7304 0 : st.total_layer_count = total_layer_count as u64;
7305 : }
7306 :
7307 0 : let mut remaining = remaining.into_iter();
7308 0 : let mut have_remaining = true;
7309 0 : let mut js = tokio::task::JoinSet::new();
7310 :
7311 0 : let cancel = task_mgr::shutdown_token();
7312 :
7313 0 : let limit = request.max_concurrent_downloads;
7314 :
7315 : loop {
7316 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
7317 0 : let Some(next) = remaining.next() else {
7318 0 : have_remaining = false;
7319 0 : break;
7320 : };
7321 :
7322 0 : let span = tracing::info_span!("download", layer = %next);
7323 :
7324 0 : let ctx = ctx.attached_child();
7325 0 : js.spawn(
7326 0 : async move {
7327 0 : let res = next.download(&ctx).await;
7328 0 : (next, res)
7329 0 : }
7330 0 : .instrument(span),
7331 : );
7332 : }
7333 :
7334 0 : while let Some(res) = js.join_next().await {
7335 0 : match res {
7336 : Ok((_, Ok(_))) => {
7337 0 : lock_status!(st);
7338 0 : st.successful_download_count += 1;
7339 : }
7340 0 : Ok((layer, Err(e))) => {
7341 0 : tracing::error!(%layer, "download failed: {e:#}");
7342 0 : lock_status!(st);
7343 0 : st.failed_download_count += 1;
7344 : }
7345 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
7346 0 : Err(je) if je.is_panic() => {
7347 0 : lock_status!(st);
7348 0 : st.failed_download_count += 1;
7349 : }
7350 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
7351 : }
7352 : }
7353 :
7354 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
7355 0 : break;
7356 0 : }
7357 : }
7358 :
7359 : {
7360 0 : lock_status!(st);
7361 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7362 : }
7363 0 : }
7364 :
7365 0 : pub(crate) fn get_download_all_remote_layers_task_info(
7366 0 : &self,
7367 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
7368 0 : self.download_all_remote_layers_task_info
7369 0 : .read()
7370 0 : .unwrap()
7371 0 : .clone()
7372 0 : }
7373 :
7374 : /* BEGIN_HADRON */
7375 0 : pub(crate) async fn compute_image_consistent_lsn(&self) -> anyhow::Result<Lsn> {
7376 0 : let guard = self
7377 0 : .layers
7378 0 : .read(LayerManagerLockHolder::ComputeImageConsistentLsn)
7379 0 : .await;
7380 0 : let layer_map = guard.layer_map()?;
7381 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
7382 :
7383 0 : Ok(layer_map.compute_image_consistent_lsn(disk_consistent_lsn))
7384 0 : }
7385 : /* END_HADRON */
7386 : }
7387 :
7388 : impl Timeline {
7389 : /// Returns non-remote layers for eviction.
7390 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
7391 0 : let guard = self.layers.read(LayerManagerLockHolder::Eviction).await;
7392 0 : let mut max_layer_size: Option<u64> = None;
7393 :
7394 0 : let resident_layers = guard
7395 0 : .likely_resident_layers()
7396 0 : .map(|layer| {
7397 0 : let file_size = layer.layer_desc().file_size;
7398 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
7399 :
7400 0 : let last_activity_ts = layer.latest_activity();
7401 :
7402 0 : EvictionCandidate {
7403 0 : layer: layer.to_owned().into(),
7404 0 : last_activity_ts,
7405 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
7406 0 : visibility: layer.visibility(),
7407 0 : }
7408 0 : })
7409 0 : .collect();
7410 :
7411 0 : DiskUsageEvictionInfo {
7412 0 : max_layer_size,
7413 0 : resident_layers,
7414 0 : }
7415 0 : }
7416 :
7417 962 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
7418 962 : ShardIndex {
7419 962 : shard_number: self.tenant_shard_id.shard_number,
7420 962 : shard_count: self.tenant_shard_id.shard_count,
7421 962 : }
7422 962 : }
7423 :
7424 : /// Persistently blocks gc for `Manual` reason.
7425 : ///
7426 : /// Returns true if no such block existed before, false otherwise.
7427 0 : pub(crate) async fn block_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<bool> {
7428 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7429 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7430 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
7431 0 : }
7432 :
7433 : /// Persistently unblocks gc for `Manual` reason.
7434 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<()> {
7435 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7436 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7437 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
7438 0 : }
7439 :
7440 : #[cfg(test)]
7441 31 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
7442 31 : self.last_record_lsn.advance(new_lsn);
7443 31 : }
7444 :
7445 : #[cfg(test)]
7446 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
7447 2 : self.disk_consistent_lsn.store(new_value);
7448 2 : }
7449 :
7450 : /// Force create an image layer and place it into the layer map.
7451 : ///
7452 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7453 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7454 : /// placed into the layer map in one run AND be validated.
7455 : #[cfg(test)]
7456 36 : pub(super) async fn force_create_image_layer(
7457 36 : self: &Arc<Timeline>,
7458 36 : lsn: Lsn,
7459 36 : mut images: Vec<(Key, Bytes)>,
7460 36 : check_start_lsn: Option<Lsn>,
7461 36 : ctx: &RequestContext,
7462 36 : ) -> anyhow::Result<()> {
7463 36 : let last_record_lsn = self.get_last_record_lsn();
7464 36 : assert!(
7465 36 : lsn <= last_record_lsn,
7466 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
7467 : );
7468 36 : if let Some(check_start_lsn) = check_start_lsn {
7469 36 : assert!(lsn >= check_start_lsn);
7470 0 : }
7471 240 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
7472 36 : let min_key = *images.first().map(|(k, _)| k).unwrap();
7473 36 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
7474 36 : let mut image_layer_writer = ImageLayerWriter::new(
7475 36 : self.conf,
7476 36 : self.timeline_id,
7477 36 : self.tenant_shard_id,
7478 36 : &(min_key..end_key),
7479 36 : lsn,
7480 36 : &self.gate,
7481 36 : self.cancel.clone(),
7482 36 : ctx,
7483 36 : )
7484 36 : .await?;
7485 312 : for (key, img) in images {
7486 276 : image_layer_writer.put_image(key, img, ctx).await?;
7487 : }
7488 36 : let (desc, path) = image_layer_writer.finish(ctx).await?;
7489 36 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7490 36 : info!("force created image layer {}", image_layer.local_path());
7491 : {
7492 36 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7493 36 : guard
7494 36 : .open_mut()
7495 36 : .unwrap()
7496 36 : .force_insert_layer(image_layer.clone());
7497 : }
7498 :
7499 : // Update remote_timeline_client state to reflect existence of this layer
7500 36 : self.remote_client
7501 36 : .schedule_layer_file_upload(image_layer)
7502 36 : .unwrap();
7503 :
7504 36 : Ok(())
7505 36 : }
7506 :
7507 : /// Force create a delta layer and place it into the layer map.
7508 : ///
7509 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7510 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7511 : /// placed into the layer map in one run AND be validated.
7512 : #[cfg(test)]
7513 50 : pub(super) async fn force_create_delta_layer(
7514 50 : self: &Arc<Timeline>,
7515 50 : mut deltas: DeltaLayerTestDesc,
7516 50 : check_start_lsn: Option<Lsn>,
7517 50 : ctx: &RequestContext,
7518 50 : ) -> anyhow::Result<()> {
7519 50 : let last_record_lsn = self.get_last_record_lsn();
7520 50 : deltas
7521 50 : .data
7522 124364 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7523 50 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
7524 50 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
7525 10464 : for (_, lsn, _) in &deltas.data {
7526 10414 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
7527 : }
7528 50 : assert!(
7529 50 : deltas.lsn_range.end <= last_record_lsn,
7530 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
7531 : deltas.lsn_range.end,
7532 : last_record_lsn
7533 : );
7534 50 : if let Some(check_start_lsn) = check_start_lsn {
7535 50 : assert!(deltas.lsn_range.start >= check_start_lsn);
7536 0 : }
7537 50 : let mut delta_layer_writer = DeltaLayerWriter::new(
7538 50 : self.conf,
7539 50 : self.timeline_id,
7540 50 : self.tenant_shard_id,
7541 50 : deltas.key_range.start,
7542 50 : deltas.lsn_range,
7543 50 : &self.gate,
7544 50 : self.cancel.clone(),
7545 50 : ctx,
7546 50 : )
7547 50 : .await?;
7548 10464 : for (key, lsn, val) in deltas.data {
7549 10414 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
7550 : }
7551 50 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
7552 50 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7553 50 : info!("force created delta layer {}", delta_layer.local_path());
7554 : {
7555 50 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7556 50 : guard
7557 50 : .open_mut()
7558 50 : .unwrap()
7559 50 : .force_insert_layer(delta_layer.clone());
7560 : }
7561 :
7562 : // Update remote_timeline_client state to reflect existence of this layer
7563 50 : self.remote_client
7564 50 : .schedule_layer_file_upload(delta_layer)
7565 50 : .unwrap();
7566 :
7567 50 : Ok(())
7568 50 : }
7569 :
7570 : /// Force create an in-memory layer and place them into the layer map.
7571 : #[cfg(test)]
7572 4 : pub(super) async fn force_create_in_memory_layer(
7573 4 : self: &Arc<Timeline>,
7574 4 : mut in_memory: InMemoryLayerTestDesc,
7575 4 : check_start_lsn: Option<Lsn>,
7576 4 : ctx: &RequestContext,
7577 4 : ) -> anyhow::Result<()> {
7578 : use utils::bin_ser::BeSer;
7579 :
7580 : // Validate LSNs
7581 4 : if let Some(check_start_lsn) = check_start_lsn {
7582 4 : assert!(in_memory.lsn_range.start >= check_start_lsn);
7583 0 : }
7584 :
7585 4 : let last_record_lsn = self.get_last_record_lsn();
7586 4 : let layer_end_lsn = if in_memory.is_open {
7587 1 : in_memory
7588 1 : .data
7589 1 : .iter()
7590 1 : .map(|(_key, lsn, _value)| lsn)
7591 1 : .max()
7592 1 : .cloned()
7593 : } else {
7594 3 : Some(in_memory.lsn_range.end)
7595 : };
7596 :
7597 4 : if let Some(end) = layer_end_lsn {
7598 4 : assert!(
7599 4 : end <= last_record_lsn,
7600 0 : "advance last record lsn before inserting a layer, end_lsn={end}, last_record_lsn={last_record_lsn}",
7601 : );
7602 0 : }
7603 :
7604 19820 : in_memory.data.iter().for_each(|(_key, lsn, _value)| {
7605 19820 : assert!(*lsn >= in_memory.lsn_range.start);
7606 19820 : assert!(*lsn < in_memory.lsn_range.end);
7607 19820 : });
7608 :
7609 : // Build the batch
7610 4 : in_memory
7611 4 : .data
7612 273384 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7613 :
7614 4 : let data = in_memory
7615 4 : .data
7616 4 : .into_iter()
7617 19820 : .map(|(key, lsn, value)| {
7618 19820 : let value_size = value.serialized_size().unwrap() as usize;
7619 19820 : (key.to_compact(), lsn, value_size, value)
7620 19820 : })
7621 4 : .collect::<Vec<_>>();
7622 :
7623 4 : let batch = SerializedValueBatch::from_values(data);
7624 :
7625 : // Create the in-memory layer and write the batch into it
7626 4 : let layer = InMemoryLayer::create(
7627 4 : self.conf,
7628 4 : self.timeline_id,
7629 4 : self.tenant_shard_id,
7630 4 : in_memory.lsn_range.start,
7631 4 : &self.gate,
7632 4 : // TODO: if we ever use this function in production code, we need to pass the real cancellation token
7633 4 : &CancellationToken::new(),
7634 4 : ctx,
7635 4 : )
7636 4 : .await
7637 4 : .unwrap();
7638 :
7639 4 : layer.put_batch(batch, ctx).await.unwrap();
7640 4 : if !in_memory.is_open {
7641 3 : layer.freeze(in_memory.lsn_range.end).await;
7642 1 : }
7643 :
7644 4 : info!("force created in-memory layer {:?}", in_memory.lsn_range);
7645 :
7646 : // Link the layer to the layer map
7647 : {
7648 4 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7649 4 : let layer_map = guard.open_mut().unwrap();
7650 4 : layer_map.force_insert_in_memory_layer(Arc::new(layer));
7651 : }
7652 :
7653 4 : Ok(())
7654 4 : }
7655 :
7656 : /// Return all keys at the LSN in the image layers
7657 : #[cfg(test)]
7658 3 : pub(crate) async fn inspect_image_layers(
7659 3 : self: &Arc<Timeline>,
7660 3 : lsn: Lsn,
7661 3 : ctx: &RequestContext,
7662 3 : io_concurrency: IoConcurrency,
7663 3 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
7664 3 : let mut all_data = Vec::new();
7665 3 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7666 17 : for layer in guard.layer_map()?.iter_historic_layers() {
7667 17 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
7668 4 : let layer = guard.get_from_desc(&layer);
7669 4 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
7670 4 : layer
7671 4 : .get_values_reconstruct_data(
7672 4 : KeySpace::single(Key::MIN..Key::MAX),
7673 4 : lsn..Lsn(lsn.0 + 1),
7674 4 : &mut reconstruct_data,
7675 4 : ctx,
7676 4 : )
7677 4 : .await?;
7678 33 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
7679 33 : let v = v.collect_pending_ios().await?;
7680 33 : all_data.push((k, v.img.unwrap().1));
7681 : }
7682 13 : }
7683 : }
7684 3 : all_data.sort();
7685 3 : Ok(all_data)
7686 3 : }
7687 :
7688 : /// Get all historic layer descriptors in the layer map
7689 : #[cfg(test)]
7690 12 : pub(crate) async fn inspect_historic_layers(
7691 12 : self: &Arc<Timeline>,
7692 12 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
7693 12 : let mut layers = Vec::new();
7694 12 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7695 57 : for layer in guard.layer_map()?.iter_historic_layers() {
7696 57 : layers.push(layer.key());
7697 57 : }
7698 12 : Ok(layers)
7699 12 : }
7700 :
7701 : #[cfg(test)]
7702 5 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
7703 5 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
7704 5 : keyspace.merge(&ks);
7705 5 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
7706 5 : }
7707 : }
7708 :
7709 : /// Tracking writes ingestion does to a particular in-memory layer.
7710 : ///
7711 : /// Cleared upon freezing a layer.
7712 : pub(crate) struct TimelineWriterState {
7713 : open_layer: Arc<InMemoryLayer>,
7714 : current_size: u64,
7715 : // Previous Lsn which passed through
7716 : prev_lsn: Option<Lsn>,
7717 : // Largest Lsn which passed through the current writer
7718 : max_lsn: Option<Lsn>,
7719 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
7720 : cached_last_freeze_at: Lsn,
7721 : }
7722 :
7723 : impl TimelineWriterState {
7724 660 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
7725 660 : Self {
7726 660 : open_layer,
7727 660 : current_size,
7728 660 : prev_lsn: None,
7729 660 : max_lsn: None,
7730 660 : cached_last_freeze_at: last_freeze_at,
7731 660 : }
7732 660 : }
7733 : }
7734 :
7735 : /// Various functions to mutate the timeline.
7736 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
7737 : // This is probably considered a bad practice in Rust and should be fixed eventually,
7738 : // but will cause large code changes.
7739 : pub(crate) struct TimelineWriter<'a> {
7740 : tl: &'a Timeline,
7741 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
7742 : }
7743 :
7744 : impl Deref for TimelineWriter<'_> {
7745 : type Target = Timeline;
7746 :
7747 4949226 : fn deref(&self) -> &Self::Target {
7748 4949226 : self.tl
7749 4949226 : }
7750 : }
7751 :
7752 : #[derive(PartialEq)]
7753 : enum OpenLayerAction {
7754 : Roll,
7755 : Open,
7756 : None,
7757 : }
7758 :
7759 : impl TimelineWriter<'_> {
7760 2402129 : async fn handle_open_layer_action(
7761 2402129 : &mut self,
7762 2402129 : at: Lsn,
7763 2402129 : action: OpenLayerAction,
7764 2402129 : ctx: &RequestContext,
7765 2402129 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
7766 2402129 : match action {
7767 : OpenLayerAction::Roll => {
7768 40 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
7769 40 : self.roll_layer(freeze_at).await?;
7770 40 : self.open_layer(at, ctx).await?;
7771 : }
7772 620 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
7773 : OpenLayerAction::None => {
7774 2401469 : assert!(self.write_guard.is_some());
7775 : }
7776 : }
7777 :
7778 2402129 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
7779 2402129 : }
7780 :
7781 660 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
7782 660 : let layer = self
7783 660 : .tl
7784 660 : .get_layer_for_write(at, &self.write_guard, ctx)
7785 660 : .await?;
7786 660 : let initial_size = layer.len();
7787 :
7788 660 : let last_freeze_at = self.last_freeze_at.load();
7789 660 : self.write_guard.replace(TimelineWriterState::new(
7790 660 : layer,
7791 660 : initial_size,
7792 660 : last_freeze_at,
7793 660 : ));
7794 :
7795 660 : Ok(())
7796 660 : }
7797 :
7798 40 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
7799 40 : let current_size = self.write_guard.as_ref().unwrap().current_size;
7800 :
7801 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
7802 : // to propagate the backpressure up into WAL ingestion.
7803 40 : let l0_count = self
7804 40 : .tl
7805 40 : .layers
7806 40 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7807 40 : .await
7808 40 : .layer_map()?
7809 40 : .level0_deltas()
7810 40 : .len();
7811 40 : let wait_thresholds = [
7812 40 : self.get_l0_flush_delay_threshold(),
7813 40 : self.get_l0_flush_stall_threshold(),
7814 40 : ];
7815 40 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
7816 :
7817 : // self.write_guard will be taken by the freezing
7818 40 : let flush_id = self
7819 40 : .tl
7820 40 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
7821 40 : .await?;
7822 :
7823 40 : assert!(self.write_guard.is_none());
7824 :
7825 40 : if let Some(wait_threshold) = wait_threshold {
7826 0 : if l0_count >= wait_threshold {
7827 0 : debug!(
7828 0 : "layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers"
7829 : );
7830 0 : self.tl.wait_flush_completion(flush_id).await?;
7831 0 : }
7832 40 : }
7833 :
7834 40 : if current_size >= self.get_checkpoint_distance() * 2 {
7835 0 : warn!("Flushed oversized open layer with size {}", current_size)
7836 40 : }
7837 :
7838 40 : Ok(())
7839 40 : }
7840 :
7841 2402129 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
7842 2402129 : let state = &*self.write_guard;
7843 2402129 : let Some(state) = &state else {
7844 620 : return OpenLayerAction::Open;
7845 : };
7846 :
7847 : #[cfg(feature = "testing")]
7848 2401509 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
7849 : // this check and assertion are not really needed because
7850 : // LayerManager::try_freeze_in_memory_layer will always clear out the
7851 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
7852 : // is no TimelineWriterState.
7853 0 : assert!(
7854 0 : state.open_layer.end_lsn.get().is_some(),
7855 0 : "our open_layer must be outdated"
7856 : );
7857 :
7858 : // this would be a memory leak waiting to happen because the in-memory layer always has
7859 : // an index
7860 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
7861 2401509 : }
7862 :
7863 2401509 : if state.prev_lsn == Some(lsn) {
7864 : // Rolling mid LSN is not supported by [downstream code].
7865 : // Hence, only roll at LSN boundaries.
7866 : //
7867 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
7868 3 : return OpenLayerAction::None;
7869 2401506 : }
7870 :
7871 2401506 : if state.current_size == 0 {
7872 : // Don't roll empty layers
7873 0 : return OpenLayerAction::None;
7874 2401506 : }
7875 :
7876 2401506 : if self.tl.should_roll(
7877 2401506 : state.current_size,
7878 2401506 : state.current_size + new_value_size,
7879 2401506 : self.get_checkpoint_distance(),
7880 2401506 : lsn,
7881 2401506 : state.cached_last_freeze_at,
7882 2401506 : state.open_layer.get_opened_at(),
7883 : ) {
7884 40 : OpenLayerAction::Roll
7885 : } else {
7886 2401466 : OpenLayerAction::None
7887 : }
7888 2402129 : }
7889 :
7890 : /// Put a batch of keys at the specified Lsns.
7891 2402128 : pub(crate) async fn put_batch(
7892 2402128 : &mut self,
7893 2402128 : batch: SerializedValueBatch,
7894 2402128 : ctx: &RequestContext,
7895 2402128 : ) -> anyhow::Result<()> {
7896 2402128 : if !batch.has_data() {
7897 0 : return Ok(());
7898 2402128 : }
7899 :
7900 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
7901 : // We don't assert this in release builds, since key ownership policies may change over
7902 : // time. Stray keys will be removed during compaction.
7903 2402128 : if cfg!(debug_assertions) {
7904 4947529 : for metadata in &batch.metadata {
7905 2545401 : if let ValueMeta::Serialized(metadata) = metadata {
7906 2545401 : let key = Key::from_compact(metadata.key);
7907 2545401 : assert!(
7908 2545401 : self.shard_identity.is_key_local(&key)
7909 12 : || self.shard_identity.is_key_global(&key),
7910 0 : "key {key} does not belong on shard {}",
7911 0 : self.shard_identity.shard_index()
7912 : );
7913 0 : }
7914 : }
7915 0 : }
7916 :
7917 2402128 : let batch_max_lsn = batch.max_lsn;
7918 2402128 : let buf_size: u64 = batch.buffer_size() as u64;
7919 :
7920 2402128 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
7921 2402128 : let layer = self
7922 2402128 : .handle_open_layer_action(batch_max_lsn, action, ctx)
7923 2402128 : .await?;
7924 :
7925 2402128 : let res = layer.put_batch(batch, ctx).await;
7926 :
7927 2402128 : if res.is_ok() {
7928 2402128 : // Update the current size only when the entire write was ok.
7929 2402128 : // In case of failures, we may have had partial writes which
7930 2402128 : // render the size tracking out of sync. That's ok because
7931 2402128 : // the checkpoint distance should be significantly smaller
7932 2402128 : // than the S3 single shot upload limit of 5GiB.
7933 2402128 : let state = self.write_guard.as_mut().unwrap();
7934 2402128 :
7935 2402128 : state.current_size += buf_size;
7936 2402128 : state.prev_lsn = Some(batch_max_lsn);
7937 2402128 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
7938 2402128 : }
7939 :
7940 2402128 : res
7941 2402128 : }
7942 :
7943 : #[cfg(test)]
7944 : /// Test helper, for tests that would like to poke individual values without composing a batch
7945 2195079 : pub(crate) async fn put(
7946 2195079 : &mut self,
7947 2195079 : key: Key,
7948 2195079 : lsn: Lsn,
7949 2195079 : value: &Value,
7950 2195079 : ctx: &RequestContext,
7951 2195079 : ) -> anyhow::Result<()> {
7952 : use utils::bin_ser::BeSer;
7953 2195079 : if !key.is_valid_key_on_write_path() {
7954 0 : bail!(
7955 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
7956 : key
7957 : );
7958 2195079 : }
7959 2195079 : let val_ser_size = value.serialized_size().unwrap() as usize;
7960 2195079 : let batch = SerializedValueBatch::from_values(vec![(
7961 2195079 : key.to_compact(),
7962 2195079 : lsn,
7963 2195079 : val_ser_size,
7964 2195079 : value.clone(),
7965 2195079 : )]);
7966 :
7967 2195079 : self.put_batch(batch, ctx).await
7968 2195079 : }
7969 :
7970 1 : pub(crate) async fn delete_batch(
7971 1 : &mut self,
7972 1 : batch: &[(Range<Key>, Lsn)],
7973 1 : ctx: &RequestContext,
7974 1 : ) -> anyhow::Result<()> {
7975 1 : if let Some((_, lsn)) = batch.first() {
7976 1 : let action = self.get_open_layer_action(*lsn, 0);
7977 1 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
7978 1 : layer.put_tombstones(batch).await?;
7979 0 : }
7980 :
7981 1 : Ok(())
7982 1 : }
7983 :
7984 : /// Track the end of the latest digested WAL record.
7985 : /// Remember the (end of) last valid WAL record remembered in the timeline.
7986 : ///
7987 : /// Call this after you have finished writing all the WAL up to 'lsn'.
7988 : ///
7989 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
7990 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
7991 : /// the latest and can be read.
7992 2639560 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
7993 2639560 : self.tl.finish_write(new_lsn);
7994 2639560 : }
7995 :
7996 135285 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
7997 135285 : self.tl.update_current_logical_size(delta)
7998 135285 : }
7999 : }
8000 :
8001 : // We need TimelineWriter to be send in upcoming conversion of
8002 : // Timeline::layers to tokio::sync::RwLock.
8003 : #[test]
8004 1 : fn is_send() {
8005 1 : fn _assert_send<T: Send>() {}
8006 1 : _assert_send::<TimelineWriter<'_>>();
8007 1 : }
8008 :
8009 : #[cfg(test)]
8010 : mod tests {
8011 : use std::sync::Arc;
8012 :
8013 : use pageserver_api::key::Key;
8014 : use postgres_ffi::PgMajorVersion;
8015 : use std::iter::Iterator;
8016 : use tracing::Instrument;
8017 : use utils::id::TimelineId;
8018 : use utils::lsn::Lsn;
8019 : use wal_decoder::models::value::Value;
8020 :
8021 : use super::HeatMapTimeline;
8022 : use crate::context::RequestContextBuilder;
8023 : use crate::tenant::harness::{TenantHarness, test_img};
8024 : use crate::tenant::layer_map::LayerMap;
8025 : use crate::tenant::storage_layer::{Layer, LayerName, LayerVisibilityHint};
8026 : use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
8027 : use crate::tenant::timeline::{DeltaLayerTestDesc, EvictionError};
8028 : use crate::tenant::{PreviousHeatmap, Timeline};
8029 :
8030 5 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
8031 5 : assert_eq!(lhs.all_layers().count(), rhs.all_layers().count());
8032 5 : let lhs_rhs = lhs.all_layers().zip(rhs.all_layers());
8033 25 : for (l, r) in lhs_rhs {
8034 20 : assert_eq!(l.name, r.name);
8035 20 : assert_eq!(l.metadata, r.metadata);
8036 : }
8037 5 : }
8038 :
8039 : #[tokio::test]
8040 1 : async fn test_heatmap_generation() {
8041 1 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
8042 :
8043 1 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
8044 1 : Lsn(0x10)..Lsn(0x20),
8045 1 : vec![(
8046 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
8047 1 : Lsn(0x11),
8048 1 : Value::Image(test_img("foo")),
8049 1 : )],
8050 : );
8051 1 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
8052 1 : Lsn(0x10)..Lsn(0x20),
8053 1 : vec![(
8054 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
8055 1 : Lsn(0x11),
8056 1 : Value::Image(test_img("foo")),
8057 1 : )],
8058 : );
8059 1 : let l0_delta = DeltaLayerTestDesc::new(
8060 1 : Lsn(0x20)..Lsn(0x30),
8061 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
8062 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
8063 1 : vec![(
8064 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
8065 1 : Lsn(0x25),
8066 1 : Value::Image(test_img("foo")),
8067 1 : )],
8068 : );
8069 1 : let delta_layers = vec![
8070 1 : covered_delta.clone(),
8071 1 : visible_delta.clone(),
8072 1 : l0_delta.clone(),
8073 : ];
8074 :
8075 1 : let image_layer = (
8076 1 : Lsn(0x40),
8077 1 : vec![(
8078 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
8079 1 : test_img("bar"),
8080 1 : )],
8081 1 : );
8082 1 : let image_layers = vec![image_layer];
8083 :
8084 1 : let (tenant, ctx) = harness.load().await;
8085 1 : let timeline = tenant
8086 1 : .create_test_timeline_with_layers(
8087 1 : TimelineId::generate(),
8088 1 : Lsn(0x10),
8089 1 : PgMajorVersion::PG14,
8090 1 : &ctx,
8091 1 : Vec::new(), // in-memory layers
8092 1 : delta_layers,
8093 1 : image_layers,
8094 1 : Lsn(0x100),
8095 1 : )
8096 1 : .await
8097 1 : .unwrap();
8098 1 : let ctx = &ctx.with_scope_timeline(&timeline);
8099 :
8100 : // Layer visibility is an input to heatmap generation, so refresh it first
8101 1 : timeline.update_layer_visibility().await.unwrap();
8102 :
8103 1 : let heatmap = timeline
8104 1 : .generate_heatmap()
8105 1 : .await
8106 1 : .expect("Infallible while timeline is not shut down");
8107 :
8108 1 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
8109 :
8110 : // L0 should come last
8111 1 : let heatmap_layers = heatmap.all_layers().collect::<Vec<_>>();
8112 1 : assert_eq!(heatmap_layers.last().unwrap().name, l0_delta.layer_name());
8113 :
8114 1 : let mut last_lsn = Lsn::MAX;
8115 5 : for layer in heatmap_layers {
8116 : // Covered layer should be omitted
8117 4 : assert!(layer.name != covered_delta.layer_name());
8118 :
8119 4 : let layer_lsn = match &layer.name {
8120 2 : LayerName::Delta(d) => d.lsn_range.end,
8121 2 : LayerName::Image(i) => i.lsn,
8122 : };
8123 :
8124 : // Apart from L0s, newest Layers should come first
8125 4 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
8126 3 : assert!(layer_lsn <= last_lsn);
8127 3 : last_lsn = layer_lsn;
8128 1 : }
8129 : }
8130 :
8131 : // Evict all the layers and stash the old heatmap in the timeline.
8132 : // This simulates a migration to a cold secondary location.
8133 :
8134 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
8135 1 : let mut all_layers = Vec::new();
8136 1 : let forever = std::time::Duration::from_secs(120);
8137 5 : for layer in guard.likely_resident_layers() {
8138 5 : all_layers.push(layer.clone());
8139 5 : layer.evict_and_wait(forever).await.unwrap();
8140 : }
8141 1 : drop(guard);
8142 :
8143 1 : timeline
8144 1 : .previous_heatmap
8145 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
8146 1 : heatmap: heatmap.clone(),
8147 1 : read_at: std::time::Instant::now(),
8148 1 : end_lsn: None,
8149 1 : })));
8150 :
8151 : // Generate a new heatmap and assert that it contains the same layers as the old one.
8152 1 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
8153 1 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
8154 :
8155 : // Download each layer one by one. Generate the heatmap at each step and check
8156 : // that it's stable.
8157 6 : for layer in all_layers {
8158 5 : if layer.visibility() == LayerVisibilityHint::Covered {
8159 1 : continue;
8160 4 : }
8161 1 :
8162 4 : eprintln!("Downloading {layer} and re-generating heatmap");
8163 1 :
8164 4 : let ctx = &RequestContextBuilder::from(ctx)
8165 4 : .download_behavior(crate::context::DownloadBehavior::Download)
8166 4 : .attached_child();
8167 1 :
8168 4 : let _resident = layer
8169 4 : .download_and_keep_resident(ctx)
8170 4 : .instrument(tracing::info_span!(
8171 4 : parent: None,
8172 1 : "download_layer",
8173 1 : tenant_id = %timeline.tenant_shard_id.tenant_id,
8174 1 : shard_id = %timeline.tenant_shard_id.shard_slug(),
8175 1 : timeline_id = %timeline.timeline_id
8176 1 : ))
8177 4 : .await
8178 4 : .unwrap();
8179 1 :
8180 4 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
8181 4 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
8182 1 : }
8183 1 :
8184 1 : // Everything from the post-migration heatmap is now resident.
8185 1 : // Check that we drop it from memory.
8186 1 : assert!(matches!(
8187 1 : timeline.previous_heatmap.load().as_deref(),
8188 1 : Some(PreviousHeatmap::Obsolete)
8189 1 : ));
8190 1 : }
8191 :
8192 : #[tokio::test]
8193 1 : async fn test_previous_heatmap_obsoletion() {
8194 1 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
8195 1 : .await
8196 1 : .unwrap();
8197 :
8198 1 : let l0_delta = DeltaLayerTestDesc::new(
8199 1 : Lsn(0x20)..Lsn(0x30),
8200 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
8201 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
8202 1 : vec![(
8203 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
8204 1 : Lsn(0x25),
8205 1 : Value::Image(test_img("foo")),
8206 1 : )],
8207 : );
8208 :
8209 1 : let image_layer = (
8210 1 : Lsn(0x40),
8211 1 : vec![(
8212 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
8213 1 : test_img("bar"),
8214 1 : )],
8215 1 : );
8216 :
8217 1 : let delta_layers = vec![l0_delta];
8218 1 : let image_layers = vec![image_layer];
8219 :
8220 1 : let (tenant, ctx) = harness.load().await;
8221 1 : let timeline = tenant
8222 1 : .create_test_timeline_with_layers(
8223 1 : TimelineId::generate(),
8224 1 : Lsn(0x10),
8225 1 : PgMajorVersion::PG14,
8226 1 : &ctx,
8227 1 : Vec::new(), // in-memory layers
8228 1 : delta_layers,
8229 1 : image_layers,
8230 1 : Lsn(0x100),
8231 1 : )
8232 1 : .await
8233 1 : .unwrap();
8234 :
8235 : // Layer visibility is an input to heatmap generation, so refresh it first
8236 1 : timeline.update_layer_visibility().await.unwrap();
8237 :
8238 1 : let heatmap = timeline
8239 1 : .generate_heatmap()
8240 1 : .await
8241 1 : .expect("Infallible while timeline is not shut down");
8242 :
8243 : // Both layers should be in the heatmap
8244 1 : assert!(heatmap.all_layers().count() > 0);
8245 :
8246 : // Now simulate a migration.
8247 1 : timeline
8248 1 : .previous_heatmap
8249 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
8250 1 : heatmap: heatmap.clone(),
8251 1 : read_at: std::time::Instant::now(),
8252 1 : end_lsn: None,
8253 1 : })));
8254 :
8255 : // Evict all the layers in the previous heatmap
8256 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
8257 1 : let forever = std::time::Duration::from_secs(120);
8258 3 : for layer in guard.likely_resident_layers() {
8259 3 : layer.evict_and_wait(forever).await.unwrap();
8260 : }
8261 1 : drop(guard);
8262 :
8263 : // Generate a new heatmap and check that the previous heatmap
8264 : // has been marked obsolete.
8265 1 : let post_eviction_heatmap = timeline
8266 1 : .generate_heatmap()
8267 1 : .await
8268 1 : .expect("Infallible while timeline is not shut down");
8269 :
8270 1 : assert_eq!(post_eviction_heatmap.all_layers().count(), 0);
8271 1 : assert!(matches!(
8272 1 : timeline.previous_heatmap.load().as_deref(),
8273 1 : Some(PreviousHeatmap::Obsolete)
8274 1 : ));
8275 1 : }
8276 :
8277 : #[tokio::test]
8278 1 : async fn two_layer_eviction_attempts_at_the_same_time() {
8279 1 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
8280 1 : .await
8281 1 : .unwrap();
8282 :
8283 1 : let (tenant, ctx) = harness.load().await;
8284 1 : let timeline = tenant
8285 1 : .create_test_timeline(
8286 1 : TimelineId::generate(),
8287 1 : Lsn(0x10),
8288 1 : PgMajorVersion::PG14,
8289 1 : &ctx,
8290 1 : )
8291 1 : .await
8292 1 : .unwrap();
8293 :
8294 1 : let layer = find_some_layer(&timeline).await;
8295 1 : let layer = layer
8296 1 : .keep_resident()
8297 1 : .await
8298 1 : .expect("no download => no downloading errors")
8299 1 : .drop_eviction_guard();
8300 :
8301 1 : let forever = std::time::Duration::from_secs(120);
8302 :
8303 1 : let first = layer.evict_and_wait(forever);
8304 1 : let second = layer.evict_and_wait(forever);
8305 :
8306 1 : let (first, second) = tokio::join!(first, second);
8307 :
8308 1 : let res = layer.keep_resident().await;
8309 1 : assert!(res.is_none(), "{res:?}");
8310 :
8311 1 : match (first, second) {
8312 1 : (Ok(()), Ok(())) => {
8313 1 : // because there are no more timeline locks being taken on eviction path, we can
8314 1 : // witness all three outcomes here.
8315 1 : }
8316 1 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
8317 0 : // if one completes before the other, this is fine just as well.
8318 0 : }
8319 1 : other => unreachable!("unexpected {:?}", other),
8320 1 : }
8321 1 : }
8322 :
8323 1 : async fn find_some_layer(timeline: &Timeline) -> Layer {
8324 1 : let layers = timeline
8325 1 : .layers
8326 1 : .read(LayerManagerLockHolder::GetLayerMapInfo)
8327 1 : .await;
8328 1 : let desc = layers
8329 1 : .layer_map()
8330 1 : .unwrap()
8331 1 : .iter_historic_layers()
8332 1 : .next()
8333 1 : .expect("must find one layer to evict");
8334 :
8335 1 : layers.get_from_desc(&desc)
8336 1 : }
8337 : }
|
